{"id": "2606.18264", "title": "Simulating Hate Speech Cascades with Multi-LLM Agents: Empirical Grounding, Modeling Fidelity, and Intervention Strategies", "abstract": "Faithful modeling of hateful content propagation on online platforms remains an open problem for moderation research. Classical cascade models that do not explicitly represent the profile, community, and content factors associated with hateful-content propagation may yield moderation strategies that behave less effectively when deployed in real-world scenarios. Multi-agent large language model (LLM) systems can, in principle, make each reshare decision depend on the user's profile, the surrounding community, and the post's content, but it remains unclear whether this added flexibility actually reproduces real hateful cascades more faithfully than classical baselines. We study three hateful Bluesky cascades and a size-matched benign control. In the empirical Bluesky data, we found that: 97.4--99.7\\% of reposters take a hostile stance; toxicity-engagement homophily is higher on the diffusion tree than on the follower graph for hateful cascades; topology is star-like for the hateful cascades (most reposts come directly from the root) versus tree-like for the benign cascade (reposts propagate through multi-hop chains). In simulation, a multi-LLM-agent simulator reproduces the stance monoculture and the toxicity-delta direction. A structured ablation identifies agent heterogeneity as the leading fidelity factor, and amplifier targeting on dense networks yields 7.5--12.9\\% reduction at 5.7\\% benign collateral.", "date": "2026-05-21", "categories": []}
{"id": "2606.18852", "title": "Aligning Implied Statements for Implicit Hate Speech Generalizability with Context-Bounded Semi-hard Negative Mining", "abstract": "Classifying implicit hate speech remains a challenge, as intent is often masked through insinuation and context rather than explicit slurs. Prior supervised contrastive approaches improve in-domain detection but can overfit surface cues and struggle to transfer across datasets. We propose ImpSH, a triplet-based framework that aligns posts with implied statements when available and uses context-bounded semi-hard negatives to focus learning on near confusions. We also examine AugSH, which forms positives via data augmentation. In controlled evaluations on IHC, SBIC, and DynaHate with BERT and HateBERT, ImpSH is a viable alternative to standard supervised contrastive baselines and often improves cross-domain performance under matched preprocessing and tuning budgets. Representation analysis using alignment and uniformity indicates tighter positive pairs with balanced global spread, and qualitative nearest-neighbor case studies illustrate typical false negatives under domain shift. These results demonstrate that aligning posts with their implied statements via context-bounded mining provides a more stable, bijective-like mapping to related insinuations, overcoming the volatility inherent in traditional clustering-based representation learning.", "date": "2026-06-17", "categories": []}
{"id": "2403.18957", "title": "Moderating Illicit Online Image Promotion for Unsafe User-Generated Content Games Using Large Vision-Language Models", "abstract": "Online user generated content games (UGCGs) are increasingly popular among children and adolescents for social interaction and more creative online entertainment. However, they pose a heightened risk of exposure to explicit content, raising growing concerns for the online safety of children and adolescents. Despite these concerns, few studies have addressed the issue of illicit image-based promotions of unsafe UGCGs on social media, which can inadvertently attract young users. This challenge arises from the difficulty of obtaining comprehensive training data for UGCG images and the unique nature of these images, which differ from traditional unsafe content. In this work, we take the first step towards studying the threat of illicit promotions of unsafe UGCGs. We collect a real-world dataset comprising 2,924 images that display diverse sexually explicit and violent content used to promote UGCGs by their game creators. Our in-depth studies reveal a new understanding of this problem and the urgent need for automatically flagging illicit UGCG promotions. We additionally create a cutting-edge system, UGCG-Guard, designed to aid social media platforms in effectively identifying images used for illicit UGCG promotions. This system leverages recently introduced large vision-language models (VLMs) and employs a novel conditional prompting strategy for zero-shot domain adaptation, along with chain-of-thought (CoT) reasoning for contextual identification. UGCG-Guard achieves outstanding results, with an accuracy rate of 94% in detecting these images used for the illicit promotion of such games in real-world scenarios.", "date": "2024-03-27", "categories": []}
{"id": "2502.08266", "title": "Dealing with Annotator Disagreement in Hate Speech Classification", "abstract": "Hate speech detection is a crucial task, especially on social media where harmful content can spread quickly. Collecting social media content (tweets etc.) to train machine learning models is easy, but detecting and categorizing hate speech can be difficult due to the inherently subjective nature. This subjectivity leads to frequent disagreement among annotators, particularly for subtle or borderline content. Traditional approaches either discard non-consensus samples or force a ''gold standard'' through expert adjudication, ignoring valuable information about uncertainty and diverse human perspectives. We examine the largely overlooked problem of annotator disagreement in hate speech classification and evaluate a range of aggregation methods, including majority voting, ordinal strategies (minimum, maximum, and mean), and analyze their impact across binary, 4-class, and 6-class classification tasks. In addition, we leverage annotators' perceived hate speech strength scores to explore regression-based and hybrid modeling approaches. Among others, we show that filtering non-consensus samples results in over-optimistic results and that the perceived strength provides a complementary signal that enhance classification performance. Finally, we establish new state-of-the-art results for hate speech detection in Turkish tweets, and demonstrate that annotator disagreement, when properly modeled, is a valuable resource for building more robust and reliable systems.", "date": "2025-02-12", "categories": []}
{"id": "2504.08609", "title": "A Survey of Machine Learning Models and Datasets for the Multi-label Classification of Textual Hate Speech in English", "abstract": "The dissemination of online hate speech can have serious negative consequences for individuals, online communities, and entire societies. This and the large volume of hateful online content prompted both practitioners', i.e., in content moderation or law enforcement, and researchers' interest in machine learning models to automatically classify instances of hate speech. Whereas most scientific works address hate speech classification as a binary task, practice often requires a differentiation into sub-types, e.g., according to target, severity, or legality, which may overlap for individual content. Hence, researchers created datasets and machine learning models that approach hate speech classification in textual data as a multi-label problem. This work presents the first systematic and comprehensive survey of scientific literature on this emerging research landscape in English (N=46). We contribute with a concise overview of 28 datasets suited for training multi-label classification models that reveals significant heterogeneity regarding label-set, size, meta-concept, annotation process, and inter-annotator agreement. Our analysis of 24 publications proposing suitable classification models further establishes inconsistency in evaluation and a preference for architectures based on Bidirectional Encoder Representation from Transformers (BERT) and Recurrent Neural Networks (RNNs). We identify imbalanced training data, reliance on crowdsourcing platforms, small and sparse datasets, and missing methodological alignment as critical open issues and formulate ten recommendations for research.", "date": "2025-04-11", "categories": []}
{"id": "2506.21613", "title": "ChildGuard: A Specialized Dataset for Combatting Child-Targeted Hate Speech", "abstract": "Mental health industry faces growing concerns regarding hate speech directed at children's on social media, as exposure to such content can contribute to adverse psychological outcomes during critical stages of development. Current hate speech datasets and detection systems provide limited support for child-focused applications because they are primarily designed for adults and lack dedicated representations of age-specific characteristics associated with hate speech directed at children's. To address this gap, we introduce ChildGuard, a large-scale English dataset for child-targeted hate speech containing 351,877 annotated instances collected from X (formerly Twitter), Reddit, and YouTube. The dataset covers three age groups such as younger children's (under 11), pre-teens (11-12), and teens (13-17). ChildGuard contains two subsets such as a contextual subset (157K) and a lexical subset (194K). Evaluation using recent transformer-based models and LLMs achieves a best Macro-F1 of 82.07%, decreasing to 79.41%, 79.24%, 76.04%, and 74.88% on younger children's, contextual, implicit hate, and cross-subset settings, respectively.", "date": "2025-06-21", "categories": []}
{"id": "2602.22391", "title": "Detecting Hate and Inflammatory Content in Bengali Memes: A New Multimodal Dataset and Co-Attention Framework", "abstract": "Internet memes have become a dominant form of expression on social media, including within the Bengali speaking community. While often humorous, memes can also be exploited to spread offensive, harmful, and inflammatory content targeting individuals and groups. Detecting this type of content is exceptionally challenging due to its satirical, subtle, and culturally specific nature. This problem is magnified for low-resource languages like Bengali, as existing research predominantly focuses on high-resource languages. To address this critical research gap, we introduce Bn-HIB (Bangla Hate Inflammatory Benign), a novel dataset containing 3,247 manually annotated Bengali memes categorized as Benign, Hate, or Inflammatory. Significantly, Bn- HIB is the first dataset to distinguish inflammatory content from direct hate speech in Bengali memes. Furthermore, we propose the MCFM (Multi-Modal Co-Attention Fusion Model), a simple yet effective architecture that mutually analyses both the visual and textual elements of a meme. MCFM employs a co-attention mechanism to identify and fuse the most critical features from each modality, leading to a more accurate classification. Our experiments show that MCFM significantly outperforms several state-of-the-art models on the Bn-HIB dataset, demonstrating its effectiveness in this nuanced task. To facilitate reproducibility and future research, the Bn-HIB dataset has been made publicly available through Mendeley Data. Warning: This work contains material that may be disturbing to some audience members. Viewer discretion is advised", "date": "2026-02-25", "categories": []}
{"id": "2606.14780", "title": "YTClickbait21K: Human-Annotated Multimodal Dataset for YouTube Clickbait Detection Across Diverse Channels and Content Categories", "abstract": "Clickbait content on video-sharing platforms poses a significant challenge to information reliability, yet progress in automated detection has been constrained by the lack of large-scale, high-quality multimodal datasets. We present YTClickbait21K, a human-annotated YouTube clickbait dataset comprising 21,238 videos collected from 40 channels across 29 countries, covering diverse content categories such as news, entertainment, education, and gaming. Each sample includes structured metadata (title, description, engagement statistics) along with associated thumbnail images, enabling comprehensive multimodal analysis. To ensure annotation quality, every video was independently labeled by three annotators using a standardized decision framework that incorporates textual, visual, and cross-modal consistency cues, with final labels determined through majority voting. The dataset exhibits substantial inter-annotator agreement (k=0.65), confirming reliable labeling despite the inherent subjectivity of clickbait detection. By combining scale, annotation rigor, and multimodal richness, this dataset provides a robust benchmark for developing and evaluating machine learning models, facilitating research in cross-modal semantic understanding, and advancing automated content moderation systems.", "date": "2026-06-10", "categories": []}
{"id": "2606.14786", "title": "MatchLM2Lite: A Scalable MLLM-to-Lite Framework for Reproduced Content Identification", "abstract": "Content moderation is critical for online video platforms to ensure content safety, protect creators, and sustain positive user experiences. Beyond filtering harmful content, platforms must guarantee content authenticity at scale so that users are exposed to diverse, original videos rather than low-value reproductions. We present MatchLM2Lite, a real-time, production-grade reproduced content identification (RCI) system that leverages the powerful understanding of a multimodal large language model (MLLM) distilled into a small and fast-inference model. Our system jointly models video, audio, and text signals, operating on pairs of videos to produce fine-grained reproduction scores. The system comprises two modules, MatchLM and MatchLite, and a two-stage training recipe. First, our high-capacity MLLM, MatchLM, serves as a teacher model to define the upper bound of RCI performance. Its capabilities are then distilled into a compact student model, MatchLite. This design allows MatchLite to deliver low-latency, high-throughput inference on video pairs while preserving much of MatchLM's accuracy, making it suitable for integration into real-time recommendation systems. MatchLM achieves an F1-score improvement of +8.57 compared to our previous production model. After knowledge distillation, MatchLite retains a +6.55 gain in F1-score while reducing computational cost by 35x. Deployed at scale, MatchLM2Lite enables efficient, pairwise multimodal RCI, stably serving online traffic at high queries per second (QPS) with an end-to-end latency below 30 seconds. This system has reduced the reproduced video view rate on our platform by 2.5% without degrading user engagement, demonstrating its effectiveness in a large-scale production environment.", "date": "2026-06-10", "categories": []}
{"id": "2606.15307", "title": "Adapting Reinforcement Learning with Chain-of-Thought Supervision for Explainable Detection of Hateful and Propagandistic Memes", "abstract": "Hateful and propagandistic memes exploit the interplay between images and text to convey harmful intent that neither modality reveals alone. Although thinking-based multimodal large language models (MLLMs) have advanced vision-language understanding, their application to meme content moderation remains underexplored. We propose a reinforcement learning-based post-training method that improves classification performance and reference-based explanation quality in thinking-based MLLMs via task-specific rewards and Group Relative Policy Optimization (GRPO). Concretely, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful and propagandistic meme understanding across English and Arabic benchmarks, (ii) extend existing meme datasets with weakly supervised chain-of-thought (CoT) rationales via distillation and multi-LLM fine-grained propaganda annotations, (iii) introduce a GRPO-based objective with thinking-length regularization that jointly optimizes classification accuracy and explanation quality, and (iv) investigate self-supervised GRPO on unlabeled memes using consensus-based pseudo-labels. Experiments on the Hateful Memes and ArMeme benchmarks show that our approach improves over previously reported results on FHM accuracy (up to +2.1%, from 79.9% to 82.0%) and on ArMeme macro-F1 (up to +7.6 points, from 0.536 to 0.612 with explanations; +6.1 compared to the original ArMeme benchmark), while also generating natural-language explanations. On ArMeme, sequence-classification baselines remain stronger in terms of raw accuracy, whereas our approach provides more balanced per-class performance along with explanations. We publicly release our code, data extensions, and evaluation resources.", "date": "2026-06-13", "categories": []}
{"id": "2606.15993", "title": "Classifying by Proxy: Explainable and Reproducible Ensemble of Proxy Tasks for Child Sexual Abuse Imagery Classification", "abstract": "Child Sexual Abuse Imagery (CSAI) classification systems are needed solutions for lessening the psychological impacts often felt by law enforcement agents responsible for evaluating these materials and for efficient removal of these materials from the web. However, due to the nature of the task, researching and developing such systems is not a trivial endeavor. The images are highly sensitive, and the related datasets are under restrictive access regimes, which means most studies in the area are not reproducible or distributable and are therefore hard to compare and validate. More concerning still, most models for this task today lack an aspect often desired by law enforcement agents: explainability. In this paper, we apply an ensemble of Proxy Tasks -- tasks that correlate to CSAI classification -- yielding improvements in reproducibility, explainability, and security for distribution. This concept is applied for the first time to real CSAI, with a novel selection of relevant Proxy Tasks (selected from the CSAI literature) and training adaptations to the original framework. Our final model achieves competitive results, yielding 91.9% balanced accuracy on the RCPD dataset with the best Proxy Task combination. We furthermore contrast these results with the best-in-class representation learning model, DINO, and show that our ensemble improves accuracy and provides explanations for its classification results, a feature that a single deep learning model can seldom provide.", "date": "2026-06-14", "categories": []}
{"id": "2606.13397", "title": "Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities", "abstract": "Language operates as a mechanism of both marginalization and resistance, especially for minority communities navigating insensitive and harmful speech online. As content moderation increasingly depends on large language models (LLMs), concerns arise about whether these systems can recognize culturally insensitive speech-language that disregards or marginalizes the cultural and religious perspectives of historically underrepresented communities, often through implicit erasure, misrepresentation, or normative framing, rather than overt hostility. Focusing on Bangladesh's Hindu and Chakma communities -- the country's largest religious and Indigenous ethnic minorities, respectively -- this paper investigates the epistemic limits of LLM-based moderation systems and explores methods for incorporating minority perspectives. We co-created a culturally grounded corpus of insensitive speech with community members and integrated their narratives into moderation pipelines using retrieval augmented generation (RAG). Our tool, Mod-Guide, improves LLM sensitivity to minority viewpoints by leveraging contextual cues derived from lived experience. Through mixed-method evaluations involving both minority and majority participants, we demonstrate that RAG-enhanced moderation responses are more contextually accurate and perceived differently across ethnic lines. This work advances research in human-computer interaction, AI ethics, and social computing by foregrounding restorative justice and hermeneutical inclusion in the design of content moderation systems.", "date": "2026-06-11", "categories": []}
{"id": "2409.12263", "title": "Detecting LGBTQ+ Instances of Cyberbullying", "abstract": "  Social media continues to have an impact on the trajectory of humanity. However, its introduction has also weaponized keyboards, allowing the abusive language normally reserved for in-person bullying to jump onto the screen, i.e., cyberbullying. Cyberbullying poses a significant threat to adolescents globally, affecting the mental health and well-being of many. A group that is particularly at risk is the LGBTQ+ community, as researchers have uncovered a strong correlation between identifying as LGBTQ+ and suffering from greater online harassment. Therefore, it is critical to develop machine learning models that can accurately discern cyberbullying incidents as they happen to LGBTQ+ members. The aim of this study is to compare the efficacy of several transformer models in identifying cyberbullying targeting LGBTQ+ individuals. We seek to determine the relative merits and demerits of these existing methods in addressing complex and subtle kinds of cyberbullying by assessing their effectiveness with real social media data.", "date": "2024-09-18", "categories": []}
{"id": "2606.08770", "title": "TeamHerald@CHIPSAL 2026: Hate Speech Detection and Sentiment Analysis of Nepali Memes using Transformer-based Architectures and Ensemble Learning", "abstract": "The analysis of internet memes in the Nepali language is complicated by frequent code-mixing and a lack of established baseline resources. While memes inherently combine visual and textual elements, this study focuses on a text-centric approach by extracting embedded text using an OCR layer and modeling it with Transformer-based architectures. We evaluate six distinct models and investigate the comparative effectiveness of Hard and Soft Voting ensemble strategies across two tasks: binary hate speech detection and three-class sentiment analysis. Experimental results show that a standalone decoder-only model achieved the highest performance for binary classification, whereas the Soft Voting ensemble performed best for the multi-class sentiment task, yielding a 15.8% relative improvement in Macro F1-score over the strongest standalone baseline. These findings suggest that ensemble strategies behave differently across binary and multi-class tasks, highlighting the importance of selecting aggregation methods suited to the classification objective.", "date": "2026-06-07", "categories": []}
{"id": "2606.09114", "title": "MAAM: Anchor-Preserving Compression and Contextual Calibration for Chinese Discriminatory Language Detection", "abstract": "Chinese discriminatory-language detection is challenging because harmful intent is often implicit and context-dependent. We propose MAAM (Myopia--Astigmatism Anchor Mechanism), a lightweight, model-agnostic framework inspired by functional visual blur: rather than preserving every token equally, MAAM retains discrimination-relevant semantic anchors and calibrates them with C--I--S contextual priors (Contextual Tone, Group Identity, and Stance Polarity). We also introduce ChLGBT, to our knowledge the first Chinese LGBT-focused discriminatory-language dataset, with 8,120 manually annotated samples and three ordinal labels: explicit bias, implicit bias, and emotional intensity. Across strong encoder baselines, MAAM improves all three prediction dimensions, with consistent gains in accuracy, F1, Brier score, and expected calibration error. Compared with frontier LLM baselines under zero-shot and few-shot prompting protocols, MAAM remains competitive while offering stronger compactness and stability. These results suggest that interpretable anchor preservation and contextual calibration provide a practical alternative to heavier model scaling for Chinese discriminatory-language assessment.", "date": "2026-06-08", "categories": []}
{"id": "2606.09700", "title": "What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks", "abstract": "Large language model (LLM)-powered content moderation systems have become a critical defense against harmful online content. However, these systems primarily operate on tokenized text and largely ignore the visual cues that humans naturally rely on when interpreting content. We show that this discrepancy creates a fundamental perceptual mismatch: content that is readily recognized as harmful by humans can become effectively invisible to automated moderation systems. To study this vulnerability, we introduce a class of Human-Perceptible Adversarial Attacks (HPAA), in which harmful expressions are embedded into otherwise benign text through visually salient typographic manipulations. Our key insight is that typographic features, including spacing, visual emphasis, and spatial arrangement, can be strategically combined to preserve human recognition of harmful content while substantially reducing machine detectability. Operating in black-box settings with only a small query budget, our attack automatically generates evasive content without requiring model access or gradient information. We evaluate the attack across multiple datasets and ten deployed moderation systems, including commercial APIs and state-of-the-art open-source guardrails. Results reveal a striking gap between human and machine perception: with only three detector queries, generated attacks achieve over 86\\% human recognition while maintaining detection rates below 1\\% across the evaluated systems. We further conduct ablation studies to identify the typographic factors driving successful evasion, analyze why current moderation architectures fail to capture these signals, and discuss practical defenses. Our findings expose a fundamental blind spot in today's LLM-based moderation ecosystem and highlight need for moderation systems that reason about content in a manner more consistent with human perceptual understanding.", "date": "2026-06-08", "categories": []}
{"id": "2606.05654", "title": "When Surface Form Changes Moderation Decisions: A Paired Study of Code-Mixed Workflow Instability", "abstract": "Hate moderation is often evaluated as classification on clean English inputs, but deployed systems must route content to actions such as ALLOW, FLAG, or REVIEW. We study how this workflow changes under code-mixed inputs using a paired evaluation setting where the same underlying content is expressed as clean English and Tamil-English code-mix. Under thresholds tuned on clean English development data, code-mixed inputs produce substantial action instability, with a paired clean- to-code-mix decision flip rate of 0.265. The main workflow effects are increased review burden and increased false-flagging of non-hateful content: review rate rises from 0.138 to 0.297 and non-hate false-flag rate rises from 0.069 to 0.104. Tamil-only inputs show stronger degradation overall, suggesting a broader language-coverage limitation rather than the same code-mixed instability pattern. A simple disagreement-based deferral rule reduces automatic errors on stressed inputs, but only by increasing review load. These results show that workflow-level evaluation reveals moderation failures that standard classification summaries can miss.", "date": "2026-06-04", "categories": []}
{"id": "2606.05748", "title": "UNIVID: Unified Vision-Language Model for Video Moderation", "abstract": "Global-scale video moderation faces a dual challenge: the need for fine-grained multi-modal reasoning and the demand for interpretable outputs to support downstream enforcement. Traditional moderation systems often rely on fragmented black-box classifiers that are difficult to maintain and lack transparency. In this paper, we present UNIVID, a UNIfied VIsion-language model for video moDeration. Unlike standard classification models, UNIVID generates policy-aware captions that serve as an interpretable intermediate representation, enabling human-verifiable decisions and multi-task reusability. While existing open-source and commercial VLMs often suffer from safety-guardrail refusals and lack fine-grained policy alignment, we develop a specialized training data recipe that combines expert human-refined labels with synthetic data to align the model with our safety guidelines. By integrating UNIVID as the core captioner, we design a novel end-to-end video moderation system that reduces violation leakage by 42.7% and overkill rate by 37.0% relatively. Meanwhile, by replacing over 1,000 policy-specific models with a single UNIVID backbone, we recycled extensive computation resources while reducing engineering maintenance overhead. To our knowledge, this is one of the first reports of a high-efficiency captioning VLM successfully supporting industrial-scale moderation and cross-functional business.", "date": "2026-06-04", "categories": []}
{"id": "2606.06266", "title": "From Self to Other: Evaluating Demographic Perspective-Taking in LLM Hate Speech Annotation", "abstract": "Hate speech detection is inherently subjective: people from different demographic groups perceive the same content very differently. Collecting enough annotations from multiple demographic groups is costly and difficult to scale. Persona-conditioned Large Language Models (models prompted to adopt a specific demographic identity) have been proposed as a way to simulate diverse perspectives at scale. But do they actually reflect how different groups disagree? We evaluate three aspects of human social judgement: (i) whether personas from different groups disagree in human-like ways (inter-group disagreement), (ii) whether they become more sensitive when content targets their own identity (in-group sensitivity), and (iii) whether they can accurately predict how another group would react (vicarious prediction). Our results show that no model consistently captures all three dimensions, and performance is highly model-dependent and does not emerge reliably from minimal identity prompts alone. However, vicarious prompting with Llama 3.1 yields the highest cross-group agreement in most demographic axes and provides the closest overall approximation to human disagreement patterns, indicating that this configuration may provide a more reliable setting for automatic annotation aligned with human judgements.", "date": "2026-06-04", "categories": []}
{"id": "2512.03553", "title": "Dynamic Content Moderation in Livestreams: Combining Supervised Classification with MLLM-Boosted Similarity Matching", "abstract": "Content moderation remains a critical yet challenging task for large-scale user-generated video platforms, especially in livestreaming environments where moderation must be timely, multimodal, and robust to evolving forms of unwanted content. We present a hybrid moderation framework deployed at production scale that combines supervised classification for known violations with reference-based similarity matching for novel or subtle cases. This hybrid design enables robust detection of both explicit violations and novel edge cases that evade traditional classifiers. Multimodal inputs (text, audio, visual) are processed through both pipelines, with a multimodal large language model (MLLM) distilling knowledge into each to boost accuracy while keeping inference lightweight. In production, the classification pipeline achieves 67% recall at 80% precision, and the similarity pipeline achieves 76% recall at 80% precision. Large-scale A/B tests show a 6-8% reduction in user views of unwanted livestreams}. These results demonstrate a scalable and adaptable approach to multimodal content governance, capable of addressing both explicit violations and emerging adversarial behaviors.", "date": "2025-12-03", "categories": []}
{"id": "2606.01298", "title": "Challenger at MultiPRIDE: Is It Hate Speech or Reclaimed?", "abstract": "The spread of hate speech has become increasingly harmful in modern digital environments, particularly on social networking platforms. While recent advances have shown promising results in automatic hate speech detection, a key challenge remains: distinguishing genuine hate speech from reclaimed language. Accurate labeling is difficult due to the nuanced and context-dependent nature of reclaimed expressions. In this paper, we present a simple and interpretable approach for distinguishing hate speech from reclaimed language, developed for the MultiPride Shared Task. Our method generates dense semantic text embeddings and incorporates a label-noise filtering stage using Cleanlab with logistic regression, followed by a Multi-layer Perceptron (MLP) neural network for final classification. The system is designed to operate under limited computational resources while maintaining strong performance. We evaluate our approach using precision, recall, and F1-score, including macro-averaged metrics. Experimental results demonstrate robust performance despite extreme class imbalance in the dataset. Overall, the findings highlight the potential for further improvements through larger embedding models and more advanced preprocessing techniques while preserving interpretability.", "date": "2026-05-31", "categories": []}
{"id": "2606.02523", "title": "FigSIM: A Dataset for Fine-grained Suicide Severity and Figurative Language in Suicide Memes", "abstract": "Suicide memes are memes used to express suicide-related thoughts or comment on suicide-related issues. Suicide memes are increasingly common on social media, yet remain poorly understood and potentially harmful. There is an urgent need to better understand their characteristics and to develop appropriate content moderation strategies that limits users' exposure to potentially harmful content. Currently, the absence of annotated datasets of suicide memes remains a key barrier to developing and evaluating automated moderation approaches. In this paper, we introduce FigSIM, the first dataset designed for fine-grained analysis of suicide memes. The dataset consists of 1049 memes, each annotated for (1) fine-grained suicide severity levels, (2) figurative phenomena (e.g., metaphors), and (3) suicide-related content (e.g., suicide method depiction). We benchmark 16 unimodal and multimodal models across three tasks: figurative language, suicide severity, and suicide-related content detection. Overall, FigSIM demonstrates that suicide memes pose unique challenges for both modeling and content moderation. Analysis revealed biases, such as underprediction of higher suicide severity levels, especially for figurative memes. The dataset (including splits used for analyses) is publicly available. Content Warning: This paper contains suicide-related content that may be triggering.", "date": "2026-06-01", "categories": []}
{"id": "2605.31349", "title": "FBHM: Functional Benchmarking and Steering of VLMs for Hateful Meme Detection", "abstract": "Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational - confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of Functionality Based Hateful Memes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (learnable steering vectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by ~30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance.", "date": "2026-05-29", "categories": []}
{"id": "2605.31563", "title": "Disagreeing Rationales: Rethinking Classification and Explainability Evaluation in Hate Speech Detection", "abstract": "Human disagreement is ubiquitous and well-known in labeling. However, variation in explanations, captured through token-level human rationales, remains far less explored. At the same time, it is unclear how to best evaluate human labels and rationales -- or even how to best aggregate rationales beyond majority vote -- in light of this variation. Yet, rationales may provide additional insights into the richness of human reasoning, that may differ in style, values and interpretations -- especially in subjective NLP tasks like hate speech detection. In this work, we unify diverse models, training strategies, loss functions, and existing evaluation metrics under a single protocol by systematically re-implementing them across different label and rationale representation spaces. Classification metrics are organized around two key properties -- predictive and distributional -- while explainability metrics through three complementary dimensions: plausibility, faithfulness, and complexity. In this unified supervision framework, we evaluate model behavior across classification and explainability metrics, as well as metric sensitivity to the choice of label (hard and soft) and rationale representation space (hard, intermediate and soft). Results show that both hard and soft metrics favor softer representations, highlighting their effectiveness in capturing variation and the need to rethink evaluation in subjective NLP.", "date": "2026-05-29", "categories": []}
{"id": "2510.10961", "title": "Obfuscation Rules for Detecting and Detoxifying Korean Toxicity", "abstract": "As language models become increasingly deployed in online environments, toxicity detection and detoxification have received growing attention. Existing studies primarily focus on non-obfuscated text, which limits robustness when users intentionally disguise toxic expressions. In particular, Korean toxic expressions can be easily disguised through agglutinative morphology and Hangeul-specific orthographic variation. However, obfuscation in Korean remains largely unexplored, which motivates us to introduce a KOTOX: Korean toxic dataset for deobfuscation and detoxification. We categorize Korean obfuscation patterns into linguistically grounded classes, define transformation rules derived from real-world examples, and provide the resulting obfuscation framework as an open transformation package. Using these rules, we provide paired neutral and toxic sentences alongside their obfuscated counterparts. Models trained on our dataset better handle obfuscated text without sacrificing performance on non-obfuscated text. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect the dataset to facilitate better understanding and mitigation of obfuscated toxic content in LLM for Korean. Our code and data are available at https://github.com/leeyejin1231/KOTOX.", "date": "2025-10-13", "categories": []}
{"id": "2601.11178", "title": "TANDEM: Temporal-Aware Neural Detection for Multimodal Hate Speech", "abstract": "Social media platforms are increasingly dominated by long-form multimodal content, where harmful narratives are constructed through a complex interplay of audio, visual, and textual cues. While automated systems can flag hate speech with high accuracy, they often function as \"black boxes\" that fail to provide the granular, interpretable evidence, such as precise timestamps and target identities, required for effective human-in-the-loop moderation. In this work, we introduce TANDEM, a unified framework that transforms audio-visual hate detection from a binary classification task into a structured reasoning problem. Our approach employs a novel tandem reinforcement learning strategy where vision-language and audio-language models optimize each other through self-constrained cross-modal context, stabilizing reasoning over extended temporal sequences without requiring dense frame-level supervision. Experiments across three benchmark datasets demonstrate that TANDEM significantly outperforms zero-shot and context-augmented baselines, achieving 0.73 F1 in target identification on HateMM (a 30% improvement over state-of-the-art) while maintaining precise temporal grounding. We further observe that while binary detection is robust, differentiating between offensive and hateful content remains challenging in multi-class settings due to inherent label ambiguity and dataset imbalance. More broadly, our findings suggest that structured, interpretable alignment is achievable even in complex multimodal settings, offering a blueprint for the next generation of transparent and actionable online safety moderation tools.", "date": "2026-01-16", "categories": []}
{"id": "2505.17654", "title": "EVADE-Bench: Multimodal Benchmark for Evaluating and Enhancing Evasive Content Detection", "abstract": "E-commerce platforms increasingly rely on Large Language Models (LLMs) and Vision Language Models (VLMs) to detect illicit or misleading product content. However, these models remain vulnerable to evasive content, which refers to inputs that have been deliberately modified through techniques such as word splitting, euphemistic language, or image cropping to conceal policy violations while still conveying prohibited claims. Crucially, detecting such content requires a model to simultaneously master two capabilities: accurately comprehending complex rules, and correctly inferring the true intent behind deliberately obfuscated multimodal inputs. While prior work has separately explored LLM reasoning over complex rules and LLM-based detection of evasive content, no existing benchmark combines both within a unified evaluation framework. This gap is particularly consequential in e-commerce, where accurate moderation demands that both capabilities operate in concert. To address this gap, we introduce EVADE-Bench, the first expert-curated Chinese multimodal benchmark specifically designed to evaluate LLMs and VLMs on evasive content detection in real-world e-commerce scenarios. Our comprehensive evaluation of 26 open- and closed-source LLMs and VLMs reveals that even state-of-the-art models frequently misclassify evasive samples. We further demonstrate that clearer rule categorization significantly improves model prediction consistency and reduces false predictions, highlighting the critical role of benchmark design in enabling reliable evaluation. To explore paths for performance improvement, we investigate the feasibility of multi-agent decomposition for multimodal reasoning, wherein visual description and logical inference are decoupled into separate agents, and find that this strategy yields notable accuracy gains.", "date": "2025-05-23", "categories": []}
{"id": "2605.27584", "title": "Cyberbullying Governance on Social Media: A Unified Framework from Content Identification to Intervention", "abstract": "The proliferation of social media platforms and online communities has inadvertently catalyzed the spread of cyberbullying, hate speech, and other forms of online toxicity, making the effective governance of such harm a critical societal and computational challenge. While significant strides have been made in automating content moderation, existing research predominantly treats cyberbullying governance as passive, isolated detection at the post level. This reductionist view overlooks the continuous behavioral dynamics of users, the structural diffusion of toxic events, and the critical need for proactive mitigation. To bridge these gaps, this paper proposes a unified full-lifecycle governance framework that shifts the paradigm of cyberbullying governance from isolated static detection toward integrated, continuous, and proactive moderation. Drawing on cyberbullying research and adjacent fields, we systematically synthesize the state-of-the-art literature across four interconnected stages: (1) Content Identification, (2) User and Behavior Modeling, (3) Diffusion Dynamics and Early Warning, and (4) Intervention and Governance. Furthermore, we review available datasets and evaluation practices, and discuss emerging challenges including multimodality, explainability, algorithmic fairness, and the dual-use risks of generative AI, providing a roadmap for future research toward a safer and more resilient digital ecosystem.", "date": "2026-05-26", "categories": []}
{"id": "2605.27313", "title": "When Does Demographic Information Help? Data and Modeling Regimes for Perspective-Aware Hate Speech Detection", "abstract": "Demographic information is often used to model annotator perspectives in subjective tasks such as hate speech detection, but its benefit is inconsistent: it improves performance in some settings and behaves as noise in others. This paper asks when demographic features help. We analyze demographic gain as a function of both data split properties and modeling frameworks. For data splits, we measure annotator disagreement, namely how often annotators assign different labels to the same example, along with training size and train-test demographic coverage. We find that demographic gains concentrate in regimes with low training disagreement, high test disagreement, fine-grained ambiguity measurement, sufficient training data, and greater demographic overlap. Motivated by these regimes, we introduce a gated demographic residual model that treats demographics as a selective adjustment to text-only predictions. Experiments on MHS and POPQUORN show that this design is effective, especially on high disagreement or low confidence examples. Overall, our results suggest that demographics should not be assumed useful by default; their value depends jointly on the data regime and the modeling framework.", "date": "2026-05-26", "categories": []}
{"id": "2605.24000", "title": "Toxicity in Twitch Chats: An LLM-Based Analysis Across Gaming Communities", "abstract": "Toxicity in online gaming communities remains a persistent challenge, manifesting across genres, platforms, and player interactions. While much research is focused on in-game toxicity, less is known about how toxic behavior varies between gaming communities on streaming platforms. To address this shortcoming, we analyze approximately 20 million chat messages from 4,452 streams, spanning seven game genres on Twitch. We categorize messages according to Twitch's toxicity taxonomy with a pre-trained Large Language Model using zero-shot classification. The taxonomy comprises four categories and eight subclasses, including harassment, discrimination, sexual content, and profanity. Our approach achieves an F1 score of 94.5% on the TextDetox dataset and demonstrates human-model agreement comparable to inter-human agreement. Our analysis reveals that 2.4% of all messages are classified as toxic, with notable differences across genres: streams of MOBA games exhibit the highest relative rate of toxicity (3.2%), and sports games show the lowest rate (2%). Furthermore, results indicate that individual games differ significantly in their toxicity distributions, even within genres, suggesting the existence of game-specific community norms and mechanics that shape toxic behavior beyond genre-level effects. These findings offer empirical insights into genre- and game-specific toxicity patterns on Twitch and can inform more targeted moderation strategies for gaming communities.", "date": "2026-05-18", "categories": []}
{"id": "2605.24247", "title": "Improving Labeling Consistency with Detailed Constitutional Definitions and AI-Driven Evaluation", "abstract": "Many automated labeling pipelines classify inputs into categories defined by a written specification, content moderation being a prominent use case. Simple category definitions are not detailed enough for labelers to produce the accurate, consistent golden labels these pipelines require. One solution is to write a prescriptive definition that settles enough real boundary cases that labelers cannot disagree with the written interpretation. In practice, definitions at that level of detail exceed what a human annotator can hold in working memory, so annotators fall back on intuition and the labels drift from the written rules, regressing on accuracy and consistency.\n  We propose and demonstrate the efficacy of an AI-driven workflow in which AI helps write a per-category constitution that defines the label in enough detail to cover edge cases, and a frontier LLM interprets it on each input to produce the golden label more consistently and accurately than humans reading the same document. We evaluate on three content moderation categories (harassment, hate speech, non-violent crime) and show that the approach reduces cross-model inconsistency by up to 57x compared to paragraph definitions, with cross-model disagreement diagnosing specification gaps and the human responsible for high-level decisions about what each category should mean rather than individual labeling calls. For the safety evaluation, we introduce a dual-axis formulation scoring intent and content independently over the full conversation, so downstream consumers can act on either axis or both.", "date": "2026-05-22", "categories": []}
{"id": "2605.24344", "title": "Distinguishing Right from Wrong in Debates: Attribution Analysis of Chinese Harmful Memes", "abstract": "Research on harmful meme detection has garnered significant attention, resulting in the development of numerous datasets and methods. However, progress in detecting Chinese harmful memes lags considerably, primarily due to two challenges: first, accurately assessing a meme's harmfulness depends heavily on understanding deep cultural context; second, many memes are semantically ambiguous, making harmfulness highly subjective. To address these issues, we focus on the interpretable detection of Chinese harmful memes by constructing the first Chinese harmful meme explanation dataset, Ex-ToxiCN-MM. This dataset offers opposing interpretations, categorized as \"harmful\" and \"non-harmful\", for each meme, aiming to rigorously evaluate a model's ability to discern and comprehend ambiguous, culturally grounded content. We built a specialized knowledge base of Chinese cultural concepts and offensive vocabulary to supply models with essential prior knowledge (C-HarmKB). To address the ambiguity and lack of background knowledge in meme attribution, we have developed a comprehensive attribution analysis framework, RIKE, which includes an Attribution Knowledge Enhancement module (AKE) and a Relative Intent Reasoning module (RIR). Extensive quantitative and qualitative experiments demonstrate that our method outperforms mainstream baseline models across multiple metrics in the task of attributing harmful memes in Chinese. The code, Ex-ToxiCN-MM dataset, and Chinese Harmful Semantic Knowledge Base (C-HarmKB) involved in this study have been open-sourced at https://github.com/wimiw123/Ex-ToxiCN-MM", "date": "2026-05-23", "categories": []}
{"id": "2605.22380", "title": "Multi-Stage Training for Abusive Comment Detection in Indic Languages", "abstract": "In recent years social media has become an increasingly popular tool for communication. People use it to share their ideas, exchange information, and discuss thoughts. Given its prevalence and widespread reach, social media must remain a safe space for people. Content generated on social media can be abusive and it has become increasingly important to detect such content. In this paper, we use a language-based preprocessing and an ensemble of several models and analyze their performance of abusive comment detection. Through extensive experimentation, we propose a pipeline that minimizes the false-positive rate (marking non-abusive as abusive) so that these systems can detect abusive comments without undermining the freedom of expression.", "date": "2026-05-21", "categories": []}
{"id": "2605.20584", "title": "QwenSafe: Multimodal Content Rating Description Identification via Preference-Aligned VLMs", "abstract": "Mobile app marketplaces require developers to disclose standardized content rating descriptors (CRDs) to inform users about potentially sensitive or restricted content. Ensuring the accuracy and consistency of these disclosures remains challenging due to the multimodal nature of app content, which spans textual descriptions and visual interfaces. In this paper, we present QwenSafe, a Vision-Language Model (VLM) designed to automatically identify the presence of Apple-defined CRDs by jointly reasoning over app metadata and screenshots. To enable scalable training for this task, we introduce metadata2CRD, a data-construction pipeline that synthesizes descriptor-aligned question-answer pairs by combining app descriptions, screenshots, and formal descriptor definitions. We adapt Qwen3-VL-8B using supervised fine-tuning followed by Direct Preference Optimization (DPO) to align model predictions with descriptor-specific evidence and explanations across visual and textual modalities. We evaluate QwenSafe on 12 Apple-defined content rating descriptors and compare it against state-of-the-art vision-language models, including Qwen3-VL, LLaVA-1.6, and Gemini-2.5-Flash. QwenSafe consistently outperforms all baselines in binary CRD classification, achieving improvements in positive-class recall of 111.8%, 36.1%, and 2.1%, respectively. Our results demonstrate that descriptor-aware multimodal alignment substantially improves automated content classification and highlights the potential of vision-language models to support scalable and consistent content rating in mobile app marketplaces.", "date": "2026-05-20", "categories": []}
{"id": "2512.14700", "title": "Context-Aware Detection and Victim-Centered Response Generation for Online Harassment in Private Messaging", "abstract": "Online harassment is a widespread social and public health concern, yet most computational approaches for detecting and addressing harassment focus on publicly visible social media content rather than private messaging environments. Private conversations present unique challenges because harmful interactions often unfold through context-dependent, multi-turn exchanges, while victims may lack timely support during moments of harassment. In this study, we investigate how large language models (LLMs) can support both the detection of and response to online harassment in private messaging. Using a dataset of 80,053 Instagram direct messages donated by 26 adolescents aged 12-18, including youth with suicide risk factors, we first construct a human-labeled dataset of online harassment in private conversations and develop a context-aware cascading LLM classification pipeline. The proposed pipeline outperforms baseline toxicity classifiers trained primarily on public social media data. We then develop a victim-centered response framework that produces context-sensitive and psychologically-grounded AI-generated responses to online harassment messages. Human evaluators perceived the AI-generated responses as significantly more helpful than the original participant responses (95% CI: 0.767--0.815, p < .001), particularly in terms of emotional support and de-escalation. Our findings highlight the potential of context-aware and victim-centered AI systems to provide just-in-time support during harassment in private messaging environments.", "date": "2025-11-28", "categories": []}
{"id": "2605.17187", "title": "PluRule: A Benchmark for Moderating Pluralistic Communities on Social Media", "abstract": "Social media are shifting towards pluralism -- community-governed platforms where groups define their own norms. What violates rules in one community may be perfectly acceptable in another. Can AI models help moderate such pluralistic communities? We formalize the task as a multiple-choice problem, mirroring how human moderators operate in the real world: given a comment and its surrounding context, identify which specific rule, if any, is violated. We introduce PluRule, a multimodal, multilingual benchmark for detecting 13,371 rule violations across 1,989 Reddit communities spanning 2,885 rules in 9 languages. Using this benchmark, we show that state-of-the-art vision-language models struggle significantly: even GPT-5.2 with high reasoning performs only slightly better than a trivial baseline. We also find that bigger models and increased context provide marginal gains, and universal rules like civility and self-promotion are easier to detect. Our results show that moderation of pluralistic communities on social media is a fundamental challenge for language models. Our code and benchmark are publicly available.", "date": "2026-05-16", "categories": []}
{"id": "2605.17610", "title": "SafeLens: Deliberate and Efficient Video Guardrails with Fast-and-Slow Screening", "abstract": "The rapid growth of online video platforms and AI-generated content has made reliable video guardrails a key challenge for safety and real-world deployment. While most videos can be screened through fast pattern recognition, a small subset requires deeper reasoning over temporally complex content and nuanced policy constraints. Existing approaches typically rely on large vision-language models applied uniformly across all inputs, resulting in high inference costs and inefficient allocation of computation. We propose SafeLens, a video guardrail framework that introduces a fast-and-slow inference architecture for efficient and accurate content moderation with variable computational cost across inputs. Additionally, we construct a high-quality dataset by applying influence-guided filtering to the SafeWatch Dataset, retaining only 2.4% of the original data. To further address limitations of training-time scaling, we enable test-time reasoning by augmenting the filtered data with structured Chain-of-Thought traces. Across real-world and AI-generated video benchmarks, SafeLens achieves state-of-the-art performance, outperforming strong open-source video guardrails (e.g., SafeWatch-8B, OmniGuard-7B) and closed-source models (e.g., GPT-5.4, Gemini-3.1-pro) while significantly reducing inference cost, demonstrating that efficient design serves to be more effective than scaling data or model size alone.", "date": "2026-05-17", "categories": []}
{"id": "2605.15984", "title": "Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues", "abstract": "Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic cues.To address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech analysis.Furthermore, we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss functions.Our experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.", "date": "2026-05-15", "categories": []}
{"id": "2503.16072", "title": "Toxicity Detection Should Measure Contextual Harm, Not Text-Intrinsic Badness", "abstract": "Toxicity detection has become core safety infrastructure for online moderation, dataset filtering, and deployed language-model systems. Yet most detectors still treat toxicity as an intrinsic property of isolated text. This position paper argues that toxicity detection should be evaluated as the contextual measurement of situated communicative harm, rather than as single-label text classification. Toxicity is not contained in words alone; it emerges when a communicative act is interpreted by an audience within a normative and social context.\n  We introduce the Contextual Stress Framework (CSF), which defines toxicity as a relation between perceived norm violation and induced stress or disruption. CSF explains why text-intrinsic detectors overflag dialectal or reclaimed language, miss coded or pragmatic abuse, and remain brittle under meaning-preserving transformations. We propose CSF-Eval, an evaluation agenda that separates text risk, norm violation, disruption, uncertainty, and policy action.", "date": "2025-03-20", "categories": []}
{"id": "2503.18273", "title": "Decoding Islamophobic Discourse: Using LLMs to Identify Tropes and Semi-Coded Hate Speech", "abstract": "In recent years, Islamophobia has gained significant traction across Western societies, fueled by the rise of digital communication networks. This paper performs a large-scale analysis of specialized, semi-coded Islamophobic terms such as (muzrat, pislam, mudslime, mohammedan, muzzies) floated on extremist social platforms, i.e., 4Chan, Gab, Telegram, etc. Many of these terms appear lexically neutral or ambiguous outside of specific contexts, making them difficult for both human moderators and automated systems to reliably identify as hate speech. First, we use Large Language Models (LLMs) to show their ability to understand these terms. Second, Google Perspective API suggests that Islamophobic posts tend to receive higher toxicity scores than other categories of hate speech like Antisemitism. Finally, we use BERT topic modeling approach to extract different topics and Islamophobic discourse on these social platforms. Our findings indicate that LLMs understand these Out-Of-Vocabulary (OOV) slurs; however, further improvements in moderation strategies and algorithmic detection are necessary to address such discourse effectively. Our topic modeling also indicates that Islamophobic text is found across various political, conspiratorial, and far-right movements and is particularly directed against Muslim immigrants. Taken altogether, we performed one of the first studies on Islamophobic semi-coded terms and shed a global light on Islamophobia.", "date": "2025-03-24", "categories": []}
{"id": "2505.20654", "title": "Chinese Cyberbullying Detection: Dataset, Method, and Validation", "abstract": "Existing cyberbullying detection benchmarks were organized by the polarity of speech, such as \"offensive\" and \"non-offensive\", which were essentially hate speech detection. However, in the real world, cyberbullying often attracted widespread social attention through incidents. To address this problem, we propose a novel annotation method to construct a cyberbullying dataset that organized by incidents. The constructed CHNCI is the first Chinese cyberbullying incident detection dataset, which consists of 220,676 comments in 91 incidents. Specifically, we first combine three cyberbullying detection methods based on explanations generation as an ensemble method to generate the pseudo labels, and then let human annotators judge these labels. Then we propose the evaluation criteria for validating whether it constitutes a cyberbullying incident. Experimental results demonstrate that the constructed dataset can be a benchmark for the tasks of cyberbullying detection and incident prediction. To the best of our knowledge, this is the first study for the Chinese cyberbullying incident detection task.", "date": "2025-05-27", "categories": []}
{"id": "2605.10394", "title": "Sens-VisualNews: A Benchmark Dataset for Sensational Image Detection", "abstract": "The detection of sensational content in media items can be a critical filtering mechanism for identifying check-worthy content and flagging potential disinformation, since such content triggers physiological arousal that often bypasses critical evaluation and accelerates viral sharing. In this paper we introduce the task of sensational image detection, which aims to determine whether an image contains shocking, provocative, or emotionally charged features to grab attention and trigger strong emotional responses. To support research on this task, we create a new benchmark dataset (called Sens-VisualNews) that contains 9,576 images from news items, annotated based on the (in-)existence of various sensational concepts and events in their visual content. Finally, using Sens-VisualNews, we study the prompt sensitivity, performance and robustness of a wide range of open SotA Multimodal LLMs, across both zero-shot and fine-tuned settings.", "date": "2026-05-11", "categories": []}
{"id": "2605.10563", "title": "ThreatCore: A Benchmark for Explicit and Implicit Threat Detection", "abstract": "Threat detection in Natural Language Processing lacks consistent definitions and standardized benchmarks, and is often conflated with broader phenomena such as toxicity, hate speech, or offensive language. In this work, we introduce ThreatCore, a public available benchmark dataset for fine-grained threat detection that distinguishes between explicit threats, implicit threats, and non-threats. The dataset is constructed by aggregating multiple publicly available resources and systematically re-annotating them under a unified operational definition of threat, revealing substantial inconsistencies across existing labels. To improve the coverage of underrepresented cases, particularly implicit threats, we further augment the dataset with synthetic examples, which are manually validated using the same annotation protocol adopted for the re-annotation of the public datasets, ensuring consistency across all data sources. We evaluate Perspective API, zero-shot classifiers, and recent language models on ThreatCore, showing that implicit threats remain substantially harder to detect than explicit ones. Our results also indicate that incorporating Semantic Role Labeling as an intermediate representation can improve performance by making the structure of harmful intent more explicit. Overall, ThreatCore provides a more consistent benchmark for studying fine-grained threat detection and highlights the challenges that current models still face in identifying indirect expressions of harmful intent.", "date": "2026-05-11", "categories": []}
{"id": "2605.07201", "title": "PSK@EEUCA 2026: Fine-Tuning Large Language Models with Synthetic Data Augmentation for Multi-Class Toxicity Detection in Gaming Chat", "abstract": "This paper describes our system for the EEUCA 2026 Shared Task on Understanding Toxic Behavior in Gaming Communities. The task involves classifying World of Tanks chat messages into six toxicity categories: Non-toxic, Insults/Flaming, Other Offensive, Hate/Harassment, Threats, and Extremism. We explore multiple approaches including encoder-based models, instruction-tuned LLMs with LoRA fine-tuning, hierarchical classification, one-vs-rest strategies, and various ensemble methods. Our best system combines Llama 3.1 8B with carefully calibrated 5\\% synthetic data augmentation, achieving an F1-macro score of 0.6234 on the test set, placing 4th out of 35 participating teams. We provide extensive analysis of the dataset's annotation patterns and their impact on model generalization, revealing a critical ''validation trap'' phenomenon where high validation performance correlates with poor test transfer.", "date": "2026-05-08", "categories": []}
{"id": "2605.07760", "title": "RuleSafe-VL: Evaluating Rule-Conditioned Decision Reasoning in Vision-Language Content Moderation", "abstract": "Platform content moderation applies explicit policy rules and context-dependent conditions to decide whether user content is allowed, restricted, or removed. A correct moderation outcome must therefore depend on which rules a case activates, how those rules interact, and whether the available evidence is sufficient. Current multimodal safety benchmarks largely reduce moderation to matching predefined final labels, leaving this underlying rule structure untested. As a result, a high benchmark score reveals little about whether a model applies the policy correctly or arrives at the correct label through superficial cues. To evaluate this rule-governed process, we introduce RuleSafe-VL, a benchmark for rule-conditioned decision reasoning in vision-language content moderation. Derived from publicly available platform moderation policies, RuleSafe-VL formalizes 93 atomic rules and 92 typed rule relations, yielding 2,166 context-sensitive image-text cases across three high-risk policy families. Its four diagnostic tasks decompose moderation into a rule-conditioned decision chain. They identify activated rules, recover rule interactions, judge decision sufficiency, and resolve outcomes once missing context is supplied. Experiments on 10 frontier, open-source, and safety-oriented VLMs reveal rule-relation recovery as the dominant bottleneck, where the best model reaches only 64.8 Macro-F1 and some safety-oriented models fall below 7 Macro-F1. Decision-state prediction also remains unreliable, peaking at 64.5 Macro-F1. RuleSafe-VL shifts moderation evaluation from final-label scoring toward diagnostic assessment of rule-conditioned decision reasoning.", "date": "2026-05-08", "categories": []}
{"id": "2605.04885", "title": "A Comparative Study of PyCaret AutoML and CNN-BiLSTM for Binary Hate Speech Detection in Indonesian Twitter", "abstract": "This paper compares a PyCaret AutoML branch and a CNN-BiLSTM branch for binary hate speech detection on Indonesian Twitter using the HS label from the corpus of Ibrohim and Budi. Both branches share the same preprocessing pipeline so that the comparison reflects modelling differences rather than inconsistent data preparation. The conventional branch uses TF-IDF with a lexicon-based abusive-word count, whereas the neural branch learns dense token representations and captures both local phrase patterns and bidirectional context. The benchmark is built from the released 13,130-row annotation table, whose HS label yields a 58:42 class ratio. On the held-out split, CNN-BiLSTM achieves the best result with 83.8% accuracy, 79.8% precision, 82.7% recall, and 81.2% F1-score. Within the PyCaret branch, Random Forest is the strongest conventional model with 77.2% accuracy and 77.0% F1-score. The neural branch therefore improves accuracy by 6.6 points and F1-score by 4.2 points. Exploratory corpus analysis, learning curves, and confusion matrices show that the dataset is short-text, moderately imbalanced, and still difficult because many decisions depend on local lexical cues plus short contextual composition. The study concludes that PyCaret AutoML is an effective conventional benchmarking framework, whereas CNN-BiLSTM is the stronger end model for the reported benchmark setting.", "date": "2026-05-06", "categories": []}
{"id": "2605.02457", "title": "Leveraging Argument Structure to Predict Content Hatefulness", "abstract": "Information disorder is a challenging phenomenon that affects society at large. This phenomenon entails the diffusion of misleading, misinforming, and hateful content online. In different contexts, one aspect of the problem may prevail, but overall, this is a broad problem that requires comprehensive solutions. While each dimension of the problem (hate speech, disinformation, misinformation, etc.) requires in-depth analysis, in this paper, we look into the possibility of argument structure to provide relevant information to link these different areas of the problem. In particular, we focus on the WSF-ARG+ dataset, which consists of white supremacy forum messages annotated in terms of argument structure (premises and conclusion). There, we leverage the checkworthiness and hatefulness annotations of the argument components to obtain insights into the hatefulness of the whole message. Our results show promising insights (up to 96% F1), indicating the possibility of extending this direction in the future to tackle hateful content identification and information disorder countering.", "date": "2026-05-04", "categories": []}
{"id": "2605.02940", "title": "PrismAgent: Illuminating Harm in Memes via a Zero-Shot Interpretable Multi-Agent Framework", "abstract": "The rapid spread of memes makes harmful content detection increasingly crucial, as effective identification can curb the circulation of misinformation. However, existing methods rely heavily on high-volume annotated data, which leads to substantial training costs and limited generalization. To address these challenges, we propose PrismAgent, a zero-shot, multi-agent, interpretable framework. PrismAgent conceptualizes this task as a criminal case investigation, employing four specialized agents responsible for the analysis, investigation, prosecution, and judgment stages within a structured collaborative workflow. In the first stage, the analyst agent paraphrases each meme under benevolent and malicious assumptions to probe its underlying intent. The investigator agent then retrieves supporting evidence from an unannotated dataset and constructs contextual interpretations for the meme and its variants. Next, the prosecutor agent performs three independent preliminary judgments by pairing the original meme with each of the three interpretations. Finally, the judge agent deliberates across all evidence to render a final verdict. Moreover, PrismAgent's explicit multi-stage reasoning chain makes the model inherently interpretable, as every intermediate step is explicitly explained rather than only producing a final detection result. Extensive experiments on three public datasets show that PrismAgent significantly outperforms existing zero-shot detection methods.", "date": "2026-05-01", "categories": []}
{"id": "2412.15251", "title": "IPS: In-Prompt Process Supervision for Short Video Content Moderation", "abstract": "Multimodal large language models (MLLMs) are effective at capturing the semantics of short video content; however, they often fail to attend to the policy-specific details required for reliable content moderation. To address this limitation, we introduce IPS, a novel framework that integrates In-prompt Process Supervision into MLLMs by introducing sequential reasoning over ancillary questions during fine-tuning. IPS consistently outperforms baseline MLLMs on public and proprietary benchmarks. Moreover, replacing human-annotated ancillary labels with MLLM-generated ones results in only marginal performance degradation, demonstrating robustness to noisy supervision and strong scalability with model-generated annotations. These findings establish IPS as a scalable and effective solution for complex multimodal classification in large-scale industrial settings.", "date": "2024-12-15", "categories": []}
{"id": "2509.00673", "title": "Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection", "abstract": "We investigate the efficacy of Large Language Models (LLMs) in detecting implicit and explicit hate speech, examining how models with minimal safety alignment (uncensored) compare with more heavily aligned (censored) counterparts in a deployed-model setting when deployed using political personas. While uncensored models are often framed as offering a less constrained perspective, our results reveal a trade-off: censored models outperform their uncensored counterparts in both accuracy and robustness, achieving 69.0\\% versus 64.1\\% strict accuracy. However, this higher performance is also associated with greater resistance to persona-based influence, while uncensored models are more malleable to ideological framing. Furthermore, we identify critical failures across all models in understanding nuanced language such as irony. We also find alarming fairness disparities in performance across different targeted groups and systemic overconfidence that renders self-reported certainty unreliable. These findings challenge the notion of LLMs as objective arbiters and highlight the need for more sophisticated auditing frameworks that account for fairness, calibration, and ideological consistency. Taken together, these results point to censorship-as-deployed rather than safety alignment in isolation as the more appropriate frame for interpreting model differences.", "date": "2025-08-31", "categories": []}
{"id": "2605.00874", "title": "Latent Space Probing for Adult Content Detection in Video Generative Models", "abstract": "The rapid proliferation of AI-powered video generation systems has introduced significant challenges in content moderation, particularly with respect to adult and sexually explicit material. Existing detection methods operate on either prompts or decoded pixel-space outputs. Therefore, both approaches are blind to the rich internal representations formed during generation. In this paper, we propose a novel latent space probing framework that intercepts the denoised latent representations produced by the CogVideoX video diffusion model during inference and attaches lightweight classifiers to perform real-time adult content detection. To support this work, we construct a large-scale binary dataset of 11039 ten-second video clips (5086 violating, 5953 non-violating) sourced from adult websites and YouTube respectively. We introduce two lightweight probing classifier architectures. We train and evaluate it on the dataset. Our work demonstrates that latent-space signals encode strong discriminative features for harmful content detection, achieving 97.29% F1 on our held-out test set with an overhead in the 4-6ms range. Our results suggest that probing the latent space results in improvements in both detection performance as well as cost.", "date": "2026-04-25", "categories": []}
{"id": "2605.01416", "title": "Who Decides What Is Harmful? Content Moderation Policy Through A Multi-Agent Personalised Inference Framework", "abstract": "The increasing scale and complexity of online platforms raises critical policy questions around harmful content, digital well-being, and user autonomy. Traditional content moderation systems rely on centralised, top-down rules, often failing to accommodate the subjective nature of harm perception. This paper proposes an LLM-based multi-agent personalised inference framework that filters content based on unique sensitivity profiles of individual users. Our architecture combines domain-specific Expert Agents, a Manager Agent for orchestrating content analysis and agent selection, and a Ghost Profile Agent for simulating user perspectives, to inform moderation decisions. Evaluated against a range of non-personalised baselines, the system demonstrates up to a 32% improvement in accuracy, showing increased alignment with individual user sensitivities. Beyond technical performance, our framework provides policy-relevant insights for platform governance, providing a scalable way to reconcile moderation policies with societal and individual digital rights", "date": "2026-05-02", "categories": []}
{"id": "2604.26229", "title": "Comparative Analysis of AutoML and BiLSTM Models for Cyberbullying Detection on Indonesian Instagram Comments", "abstract": "This study compares machine learning and deep learning approaches for cyberbullying detection in Indonesian-language Instagram comments. Using a balanced dataset of 650 comments labeled as Bullying and Non-Bullying, the study evaluates Naive Bayes, Logistic Regression, and Support Vector Machine with TF-IDF features, as well as BiLSTM and BiLSTM with Bahdanau Attention. A preprocessing pipeline tailored to informal Indonesian text is applied, including slang normalization, stopword removal, and stemming. The results show that Logistic Regression performs best among the machine learning models, while BiLSTM with Attention achieves the strongest overall deep learning performance. The findings highlight the value of domain-specific preprocessing and show that although deep learning captures contextual patterns more effectively, machine learning remains a competitive option for resource-constrained deployments.", "date": "2026-04-29", "categories": []}
{"id": "2604.23214", "title": "DARC-CLIP: Dynamic Adaptive Refinement with Cross-Attention for Meme Understanding", "abstract": "Memes convey meaning through the interaction of visual and textual signals, often combining humor, irony, and offense in subtle ways. Detecting harmful or sensitive content in memes requires accurate modeling of these multimodal cues. Existing CLIP-based approaches rely on static fusion, which struggles to capture fine grained dependencies between modalities. We propose DARC-CLIP, a CLIP-based framework for adaptive multimodal fusion with a hierarchical refinement stack. DARC-CLIP introduces Adaptive Cross-Attention Refiners to for bidirectional information alignment and Dynamic Feature Adapters for task-sensitive signal adaptation. We evaluate DARC-CLIP on the PrideMM benchmark, which includes hate, target, stance, and humor classification, and further test generalization on the CrisisHateMM dataset. DARC-CLIP achieves highly competitive classification accuracy across tasks, with significant gains of +4.18 AUROC and +6.84 F1 in hate detection over the strongest baseline. Ablation studies confirm that ACAR and DFA are the main contributors to these gains. These results show that adaptive cross-signal refinement is an effective strategy for multimodal content analysis in socially sensitive classification.", "date": "2026-04-25", "categories": []}
{"id": "2604.24179", "title": "MemeScouts@LT-EDI 2026: Asking the Right Questions -- Prompted Weak Supervision for Meme Hate Speech Detection", "abstract": "Detecting hate speech in memes is challenging due to their multimodal nature and subtle, culturally grounded cues such as sarcasm and context. While recent vision-language models (VLMs) enable joint reasoning over text and images, end-to-end prompting can be brittle, as a single prediction must resolve target, stance, implicitness, and irony. These challenges are amplified in multilingual settings. We propose a prompted weak supervision (PWS) approach that decomposes meme understanding into targeted, question-based labeling functions with constrained answer options for homophobia and transphobia detection in the LT-EDI 2026 shared task. Using a quantized Qwen3-VLM to extract features by answering targeted questions, our method outperforms direct VLM classification, with substantial gains for Chinese and Hindi, ranking 1st in English, 2nd in Chinese, and 3rd in Hindi. Iterative refinement via error-driven LF expansion and feature pruning reduces redundancy and improves generalization. Our results highlight the effectiveness of prompted weak supervision for multilingual multimodal hate speech detection.", "date": "2026-04-27", "categories": []}
{"id": "2604.25580", "title": "Bye Bye Perspective API: Lessons for Measurement Infrastructure in NLP, CSS and LLM Evaluation", "abstract": "The closure of Perspective API at the end of 2026 discards what has functioned as the de facto standard for automated toxicity measurement in NLP, CSS, and LLM evaluation research. We document the structural dependence that the communities built on this single proprietary tool and discuss how this dependence caused epistemic problems that have affected - and will likely continue to affect - collective research efforts. Perspective's model was periodically updated without versioning or disclosure, its annotation structure reflected a single corporate operationalisation of a contested concept, and its scores were used simultaneously as an evaluation target and an evaluation standard. Its closure leaves behind non-updatable benchmarks, irreproducible results, and ultimately a field at risk of perpetuating these issues by turning to closed-source LLMs. We use Perspective's announced termination as an opportunity to call for an independent, valid, adaptable, and reproducible toxicity and hate speech measurement infrastructure, with the technical and governance requirements outlined in this paper.", "date": "2026-04-28", "categories": []}
{"id": "2510.10971", "title": "RV-HATE: Reinforced Multi-Module Voting for Implicit Hate Speech Detection", "abstract": "Hate speech remains prevalent in human society and continues to evolve in its forms and expressions. Modern advancements in internet and online anonymity accelerate its rapid spread and complicate its detection. However, hate speech datasets exhibit diverse characteristics primarily because they are constructed from different sources and platforms, each reflecting different linguistic styles and social contexts. Despite this diversity, prior studies on hate speech detection often rely on fixed methodologies without adapting to data-specific features. We introduce RV-HATE, a detection framework designed to account for the dataset-specific characteristics of each hate speech dataset. RV-HATE consists of multiple specialized modules, where each module focuses on distinct linguistic or contextual features of hate speech. The framework employs reinforcement learning to optimize weights that determine the contribution of each module for a given dataset. A voting mechanism then aggregates the module outputs to produce the final decision. RV-HATE offers two primary advantages: (1)~it improves detection accuracy by tailoring the detection process to dataset-specific attributes, and (2)~it also provides interpretable insights into the distinctive features of each dataset. Consequently, our approach effectively addresses implicit hate speech and achieves superior performance compared to conventional static methods. Our code is available at https://github.com/leeyejin1231/RV-HATE.", "date": "2025-10-13", "categories": []}
{"id": "2604.20511", "title": "CHASM: Unveiling Covert Advertisements on Chinese Social Media", "abstract": "Current benchmarks for evaluating large language models (LLMs) in social media moderation completely overlook a serious threat: covert advertisements, which disguise themselves as regular posts to deceive and mislead consumers into making purchases, leading to significant ethical and legal concerns. In this paper, we present the CHASM, a first-of-its-kind dataset designed to evaluate the capability of Multimodal Large Language Models (MLLMs) in detecting covert advertisements on social media. CHASM is a high-quality, anonymized, manually curated dataset consisting of 4,992 instances, based on real-world scenarios from the Chinese social media platform Rednote. The dataset was collected and annotated under strict privacy protection and quality control protocols. It includes many product experience sharing posts that closely resemble covert advertisements, making the dataset particularly challenging.The results show that under both zero-shot and in-context learning settings, none of the current MLLMs are sufficiently reliable for detecting covert advertisements.Our further experiments revealed that fine-tuning open-source MLLMs on our dataset yielded noticeable performance gains. However, significant challenges persist, such as detecting subtle cues in comments and differences in visual and textual structures.We provide in-depth error analysis and outline future research directions. We hope our study can serve as a call for the research community and platform moderators to develop more precise defenses against this emerging threat.", "date": "2026-04-22", "categories": []}
{"id": "2603.21298", "title": "More Than Sum of Its Parts: Deciphering Intent Shifts in Multimodal Hate Speech Detection", "abstract": "Combating hate speech on social media is critical for securing cyberspace, yet relies heavily on the efficacy of automated detection systems. As content formats evolve, hate speech is transitioning from solely plain text to complex multimodal expressions, making implicit attacks harder to spot. Current systems, however, often falter on these subtle cases, as they struggle with multimodal content where the emergent meaning transcends the aggregation of individual modalities. To bridge this gap, we move beyond binary classification to characterize semantic intent shifts where modalities interact to construct implicit hate from benign cues or neutralize toxicity through semantic inversion. Guided by this fine-grained formulation, we curate the Hate via Vision-Language Interplay (H-VLI) benchmark where the true intent hinges on the intricate interplay of modalities rather than overt visual or textual slurs. To effectively decipher these complex cues, we further propose the Asymmetric Reasoning via Courtroom Agent DEbate (ARCADE) framework. By simulating a judicial process where agents actively argue for accusation and defense, ARCADE forces the model to scrutinize deep semantic cues before reaching a verdict. Extensive experiments demonstrate that ARCADE significantly outperforms state-of-the-art baselines on H-VLI, particularly for challenging implicit cases, while maintaining competitive performance on established benchmarks. Our code and data are available at: https://github.com/Sayur1n/H-VLI", "date": "2026-03-22", "categories": []}
{"id": "2604.16654", "title": "IYKYK (But AI Doesn't): Automated Content Moderation Does Not Capture Communities' Heterogeneous Attitudes Towards Reclaimed Language", "abstract": "Reclaimed slur usage is a common and meaningful practice online for many marginalized communities. It serves as a source of solidarity, identity, and shared experience. However, contemporary automated and AI-based moderation tools for online content largely fail to distinguish between reclaimed and hateful uses of slurs, resulting in the suppression of marginalized voices. In this work, we use quantitative and qualitative methods to examine the attitudes of social media users in LGBTQIA+, Black, and women communities around reclaimed slurs targeting our focus groups including the f-word, n-word, and b-word. With social media users from these communities, we collect and analyze an annotated online slur usage corpus. The corpus includes annotators' perceptions of whether an online text containing a slur should be flagged as hate speech, as well as contextual features of the slur usage. Across all communities and annotation questions, we observe low inter-annotator agreement, indicating substantial disagreement among in-group annotators. This is compounded by the fact that, absent clear contextual signals of identity and intent, even in-group members may disagree on how to interpret reclaimed slur usage online. Semi-structured interviews with annotators suggest that differences in lived experience and personal history contribute to this variation as well. We find poor alignment between annotator judgments and automated hate speech assessments produced by Perspective API. We further observe that certain features of a text such as whether the slur usage was derogatory and if the slur was targeted at oneself are more associated with whether annotators report the text as hate speech. Together, these findings highlight the inherent subjectivity and contextual nature of how marginalized communities interpret slurs online.", "date": "2026-04-17", "categories": []}
{"id": "2508.11281", "title": "ToxiFrench: Benchmarking and Enhancing Language Models via CoT Fine-Tuning for French Toxicity Detection", "abstract": "Detecting toxic content using language models is crucial yet challenging. While substantial progress has been made in English, toxicity detection in French remains underdeveloped, primarily due to the lack of culturally relevant, human-annotated, large-scale datasets. In this work, we release ToxiFrench, a dataset of 53,622 French online comments together with a balanced benchmark split for systematic evaluation. The dataset is constructed via a semi-automated annotation pipeline that reduces manual labeling to only 10% through high-confidence LLM-based pre-annotation and human verification, while ensuring statistical alignment with human-only annotation. We then benchmark a broad range of models and uncover a counterintuitive finding: Small Language Models (SLMs) often surpass larger models in robustness and generalization on this task. Motivated by this finding, we propose a novel Chain-of-Thought (CoT) fine-tuning strategy using a Dynamic Weighted Loss (DWL) that progressively emphasizes the model's final decision and significantly improves faithfulness. Our fine-tuned 4B model (Qwen3-4B) achieves state-of-the-art performance on the benchmark. It improves its balanced accuracy by 10% over its baseline and achieves better performance than GPT-4o and DeepSeek-R1 on our benchmark, while successfully retaining cross-lingual capabilities.", "date": "2025-08-15", "categories": []}
{"id": "2602.07954", "title": "Bielik Guard: Efficient Polish Language Safety Classifiers for LLM Content Moderation", "abstract": "As Large Language Models (LLMs) become increasingly deployed in Polish language applications, the need for efficient and accurate content safety classifiers has become paramount. We present Bielik Guard, a family of compact Polish language safety classifiers comprising two model variants: a 0.1B parameter model based on MMLW-RoBERTa-base and a 0.5B parameter model based on PKOBP/polish-roberta-8k. Fine-tuned on a community-annotated dataset of 6,885 Polish texts, these models classify content across five safety categories: Hate/Aggression, Vulgarities, Sexual Content, Crime, and Self-Harm. Our evaluation demonstrates that both models achieve strong performance on multiple benchmarks. The 0.5B variant offers the best overall discrimination capability with F1 scores of 0.791 (micro) and 0.785 (macro) on the test set, while the 0.1B variant demonstrates exceptional efficiency. Notably, Bielik Guard 0.1B v1.1 achieves superior precision (77.65%) and very low false positive rate (0.63%) on real user prompts, outperforming HerBERT-PL-Guard (31.55% precision, 4.70% FPR) despite identical model size. The models are publicly available and designed to provide appropriate responses rather than simple content blocking, particularly for sensitive categories like self-harm.", "date": "2026-02-08", "categories": []}
{"id": "2601.04567", "title": "All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction", "abstract": "Harmful memes are ever-shifting in the Internet communities, which are difficult to analyze due to their type-shifting and temporal-evolving nature. Although these memes are shifting, we find that different memes may share invariant principles, i.e., the underlying design concept of malicious users, which can help us analyze why these memes are harmful. In this paper, we propose RepMD, an ever-shifting harmful meme detection method based on the design concept reproduction. We first refer to the attack tree to define the Design Concept Graph (DCG), which describes steps that people may take to design a harmful meme. Then, we derive the DCG from historical memes with design step reproduction and graph pruning. Finally, we use DCG to guide the Multimodal Large Language Model (MLLM) to detect harmful memes. The evaluation results show that RepMD achieves the highest accuracy with 81.1% and has slight accuracy decreases when generalized to type-shifting and temporal-evolving memes. Human evaluation shows that RepMD can improve the efficiency of human discovery on harmful memes, with 15$\\sim$30 seconds per meme.", "date": "2026-01-08", "categories": []}
{"id": "2604.14218", "title": "MEME-Fusion@CHiPSAL 2026: Multimodal Ablation Study of Hate Detection and Sentiment Analysis on Nepali Memes", "abstract": "Hate speech detection in Devanagari-scripted social media memes presents compounded challenges: multimodal content structure, script-specific linguistic complexity, and extreme data scarcity in low-resource settings. This paper presents our system for the CHiPSAL 2026 shared task, addressing both Subtask A (binary hate speech detection) and Subtask B (three-class sentiment classification: positive, neutral, negative). We propose a hybrid cross-modal attention fusion architecture that combines CLIP (ViT-B/32) for visual encoding with BGE-M3 for multilingual text representation, connected through 4-head self-attention and a learnable gating network that dynamically weights modality contributions on a per-sample basis. Systematic evaluation across eight model configurations demonstrates that explicit cross-modal reasoning achieves a 5.9% F1-macro improvement over text-only baselines on Subtask A, while uncovering two unexpected but critical findings: English-centric vision models exhibit near-random performance on Devanagari script, and standard ensemble methods catastrophically degrade under data scarcity (N nearly equal to 850 per fold) due to correlated overfitting. The code can be accessed at https://github.com/Tri-Yantra-Technologies/MEME-Fusion/", "date": "2026-04-13", "categories": []}
{"id": "2604.14907", "title": "Comparison of Modern Multilingual Text Embedding Techniques for Hate Speech Detection Task", "abstract": "Online hate speech and abusive language pose a growing challenge for content moderation, especially in multilingual settings and for low-resource languages such as Lithuanian. This paper investigates to what extent modern multilingual sentence embedding models can support accurate hate speech detection in Lithuanian, Russian, and English, and how their performance depends on downstream modeling choices and feature dimensionality. We introduce LtHate, a new Lithuanian hate speech corpus derived from news portals and social networks, and benchmark six modern multilingual encoders (potion, gemma, bge, snow, jina, e5) on LtHate, RuToxic, and EnSuperset using a unified Python pipeline. For each embedding, we train both a one class HBOS anomaly detector and a two class CatBoost classifier, with and without principal component analysis (PCA) compression to 64-dimensional feature vectors. Across all datasets, two class supervised models consistently and substantially outperform one class anomaly detection, with the best configurations achieving up to 80.96% accuracy and AUC ROC of 0.887 in Lithuanian (jina), 92.19% accuracy and AUC ROC of 0.978 in Russian (e5), and 77.21% accuracy and AUC ROC of 0.859 in English (e5 with PCA). PCA compression preserves almost all discriminative power in the supervised setting, while showing some negative impact for the unsupervised anomaly detection case. These results demonstrate how modern multilingual sentence embeddings combined with gradient boosted decision trees provide robust soft-computing solutions for multilingual hate speech detection applications.", "date": "2026-04-16", "categories": []}
{"id": "2604.14970", "title": "Explain the Flag: Contextualizing Hate Speech Beyond Censorship", "abstract": "Hate, derogatory, and offensive speech remains a persistent challenge in online platforms and public discourse. While automated detection systems are widely used, most focus on censorship or removal, raising concerns for transparency and freedom of expression, and limiting opportunities to explain why content is harmful. To address these issues, explanatory approaches have emerged as a promising solution, aiming to make hate speech detection more transparent, accountable, and informative. In this paper, we present a hybrid approach that combines Large Language Models (LLMs) with three newly created and curated vocabularies to detect and explain hate speech in English, French, and Greek. Our system captures both inherently derogatory expressions tied to identity characteristics and direct group-targeted content through two complementary pipelines: one that detects and disambiguates problematic terms using the curated vocabularies, and one that leverages LLMs as context-aware evaluators of group-targeting content. The outputs are fused into grounded explanations that clarify why content is flagged. Human evaluation shows that our hybrid approach is accurate, with high-quality explanations, outperforming LLM-only baselines.", "date": "2026-04-16", "categories": []}
{"id": "2604.12289", "title": "The Enforcement and Feasibility of Hate Speech Moderation on Twitter", "abstract": "Online hate speech is associated with substantial social harms, yet it remains unclear how consistently platforms enforce hate speech policies or whether enforcement is feasible at scale. We address these questions through a global audit of hate speech moderation on Twitter (now X). Using a complete 24-hour snapshot of public tweets, we construct representative samples comprising 540,000 tweets annotated for hate speech by trained annotators across eight major languages. Five months after posting, 80% of hateful tweets remain online, including explicitly violent hate speech. Such tweets are no more likely to be removed than non-hateful tweets, with neither severity nor visibility increasing the likelihood of removal. We then examine whether these enforcement gaps reflect technical limits of large-scale moderation systems. While fully automated detection systems cannot reliably identify hate speech without generating large numbers of false positives, they effectively prioritize likely violations for human review. Simulations of a human-AI moderation pipeline indicate that substantially reducing user exposure to hate speech is economically feasible at a cost below existing regulatory penalties. These results suggest that the persistence of online hate cannot be explained by technical constraints alone but also reflects institutional choices in the allocation of moderation resources.", "date": "2026-04-14", "categories": []}
{"id": "2604.12321", "title": "ToxiTrace: Gradient-Aligned Training for Explainable Chinese Toxicity Detection", "abstract": "Existing Chinese toxic content detection methods mainly target sentence-level classification but often fail to provide readable and contiguous toxic evidence spans. We propose \\textbf{ToxiTrace}, an explainability-oriented method for BERT-style encoders with three components: (1) \\textbf{CuSA}, which refines encoder-derived saliency cues into fine-grained toxic spans with lightweight LLM guidance; (2) \\textbf{GCLoss}, a gradient-constrained objective that concentrates token-level saliency on toxic evidence while suppressing irrelevant activations; and (3) \\textbf{ARCL}, which constructs sample-specific contrastive reasoning pairs to sharpen the semantic boundary between toxic and non-toxic content. Experiments show that ToxiTrace improves classification accuracy and toxic span extraction while preserving efficient encoder-based inference and producing more coherent, human-readable explanations. We have released the model at https://huggingface.co/ArdLi/ToxiTrace.", "date": "2026-04-14", "categories": []}
{"id": "2604.09625", "title": "Toward Generalized Cross-Lingual Hateful Language Detection with Web-Scale Data and Ensemble LLM Annotations", "abstract": "We study whether large-scale unlabelled web data and LLM-based synthetic annotations can improve multilingual hate speech detection. Starting from texts crawled via OpenWebSearch.eu~(OWS) in four languages (English, German, Spanish, Vietnamese), we pursue two complementary strategies. First, we apply continued pre-training to BERT models by continuing masked language modelling on unlabelled OWS texts before supervised fine-tuning, and show that this yields an average macro-F1 gain of approximately 3% over standard baselines across sixteen benchmarks, with stronger gains in low-resource settings. Second, we use four open-source LLMs (Mistral-7B, Llama3.1-8B, Gemma2-9B, Qwen2.5-14B) to produce synthetic annotations through three ensemble strategies: mean averaging, majority voting, and a LightGBM meta-learner. The LightGBM ensemble consistently outperforms the other strategies. Fine-tuning on these synthetic labels substantially benefits a small model (Llama3.2-1B: +11% pooled F1), but provides only a modest gain for the larger Qwen2.5-14B (+0.6%). Our results indicate that the combination of web-scale unlabelled data and LLM-ensemble annotations is the most valuable for smaller models and low-resource languages.", "date": "2026-03-18", "categories": []}
{"id": "2604.10175", "title": "\"bot lane noob\" Towards Deployment of NLP-based Toxicity Detectors in Video Games", "abstract": "Toxicity and harassment are widespread in the video-gaming context. Especially in competitive online multiplayer scenarios, gamers oftentimes send harmful messages to other players (teammates or opponents) whose consequences span from mild annoyance to withdrawal and depression. Abundant prior work tackled these problems, e.g., pointing out the negative effects of toxic interactions. However, few works proposed countermeasures specifically developed and tested on textual messages sent during a match -- i.e., when the \"harassment\" actually occurs. We posit that such a scarcity stems from the lack of high-quality datasets that can be used to devise \"automated\" detectors based on natural-language processing (NLP) and machine learning (ML), and which can -- ideally -- mitigate the harm of toxic comments during a gaming session. This work provides a foundation for addressing the problem of toxicity and harassment in video games. First, through a systematic literature review (n=1,039), we provide evidence that only few works proposed ML/NLP-based detectors of toxicity/harassment during live matches. Then, we partner-up with 8 expert League of Legend (LoL) players and create a fine-grained labelled dataset, L2DTnH, containing 1.4k toxic and 13.8k non-toxic messages exchanged during LoL matches. We use L2DTnH to develop a detector that we then empirically show outperforms general-purpose and state-of-the-art toxicity detectors reliant on NLP. To further demonstrate the practicality of our resources, we test our detector on game-related data beyond that included in L2DTnH; and we develop a Web-browser extension that flags toxic content in Webpages -- without querying third-party servers owned by AI companies. We publicly release all of our resources. Our contributions pave the way for more applied research devoted to fighting the spread of toxicity and harassment in video games.", "date": "2026-04-11", "categories": []}
{"id": "2604.10460", "title": "Toward Accountable AI-Generated Content on Social Platforms: Steganographic Attribution and Multimodal Harm Detection", "abstract": "The rapid growth of generative AI has introduced new challenges in content moderation and digital forensics. In particular, benign AI-generated images can be paired with harmful or misleading text, creating difficult-to-detect misuse. This contextual misuse undermines the traditional moderation framework and complicates attribution, as synthetic images typically lack persistent metadata or device signatures. We introduce a steganography enabled attribution framework that embeds cryptographically signed identifiers into images at creation time and uses multimodal harmful content detection as a trigger for attribution verification. Our system evaluates five watermarking methods across spatial, frequency, and wavelet domains. It also integrates a CLIP-based fusion model for multimodal harmful-content detection. Experiments demonstrate that spread-spectrum watermarking, especially in the wavelet domain, provides strong robustness under blur distortions, and our multimodal fusion detector achieves an AUC-ROC of 0.99, enabling reliable cross-modal attribution verification. These components form an end-to-end forensic pipeline that enables reliable tracing of harmful deployments of AI-generated imagery, supporting accountability in modern synthetic media environments. Our code is available at GitHub: https://github.com/bli1/steganography", "date": "2026-04-12", "categories": []}
{"id": "2604.10502", "title": "CHAIRO: Contextual Hierarchical Analogical Induction and Reasoning Optimization for LLMs", "abstract": "Content moderation in online platforms faces persistent challenges due to the evolving complexity of user-generated content and the limitations of traditional rule-based and machine learning approaches. While recent advances in large language models (LLMs) have enabled more sophisticated moderation via direct prompting or fine-tuning, these approaches often exhibit limited generalization, interpretability, and adaptability to unseen or ambiguous cases.\n  In this work, we propose a novel moderation framework that leverages analogical examples to enhance rule induction and decision reliability. Our approach integrates end-to-end optimization of analogical retrieval, rule generation, and moderation classification, enabling the dynamic adaptation of moderation rules to diverse content scenarios. Through comprehensive experiments, we demonstrate that our method significantly outperforms both rule-injected fine-tuning baselines and multi-stage static RAG pipelines in terms of moderation accuracy and rule quality. Further evaluations, including human assessments and external model generalization tests, confirm that our framework produces rules with better clarity, interpretability, and applicability. These findings show that analogical example-driven methods can advance robust, explainable, and generalizable content moderation in real-world applications.", "date": "2026-04-12", "categories": []}
{"id": "2604.10504", "title": "CARO: Chain-of-Analogy Reasoning Optimization for Robust Content Moderation", "abstract": "Current large language models (LLMs), even those explicitly trained for reasoning, often struggle with ambiguous content moderation cases due to misleading \"decision shortcuts\" embedded in context. Inspired by cognitive psychology insights into expert moderation, we introduce \\caro (Chain-of-Analogy Reasoning Optimization), a novel two-stage training framework to induce robust analogical reasoning in LLMs. First, \\caro bootstraps analogical reasoning chains via retrieval-augmented generation (RAG) on moderation data and performs supervised fine-tuning (SFT). Second, we propose a customized direct preference optimization (DPO) approach to reinforce analogical reasoning behaviors explicitly. Unlike static retrieval methods, \\caro dynamically generates tailored analogical references during inference, effectively mitigating harmful decision shortcuts. Extensive experiments demonstrate that \\caro substantially outperforms state-of-the-art reasoning models (DeepSeek R1, QwQ), specialized moderation models (LLaMA Guard), and advanced fine-tuning and retrieval-augmented methods, achieving an average F1 score improvement of 24.9\\% on challenging ambiguous moderation benchmarks.", "date": "2026-04-12", "categories": []}
{"id": "2604.08609", "title": "Detection of Hate and Threat in Digital Forensics: A Case-Driven Multimodal Approach", "abstract": "Digital forensic investigations increasingly rely on heterogeneous evidence such as images, scanned documents, and contextual reports. These artifacts may contain explicit or implicit expressions of harm, hate, threat, violence, or intimidation, yet existing automated approaches often assume clean text input or apply vision models without forensic justification. This paper presents a case-driven multimodal approach for hate and threat detection in forensic analysis. The proposed framework explicitly determines the presence and source of textual evidence, distinguishing between embedded text, associated contextual text, and image-only evidence. Based on the identified evidence configuration, the framework selectively applies text analysis, multimodal fusion, or image-only semantic reasoning using vision language models with vision transformer backbones (ViT). By conditioning inference on evidence availability, the approach mirrors forensic decision-making, improves evidentiary traceability, and avoids unjustified modality assumptions. Experimental evaluation on forensic-style image evidence demonstrates consistent and interpretable behavior across heterogeneous evidence scenarios.", "date": "2026-04-08", "categories": []}
{"id": "2604.09094", "title": "Few-Shot Contrastive Adaptation for Audio Abuse Detection in Low-Resource Indic Languages", "abstract": "Abusive speech detection is becoming increasingly important as social media shifts towards voice-based interaction, particularly in multilingual and low-resource settings. Most current systems rely on automatic speech recognition (ASR) followed by text-based hate speech classification, but this pipeline is vulnerable to transcription errors and discards prosodic information carried in speech. We investigate whether Contrastive Language-Audio Pre-training (CLAP) can support abusive speech detection directly from audio. Using the ADIMA dataset, we evaluate CLAP-based representations under few-shot supervised contrastive adaptation in cross-lingual and leave-one-language-out settings, with zero-shot prompting included as an auxiliary analysis. Our results show that CLAP yields strong cross-lingual audio representations across ten Indic languages, and that lightweight projection-only adaptation achieves competitive performance with respect to fully supervised systems trained on complete training data. However, the benefits of few-shot adaptation are language-dependent and not monotonic with shot size. These findings suggest that contrastive audio-text models provide a promising basis for cross-lingual audio abuse detection in low-resource settings, while also indicating that transfer remains incomplete and language-specific in important ways.", "date": "2026-04-10", "categories": []}
{"id": "2604.06205", "title": "Tool-MCoT: Tool Augmented Multimodal Chain-of-Thought for Content Safety Moderation", "abstract": "The growth of online platforms and user content requires strong content moderation systems that can handle complex inputs from various media types. While large language models (LLMs) are effective, their high computational cost and latency present significant challenges for scalable deployment. To address this, we introduce Tool-MCoT, a small language model (SLM) fine-tuned for content safety moderation leveraging external framework. By training our model on tool-augmented chain-of-thought data generated by LLM, we demonstrate that the SLM can learn to effectively utilize these tools to improve its reasoning and decision-making. Our experiments show that the fine-tuned SLM achieves significant performance gains. Furthermore, we show that the model can learn to use these tools selectively, achieving a balance between moderation accuracy and inference efficiency by calling tools only when necessary.", "date": "2026-03-15", "categories": []}
{"id": "2510.17018", "title": "CoGate-LSTM: Prototype-Guided Feature-Space Gating for Mitigating Gradient Dilution in Imbalanced Toxic Comment Classification", "abstract": "Toxic text classification for online moderation remains challenging under extreme class imbalance, where rare but high-risk labels such as threat and severe_toxic are consistently underdetected by conventional models. We propose CoGate-LSTM, a parameter-efficient recurrent architecture built around a novel cosine-similarity feature gating mechanism that adaptively rescales token embeddings by their directional similarity to a learned toxicity prototype. Unlike token-position attention, the gate emphasizes feature directions most informative for minority toxic classes. The model combines frozen multi-source embeddings (GloVe, FastText, and BERT-CLS), a character-level BiLSTM, embedding-space SMOTE, and weighted focal loss. On the Jigsaw Toxic Comment benchmark, CoGate-LSTM achieves 0.881 macro-F1 (95% CI: [0.873, 0.889]) and 96.0% accuracy, outperforming fine-tuned BERT by 6.9 macro-F1 points (p < 0.001) and XGBoost by 4.7, while using only 7.3M parameters (about 15$\\times$ fewer than BERT) and 48 ms CPU inference latency. Gains are strongest on minority labels, with F1 improvements of +71% for severe_toxic, +33% for threat, and +28% for identity_hate relative to fine-tuned BERT. Ablations identify cosine gating as the primary driver of performance (-4.8 macro-F1 when removed), with additional benefits from character-level fusion (-2.4) and multi-head attention (-2.9). CoGate-LSTM also transfers reasonably across datasets, reaching a 0.71 macro-F1 zero-shot on the Contextual Abuse Dataset and 0.73 with lightweight threshold adaptation. These results show that direction-aware feature gating offers an effective and efficient alternative to large, fully fine-tuned transformers for classifying imbalanced toxic comments.", "date": "2025-10-19", "categories": []}
{"id": "2511.10287", "title": "OutSafe-Bench: A Benchmark for Multimodal Offensive Content Detection in Large Language Models", "abstract": "Since Multimodal Large Language Models (MLLMs) are increasingly being integrated into everyday tools and intelligent agents, growing concerns have arisen regarding their possible output of unsafe contents, ranging from toxic language and biased imagery to privacy violations and harmful misinformation. Current safety benchmarks remain highly limited in both modality coverage and performance evaluations, often neglecting the extensive landscape of content safety. In this work, we introduce OutSafe-Bench, the first most comprehensive content safety evaluation test suite designed for the multimodal era. OutSafe-Bench includes a large-scale dataset that spans four modalities, featuring over 18,000 bilingual (Chinese and English) text prompts, 4,500 images, 450 audio clips and 450 videos, all systematically annotated across nine critical content risk categories. In addition to the dataset, we introduce a Multidimensional Cross Risk Score (MCRS), a novel metric designed to model and assess overlapping and correlated content risks across different categories. To ensure fair and robust evaluation, we propose FairScore, an explainable automated multi-reviewer weighted aggregation framework. FairScore selects top-performing models as adaptive juries, thereby mitigating biases from single-model judgments and enhancing overall evaluation reliability. Our evaluation of nine state-of-the-art MLLMs reveals persistent and substantial safety vulnerabilities, underscoring the pressing need for robust safeguards in MLLMs.", "date": "2025-11-13", "categories": []}
{"id": "2604.05318", "title": "DIA-HARM: Dialectal Disparities in Harmful Content Detection Across 50 English Dialects", "abstract": "Harmful content detectors-particularly disinformation classifiers-are predominantly developed and evaluated on Standard American English (SAE), leaving their robustness to dialectal variation unexplored. We present DIA-HARM, the first benchmark for evaluating disinformation detection robustness across 50 English dialects spanning U.S., British, African, Caribbean, and Asia-Pacific varieties. Using Multi-VALUE's linguistically grounded transformations, we introduce D3 (Dialectal Disinformation Detection), a corpus of 195K samples derived from established disinformation benchmarks. Our evaluation of 16 detection models reveals systematic vulnerabilities: human-written dialectal content degrades detection by 1.4-3.6% F1, while AI-generated content remains stable. Fine-tuned transformers substantially outperform zero-shot LLMs (96.6% vs. 78.3% best-case F1), with some models exhibiting catastrophic failures exceeding 33% degradation on mixed content. Cross-dialectal transfer analysis across 2,450 dialect pairs shows that multilingual models (mDeBERTa: 97.2% average F1) generalize effectively, while monolingual models like RoBERTa and XLM-RoBERTa fail on dialectal inputs. These findings demonstrate that current disinformation detectors may systematically disadvantage hundreds of millions of non-SAE speakers worldwide. We release the DIA-HARM framework, D3 corpus, and evaluation tools: https://github.com/jsl5710/dia-harm", "date": "2026-04-07", "categories": []}
{"id": "2511.06391", "title": "HatePrototypes: Interpretable and Transferable Representations for Implicit and Explicit Hate Speech Detection", "abstract": "Optimization of offensive content moderation models for different types of hateful messages is typically achieved through continued pre-training or fine-tuning on new hate speech benchmarks. However, existing benchmarks mainly address explicit hate toward protected groups and often overlook implicit or indirect hate, such as demeaning comparisons, calls for exclusion or violence, and subtle discriminatory language that still causes harm. While explicit hate can often be captured through surface features, implicit hate requires deeper, full-model semantic processing. In this work, we question the need for repeated fine-tuning and analyze the role of HatePrototypes, class-level vector representations derived from language models optimized for hate speech detection and safety moderation. We find that these prototypes, built from as few as 50 examples per class, enable cross-task transfer between explicit and implicit hate, with interchangeable prototypes across benchmarks. Moreover, we show that parameter-free early exiting with prototypes is effective for both hate types. We release the code, prototype resources, and evaluation scripts to support future research on efficient and transferable hate speech detection.", "date": "2025-11-09", "categories": []}
{"id": "2604.01354", "title": "Open-Domain Safety Policy Construction", "abstract": "Moderation layers are increasingly a core component of many products built on user- or model-generated content. However, drafting and maintaining domain-specific safety policies remains costly. We present Deep Policy Research (DPR), a minimal agentic system that drafts a full content moderation policy based on only human-written seed domain information. DPR uses a single web search tool and lightweight scaffolding to iteratively propose search queries, distill diverse web sources into policy rules, and organize rules into an indexed document. We evaluate DPR on (1) the OpenAI undesired content benchmark across five domains with two compact reader LLMs and (2) an in-house multimodal advertisement moderation benchmark. DPR consistently outperforms definition-only and in-context learning baselines, and in our end-to-end setting it is competitive with expert-written policy sections in several domains. Moreover, under the same seed specification and evaluation protocol, DPR outperforms a general-purpose deep research system, suggesting that a task-specific, structured research loop can be more effective than generic web research for policy drafting. We release our experiment code at https://github.com/xiaowu0162/deep-policy-research.", "date": "2026-04-01", "categories": []}
{"id": "2604.01745", "title": "Detecting Toxic Language: Ontology and BERT-based Approaches for Bulgarian Text", "abstract": "Toxic content detection in online communication remains a significant challenge, with current solutions often inadvertently blocking valuable information, including medical terms and text related to minority groups. This paper presents a more nu-anced approach to identifying toxicity in Bulgarian text while preserving access to essential information. The research explores two distinct methodologies for detecting toxic content. The developed methodologies have po-tential applications across diverse online platforms and content moderation systems. First, we propose an ontology that models the potentially toxic words in Bulgarian language. Then, we compose a dataset that comprises 4,384 manually anno-tated sentences from Bulgarian online forums across four categories: toxic language, medical terminology, non-toxic lan-guage, and terms related to minority communities. We then train a BERT-based model for toxic language classification, which reaches a 0.89 F1 macro score. The trained model is directly applicable in a real environment and can be integrated as a com-ponent of toxic content detection systems.", "date": "2026-04-02", "categories": []}
{"id": "2401.09244", "title": "Cross-lingual Offensive Language Detection: A Systematic Review of Datasets, Transfer Approaches and Challenges", "abstract": "The growing prevalence and rapid evolution of offensive language in social media amplify the complexities of detection, particularly highlighting the challenges in identifying such content across diverse languages. This survey presents a systematic and comprehensive exploration of Cross-Lingual Transfer Learning (CLTL) techniques in offensive language detection in social media. Our study stands as the first holistic overview to focus exclusively on the cross-lingual scenario in this domain. We analyse 67 relevant papers and categorise these studies across various dimensions, including the characteristics of multilingual datasets used, the cross-lingual resources employed, and the specific CLTL strategies implemented. According to \"what to transfer\", we also summarise three main CLTL transfer approaches: instance, feature, and parameter transfer. Additionally, we shed light on the current challenges and future research opportunities in this field. Furthermore, we have made our survey resources available online, including two comprehensive tables that provide accessible references to the multilingual datasets and CLTL methods used in the reviewed literature.", "date": "2024-01-17", "categories": []}
{"id": "2506.18919", "title": "MemeMind: A Large-Scale Multimodal Dataset with Chain-of-Thought Reasoning for Harmful Meme Detection", "abstract": "As a multimodal medium combining images and text, memes frequently convey implicit harmful content through metaphors and humor, rendering the detection of harmful memes a complex and challenging task. Although recent studies have made progress in detection accuracy and interpretability, large-scale, high-quality datasets for harmful memes remain scarce, and current methods still struggle to capture implicit risks and nuanced semantics. Thus, we construct MemeMind, a large-scale harmful meme dataset. Aligned with the international standards and the context of internet, MemeMind provides detailed Chain-of-Thought (CoT) reasoning annotations to support fine-grained analysis of implicit intentions in memes. Based on this dataset, we further propose MemeGuard, a reasoning-oriented multimodal detection framework that significantly improves both the accuracy of harmful meme detection and the interpretability of model decisions. Extensive experimental results demonstrate that MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, establishing a solid foundation for future research in harmful meme detection. The complete dataset and code will be released upon acceptance.", "date": "2025-06-15", "categories": []}
{"id": "2603.29211", "title": "Xuanwu: Evolving General Multimodal Models into an Industrial-Grade Foundation for Content Ecosystems", "abstract": "In recent years, multimodal large models have continued to improve on general benchmarks. However, in real-world content moderation and adversarial settings, mainstream models still suffer from degraded generalization and catastrophic forgetting because of limited fine-grained visual perception and insufficient modeling of long-tail noise. In this paper, we present Xuanwu VL-2B as a case study of how general multimodal models can be developed into an industrial-grade foundation model for content ecosystems. The model adopts a compact InternViT-300M + MLP + Qwen3 1.7B architecture, balancing fine-grained visual perception, language-semantic alignment, and deployment cost within an approximately 2B-parameter budget. To balance business specialization with the retention of general capabilities, we developed a data iteration and curation mechanism and trained the model through a progressive three-stage pipeline: pre-training, mid-training, and post-training. Ablation studies and offline business evaluations show that Xuanwu VL-2B achieves an average score of 67.90 across seven OpenCompass multimodal metrics (vs. 64.27 for InternVL 3.5 2B), an average recall of 94.38% over seven independent business moderation tasks, and a weighted overall recall of 82.82% on policy-violating text in challenging adversarial OCR scenarios, outperforming Gemini-2.5-Pro (76.72%). These results show that, under a limited parameter budget, Xuanwu VL-2B achieves a practical balance among business alignment, visual perception, general capability retention, and deployment cost.", "date": "2026-03-31", "categories": []}
{"id": "2603.25269", "title": "When Hate Meets Facts: LLMs-in-the-Loop for Check-worthiness Detection in Hate Speech", "abstract": "Hateful content online is often expressed using fact-like, not necessarily correct information, especially in coordinated online harassment campaigns and extremist propaganda. Failing to jointly address hate speech (HS) and misinformation can deepen prejudice, reinforce harmful stereotypes, and expose bystanders to psychological distress, while polluting public debate. Moreover, these messages require more effort from content moderators because they must assess both harmfulness and veracity, i.e., fact-check them. To address this challenge, we release WSF-ARG+, the first dataset which combines hate speech with check-worthiness information. We also introduce a novel LLM-in-the-loop framework to facilitate the annotation of check-worthy claims. We run our framework, testing it with 12 open-weight LLMs of different sizes and architectures. We validate it through extensive human evaluation, and show that our LLM-in-the-loop framework reduces human effort without compromising the annotation quality of the data. Finally, we show that HS messages with check-worthy claims show significantly higher harassment and hate, and that incorporating check-worthiness labels improves LLM-based HS detection up to 0.213 macro-F1 and to 0.154 macro-F1 on average for large models.", "date": "2026-03-26", "categories": []}
{"id": "2603.22985", "title": "Beyond Hate: Differentiating Uncivil and Intolerant Speech in Multimodal Content Moderation", "abstract": "Current multimodal toxicity benchmarks typically use a single binary hatefulness label. This coarse approach conflates two fundamentally different characteristics of expression: tone and content. Drawing on communication science theory, we introduce a fine-grained annotation scheme that distinguishes two separable dimensions: incivility (rude or dismissive tone) and intolerance (content that attacks pluralism and targets groups or identities) and apply it to 2,030 memes from the Hateful Memes dataset. We evaluate different vision-language models under coarse-label training, transfer learning across label schemes and a joint learning approach that combines the coarse hatefulness label with our fine-grained annotations. Our results show that fine-grained annotations complement existing coarse labels and, when used jointly, improve overall model performance. Moreover, models trained with the fine-grained scheme exhibit more balanced moderation-relevant error profiles and are less prone to under-detection of harmful content than models trained on hatefulness labels alone (FNR-FPR, the difference between false negative and false positive rates: 0.74 to 0.42 for LLaVA-1.6-Mistral-7B; 0.54 to 0.28 for Qwen2.5-VL-7B). This work contributes to data-centric approaches in content moderation by improving the reliability and accuracy of moderation systems through enhanced data quality. Overall, combining both coarse and fine-grained labels provides a practical route to more reliable multimodal moderation.", "date": "2026-03-24", "categories": []}
{"id": "2603.19279", "title": "Multilingual Hate Speech Detection and Counterspeech Generation: A Comprehensive Survey and Practical Guide", "abstract": "Combating online hate speech in multilingual settings requires approaches that go beyond English-centric models and capture the cultural and linguistic diversity of global online discourse. This paper presents a comprehensive survey and practical guide to multilingual hate speech detection and counterspeech generation, integrating recent advances in natural language processing. We analyze why monolingual systems often fail in non-English and code-mixed contexts, missing implicit hate and culturally specific expressions. To address these challenges, we outline a structured three-phase framework - task design, data curation, and evaluation - drawing on state-of-the-art datasets, models, and metrics. The survey consolidates progress in multilingual resources and techniques while highlighting persistent obstacles, including data scarcity in low-resource languages, fairness and bias in system development, and the need for multimodal solutions. By bridging technical progress with ethical and cultural considerations, we provide researchers, practitioners, and policymakers with scalable guidelines for building context-aware, inclusive systems. Our roadmap contributes to advancing online safety through fairer, more effective detection and counterspeech generation across diverse linguistic environments.", "date": "2026-03-01", "categories": []}
{"id": "2511.11599", "title": "SynBullying: A Multi LLM Synthetic Conversational Dataset for Cyberbullying Detection", "abstract": "We introduce SynBullying, a synthetic multi-LLM conversational dataset for studying and detecting cyberbullying (CB). SynBullying provides a scalable and ethically safe alternative to human data collection by leveraging large language models (LLMs) to simulate realistic bullying interactions. The dataset offers (i) conversational structure, capturing multi-turn exchanges rather than isolated posts; (ii) context-aware annotations, where harmfulness is assessed within the conversational flow considering context, intent, and discourse dynamics; and (iii) fine-grained labeling, covering various CB categories for detailed linguistic and behavioral analysis. We evaluate SynBullying across five dimensions, including conversational structure, lexical patterns, sentiment/toxicity, role dynamics, harm intensity, and CB-type distribution. We further examine its utility by testing its performance as standalone training data and as an augmentation source for CB classification.", "date": "2025-10-30", "categories": []}
{"id": "2603.18015", "title": "Beyond Accuracy: An Explainability-Driven Analysis of Harmful Content Detection", "abstract": "Although automated harmful content detection systems are frequently used to monitor online platforms, moderators and end users frequently cannot understand the logic underlying their predictions. While recent studies have focused on increasing classification accuracy, little focus has been placed on comprehending why neural models identify content as harmful, especially when it comes to borderline, contextual, and politically sensitive situations. In this work, a neural harmful content detection model trained on the Civil Comments dataset is analyzed explainability-drivenly. Two popular post-hoc explanation methods, Shapley Additive Explanations and Integrated Gradients, are used to analyze the behavior of a RoBERTa-based classifier in both correct predictions and systematic failure cases. Despite strong overall performance, with an area under the curve of 0.93 and an accuracy of 0.94, the analysis reveals limitations that are not observable from aggregate evaluation metrics alone. Integrated Gradients appear to extract more diffuse contextual attributions while Shapley Additive Explanations extract more focused attributions on explicit lexical cues. The consequent divergence in their outputs manifests in both false negatives and false positives. Qualitative case studies reveal recurring failure modes such as indirect toxicity, lexical over-attribution, or political discourse. The results suggest that explainable AI can foster human-in-the-loop moderation by exposing model uncertainty and increasing the interpretable rationale behind automated decisions. Most importantly, this work highlights the role of explainability as a transparency and diagnostic resource for online harmful content detection systems rather than as a performance-enhancing lever.", "date": "2026-02-24", "categories": []}
{"id": "2603.16070", "title": "SEAHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource Languages of Southeast Asia", "abstract": "Hate speech detection relies heavily on linguistic resources, which are primarily available in high-resource languages such as English and Chinese, creating barriers for researchers and platforms developing tools for low-resource languages in Southeast Asia, where diverse socio-linguistic contexts complicate online hate moderation. To address this, we introduce SEAHateCheck, a pioneering dataset tailored to Indonesia, Thailand, the Philippines, and Vietnam, covering Indonesian, Tagalog, Thai, and Vietnamese. Building on HateCheck's functional testing framework and refining SGHateCheck's methods, SEAHateCheck provides culturally relevant test cases, augmented by large language models and validated by local experts for accuracy. Experiments with state-of-the-art and multilingual models revealed limitations in detecting hate speech in specific low-resource languages. In particular, Tagalog test cases showed the lowest model accuracy, likely due to linguistic complexity and limited training data. In contrast, slang-based functional tests proved the hardest, as models struggled with culturally nuanced expressions. The diagnostic insights of SEAHateCheck further exposed model weaknesses in implicit hate detection and models' struggles with counter-speech expression. As the first functional test suite for these Southeast Asian languages, this work equips researchers with a robust benchmark, advancing the development of practical, culturally attuned hate speech detection tools for inclusive online content moderation.", "date": "2026-03-17", "categories": []}
{"id": "2603.16181", "title": "KidsNanny: A Two-Stage Multimodal Content Moderation Pipeline Integrating Visual Classification, Object Detection, OCR, and Contextual Reasoning for Child Safety", "abstract": "We present KidsNanny, a two-stage multimodal content moderation architecture for child safety. Stage 1 combines a vision transformer (ViT) with an object detector for visual screening (11.7 ms); outputs are routed as text not raw pixels to Stage 2, which applies OCR and a text based 7B language model for contextual reasoning (120 ms total pipeline). We evaluate on the UnsafeBench Sexual category (1,054 images) under two regimes: vision-only, isolating Stage 1, and multimodal, evaluating the full Stage 1+2 pipeline. Stage 1 achieves 80.27% accuracy and 85.39% F1 at 11.7 ms; vision-only baselines range from 59.01% to 77.04% accuracy. The full pipeline achieves 81.40% accuracy and 86.16% F1 at 120 ms, compared to ShieldGemma-2 (64.80% accuracy, 1,136 ms) and LlavaGuard (80.36% accuracy, 4,138 ms). To evaluate text-awareness, we filter two subsets: a text+visual subset (257 images) and a text-only subset (44 images where safety depends primarily on embedded text). On text-only images, KidsNanny achieves 100% recall (25/25 positives; small sample) and 75.76% precision; ShieldGemma-2 achieves 84% recall and 60% precision at 1,136 ms. Results suggest that dedicated OCR-based reasoning may offer recall-precision advantages on text-embedded threats at lower latency, though the small text-only subset limits generalizability. By documenting this architecture and evaluation methodology, we aim to contribute to the broader research effort on efficient multimodal content moderation for child safety.", "date": "2026-03-17", "categories": []}
{"id": "2603.13288", "title": "Agent-Based User-Adaptive Filtering for Categorized Harassing Communication", "abstract": "We propose an agent-based framework for personalized filtering of categorized harassing communication in online social networks. Unlike global moderation systems that apply uniform filtering rules, our approach models user-specific tolerance levels and preferences through adaptive filtering agents. These agents learn from user feedback and dynamically adjust filtering thresholds across multiple harassment categories, including offensive, abusive, and hateful content. We implement and evaluate the framework using supervised classification techniques and simulated user interaction data. Experimental results demonstrate that adaptive agents improve filtering precision and user satisfaction compared to static models. The proposed system illustrates how agent-based personalization can enhance content moderation while preserving user autonomy in digital social environments.", "date": "2026-02-28", "categories": []}
{"id": "2603.14843", "title": "ContiGuard: A Framework for Continual Toxicity Detection Against Evolving Evasive Perturbations", "abstract": "Toxicity detection mitigates the dissemination of toxic content (e.g., hateful comments, posts, and messages within online social actions) to safeguard a healthy online social environment. However, malicious users persistently develop evasive perturbations to disguise toxic content and evade detectors. Traditional detectors or methods are static over time and are inadequate in addressing these evolving evasion tactics. Thus, continual learning emerges as a logical approach to dynamically update detection ability against evolving perturbations. Nevertheless, disparities across perturbations hinder the detector's continual learning on perturbed text. More importantly, perturbation-induced noises distort semantics to degrade comprehension and also impair critical feature learning to render detection sensitive to perturbations. These amplify the challenge of continual learning against evolving perturbations. In this work, we present ContiGuard, the first framework tailored for continual learning of the detector on time-evolving perturbed text (termed continual toxicity detection) to enable the detector to continually update capability and maintain sustained resilience against evolving perturbations. Specifically, to boost the comprehension, we present an LLM-powered semantic enriching strategy, where we dynamically incorporate possible meaning and toxicity-related clues excavated by LLM into the perturbed text to improve the comprehension. To mitigate non-critical features and amplify critical ones, we propose a discriminability-driven feature learning strategy, where we strengthen discriminative features while suppressing the less-discriminative ones to shape a robust classification boundary for detection...", "date": "2026-03-16", "categories": []}
{"id": "2603.12920", "title": "HMS-BERT: Hybrid Multi-Task Self-Training for Multilingual and Multi-Label Cyberbullying Detection", "abstract": "Cyberbullying on social media is inherently multilingual and multi-faceted, where abusive behaviors often overlap across multiple categories. Existing methods are commonly limited by monolingual assumptions or single-task formulations, which restrict their effectiveness in realistic multilingual and multi-label scenarios. In this paper, we propose HMS-BERT, a hybrid multi-task self-training framework for multilingual and multi-label cyberbullying detection. Built upon a pretrained multilingual BERT backbone, HMS-BERT integrates contextual representations with handcrafted linguistic features and jointly optimizes a fine-grained multi-label abuse classification task and a three-class main classification task. To address labeled data scarcity in low-resource languages, an iterative self-training strategy with confidence-based pseudo-labeling is introduced to facilitate cross-lingual knowledge transfer. Experiments on four public datasets demonstrate that HMS-BERT achieves strong performance, attaining a macro F1-score of up to 0.9847 on the multi-label task and an accuracy of 0.6775 on the main classification task. Ablation studies further verify the effectiveness of the proposed components.", "date": "2026-03-13", "categories": []}
{"id": "2603.09984", "title": "An Efficient Hybrid Deep Learning Approach for Detecting Online Abusive Language", "abstract": "The digital age has expanded social media and online forums, allowing free expression for nearly 45% of the global population. Yet, it has also fueled online harassment, bullying, and harmful behaviors like hate speech and toxic comments across social networks, messaging apps, and gaming communities. Studies show 65% of parents notice hostile online behavior, and one-third of adolescents in mobile games experience bullying. A substantial volume of abusive content is generated and shared daily, not only on the surface web but also within dark web forums. Creators of abusive comments often employ specific words or coded phrases to evade detection and conceal their intentions. To address these challenges, we propose a hybrid deep learning model that integrates BERT, CNN, and LSTM architectures with a ReLU activation function to detect abusive language across multiple online platforms, including YouTube comments, online forum discussions, and dark web posts. The model demonstrates strong performance on a diverse and imbalanced dataset containing 77,620 abusive and 272,214 non-abusive text samples (ratio 1:3.5), achieving approximately 99% across evaluation metrics such as Precision, Recall, Accuracy, F1-score, and AUC. This approach effectively captures semantic, contextual, and sequential patterns in text, enabling robust detection of abusive content even in highly skewed datasets, as encountered in real-world scenarios.", "date": "2026-02-12", "categories": []}
{"id": "2603.04414", "title": "Multiclass Hate Speech Detection with RoBERTa-OTA: Integrating Transformer Attention and Graph Convolutional Networks", "abstract": "Multiclass hate speech detection across demographic categories remains computationally challenging due to implicit targeting strategies and linguistic variability in social media content. Existing approaches rely solely on learned representations from training data, without explicitly incorporating structured ontological frameworks that can enhance classification through formal domain knowledge integration. We propose RoBERTa-OTA, which introduces ontology-guided attention mechanisms that process textual features alongside structured knowledge representations through enhanced Graph Convolutional Networks. The architecture combines RoBERTa embeddings with scaled attention layers and graph neural networks to integrate contextual language understanding with domain-specific semantic knowledge. Evaluation across 39,747 balanced samples using 5-fold cross-validation demonstrates significant performance gains over baseline RoBERTa implementations and existing state-of-the-art methods. RoBERTa-OTA achieves 96.04\\% accuracy compared to 95.02\\% for standard RoBERTa, with substantial improvements for challenging categories: gender-based hate speech detection improves by 2.36 percentage points while other hate speech categories improve by 2.38 percentage points. The enhanced architecture maintains computational efficiency with only 0.33\\% parameter overhead, providing practical advantages for large-scale content moderation applications requiring fine-grained demographic hate speech classification.", "date": "2026-02-03", "categories": []}
{"id": "2603.04698", "title": "Hate Speech Detection using Large Language Models with Data Augmentation and Feature Enhancement", "abstract": "This paper evaluates data augmentation and feature enhancement techniques for hate speech detection, comparing traditional classifiers, e.g., Delta Term Frequency-Inverse Document Frequency (Delta TF-IDF), with transformer-based models (DistilBERT, RoBERTa, DeBERTa, Gemma-7B, gpt-oss-20b) across diverse datasets. It examines the impact of Synthetic Minority Over-sampling Technique (SMOTE), weighted loss determined by inverse class proportions, Part-of-Speech (POS) tagging, and text data augmentation on model performance. The open-source gpt-oss-20b consistently achieves the highest results. On the other hand, Delta TF-IDF responds strongly to data augmentation, reaching 98.2% accuracy on the Stormfront dataset. The study confirms that implicit hate speech is more difficult to detect than explicit hateful content and that enhancement effectiveness depends on dataset, model, and technique interaction. Our research informs the development of hate speech detection by highlighting how dataset properties, model architectures, and enhancement strategies interact, supporting more accurate and context-aware automated detection.", "date": "2026-03-05", "categories": []}
{"id": "2603.05057", "title": "MUTEX: Leveraging Multilingual Transformers and Conditional Random Fields for Enhanced Urdu Toxic Span Detection", "abstract": "Urdu toxic span detection remains limited because most existing systems rely on sentence-level classification and fail to identify the specific toxic spans within those text. It is further exacerbated by the multiple factors i.e. lack of token-level annotated resources, linguistic complexity of Urdu, frequent code-switching, informal expressions, and rich morphological variations. In this research, we propose MUTEX: a multilingual transformer combined with conditional random fields (CRF) for Urdu toxic span detection framework that uses manually annotated token-level toxic span dataset to improve performance and interpretability. MUTEX uses XLM RoBERTa with CRF layer to perform sequence labeling and is tested on multi-domain data extracted from social media, online news, and YouTube reviews using token-level F1 to evaluate fine-grained span detection. The results indicate that MUTEX achieves 60% token-level F1 score that is the first supervised baseline for Urdu toxic span detection. Further examination reveals that transformer-based models are more effective at implicitly capturing the contextual toxicity and are able to address the issues of code-switching and morphological variation than other models.", "date": "2026-03-05", "categories": []}
{"id": "2603.03315", "title": "M-QUEST -- Meme Question-Understanding Evaluation on Semantics and Toxicity", "abstract": "Internet memes are a powerful form of online communication, yet their nature and reliance on commonsense knowledge make toxicity detection challenging. Identifying key features for meme interpretation and understanding, is a crucial task. Previous work has been focused on some elements contributing to the meaning, such as the Textual dimension via OCR, the Visual dimension via object recognition, upper layers of meaning like the Emotional dimension, Toxicity detection via proxy variables, such as hate speech detection, and sentiment analysis. Nevertheless, there is still a lack of an overall architecture able to formally identify elements contributing to the meaning of a meme, and be used in the sense-making process. In this work, we present a semantic framework and a corresponding benchmark for automatic knowledge extraction from memes. First, we identify the necessary dimensions to understand and interpret a meme: Textual material, Visual material, Scene, Background Knowledge, Emotion, Semiotic Projection, Analogical Mapping, Overall Intent, Target Community, and Toxicity Assessment. Second, the framework guides a semi-automatic process of generating a benchmark with commonsense question-answer pairs about meme toxicity assessment and its underlying reason. The resulting benchmark M-QUEST consists of 609 question-answer pairs for 307 memes. Thirdly, we evaluate eight open-source large language models on their ability to correctly solve M-QUEST. Our results show that current models' commonsense reasoning capabilities for toxic meme interpretation vary depending on the dimension and architecture. Models with instruction tuning and reasoning capabilities significantly outperform the others, though pragmatic inference questions remain challenging. We release code, benchmark, and prompts to support future research intersecting multimodal content safety and commonsense reasoning.", "date": "2026-02-09", "categories": []}
{"id": "2309.11896", "title": "Focal Inferential Infusion Coupled with Tractable Density Discrimination for Implicit Hate Detection", "abstract": "Although pretrained large language models (PLMs) have achieved state-of-the-art on many natural language processing (NLP) tasks, they lack an understanding of subtle expressions of implicit hate speech. Various attempts have been made to enhance the detection of implicit hate by augmenting external context or enforcing label separation via distance-based metrics. Combining these two approaches, we introduce FiADD, a novel Focused Inferential Adaptive Density Discrimination framework. FiADD enhances the PLM finetuning pipeline by bringing the surface form/meaning of an implicit hate speech closer to its implied form while increasing the inter-cluster distance among various labels. We test FiADD on three implicit hate datasets and observe significant improvement in the two-way and three-way hate classification tasks. We further experiment on the generalizability of FiADD on three other tasks, detecting sarcasm, irony, and stance, in which surface and implied forms differ, and observe similar performance improvements. Consequently, we analyze the generated latent space to understand its evolution under FiADD, which corroborates the advantage of employing FiADD for implicit hate speech detection.", "date": "2023-09-21", "categories": []}
{"id": "2603.02684", "title": "HateMirage: An Explainable Multi-Dimensional Dataset for Decoding Faux Hate and Subtle Online Abuse", "abstract": "Subtle and indirect hate speech remains an underexplored challenge in online safety research, particularly when harmful intent is embedded within misleading or manipulative narratives. Existing hate speech datasets primarily capture overt toxicity, underrepresenting the nuanced ways misinformation can incite or normalize hate. To address this gap, we present HateMirage, a novel dataset of Faux Hate comments designed to advance reasoning and explainability research on hate emerging from fake or distorted narratives. The dataset was constructed by identifying widely debunked misinformation claims from fact-checking sources and tracing related YouTube discussions, resulting in 4,530 user comments. Each comment is annotated along three interpretable dimensions: Target (who is affected), Intent (the underlying motivation or goal behind the comment), and Implication (its potential social impact). Unlike prior explainability datasets such as HateXplain and HARE, which offer token-level or single-dimensional reasoning, HateMirage introduces a multi-dimensional explanation framework that captures the interplay between misinformation, harm, and social consequence. We benchmark multiple open-source language models on HateMirage using ROUGE-L F1 and Sentence-BERT similarity to assess explanation coherence. Results suggest that explanation quality may depend more on pretraining diversity and reasoning-oriented data rather than on model scale alone. By coupling misinformation reasoning with harm attribution, HateMirage establishes a new benchmark for interpretable hate detection and responsible AI research.", "date": "2026-03-03", "categories": []}
{"id": "2502.13061", "title": "Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection", "abstract": "Hateful memes have become a significant concern on the Internet, necessitating robust automated detection systems. While Large Multimodal Models (LMMs) have shown promise in hateful meme detection, they face notable challenges like sub-optimal performance and limited out-of-domain generalization capabilities. Recent studies further reveal the limitations of both supervised fine-tuning (SFT) and in-context learning when applied to LMMs in this setting. To address these issues, we propose a robust adaptation framework for hateful meme detection that enhances in-domain accuracy and cross-domain generalization while preserving the general vision-language capabilities of LMMs. Analysis reveals that our approach achieves improved robustness under adversarial attacks compared to SFT models. Experiments on six meme classification datasets show that our approach achieves state-of-the-art performance, outperforming larger agentic systems. Moreover, our method generates higher-quality rationales for explaining hateful content compared to standard SFT, enhancing model interpretability. Code available at https://github.com/JingbiaoMei/RGCL", "date": "2025-02-18", "categories": []}
{"id": "2502.16612", "title": "MemeIntel: Explainable Detection of Propagandistic and Hateful Memes", "abstract": "The proliferation of multimodal content on social media presents significant challenges in understanding and moderating complex, context-dependent issues such as misinformation, hate speech, and propaganda. While efforts have been made to develop resources and propose new methods for automatic detection, limited attention has been given to jointly modeling label detection and the generation of explanation-based rationales, which often leads to degraded classification performance when trained simultaneously. To address this challenge, we introduce MemeXplain, an explanation-enhanced dataset for propagandistic memes in Arabic and hateful memes in English, making it the first large-scale resource for these tasks. To solve these tasks, we propose a multi-stage optimization approach and train Vision-Language Models (VLMs). Our results show that this strategy significantly improves both label detection and explanation generation quality over the base model, outperforming the current state-of-the-art with an absolute improvement of ~1.4% (Acc) on ArMeme and ~2.2% (Acc) on Hateful Memes. For reproducibility and future research, we aim to make the MemeXplain dataset and scripts publicly available (https://github.com/MohamedBayan/MemeIntel).", "date": "2025-02-23", "categories": []}
{"id": "2510.08630", "title": "ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection", "abstract": "Hateful memes have emerged as a particularly challenging form of online abuse, motivating the development of automated detection systems. Most prior approaches rely on direct detection, producing only binary predictions. Such models fail to provide the context and explanations that real-world moderation requires. Recent Explain-then-Detect approaches, using Chain-of-Thought prompting or LMM agents, perform worse than simple SFT baselines, and even advanced post-training methods such as GRPO fail to close the gap. Our analysis identifies two key issues of such systems: important policy-relevant cues such as targets and attack types are not hypothesized by the model as a likely explanation; and the binary reward signal is insufficient to guide reasoning. To address these challenges, we propose ExPO-HM (Explain-then-Detect Policy Optimization for Hateful Memes), inspired by the training and evaluation process of human annotators. ExPO-HM combines SFT warmup, GRPO with curriculum learning, and Conditional Decision Entropy (CDE) as both metric and reward for reasoning quality. Across three hateful meme benchmarks, ExPO-HM achieves state-of-the-art performance on binary detection, fine-grained classification, and reasoning quality, with up to 15\\% and 17\\% F1 improvement over the GRPO and DPO baselines, respectively. By moving hateful meme detection from simple binary alarms to explanation-driven detection, ExPO-HM provides accurate, interpretable, and actionable moderation support. Code available at https://github.com/JingbiaoMei/ExPO-HM", "date": "2025-10-08", "categories": []}
{"id": "2603.01225", "title": "Can Thinking Models Think to Detect Hateful Memes?", "abstract": "Hateful memes often require compositional multimodal reasoning: the image and text may appear benign in isolation, yet their interaction conveys harmful intent. Although thinking-based multimodal large language models (MLLMs) have recently advanced vision-language understanding, their capabilities remain underexplored for hateful meme analysis. We propose a reinforcement learning based post-training framework that improves reasoning in thinking-based MLLMs through task-specific rewards and a novel Group Relative Policy Optimization (GRPO) objective. Specifically, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful meme understanding, (ii) extend an existing hateful meme dataset by generating weakly or pseudo-supervised chain-of-thought rationales via distillation, and (iii) introduce a GRPO-based objective that jointly optimizes meme classification and explanation quality to encourage fine-grained, step-by-step reasoning. Experiments on the Hateful Memes benchmark show that our approach achieves state-of-the-art performance, improving accuracy and F1 by approximately 1 percent and explanation quality by approximately 3 percent. We will publicly release our code, dataset extensions, and evaluation resources to support reproducibility.", "date": "2026-03-01", "categories": []}
{"id": "2603.01724", "title": "GMP: A Benchmark for Content Moderation under Co-occurring Violations and Dynamic Rules", "abstract": "Online content moderation is essential for maintaining a healthy digital environment, and reliance on AI for this task continues to grow. Consider a user comment using national stereotypes to insult a politician. This example illustrates two critical challenges in real-world scenarios: (1) Co-occurring Violations, where a single post violates multiple policies (e.g., prejudice and personal attacks); (2) Dynamic rules of moderation, where determination of a violation depends on platform-specific guidelines that evolve across contexts . The intersection of co-occurring harms and dynamically changing rules highlights a core limitation of current AI systems: although large language models (LLMs) are adept at following fixed guidelines, their judgment capabilities degrade when policies are unstable or context-dependent . In practice, such shortcomings lead to inconsistent moderation: either erroneously restricting legitimate expression or allowing harmful content to remain online . This raises a critical question for evaluation: Does high performance on existing static benchmarks truly guarantee robust generalization of AI judgment to real-world scenarios involving co-occurring violations and dynamically changing rules?", "date": "2026-03-02", "categories": []}
{"id": "2602.22449", "title": "A Fusion of context-aware based BanglaBERT and Two-Layer Stacked LSTM Framework for Multi-Label Cyberbullying Detection", "abstract": "Cyberbullying has become a serious and growing concern in todays virtual world. When left unnoticed, it can have adverse consequences for social and mental health. Researchers have explored various types of cyberbullying, but most approaches use single-label classification, assuming that each comment contains only one type of abuse. In reality, a single comment may include overlapping forms such as threats, hate speech, and harassment. Therefore, multilabel detection is both realistic and essential. However, multilabel cyberbullying detection has received limited attention, especially in low-resource languages like Bangla, where robust pre-trained models are scarce. Developing a generalized model with moderate accuracy remains challenging. Transformers offer strong contextual understanding but may miss sequential dependencies, while LSTM models capture temporal flow but lack semantic depth. To address these limitations, we propose a fusion architecture that combines BanglaBERT-Large with a two-layer stacked LSTM. We analyze their behavior to jointly model context and sequence. The model is fine-tuned and evaluated on a publicly available multilabel Bangla cyberbullying dataset covering cyberbully, sexual harassment, threat, and spam. We apply different sampling strategies to address class imbalance. Evaluation uses multiple metrics, including accuracy, precision, recall, F1-score, Hamming loss, Cohens kappa, and AUC-ROC. We employ 5-fold cross-validation to assess the generalization of the architecture.", "date": "2026-02-25", "categories": []}
{"id": "2602.20634", "title": "Enhancing Hate Speech Detection on Social Media: A Comparative Analysis of Machine Learning Models and Text Transformation Approaches", "abstract": "The proliferation of hate speech on social media platforms has necessitated the development of effective detection and moderation tools. This study evaluates the efficacy of various machine learning models in identifying hate speech and offensive language and investigates the potential of text transformation techniques to neutralize such content. We compare traditional models like CNNs and LSTMs with advanced neural network models such as BERT and its derivatives, alongside exploring hybrid models that combine different architectural features. Our results indicate that while advanced models like BERT show superior accuracy due to their deep contextual understanding, hybrid models exhibit improved capabilities in certain scenarios. Furthermore, we introduce innovative text transformation approaches that convert negative expressions into neutral ones, thereby potentially mitigating the impact of harmful content. The implications of these findings are discussed, highlighting the strengths and limitations of current technologies and proposing future directions for more robust hate speech detection systems.", "date": "2026-02-24", "categories": []}
{"id": "2602.20818", "title": "GatedCLIP: Gated Multimodal Fusion for Hateful Memes Detection", "abstract": "Detecting hateful content in multimodal memes presents unique challenges, as harmful messages often emerge from the complex interplay between benign images and text. We propose GatedCLIP, a Vision-Language model that enhances CLIP's multimodal capabilities with specialized architectural improvements for hateful memes detection. Our approach introduces learned projection heads that map CLIP embeddings to a task-optimized semantic space, a dynamic gated fusion mechanism that adaptively weights visual and textual features, and a contrastive learning objective that maintains cross-modal semantic alignment. Experiments on the Hateful Memes dataset demonstrate that GatedCLIP achieves an AUROC of 0.66, substantially outperforming the CLIP baseline (AUROC 0.49) while maintaining computational efficiency with only 350K trainable parameters.", "date": "2026-02-24", "categories": []}
{"id": "2510.16985", "title": "Parameter-Efficient Fine-Tuning for Low-Resource Languages: A Comparative Study of LLMs for Bengali Hate Speech Detection", "abstract": "Bengali social media platforms have witnessed a sharp increase in hate speech, disproportionately affecting women and adolescents. While datasets such as BD-SHS provide a basis for structured evaluation, most prior approaches rely on either computationally costly full-model fine-tuning or proprietary APIs. This paper presents the first application of Parameter-Efficient Fine-Tuning (PEFT) for Bengali hate speech detection using LoRA and QLoRA. Three instruction-tuned large language models - Gemma-3-4B, Llama-3.2-3B, and Mistral-7B - were fine-tuned on the BD-SHS dataset of 50,281 annotated comments. Each model was adapted by training fewer than 1% of its parameters, enabling experiments on a single consumer-grade GPU. The results show that Llama-3.2-3B achieved the highest F1-score of 92.23%, followed by Mistral-7B at 88.94% and Gemma-3-4B at 80.25%. These findings establish PEFT as a practical and replicable strategy for Bengali and related low-resource languages.", "date": "2025-10-19", "categories": []}
{"id": "2602.18193", "title": "BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards", "abstract": "Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle-speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.", "date": "2026-02-20", "categories": []}
{"id": "2602.18518", "title": "Measuring the Prevalence of Policy Violating Content with ML Assisted Sampling and LLM Labeling", "abstract": "Content safety teams need metrics that reflect what users actually experience, not only what is reported. We study prevalence: the fraction of user views (impressions) that went to content violating a given policy on a given day. Accurate prevalence measurement is challenging because violations are often rare and human labeling is costly, making frequent, platform-representative studies slow. We present a design-based measurement system that (i) draws daily probability samples from the impression stream using ML-assisted weights to concentrate label budget on high-exposure and high-risk content while preserving unbiasedness, (ii) labels sampled items with a multimodal LLM governed by policy prompts and gold-set validation, and (iii) produces design-consistent prevalence estimates with confidence intervals and dashboard drilldowns. A key design goal is one global sample with many pivots: the same daily sample supports prevalence by surface, viewer geography, content age, and other segments through post-stratified estimation. We describe the statistical estimators, variance and confidence interval construction, label-quality monitoring, and an engineering workflow that makes the system configurable across policies.", "date": "2026-02-19", "categories": []}
{"id": "2602.19212", "title": "Retrieval Augmented Enhanced Dual Co-Attention Framework for Target Aware Multimodal Bengali Hateful Meme Detection", "abstract": "Hateful content on social media increasingly appears as multimodal memes that combine images and text to convey harmful narratives. In low-resource languages such as Bengali, automated detection remains challenging due to limited annotated data, class imbalance, and pervasive code-mixing. To address these issues, we augment the Bengali Hateful Memes (BHM) dataset with semantically aligned samples from the Multimodal Aggression Dataset in Bengali (MIMOSA), improving both class balance and semantic diversity. We propose the Enhanced Dual Co-attention Framework (xDORA), integrating vision encoders (CLIP, DINOv2) and multilingual text encoders (XGLM, XLM-R) via weighted attention pooling to learn robust cross-modal representations. Building on these embeddings, we develop a FAISS-based k-nearest neighbor classifier for non-parametric inference and introduce RAG-Fused DORA, which incorporates retrieval-driven contextual reasoning. We further evaluate LLaVA under zero-shot, few-shot, and retrieval-augmented prompting settings. Experiments on the extended dataset show that xDORA (CLIP + XLM-R) achieves macro-average F1-scores of 0.78 for hateful meme identification and 0.71 for target entity detection, while RAG-Fused DORA improves performance to 0.79 and 0.74, yielding gains over the DORA baseline. The FAISS-based classifier performs competitively and demonstrates robustness for rare classes through semantic similarity modeling. In contrast, LLaVA exhibits limited effectiveness in few-shot settings, with only modest improvements under retrieval augmentation, highlighting constraints of pretrained vision-language models for code-mixed Bengali content without fine-tuning. These findings demonstrate the effectiveness of supervised, retrieval-augmented, and non-parametric multimodal frameworks for addressing linguistic and cultural complexities in low-resource hate speech detection.", "date": "2026-02-22", "categories": []}
{"id": "2602.17467", "title": "PEACE 2.0: Grounded Explanations and Counter-Speech for Combating Hate Expressions", "abstract": "The increasing volume of hate speech on online platforms poses significant societal challenges. While the Natural Language Processing community has developed effective methods to automatically detect the presence of hate speech, responses to it, called counter-speech, are still an open challenge. We present PEACE 2.0, a novel tool that, besides analysing and explaining why a message is considered hateful or not, also generates a response to it. More specifically, PEACE 2.0 has three main new functionalities: leveraging a Retrieval-Augmented Generation (RAG) pipeline i) to ground HS explanations into evidence and facts, ii) to automatically generate evidence-grounded counter-speech, and iii) exploring the characteristics of counter-speech replies. By integrating these capabilities, PEACE 2.0 enables in-depth analysis and response generation for both explicit and implicit hateful messages.", "date": "2026-02-19", "categories": []}
{"id": "2602.15757", "title": "Beyond Binary Classification: Detecting Fine-Grained Sexism in Social Media Videos", "abstract": "Online sexism appears in various forms, which makes its detection challenging. Although automated tools can enhance the identification of sexist content, they are often restricted to binary classification. Consequently, more subtle manifestations of sexism may remain undetected due to the lack of fine-grained, context-sensitive labels. To address this issue, we make the following contributions: (1) we present FineMuSe, a new multimodal sexism detection dataset in Spanish that includes both binary and fine-grained annotations; (2) we introduce a comprehensive hierarchical taxonomy that encompasses forms of sexism, non-sexism, and rhetorical devices of irony and humor; and (3) we evaluate a wide range of LLMs for both binary and fine-grained sexism detection. Our findings indicate that multimodal LLMs perform competitively with human annotators in identifying nuanced forms of sexism; however, they struggle to capture co-occurring sexist types when these are conveyed through visual cues.", "date": "2026-02-17", "categories": []}
{"id": "2602.15809", "title": "Decision Quality Evaluation Framework at Pinterest", "abstract": "Online platforms require robust systems to enforce content safety policies at scale. A critical component of these systems is the ability to evaluate the quality of moderation decisions made by both human agents and Large Language Models (LLMs). However, this evaluation is challenging due to the inherent trade-offs between cost, scale, and trustworthiness, along with the complexity of evolving policies. To address this, we present a comprehensive Decision Quality Evaluation Framework developed and deployed at Pinterest. The framework is centered on a high-trust Golden Set (GDS) curated by subject matter experts (SMEs), which serves as a ground truth benchmark. We introduce an automated intelligent sampling pipeline that uses propensity scores to efficiently expand dataset coverage. We demonstrate the framework's practical application in several key areas: benchmarking the cost-performance trade-offs of various LLM agents, establishing a rigorous methodology for data-driven prompt optimization, managing complex policy evolution, and ensuring the integrity of policy content prevalence metrics via continuous validation. The framework enables a shift from subjective assessments to a data-driven and quantitative practice for managing content safety systems.", "date": "2026-02-17", "categories": []}
{"id": "2602.13455", "title": "Using Machine Learning to Enhance the Detection of Obfuscated Abusive Words in Swahili: A Focus on Child Safety", "abstract": "The rise of digital technology has dramatically increased the potential for cyberbullying and online abuse, necessitating enhanced measures for detection and prevention, especially among children. This study focuses on detecting abusive obfuscated language in Swahili, a low-resource language that poses unique challenges due to its limited linguistic resources and technological support. Swahili is chosen due to its popularity and being the most widely spoken language in Africa, with over 16 million native speakers and upwards of 100 million speakers in total, spanning regions in East Africa and some parts of the Middle East.\n  We employed machine learning models including Support Vector Machines (SVM), Logistic Regression, and Decision Trees, optimized through rigorous parameter tuning and techniques like Synthetic Minority Over-sampling Technique (SMOTE) to handle data imbalance. Our analysis revealed that, while these models perform well in high-dimensional textual data, our dataset's small size and imbalance limit our findings' generalizability. Precision, recall, and F1 scores were thoroughly analyzed, highlighting the nuanced performance of each model in detecting obfuscated language.\n  This research contributes to the broader discourse on ensuring safer online environments for children, advocating for expanded datasets and advanced machine-learning techniques to improve the effectiveness of cyberbullying detection systems. Future work will focus on enhancing data robustness, exploring transfer learning, and integrating multimodal data to create more comprehensive and culturally sensitive detection mechanisms.", "date": "2026-02-13", "categories": []}
{"id": "2602.12818", "title": "AIWizards at MULTIPRIDE: A Hierarchical Approach to Slur Reclamation Detection", "abstract": "Detecting reclaimed slurs represents a fundamental challenge for hate speech detection systems, as the same lexcal items can function either as abusive expressions or as in-group affirmations depending on social identity and context. In this work, we address Subtask B of the MultiPRIDE shared task at EVALITA 2026 by proposing a hierarchical approach to modeling the slur reclamation process. Our core assumption is that members of the LGBTQ+ community are more likely, on average, to employ certain slurs in a eclamatory manner. Based on this hypothesis, we decompose the task into two stages. First, using a weakly supervised LLM-based annotation, we assign fuzzy labels to users indicating the likelihood of belonging to the LGBTQ+ community, inferred from the tweet and the user bio. These soft labels are then used to train a BERT-like model to predict community membership, encouraging the model to learn latent representations associated with LGBTQ+ identity. In the second stage, we integrate this latent space with a newly initialized model for the downstream slur reclamation detection task. The intuition is that the first model encodes user-oriented sociolinguistic signals, which are then fused with representations learned by a model pretrained for hate speech detection. Experimental results on Italian and Spanish show that our approach achieves performance statistically comparable to a strong BERT-based baseline, while providing a modular and extensible framework for incorporating sociolinguistic context into hate speech modeling. We argue that more fine-grained hierarchical modeling of user identity and discourse context may further improve the detection of reclaimed language. We release our code at https://github.com/LucaTedeschini/multipride.", "date": "2026-02-13", "categories": []}
{"id": "2602.07497", "title": "From Native Memes to Global Moderation: Cross-Cultural Evaluation of Vision-Language Models for Hateful Meme Detection", "abstract": "Cultural context profoundly shapes how people interpret online content, yet vision-language models (VLMs) remain predominantly trained through Western or English-centric lenses. This limits their fairness and cross-cultural robustness in tasks like hateful meme detection. We introduce a systematic evaluation framework designed to diagnose and quantify the cross-cultural robustness of state-of-the-art VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection. Results show that the common ``translate-then-detect'' approach deteriorate performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Our findings reveal systematic convergence toward Western safety norms and provide actionable strategies to mitigate such bias, guiding the design of globally robust multimodal moderation systems.", "date": "2026-02-07", "categories": []}
{"id": "2509.22291", "title": "Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?", "abstract": "Natural language processing (NLP) models often replicate or amplify social bias from training data, raising concerns about fairness. At the same time, their black-box nature makes it difficult for users to recognize biased predictions and for developers to effectively mitigate them. While some studies suggest that input-based explanations can help detect and mitigate bias, others question their reliability in ensuring fairness. Existing research on explainability in fair NLP has been predominantly qualitative, with limited large-scale quantitative analysis. In this work, we conduct the first systematic study of the relationship between explainability and fairness in hate speech detection, focusing on both encoder- and decoder-only models. We examine three key dimensions: (1) identifying biased predictions, (2) selecting fair models, and (3) mitigating bias during model training. Our findings show that input-based explanations can effectively detect biased predictions and serve as useful supervision for reducing bias during training, but they are unreliable for selecting fair models among candidates.Our code is available at https://github.com/Ewanwong/fairness_x_explainability.", "date": "2025-09-26", "categories": []}
{"id": "2602.08688", "title": "Old wine in old glasses: Comparing computational and qualitative methods in identifying incivility on Persian Twitter during the #MahsaAmini movement", "abstract": "This paper compares three approaches to detecting incivility in Persian tweets: human qualitative coding, supervised learning with ParsBERT, and large language models (ChatGPT). Using 47,278 tweets from the #MahsaAmini movement in Iran, we evaluate the accuracy and efficiency of each method. ParsBERT substantially outperforms seven evaluated ChatGPT models in identifying hate speech. We also find that ChatGPT struggles not only with subtle cases but also with explicitly uncivil content, and that prompt language (English vs. Persian) does not meaningfully affect its outputs. The study provides a detailed comparison of these approaches and clarifies their strengths and limitations for analyzing hate speech in a low-resource language context.", "date": "2026-02-09", "categories": []}
{"id": "2602.08945", "title": "GitSearch: Enhancing Community Notes Generation with Gap-Informed Targeted Search", "abstract": "Community-based moderation offers a scalable alternative to centralized fact-checking, yet it faces significant structural challenges, and existing AI-based methods fail in \"cold start\" scenarios. To tackle these challenges, we introduce GitSearch (Gap-Informed Targeted Search), a framework that treats human-perceived quality gaps, such as missing context, etc., as first-class signals. GitSearch has a three-stage pipeline: identifying information deficits, executing real-time targeted web-retrieval to resolve them, and synthesizing platform-compliant notes. To facilitate evaluation, we present PolBench, a benchmark of 78,698 U.S. political tweets with their associated Community Notes. We find GitSearch achieves 99% coverage, almost doubling coverage over the state-of-the-art. GitSearch surpasses human-authored helpful notes with a 69% win rate and superior helpfulness scores (3.87 vs. 3.36), demonstrating retrieval effectiveness that balanced the trade-off between scale and quality.", "date": "2026-02-09", "categories": []}
{"id": "2602.05189", "title": "Are Open-Weight LLMs Ready for Social Media Moderation? A Comparative Study on Bluesky", "abstract": "As internet access expands, so does exposure to harmful content, increasing the need for effective moderation. Research has demonstrated that large language models (LLMs) can be effectively utilized for social media moderation tasks, including harmful content detection. While proprietary LLMs have been shown to zero-shot outperform traditional machine learning models, the out-of-the-box capability of open-weight LLMs remains an open question.\n  Motivated by recent developments of reasoning LLMs, we evaluate seven state-of-the-art models: four proprietary and three open-weight. Testing with real-world posts on Bluesky, moderation decisions by Bluesky Moderation Service, and annotations by two authors, we find a considerable degree of overlap between the sensitivity (81%--97%) and specificity (91%--100%) of the open-weight LLMs and those (72%--98%, and 93%--99%) of the proprietary ones. Additionally, our analysis reveals that specificity exceeds sensitivity for rudeness detection, but the opposite holds for intolerance and threats. Lastly, we identify inter-rater agreement across human moderators and the LLMs, highlighting considerations for deploying LLMs in both platform-scale and personalized moderation contexts. These findings show open-weight LLMs can support privacy-preserving moderation on consumer-grade hardware and suggest new directions for designing moderation systems that balance community values with individual user preferences.", "date": "2026-02-05", "categories": []}
{"id": "2602.05874", "title": "xList-Hate: A Checklist-Based Framework for Interpretable and Generalizable Hate Speech Detection", "abstract": "Hate speech detection is commonly framed as a direct binary classification problem despite being a composite concept defined through multiple interacting factors that vary across legal frameworks, platform policies, and annotation guidelines. As a result, supervised models often overfit dataset-specific definitions and exhibit limited robustness under domain shift and annotation noise.\n  We introduce xList-Hate, a diagnostic framework that decomposes hate speech detection into a checklist of explicit, concept-level questions grounded in widely shared normative criteria. Each question is independently answered by a large language model (LLM), producing a binary diagnostic representation that captures hateful content features without directly predicting the final label. These diagnostic signals are then aggregated by a lightweight, fully interpretable decision tree, yielding transparent and auditable predictions.\n  We evaluate it across multiple hate speech benchmarks and model families, comparing it against zero-shot LLM classification and in-domain supervised fine-tuning. While supervised methods typically maximize in-domain performance, we consistently improves cross-dataset robustness and relative performance under domain shift. In addition, qualitative analysis of disagreement cases provides evidence that the framework can be less sensitive to certain forms of annotation inconsistency and contextual ambiguity. Crucially, the approach enables fine-grained interpretability through explicit decision paths and factor-level analysis.\n  Our results suggest that reframing hate speech detection as a diagnostic reasoning task, rather than a monolithic classification problem, provides a robust, explainable, and extensible alternative for content moderation.", "date": "2026-02-05", "categories": []}
{"id": "2301.12534", "title": "Vicarious Offense and Noise Audit of Offensive Speech Classifiers: Unifying Human and Machine Disagreement on What is Offensive", "abstract": "Offensive speech detection is a key component of content moderation. However, what is offensive can be highly subjective. This paper investigates how machine and human moderators disagree on what is offensive when it comes to real-world social web political discourse. We show that (1) there is extensive disagreement among the moderators (humans and machines); and (2) human and large-language-model classifiers are unable to predict how other human raters will respond, based on their political leanings. For (1), we conduct a noise audit at an unprecedented scale that combines both machine and human responses. For (2), we introduce a first-of-its-kind dataset of vicarious offense. Our noise audit reveals that moderation outcomes vary wildly across different machine moderators. Our experiments with human moderators suggest that political leanings combined with sensitive issues affect both first-person and vicarious offense. The dataset is available through https://github.com/Homan-Lab/voiced.", "date": "2023-01-29", "categories": []}
{"id": "2601.15605", "title": "ToxiTwitch: Toward Emote-Aware Hybrid Moderation for Live Streaming Platforms", "abstract": "The rapid growth of live-streaming platforms such as Twitch has introduced complex challenges in moderating toxic behavior. Traditional moderation approaches, such as human annotation and keyword-based filtering, have demonstrated utility, but human moderators on Twitch constantly struggle to scale effectively in the fast-paced, high-volume, and context-rich chat environment of the platform while also facing harassment themselves. Recent advances in large language models (LLMs), such as DeepSeek-R1-Distill and Llama-3-8B-Instruct, offer new opportunities for toxicity detection, especially in understanding nuanced, multimodal communication involving emotes. In this work, we present an exploratory comparison of toxicity detection approaches tailored to Twitch. Our analysis reveals that incorporating emotes improves the detection of toxic behavior. To this end, we introduce ToxiTwitch, a hybrid model that combines LLM-generated embeddings of text and emotes with traditional machine learning classifiers, including Random Forest and SVM. In our case study, the proposed hybrid approach reaches up to 80 percent accuracy under channel-specific training (with 13 percent improvement over BERT and F1-score of 76 percent). This work is an exploratory study intended to surface challenges and limits of emote-aware toxicity detection on Twitch.", "date": "2026-01-22", "categories": []}
{"id": "2602.04693", "title": "LinGO: A Linguistic Graph Optimization Framework with LLMs for Interpreting Intents of Online Uncivil Discourse", "abstract": "Detecting uncivil language is crucial for maintaining safe, inclusive, and democratic online spaces. Yet existing classifiers often misinterpret posts containing uncivil cues but expressing civil intents, leading to inflated estimates of harmful incivility online. We introduce LinGO, a linguistic graph optimization framework for large language models (LLMs) that leverages linguistic structures and optimization techniques to classify multi-class intents of incivility that use various direct and indirect expressions. LinGO decomposes language into multi-step linguistic components, identifies targeted steps that cause the most errors, and iteratively optimizes prompt and/or example components for targeted steps. We evaluate it using a dataset collected during the 2022 Brazilian presidential election, encompassing four forms of political incivility: Impoliteness (IMP), Hate Speech and Stereotyping (HSST), Physical Harm and Violent Political Rhetoric (PHAVPR), and Threats to Democratic Institutions and Values (THREAT). Each instance is annotated with six types of civil/uncivil intent. We benchmark LinGO using three cost-efficient LLMs: GPT-5-mini, Gemini 2.5 Flash-Lite, and Claude 3 Haiku, and four optimization techniques: TextGrad, AdalFlow, DSPy, and Retrieval-Augmented Generation (RAG). The results show that, across all models, LinGO consistently improves accuracy and weighted F1 compared with zero-shot, chain-of-thought, direct optimization, and fine-tuning baselines. RAG is the strongest optimization technique and, when paired with Gemini model, achieves the best overall performance. These findings demonstrate that incorporating multi-step linguistic components into LLM instructions and optimize targeted components can help the models explain complex semantic meanings, which can be extended to other complex semantic explanation tasks in the future.", "date": "2026-02-04", "categories": []}
{"id": "2510.07707", "title": "Causality Guided Representation Learning for Cross-Style Hate Speech Detection", "abstract": "The proliferation of online hate speech poses a significant threat to the harmony of the web. While explicit hate is easily recognized through overt slurs, implicit hate speech is often conveyed through sarcasm, irony, stereotypes, or coded language -- making it harder to detect. Existing hate speech detection models, which predominantly rely on surface-level linguistic cues, fail to generalize effectively across diverse stylistic variations. Moreover, hate speech spread on different platforms often targets distinct groups and adopts unique styles, potentially inducing spurious correlations between them and labels, further challenging current detection approaches. Motivated by these observations, we hypothesize that the generation of hate speech can be modeled as a causal graph involving key factors: contextual environment, creator motivation, target, and style. Guided by this graph, we propose CADET, a causal representation learning framework that disentangles hate speech into interpretable latent factors and then controls confounders, thereby isolating genuine hate intent from superficial linguistic cues. Furthermore, CADET allows counterfactual reasoning by intervening on style within the latent space, naturally guiding the model to robustly identify hate speech in varying forms. CADET demonstrates superior performance in comprehensive experiments, highlighting the potential of causal priors in advancing generalizable hate speech detection.", "date": "2025-10-09", "categories": []}
{"id": "2512.07684", "title": "When Large Language Models Do Not Work: Online Incivility Prediction through Graph Neural Networks", "abstract": "Online incivility has emerged as a widespread and persistent problem in digital communities, imposing substantial social and psychological burdens on users. Although many platforms attempt to curb incivility through moderation and automated detection, the performance of existing approaches often remains limited in both accuracy and efficiency. To address this challenge, we propose a Graph Neural Network (GNN) framework for detecting three types of uncivil behavior (i.e., toxicity, aggression, and personal attacks) within the English Wikipedia community. Our model represents each user comment as a node, with textual similarity between comments defining the edges, allowing the network to jointly learn from both linguistic content and relational structures among comments. We also introduce a dynamically adjusted attention mechanism that adaptively balances nodal and topological features during information aggregation. Empirical evaluations demonstrate that our proposed architecture outperforms 12 state-of-the-art Large Language Models (LLMs) across multiple metrics while requiring significantly lower inference cost. These findings highlight the crucial role of structural context in detecting online incivility and address the limitations of text-only LLM paradigms in behavioral prediction. All datasets and comparative outputs will be publicly available in our repository to support further research and reproducibility.", "date": "2025-12-08", "categories": []}
{"id": "2602.00132", "title": "Shedding the Facades, Connecting the Domains: Detecting Shifting Multimodal Hate Video with Test-Time Adaptation", "abstract": "Hate Video Detection (HVD) is crucial for online ecosystems. Existing methods assume identical distributions between training (source) and inference (target) data. However, hateful content often evolves into irregular and ambiguous forms to evade censorship, resulting in substantial semantic drift and rendering previously trained models ineffective. Test-Time Adaptation (TTA) offers a solution by adapting models during inference to narrow the cross-domain gap, while conventional TTA methods target mild distribution shifts and struggle with the severe semantic drift in HVD. To tackle these challenges, we propose SCANNER, the first TTA framework tailored for HVD. Motivated by the insight that, despite the evolving nature of hateful manifestations, their underlying cores remain largely invariant (i.e., targeting is still based on characteristics like gender, race, etc), we leverage these stable cores as a bridge to connect the source and target domains. Specifically, SCANNER initially reveals the stable cores from the ambiguous layout in evolving hateful content via a principled centroid-guided alignment mechanism. To alleviate the impact of outlier-like samples that are weakly correlated with centroids during the alignment process, SCANNER enhances the prior by incorporating a sample-level adaptive centroid alignment strategy, promoting more stable adaptation. Furthermore, to mitigate semantic collapse from overly uniform outputs within clusters, SCANNER introduces an intra-cluster diversity regularization that encourages the cluster-wise semantic richness. Experiments show that SCANNER outperforms all baselines, with an average gain of 4.69% in Macro-F1 over the best.", "date": "2026-01-28", "categories": []}
{"id": "2602.01101", "title": "Robust Harmful Meme Detection under Missing Modalities via Shared Representation Learning", "abstract": "Internet memes are powerful tools for communication, capable of spreading political, psychological, and sociocultural ideas. However, they can be harmful and can be used to disseminate hate toward targeted individuals or groups. Although previous studies have focused on designing new detection methods, these often rely on modal-complete data, such as text and images. In real-world settings, however, modalities like text may be missing due to issues like poor OCR quality, making existing methods sensitive to missing information and leading to performance deterioration. To address this gap, in this paper, we present the first-of-its-kind work to comprehensively investigate the behavior of harmful meme detection methods in the presence of modal-incomplete data. Specifically, we propose a new baseline method that learns a shared representation for multiple modalities by projecting them independently. These shared representations can then be leveraged when data is modal-incomplete. Experimental results on two benchmark datasets demonstrate that our method outperforms existing approaches when text is missing. Moreover, these results suggest that our method allows for better integration of visual features, reducing dependence on text and improving robustness in scenarios where textual information is missing. Our work represents a significant step forward in enabling the real-world application of harmful meme detection, particularly in situations where a modality is absent.", "date": "2026-02-01", "categories": []}
{"id": "2601.21796", "title": "KID: Knowledge-Injected Dual-Head Learning for Knowledge-Grounded Harmful Meme Detection", "abstract": "Internet memes have become pervasive carriers of digital culture on social platforms. However, their heavy reliance on metaphors and sociocultural context also makes them subtle vehicles for harmful content, posing significant challenges for automated content moderation. Existing approaches primarily focus on intra-modal and inter-modal signal analysis, while the understanding of implicit toxicity often depends on background knowledge that is not explicitly present in the meme itself. To address this challenge, we propose KID, a Knowledge-Injected Dual-Head Learning framework for knowledge-grounded harmful meme detection. KID adopts a label-constrained distillation paradigm to decompose complex meme understanding into structured reasoning chains that explicitly link visual evidence, background knowledge, and classification labels. These chains guide the learning process by grounding external knowledge in meme-specific contexts. In addition, KID employs a dual-head architecture that jointly optimizes semantic generation and classification objectives, enabling aligned linguistic reasoning while maintaining stable decision boundaries. Extensive experiments on five multilingual datasets spanning English, Chinese, and low-resource Bengali demonstrate that KID achieves SOTA performance on both binary and multi-label harmful meme detection tasks, improving over previous best methods by 2.1%--19.7% across primary evaluation metrics. Ablation studies further confirm the effectiveness of knowledge injection and dual-head joint learning, highlighting their complementary contributions to robust and generalizable meme understanding. The code and data are available at https://github.com/PotatoDog1669/KID.", "date": "2026-01-29", "categories": []}
{"id": "2506.08477", "title": "Read as You See: Guiding Unimodal LLMs for Low-Resource Explainable Harmful Meme Detection", "abstract": "Detecting harmful memes is crucial for safeguarding the integrity and harmony of online environments, yet existing detection methods are often resource-intensive, inflexible, and lacking explainability, limiting their applicability in assisting real-world web content moderation. We propose U-CoT+, a resource-efficient framework that prioritizes accessibility, flexibility and transparency in harmful meme detection by fully harnessing the capabilities of lightweight unimodal large language models (LLMs). Instead of directly prompting or fine-tuning large multimodal models (LMMs) as black-box classifiers, we avoid immediate reasoning over complex visual inputs but decouple meme content recognition from meme harmfulness analysis through a high-fidelity meme-to-text pipeline, which collaborates lightweight LMMs and LLMs to convert multimodal memes into natural language descriptions that preserve critical visual information, thus enabling text-only LLMs to \"see\" memes by \"reading\". Grounded in textual inputs, we further guide unimodal LLMs' reasoning under zero-shot Chain-of-Thoughts (CoT) prompting with targeted, interpretable, context-aware, and easily obtained human-crafted guidelines, thus providing accountable step-by-step rationales, while enabling flexible and efficient adaptation to diverse sociocultural criteria of harmfulness. Extensive experiments on seven benchmark datasets show that U-CoT+ achieves performance comparable to resource-intensive baselines, highlighting its effectiveness and potential as a scalable, explainable, and low-resource solution to support harmful meme detection.", "date": "2025-06-10", "categories": []}
{"id": "2601.20129", "title": "BengaliSent140: A Large-Scale Bengali Binary Sentiment Dataset for Hate and Non-Hate Speech Classification", "abstract": "Sentiment analysis for the Bengali language has attracted increasing research interest in recent years. However, progress remains constrained by the scarcity of large-scale and diverse annotated datasets. Although several Bengali sentiment and hate speech datasets are publicly available, most are limited in size or confined to a single domain, such as social media comments. Consequently, these resources are often insufficient for training modern deep learning based models, which require large volumes of heterogeneous data to learn robust and generalizable representations. In this work, we introduce BengaliSent140, a large-scale Bengali binary sentiment dataset constructed by consolidating seven existing Bengali text datasets into a unified corpus. To ensure consistency across sources, heterogeneous annotation schemes are systematically harmonized into a binary sentiment formulation with two classes: Not Hate (0) and Hate (1). The resulting dataset comprises 139,792 unique text samples, including 68,548 hate and 71,244 not-hate instances, yielding a relatively balanced class distribution. By integrating data from multiple sources and domains, BengaliSent140 offers broader linguistic and contextual coverage than existing Bengali sentiment datasets and provides a strong foundation for training and benchmarking deep learning models. Baseline experimental results are also reported to demonstrate the practical usability of the dataset. The dataset is publicly available at https://www.kaggle.com/datasets/akifislam/bengalisent140/", "date": "2026-01-27", "categories": []}
{"id": "2601.20256", "title": "SoftHateBench: Evaluating Moderation Models Against Reasoning-Driven, Policy-Compliant Hostility", "abstract": "Online hate on social media ranges from overt slurs and threats (\\emph{hard hate speech}) to \\emph{soft hate speech}: discourse that appears reasonable on the surface but uses framing and value-based arguments to steer audiences toward blaming or excluding a target group. We hypothesize that current moderation systems, largely optimized for surface toxicity cues, are not robust to this reasoning-driven hostility, yet existing benchmarks do not measure this gap systematically. We introduce \\textbf{\\textsc{SoftHateBench}}, a generative benchmark that produces soft-hate variants while preserving the underlying hostile standpoint. To generate soft hate, we integrate the \\emph{Argumentum Model of Topics} (AMT) and \\emph{Relevance Theory} (RT) in a unified framework: AMT provides the backbone argument structure for rewriting an explicit hateful standpoint into a seemingly neutral discussion while preserving the stance, and RT guides generation to keep the AMT chain logically coherent. The benchmark spans \\textbf{7} sociocultural domains and \\textbf{28} target groups, comprising \\textbf{4,745} soft-hate instances. Evaluations across encoder-based detectors, general-purpose LLMs, and safety models show a consistent drop from hard to soft tiers: systems that detect explicit hostility often fail when the same stance is conveyed through subtle, reasoning-based language. \\textcolor{red}{\\textbf{Disclaimer.} Contains offensive examples used solely for research.}", "date": "2026-01-28", "categories": []}
{"id": "2601.09342", "title": "Improving Implicit Hate Speech Detection via a Community-Driven Multi-Agent Framework", "abstract": "This work proposes a contextualised detection framework for implicitly hateful speech, implemented as a multi-agent system comprising a central Moderator Agent and dynamically constructed Community Agents representing specific demographic groups. Our approach explicitly integrates socio-cultural context from publicly available knowledge sources, enabling identity-aware moderation that surpasses state-of-the-art prompting methods (zero-shot prompting, few-shot prompting, chain-of-thought prompting) and alternative approaches on a challenging ToxiGen dataset. We enhance the technical rigour of performance evaluation by incorporating balanced accuracy as a central metric of classification fairness that accounts for the trade-off between true positive and true negative rates. We demonstrate that our community-driven consultative framework significantly improves both classification accuracy and fairness across all target groups.", "date": "2026-01-14", "categories": []}
{"id": "2503.21679", "title": "JiraiBench: A Bilingual Benchmark for Evaluating Large Language Models' Detection of Human Self-Destructive Behavior Content in Jirai Community", "abstract": "This paper introduces JiraiBench, the first bilingual benchmark for evaluating large language models' effectiveness in detecting self-destructive content across Chinese and Japanese social media communities. Focusing on the transnational \"Jirai\" (landmine) online subculture that encompasses multiple forms of self-destructive behaviors including drug overdose, eating disorders, and self-harm, we present a comprehensive evaluation framework incorporating both linguistic and cultural dimensions. Our dataset comprises 10,419 Chinese posts and 5,000 Japanese posts with multidimensional annotation along three behavioral categories, achieving substantial inter-annotator agreement. Experimental evaluations across four state-of-the-art models reveal significant performance variations based on instructional language, with Japanese prompts unexpectedly outperforming Chinese prompts when processing Chinese content. This emergent cross-cultural transfer suggests that cultural proximity can sometimes outweigh linguistic similarity in detection tasks. Cross-lingual transfer experiments with fine-tuned models further demonstrate the potential for knowledge transfer between these language systems without explicit target language training. These findings highlight the need for culturally-informed approaches to multilingual content moderation and provide empirical evidence for the importance of cultural context in developing more effective detection systems for vulnerable online communities.", "date": "2025-03-27", "categories": []}
{"id": "2508.06515", "title": "BigTokDetect: A Clinically-Informed Vision-Language Modeling Framework for Detecting Pro-Bigorexia Videos on TikTok", "abstract": "Social media platforms face escalating challenges in detecting harmful content that promotes muscle dysmorphic behaviors and cognitions (bigorexia). This content can evade moderation by camouflaging as legitimate fitness advice and disproportionately affects adolescent males. We address this challenge with BigTokDetect, a clinically informed framework for identifying pro-bigorexia content on TikTok. We introduce BigTok, the first expert-annotated multimodal benchmark dataset of over 2,200 TikTok videos labeled by clinical psychiatrists across five categories and eighteen fine-grained subcategories. Comprehensive evaluation of state-of-the-art vision-language models reveals that while commercial zero-shot models achieve the highest accuracy on broad primary categories, supervised fine-tuning enables smaller open-source models to perform better on fine-grained subcategory detection. Ablation studies show that multimodal fusion improves performance by 5 to 15 percent, with video features providing the most discriminative signals. These findings support a grounded moderation approach that automates detection of explicit harms while flagging ambiguous content for human review, and they establish a scalable framework for harm mitigation in emerging mental health domains.", "date": "2025-07-30", "categories": []}
{"id": "2508.11808", "title": "Labels or Input? Rethinking Augmentation in Multimodal Hate Detection", "abstract": "Online hate remains a significant societal challenge, especially as multimodal content enables subtle, culturally grounded, and implicit forms of harm. Hateful memes embed hostility through text-image interactions and humor, making them difficult for automated systems to interpret. Although recent Vision-Language Models (VLMs) perform well on explicit cases, their deployment is limited by high inference costs and persistent failures on nuanced content. This work examines how far small models can be improved through prompt optimization, fine-tuning, and automated data augmentation. We introduce an end-to-end pipeline that varies prompt structure, label granularity, and training modality, showing that structured prompts and scaled supervision significantly strengthen compact VLMs. We also develop a multimodal augmentation framework that generates counterfactually neutral memes via a coordinated LLM-VLM setup, reducing spurious correlations and improving the detection of implicit hate. Ablation studies quantify the contribution of each component, demonstrating that prompt design, granular labels, and targeted augmentation collectively narrow the gap between small and large models. The results offer a practical path toward more robust and deployable multimodal hate-detection systems without relying on costly large-model inference.", "date": "2025-08-15", "categories": []}
{"id": "2601.15115", "title": "Training-Free and Interpretable Hateful Video Detection via Multi-stage Adversarial Reasoning", "abstract": "Hateful videos pose serious risks by amplifying discrimination, inciting violence, and undermining online safety. Existing training-based hateful video detection methods are constrained by limited training data and lack of interpretability, while directly prompting large vision-language models often struggle to deliver reliable hate detection. To address these challenges, this paper introduces MARS, a training-free Multi-stage Adversarial ReaSoning framework that enables reliable and interpretable hateful content detection. MARS begins with the objective description of video content, establishing a neutral foundation for subsequent analysis. Building on this, it develops evidence-based reasoning that supports potential hateful interpretations, while in parallel incorporating counter-evidence reasoning to capture plausible non-hateful perspectives. Finally, these perspectives are synthesized into a conclusive and explainable decision. Extensive evaluation on two real-world datasets shows that MARS achieves up to 10% improvement under certain backbones and settings compared to other training-free approaches and outperforms state-of-the-art training-based methods on one dataset. In addition, MARS produces human-understandable justifications, thereby supporting compliance oversight and enhancing the transparency of content moderation workflows. The code is available at https://github.com/Multimodal-Intelligence-Lab-MIL/MARS.", "date": "2026-01-21", "categories": []}
{"id": "2601.11634", "title": "When Rules Fall Short: Agent-Driven Discovery of Emerging Content Issues in Short Video Platforms", "abstract": "Trends on short-video platforms evolve at a rapid pace, with new content issues emerging every day that fall outside the coverage of existing annotation policies. However, traditional human-driven discovery of emerging issues is too slow, which leads to delayed updates of annotation policies and poses a major challenge for effective content governance. In this work, we propose an automatic issue discovery method based on multimodal LLM agents. Our approach automatically recalls short videos containing potential new issues and applies a two-stage clustering strategy to group them, with each cluster corresponding to a newly discovered issue. The agent then generates updated annotation policies from these clusters, thereby extending coverage to these emerging issues. Our agent has been deployed in the real system. Both offline and online experiments demonstrate that this agent-based method significantly improves the effectiveness of emerging-issue discovery (with an F1 score improvement of over 20%) and enhances the performance of subsequent issue governance (reducing the view count of problematic videos by approximately 15%). More importantly, compared to manual issue discovery, it greatly reduces time costs and substantially accelerates the iteration of annotation policies.", "date": "2026-01-14", "categories": []}
{"id": "2601.13018", "title": "Bi-Attention HateXplain : Taking into account the sequential aspect of data during explainability in a multi-task context", "abstract": "Technological advances in the Internet and online social networks have brought many benefits to humanity. At the same time, this growth has led to an increase in hate speech, the main global threat. To improve the reliability of black-box models used for hate speech detection, post-hoc approaches such as LIME, SHAP, and LRP provide the explanation after training the classification model. In contrast, multi-task approaches based on the HateXplain benchmark learn to explain and classify simultaneously. However, results from HateXplain-based algorithms show that predicted attention varies considerably when it should be constant. This attention variability can lead to inconsistent interpretations, instability of predictions, and learning difficulties. To solve this problem, we propose the BiAtt-BiRNN-HateXplain (Bidirectional Attention BiRNN HateXplain) model which is easier to explain compared to LLMs which are more complex in view of the need for transparency, and will take into account the sequential aspect of the input data during explainability thanks to a BiRNN layer. Thus, if the explanation is correctly estimated, thanks to multi-task learning (explainability and classification task), the model could classify better and commit fewer unintentional bias errors related to communities. The experimental results on HateXplain data show a clear improvement in detection performance, explainability and a reduction in unintentional bias.", "date": "2026-01-19", "categories": []}
{"id": "2601.13547", "title": "HateXScore: A Metric Suite for Evaluating Reasoning Quality in Hate Speech Explanations", "abstract": "Hateful speech detection is a key component of content moderation, yet current evaluation frameworks rarely assess why a text is deemed hateful. We introduce \\textsf{HateXScore}, a four-component metric suite designed to evaluate the reasoning quality of model explanations. It assesses (i) conclusion explicitness, (ii) faithfulness and causal grounding of quoted spans, (iii) protected group identification (policy-configurable), and (iv) logical consistency among these elements. Evaluated on six diverse hate speech datasets, \\textsf{HateXScore} is intended as a diagnostic complement to reveal interpretability failures and annotation inconsistencies that are invisible to standard metrics like Accuracy or F1. Moreover, human evaluation shows strong agreement with \\textsf{HateXScore}, validating it as a practical tool for trustworthy and transparent moderation.\n  \\textcolor{red}{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}", "date": "2026-01-20", "categories": []}
{"id": "2601.08457", "title": "An Under-Explored Application for Explainable Multimodal Misogyny Detection in code-mixed Hindi-English", "abstract": "Digital platforms have an ever-expanding user base, and act as a hub for communication, business, and connectivity. However, this has also allowed for the spread of hate speech and misogyny. Artificial intelligence models have emerged as an effective solution for countering online hate speech but are under explored for low resource and code-mixed languages and suffer from a lack of interpretability. Explainable Artificial Intelligence (XAI) can enhance transparency in the decisions of deep learning models, which is crucial for a sensitive domain such as hate speech detection. In this paper, we present a multi-modal and explainable web application for detecting misogyny in text and memes in code-mixed Hindi and English. The system leverages state-of-the-art transformer-based models that support multilingual and multimodal settings. For text-based misogyny identification, the system utilizes XLM-RoBERTa (XLM-R) and multilingual Bidirectional Encoder Representations from Transformers (mBERT) on a dataset of approximately 4,193 comments. For multimodal misogyny identification from memes, the system utilizes mBERT + EfficientNet, and mBERT + ResNET trained on a dataset of approximately 4,218 memes. It also provides feature importance scores using explainability techniques including Shapley Additive Values (SHAP) and Local Interpretable Model Agnostic Explanations (LIME). The application aims to serve as a tool for both researchers and content moderators, to promote further research in the field, combat gender based digital violence, and ensure a safe digital space. The system has been evaluated using human evaluators who provided their responses on Chatbot Usability Questionnaire (CUQ) and User Experience Questionnaire (UEQ) to determine overall usability.", "date": "2026-01-13", "categories": []}
{"id": "2601.08684", "title": "MEMEWEAVER: Inter-Meme Graph Reasoning for Sexism and Misogyny Detection", "abstract": "Women are twice as likely as men to face online harassment due to their gender. Despite recent advances in multimodal content moderation, most approaches still overlook the social dynamics behind this phenomenon, where perpetrators reinforce prejudices and group identity within like-minded communities. Graph-based methods offer a promising way to capture such interactions, yet existing solutions remain limited by heuristic graph construction, shallow modality fusion, and instance-level reasoning. In this work, we present MemeWeaver, an end-to-end trainable multimodal framework for detecting sexism and misogyny through a novel inter-meme graph reasoning mechanism. We systematically evaluate multiple visual--textual fusion strategies and show that our approach consistently outperforms state-of-the-art baselines on the MAMI and EXIST benchmarks, while achieving faster training convergence. Further analyses reveal that the learned graph structure captures semantically meaningful patterns, offering valuable insights into the relational nature of online hate.", "date": "2026-01-13", "categories": []}
{"id": "2508.07063", "title": "Towards Safer AI Moderation: Evaluating LLM Moderators Through a Unified Benchmark Dataset and Advocating a Human-First Approach", "abstract": "As AI systems become more integrated into daily life, the need for safer and more reliable moderation has never been greater. Large Language Models (LLMs) have demonstrated remarkable capabilities, surpassing earlier models in complexity and performance. Their evaluation across diverse tasks has consistently showcased their potential, enabling the development of adaptive and personalized agents. However, despite these advancements, LLMs remain prone to errors, particularly in areas requiring nuanced moral reasoning. They struggle with detecting implicit hate, offensive language, and gender biases due to the subjective and context-dependent nature of these issues. Moreover, their reliance on training data can inadvertently reinforce societal biases, leading to inconsistencies and ethical concerns in their outputs. To explore the limitations of LLMs in this role, we developed an experimental framework based on state-of-the-art (SOTA) models to assess human emotions and offensive behaviors. The framework introduces a unified benchmark dataset encompassing 49 distinct categories spanning the wide spectrum of human emotions, offensive and hateful text, and gender and racial biases. Furthermore, we introduced SafePhi, a QLoRA fine-tuned version of Phi-4, adapting diverse ethical contexts and outperforming benchmark moderators by achieving a Macro F1 score of 0.89, where OpenAI Moderator and Llama Guard score 0.77 and 0.74, respectively. This research also highlights the critical domains where LLM moderators consistently underperformed, pressing the need to incorporate more heterogeneous and representative data with human-in-the-loop, for better model robustness and explainability.", "date": "2025-08-09", "categories": []}
{"id": "2601.06306", "title": "SyntaxMind at BLP-2025 Task 1: Leveraging Attention Fusion of CNN and GRU for Hate Speech Detection", "abstract": "This paper describes our system used in the BLP-2025 Task 1: Hate Speech Detection. We participated in Subtask 1A and Subtask 1B, addressing hate speech classification in Bangla text. Our approach employs a unified architecture that integrates BanglaBERT embeddings with multiple parallel processing branches based on GRUs and CNNs, followed by attention and dense layers for final classification. The model is designed to capture both contextual semantics and local linguistic cues, enabling robust performance across subtasks. The proposed system demonstrated high competitiveness, obtaining 0.7345 micro F1-Score (2nd place) in Subtask 1A and 0.7317 micro F1-Score (5th place) in Subtask 1B.", "date": "2026-01-09", "categories": []}
{"id": "2508.03296", "title": "Towards Trustworthy Multimodal Moderation via Policy-Aligned Reasoning and Hierarchical Labeling", "abstract": "Social platforms have revolutionized information sharing, but also accelerated the dissemination of harmful and policy-violating content. To ensure safety and compliance at scale, moderation systems must go beyond efficiency and offer accuracy and interpretability. However, current approaches largely rely on noisy, label-driven learning, lacking alignment with moderation rules and producing opaque decisions that hinder human review. Therefore, we propose Hierarchical Guard (Hi-Guard), a multimodal moderation framework that introduces a new policy-aligned decision paradigm. The term \"Hierarchical\" reflects two key aspects of our system design: (1) a hierarchical moderation pipeline, where a lightweight binary model first filters safe content and a stronger model handles fine-grained risk classification; and (2) a hierarchical taxonomy in the second stage, where the model performs path-based classification over a hierarchical taxonomy ranging from coarse to fine-grained levels. To ensure alignment with evolving moderation policies, Hi-Guard directly incorporates rule definitions into the model prompt. To further enhance structured prediction and reasoning, we introduce a multi-level soft-margin reward and optimize with Group Relative Policy Optimization (GRPO), penalizing semantically adjacent misclassifications and improving explanation quality. Extensive experiments and real-world deployment demonstrate that Hi-Guard achieves superior classification accuracy, generalization, and interpretability, paving the way toward scalable, transparent, and trustworthy content safety systems. Code is available at: https://github.com/lianqi1008/Hi-Guard.", "date": "2025-08-05", "categories": []}
{"id": "2601.04211", "title": "Qwerty AI: Explainable Automated Age Rating and Content Safety Assessment for Russian-Language Screenplays", "abstract": "We present Qwerty AI, an end-to-end system for automated age-rating and content-safety assessment of Russian-language screenplays according to Federal Law No. 436-FZ. The system processes full-length scripts (up to 700 pages in under 2 minutes), segments them into narrative units, detects content violations across five categories (violence, sexual content, profanity, substances, frightening elements), and assigns age ratings (0+, 6+, 12+, 16+, 18+) with explainable justifications. Our implementation leverages a fine-tuned Phi-3-mini model with 4-bit quantization, achieving 80% rating accuracy and 80-95% segmentation precision (format-dependent). The system was developed under strict constraints: no external API calls, 80GB VRAM limit, and <5 minute processing time for average scripts. Deployed on Yandex Cloud with CUDA acceleration, Qwerty AI demonstrates practical applicability for production workflows. We achieved these results during the Wink hackathon (November 2025), where our solution addressed real editorial challenges in the Russian media industry.", "date": "2025-12-10", "categories": []}
{"id": "2601.04692", "title": "See, Explain, and Intervene: A Few-Shot Multimodal Agent Framework for Hateful Meme Moderation", "abstract": "In this work, we examine hateful memes from three complementary angles - how to detect them, how to explain their content and how to intervene them prior to being posted - by applying a range of strategies built on top of generative AI models. To the best of our knowledge, explanation and intervention have typically been studied separately from detection, which does not reflect real-world conditions. Further, since curating large annotated datasets for meme moderation is prohibitively expensive, we propose a novel framework that leverages task-specific generative multimodal agents and the few-shot adaptability of large multimodal models to cater to different types of memes. We believe this is the first work focused on generalizable hateful meme moderation under limited data conditions, and has strong potential for deployment in real-world production scenarios. Warning: Contains potentially toxic contents.", "date": "2026-01-08", "categories": []}
{"id": "2512.23732", "title": "When in Doubt, Consult: Expert Debate for Sexism Detection via Confidence-Based Routing", "abstract": "Online sexism increasingly appears in subtle, context-dependent forms that evade traditional detection methods. Its interpretation often depends on overlapping linguistic, psychological, legal, and cultural dimensions, which produce mixed and sometimes contradictory signals in annotated datasets. These inconsistencies, combined with label scarcity and class imbalance, result in unstable decision boundaries and cause fine-tuned models to overlook subtler, underrepresented forms of harm. To address these challenges, we propose a two-stage framework that unifies (i) targeted training procedures to better regularize supervision to scarce and noisy data with (ii) selective, reasoning-based inference to handle ambiguous or borderline cases. First, we stabilize the training combining class-balanced focal loss, class-aware batching, and post-hoc threshold calibration, strategies for the firs time adapted for this domain to mitigate label imbalance and noisy supervision. Second, we bridge the gap between efficiency and reasoning with a a dynamic routing mechanism that distinguishes between unambiguous instances and complex cases requiring a deliberative process. This reasoning process results in the novel Collaborative Expert Judgment (CEJ) module which prompts multiple personas and consolidates their reasoning through a judge model. Our approach outperforms existing approaches across several public benchmarks, with F1 gains of +4.48% and +1.30% on EDOS Tasks A and B, respectively, and a +2.79% improvement in ICM on EXIST 2025 Task 1.1.", "date": "2025-12-21", "categories": []}
{"id": "2601.01620", "title": "The Gray Area: Characterizing Moderator Disagreement on Reddit", "abstract": "Volunteer moderators play a crucial role in sustaining online dialogue, but they often disagree about what should or should not be allowed. In this paper, we study the complexity of content moderation with a focus on disagreements between moderators, which we term the ``gray area'' of moderation. Leveraging 5 years and 4.3 million moderation log entries from 24 subreddits of different topics and sizes, we characterize how gray area, or disputed cases, differ from undisputed cases. We show that one-in-seven moderation cases are disputed among moderators, often addressing transgressions where users' intent is not directly legible, such as in trolling and brigading, as well as tensions around community governance. This is concerning, as almost half of all gray area cases involved automated moderation decisions. Through information-theoretic evaluations, we demonstrate that gray area cases are inherently harder to adjudicate than undisputed cases and show that state-of-the-art language models struggle to adjudicate them. We highlight the key role of expert human moderators in overseeing the moderation process and provide insights about the challenges of current moderation processes and tools.", "date": "2026-01-04", "categories": []}
{"id": "2601.03481", "title": "Self-Explaining Hate Speech Detection with Moral Rationales", "abstract": "Hate speech detection models rely on surface-level lexical features, increasing vulnerability to spurious correlations and limiting robustness, cultural contextualization, and interpretability. We propose Supervised Moral Rationale Attention (SMRA), the first self-explaining hate speech detection framework to incorporate moral rationales as direct supervision for attention alignment. Based on Moral Foundations Theory, SMRA aligns token-level attention with expert-annotated moral rationales, guiding models to attend to morally salient spans rather than spurious lexical patterns. Unlike prior rationale-supervised or post-hoc approaches, SMRA integrates moral rationale supervision directly into the training objective, producing inherently interpretable and contextualized explanations. To support our framework, we also introduce HateBRMoralXplain, a Brazilian Portuguese benchmark dataset annotated with hate labels, moral categories, token-level moral rationales, and socio-political metadata. Across binary hate speech detection and multi-label moral sentiment classification, SMRA consistently improves performance (e.g., +0.9 and +1.5 F1, respectively) while substantially enhancing explanation faithfulness, increasing IoU F1 (+7.4 pp) and Token F1 (+5.0 pp). Although explanations become more concise, sufficiency improves (+2.3 pp) and fairness remains stable, indicating more faithful rationales without performance or bias trade-offs", "date": "2026-01-07", "categories": []}
{"id": "2601.02697", "title": "Boosting Accuracy and Interpretability in Multilingual Hate Speech Detection Through Layer Freezing and Explainable AI", "abstract": "Sentiment analysis focuses on identifying the emotional polarity expressed in textual data, typically categorized as positive, negative, or neutral. Hate speech detection, on the other hand, aims to recognize content that incites violence, discrimination, or hostility toward individuals or groups based on attributes such as race, gender, sexual orientation, or religion. Both tasks play a critical role in online content moderation by enabling the detection and mitigation of harmful or offensive material, thereby contributing to safer digital environments. In this study, we examine the performance of three transformer-based models: BERT-base-multilingual-cased, RoBERTa-base, and XLM-RoBERTa-base with the first eight layers frozen, for multilingual sentiment analysis and hate speech detection. The evaluation is conducted across five languages: English, Korean, Japanese, Chinese, and French. The models are compared using standard performance metrics, including accuracy, precision, recall, and F1-score. To enhance model interpretability and provide deeper insight into prediction behavior, we integrate the Local Interpretable Model-agnostic Explanations (LIME) framework, which highlights the contribution of individual words to the models decisions. By combining state-of-the-art transformer architectures with explainability techniques, this work aims to improve both the effectiveness and transparency of multilingual sentiment analysis and hate speech detection systems.", "date": "2026-01-06", "categories": []}
{"id": "2601.03194", "title": "X-MuTeST: A Multilingual Benchmark for Explainable Hate Speech Detection and A Novel LLM-consulted Explanation Framework", "abstract": "Hate speech detection on social media faces challenges in both accuracy and explainability, especially for underexplored Indic languages. We propose a novel explainability-guided training framework, X-MuTeST (eXplainable Multilingual haTe Speech deTection), for hate speech detection that combines high-level semantic reasoning from large language models (LLMs) with traditional attention-enhancing techniques. We extend this research to Hindi and Telugu alongside English by providing benchmark human-annotated rationales for each word to justify the assigned class label. The X-MuTeST explainability method computes the difference between the prediction probabilities of the original text and those of unigrams, bigrams, and trigrams. Final explanations are computed as the union between LLM explanations and X-MuTeST explanations. We show that leveraging human rationales during training enhances both classification performance and explainability. Moreover, combining human rationales with our explainability method to refine the model attention yields further improvements. We evaluate explainability using Plausibility metrics such as Token-F1 and IOU-F1 and Faithfulness metrics such as Comprehensiveness and Sufficiency. By focusing on under-resourced languages, our work advances hate speech detection across diverse linguistic contexts. Our dataset includes token-level rationale annotations for 6,004 Hindi, 4,492 Telugu, and 6,334 English samples. Data and code are available on https://github.com/ziarehman30/X-MuTeST", "date": "2026-01-06", "categories": []}
{"id": "2512.17367", "title": "Adversarially Robust Detection of Harmful Online Content: A Computational Design Science Approach", "abstract": "Social media platforms are plagued by harmful content such as hate speech, misinformation, and extremist rhetoric. Machine learning (ML) models are widely adopted to detect such content; however, they remain highly vulnerable to adversarial attacks, wherein malicious users subtly modify text to evade detection. Enhancing adversarial robustness is therefore essential, requiring detectors that can defend against diverse attacks (generalizability) while maintaining high overall accuracy. However, simultaneously achieving both optimal generalizability and accuracy is challenging. Following the computational design science paradigm, this study takes a sequential approach that first proposes a novel framework (Large Language Model-based Sample Generation and Aggregation, LLM-SGA) by identifying the key invariances of textual adversarial attacks and leveraging them to ensure that a detector instantiated within the framework has strong generalizability. Second, we instantiate our detector (Adversarially Robust Harmful Online Content Detector, ARHOCD) with three novel design components to improve detection accuracy: (1) an ensemble of multiple base detectors that exploits their complementary strengths; (2) a novel weight assignment method that dynamically adjusts weights based on each sample's predictability and each base detector's capability, with weights initialized using domain knowledge and updated via Bayesian inference; and (3) a novel adversarial training strategy that iteratively optimizes both the base detectors and the weight assignor. We addressed several limitations of existing adversarial robustness enhancement research and empirically evaluated ARHOCD across three datasets spanning hate speech, rumor, and extremist content. Results show that ARHOCD offers strong generalizability and improves detection accuracy under adversarial conditions.", "date": "2025-12-19", "categories": []}
{"id": "2512.21598", "title": "From Shallow Humor to Metaphor: Towards Label-Free Harmful Meme Detection via LMM Agent Self-Improvement", "abstract": "The proliferation of harmful memes on online media poses significant risks to public health and stability. Existing detection methods heavily rely on large-scale labeled data for training, which necessitates substantial manual annotation efforts and limits their adaptability to the continually evolving nature of harmful content. To address these challenges, we present ALARM, the first lAbeL-free hARmful Meme detection framework powered by Large Multimodal Model (LMM) agent self-improvement. The core innovation of ALARM lies in exploiting the expressive information from \"shallow\" memes to iteratively enhance its ability to tackle more complex and subtle ones. ALARM consists of a novel Confidence-based Explicit Meme Identification mechanism that isolates the explicit memes from the original dataset and assigns them pseudo-labels. Besides, a new Pairwise Learning Guided Agent Self-Improvement paradigm is introduced, where the explicit memes are reorganized into contrastive pairs (positive vs. negative) to refine a learner LMM agent. This agent autonomously derives high-level detection cues from these pairs, which in turn empower the agent itself to handle complex and challenging memes effectively. Experiments on three diverse datasets demonstrate the superior performance and strong adaptability of ALARM to newly evolved memes. Notably, our method even outperforms label-driven methods. These results highlight the potential of label-free frameworks as a scalable and promising solution for adapting to novel forms and topics of harmful memes in dynamic online environments.", "date": "2025-12-25", "categories": []}
{"id": "2512.18027", "title": "CoPE: A Small Language Model for Steerable and Scalable Content Labeling", "abstract": "This paper details the methodology behind CoPE, a policy-steerable small language model capable of fast and accurate content labeling. We present a novel training curricula called Contradictory Example Training that enables the model to learn policy interpretation rather than mere policy memorization. We also present a novel method for generating content policies, called Binocular Labeling, which enables rapid construction of unambiguous training datasets. When evaluated across seven different harm areas, CoPE exhibits equal or superior accuracy to frontier models at only 1% of their size. We openly release a 9 billion parameter version of the model that can be run on a single consumer-grade GPU. Models like CoPE represent a paradigm shift for classifier systems. By turning an ML task into a policy writing task, CoPE opens up new design possibilities for the governance of online platforms.", "date": "2025-12-19", "categories": []}
{"id": "2512.16147", "title": "Decoding Fake Narratives in Spreading Hateful Stories: A Dual-Head RoBERTa Model with Multi-Task Learning", "abstract": "Social media platforms, while enabling global connectivity, have become hubs for the rapid spread of harmful content, including hate speech and fake narratives \\cite{davidson2017automated, shu2017fake}. The Faux-Hate shared task focuses on detecting a specific phenomenon: the generation of hate speech driven by fake narratives, termed Faux-Hate. Participants are challenged to identify such instances in code-mixed Hindi-English social media text. This paper describes our system developed for the shared task, addressing two primary sub-tasks: (a) Binary Faux-Hate detection, involving fake and hate speech classification, and (b) Target and Severity prediction, categorizing the intended target and severity of hateful content. Our approach combines advanced natural language processing techniques with domain-specific pretraining to enhance performance across both tasks. The system achieved competitive results, demonstrating the efficacy of leveraging multi-task learning for this complex problem.", "date": "2025-12-18", "categories": []}
{"id": "2512.09563", "title": "System Report for CCL25-Eval Task 10: Prompt-Driven Large Language Model Merge for Fine-Grained Chinese Hate Speech Detection", "abstract": "The proliferation of hate speech on Chinese social media poses urgent societal risks, yet traditional systems struggle to decode context-dependent rhetorical strategies and evolving slang. To bridge this gap, we propose a novel three-stage LLM-based framework: Prompt Engineering, Supervised Fine-tuning, and LLM Merging. First, context-aware prompts are designed to guide LLMs in extracting implicit hate patterns. Next, task-specific features are integrated during supervised fine-tuning to enhance domain adaptation. Finally, merging fine-tuned LLMs improves robustness against out-of-distribution cases. Evaluations on the STATE-ToxiCN benchmark validate the framework's effectiveness, demonstrating superior performance over baseline methods in detecting fine-grained hate speech.", "date": "2025-12-10", "categories": []}
{"id": "2512.09662", "title": "Can LLMs Evaluate What They Cannot Annotate? Revisiting LLM Reliability in Hate Speech Detection", "abstract": "Hate speech spreads widely online, harming individuals and communities, making automatic detection essential for large-scale moderation, yet detecting it remains difficult. Part of the challenge lies in subjectivity: what one person flags as hate speech, another may see as benign. Traditional annotation agreement metrics, such as Cohen's $\\kappa$, oversimplify this disagreement, treating it as an error rather than meaningful diversity. Meanwhile, Large Language Models (LLMs) promise scalable annotation, but prior studies demonstrate that they cannot fully replace human judgement, especially in subjective tasks. In this work, we reexamine LLM reliability using a subjectivity-aware framework, cross-Rater Reliability (xRR), revealing that even under fairer lens, LLMs still diverge from humans. Yet this limitation opens an opportunity: we find that LLM-generated annotations can reliably reflect performance trends across classification models, correlating with human evaluations. We test this by examining whether LLM-generated annotations preserve the relative ordering of model performance derived from human evaluation (i.e. whether models ranked as more reliable by human annotators preserve the same order when evaluated with LLM-generated labels). Our results show that, although LLMs differ from humans at the instance level, they reproduce similar ranking and classification patterns, suggesting their potential as proxy evaluators. While not a substitute for human annotators, they might serve as a scalable proxy for evaluation in subjective NLP tasks.", "date": "2025-12-10", "categories": []}
{"id": "2505.23856", "title": "OMNIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and Modalities", "abstract": "The emerging capabilities of large language models (LLMs) have sparked concerns about their immediate potential for harmful misuse. The core approach to mitigate these concerns is the detection of harmful queries to the model. Current detection approaches are fallible, and are particularly susceptible to attacks that exploit mismatched generalization of model capabilities (e.g., prompts in low-resource languages or prompts provided in non-text modalities such as image and audio). To tackle this challenge, we propose Omniguard, an approach for detecting harmful prompts across languages and modalities. Our approach (i) identifies internal representations of an LLM/MLLM that are aligned across languages or modalities and then (ii) uses them to build a language-agnostic or modality-agnostic classifier for detecting harmful prompts. Omniguard improves harmful prompt classification accuracy by 11.57\\% over the strongest baseline in a multilingual setting, by 20.44\\% for image-based prompts, and sets a new SOTA for audio-based prompts. By repurposing embeddings computed during generation, Omniguard is also very efficient ($\\approx\\!120 \\times$ faster than the next fastest baseline). Code and data are available at: https://github.com/vsahil/OmniGuard.", "date": "2025-05-29", "categories": []}
{"id": "2512.02530", "title": "Aetheria: A multimodal interpretable content safety framework based on multi-agent debate and collaboration", "abstract": "The exponential growth of digital content presents significant challenges for content safety. Current moderation systems, often based on single models or fixed pipelines, exhibit limitations in identifying implicit risks and providing interpretable judgment processes. To address these issues, we propose Aetheria, a multimodal interpretable content safety framework based on multi-agent debate and collaboration.Employing a collaborative architecture of five core agents, Aetheria conducts in-depth analysis and adjudication of multimodal content through a dynamic, mutually persuasive debate mechanism, which is grounded by RAG-based knowledge retrieval.Comprehensive experiments on our proposed benchmark (AIR-Bench) validate that Aetheria not only generates detailed and traceable audit reports but also demonstrates significant advantages over baselines in overall content safety accuracy, especially in the identification of implicit risks. This framework establishes a transparent and interpretable paradigm, significantly advancing the field of trustworthy AI content moderation.", "date": "2025-12-02", "categories": []}
{"id": "2512.07838", "title": "Detection of Cyberbullying in GIF using AI", "abstract": "Cyberbullying is a well-known social issue, and it is escalating day by day. Due to the vigorous development of the internet, social media provide many different ways for the user to express their opinions and exchange information. Cyberbullying occurs on social media using text messages, comments, sharing images and GIFs or stickers, and audio and video. Much research has been done to detect cyberbullying on textual data; some are available for images. Very few studies are available to detect cyberbullying on GIFs/stickers. We collect a GIF dataset from Twitter and Applied a deep learning model to detect cyberbullying from the dataset. Firstly, we extracted hashtags related to cyberbullying using Twitter. We used these hashtags to download GIF file using publicly available API GIPHY. We collected over 4100 GIFs including cyberbullying and non cyberbullying. we applied deep learning pre-trained model VGG16 for the detection of the cyberbullying. The deep learning model achieved the accuracy of 97%. Our work provides the GIF dataset for researchers working in this area.", "date": "2025-11-15", "categories": []}
{"id": "2512.08480", "title": "Soft Inductive Bias Approach via Explicit Reasoning Perspectives in Inappropriate Utterance Detection Using Large Language Models", "abstract": "Recent incidents in certain online games and communities, where anonymity is guaranteed, show that unchecked inappropriate remarks frequently escalate into verbal abuse and even criminal behavior, raising significant social concerns. Consequently, there is a growing need for research on techniques that can detect inappropriate utterances within conversational texts to help build a safer communication environment. Although large-scale language models trained on Korean corpora and chain-of-thought reasoning have recently gained attention, research applying these approaches to inappropriate utterance detection remains limited. In this study, we propose a soft inductive bias approach that explicitly defines reasoning perspectives to guide the inference process, thereby promoting rational decision-making and preventing errors that may arise during reasoning. We fine-tune a Korean large language model using the proposed method and conduct both quantitative performance comparisons and qualitative evaluations across different training strategies. Experimental results show that the Kanana-1.5 model achieves an average accuracy of 87.0046, improving by approximately 3.89 percent over standard supervised learning. These findings indicate that the proposed method goes beyond simple knowledge imitation by large language models and enables more precise and consistent judgments through constrained reasoning perspectives, demonstrating its effectiveness for inappropriate utterance detection.", "date": "2025-12-09", "categories": []}
{"id": "2512.02141", "title": "Feature Selection Empowered BERT for Detection of Hate Speech with Vocabulary Augmentation", "abstract": "Abusive speech on social media poses a persistent and evolving challenge, driven by the continuous emergence of novel slang and obfuscated terms designed to circumvent detection systems. In this work, we present a data efficient strategy for fine tuning BERT on hate speech classification by significantly reducing training set size without compromising performance. Our approach employs a TF IDF-based sample selection mechanism to retain only the most informative 75 percent of examples, thereby minimizing training overhead. To address the limitations of BERT's native vocabulary in capturing evolving hate speech terminology, we augment the tokenizer with domain-specific slang and lexical variants commonly found in abusive contexts. Experimental results on a widely used hate speech dataset demonstrate that our method achieves competitive performance while improving computational efficiency, highlighting its potential for scalable and adaptive abusive content moderation.", "date": "2025-12-01", "categories": []}
{"id": "2512.02845", "title": "Bangla Hate Speech Classification with Fine-tuned Transformer Models", "abstract": "Hate speech recognition in low-resource languages remains a difficult problem due to insufficient datasets, orthographic heterogeneity, and linguistic variety. Bangla is spoken by more than 230 million people of Bangladesh and India (West Bengal). Despite the growing need for automated moderation on social media platforms, Bangla is significantly under-represented in computational resources. In this work, we study Subtask 1A and Subtask 1B of the BLP 2025 Shared Task on hate speech detection. We reproduce the official baselines (e.g., Majority, Random, Support Vector Machine) and also produce and consider Logistic Regression, Random Forest, and Decision Tree as baseline methods. We also utilized transformer-based models such as DistilBERT, BanglaBERT, m-BERT, and XLM-RoBERTa for hate speech classification. All the transformer-based models outperformed baseline methods for the subtasks, except for DistilBERT. Among the transformer-based models, BanglaBERT produces the best performance for both subtasks. Despite being smaller in size, BanglaBERT outperforms both m-BERT and XLM-RoBERTa, which suggests language-specific pre-training is very important. Our results highlight the potential and need for pre-trained language models for the low-resource Bangla language.", "date": "2025-12-02", "categories": []}
{"id": "2511.17955", "title": "MTikGuard System: A Transformer-Based Multimodal System for Child-Safe Content Moderation on TikTok", "abstract": "With the rapid rise of short-form videos, TikTok has become one of the most influential platforms among children and teenagers, but also a source of harmful content that can affect their perception and behavior. Such content, often subtle or deceptive, challenges traditional moderation methods due to the massive volume and real-time nature of uploads. This paper presents MTikGuard, a real-time multimodal harmful content detection system for TikTok, with three key contributions: (1) an extended TikHarm dataset expanded to 4,723 labeled videos by adding diverse real-world samples, (2) a multimodal classification framework integrating visual, audio, and textual features to achieve state-of-the-art performance with 89.37% accuracy and 89.45% F1-score, and (3) a scalable streaming architecture built on Apache Kafka and Apache Spark for real-time deployment. The results demonstrate the effectiveness of combining dataset expansion, advanced multimodal fusion, and robust deployment for practical large-scale social media content moderation. The dataset is available at https://github.com/ntdat-8324/MTikGuard-System.git.", "date": "2025-11-22", "categories": []}
{"id": "2511.18324", "title": "Gradient Masters at BLP-2025 Task 1: Advancing Low-Resource NLP for Bengali using Ensemble-Based Adversarial Training for Hate Speech Detection", "abstract": "This paper introduces the approach of \"Gradient Masters\" for BLP-2025 Task 1: \"Bangla Multitask Hate Speech Identification Shared Task\". We present an ensemble-based fine-tuning strategy for addressing subtasks 1A (hate-type classification) and 1B (target group classification) in YouTube comments. We propose a hybrid approach on a Bangla Language Model, which outperformed the baseline models and secured the 6th position in subtask 1A with a micro F1 score of 73.23% and the third position in subtask 1B with 73.28%. We conducted extensive experiments that evaluated the robustness of the model throughout the development and evaluation phases, including comparisons with other Language Model variants, to measure generalization in low-resource Bangla hate speech scenarios and data set coverage. In addition, we provide a detailed analysis of our findings, exploring misclassification patterns in the detection of hate speech.", "date": "2025-11-23", "categories": []}
{"id": "2511.19168", "title": "RAVEN++: Pinpointing Fine-Grained Violations in Advertisement Videos with Active Reinforcement Reasoning", "abstract": "Advertising (Ad) is a cornerstone of the digital economy, yet the moderation of video advertisements remains a significant challenge due to their complexity and the need for precise violation localization. While recent advancements, such as the RAVEN model, have improved coarse-grained violation detection, critical gaps persist in fine-grained understanding, explainability, and generalization. To address these limitations, we propose RAVEN++, a novel framework that introduces three key innovations: 1) Active Reinforcement Learning (RL), which dynamically adapts training to samples of varying difficulty; 2) Fine-Grained Violation Understanding, achieved through hierarchical reward functions and reasoning distillation; and 3) Progressive Multi-Stage Training, which systematically combines knowledge injection, curriculum-based passive RL, and active RL. Extensive experiments on both public and proprietary datasets, on both offline scenarios and online deployed A/B Testing, demonstrate that RAVEN++ outperforms general-purpose LLMs and specialized models like RAVEN in terms of fine-grained violation understanding, reasoning capabilities, and generalization ability.", "date": "2025-11-24", "categories": []}
{"id": "2510.09771", "title": "PromptGuard at BLP-2025 Task 1: A Few-Shot Classification Framework Using Majority Voting and Keyword Similarity for Bengali Hate Speech Detection", "abstract": "The BLP-2025 Task 1A requires Bengali hate speech classification into six categories. Traditional supervised approaches need extensive labeled datasets that are expensive for low-resource languages. We developed PromptGuard, a few-shot framework combining chi-square statistical analysis for keyword extraction with adaptive majority voting for decision-making. We explore statistical keyword selection versus random approaches and adaptive voting mechanisms that extend classification based on consensus quality. Chi-square keywords provide consistent improvements across categories, while adaptive voting benefits ambiguous cases requiring extended classification rounds. PromptGuard achieves a micro-F1 of 67.61, outperforming n-gram baselines (60.75) and random approaches (14.65). Ablation studies confirm chi-square-based keywords show the most consistent impact across all categories.", "date": "2025-10-10", "categories": []}
{"id": "2506.10245", "title": "ToxSyn: Reducing Bias in Hate Speech Detection via Synthetic Minority Data in Brazilian Portuguese", "abstract": "The development of robust hate speech detection systems remains limited by the lack of large-scale, fine-grained training data, especially for languages beyond English. Existing corpora typically rely on coarse toxic/non-toxic labels, and the few that capture hate directed at specific minority groups critically lack the non-toxic counterexamples (i.e., benign text about minorities) required to distinguish genuine hate from mere discussion. We introduce ToxSyn, the first Portuguese large-scale corpus explicitly designed for multi-label hate speech detection across nine protected minority groups. Generated via a controllable four-stage pipeline, ToxSyn includes discourse-type annotations to capture rhetorical strategies of toxic language, such as sarcasm or dehumanization. Crucially, it systematically includes the non-toxic counterexamples absent in all other public datasets. Our experiments reveal a catastrophic, mutual generalization failure between social-media domains and ToxSyn: models trained on social media struggle to generalize to minority-specific contexts, and vice-versa. This finding indicates they are distinct tasks and exposes summary metrics like Macro F1 can be unreliable indicators of true model behavior, as they completely mask model failure. We publicly release ToxSyn at HuggingFace to foster reproducible research on synthetic data generation and benchmark progress in hate-speech detection for low- and mid-resource languages.", "date": "2025-06-11", "categories": []}
{"id": "2409.07958", "title": "Enhanced Online Grooming Detection Employing Context Determination and Message-Level Analysis", "abstract": "Online Grooming (OG) is a prevalent threat facing predominately children online, with groomers using deceptive methods to prey on the vulnerability of children on social media/messaging platforms. These attacks can have severe psychological and physical impacts, including a tendency towards revictimization. Current technical measures are inadequate, especially with the advent of end-to-end encryption which hampers message monitoring. Existing solutions focus on the signature analysis of child abuse media, which does not effectively address real-time OG detection. This paper proposes that OG attacks are complex, requiring the identification of specific communication patterns between adults and children. It introduces a novel approach leveraging advanced models such as BERT and RoBERTa for Message-Level Analysis and a Context Determination approach for classifying actor interactions, including the introduction of Actor Significance Thresholds and Message Significance Thresholds. The proposed method aims to enhance accuracy and robustness in detecting OG by considering the dynamic and multi-faceted nature of these attacks. Cross-dataset experiments evaluate the robustness and versatility of our approach. This paper's contributions include improved detection methodologies and the potential for application in various scenarios, addressing gaps in current literature and practices.", "date": "2024-09-12", "categories": []}
{"id": "2509.21486", "title": "Reasoning-Enhanced Domain-Adaptive Pretraining of Multimodal Large Language Models for Short Video Content Governance", "abstract": "Short video platforms are evolving rapidly, making the identification of inappropriate content increasingly critical. Existing approaches typically train separate and small classification models for each type of issue, which requires extensive human-labeled data and lacks cross-issue generalization. We propose a reasoning-enhanced multimodal large language model (MLLM) pretraining paradigm for unified inappropriate content detection. To address the distribution gap between short video content and the original pretraining data of MLLMs, as well as the complex issue definitions, we introduce three targeted pretraining tasks: (1) \\textit{Caption}, to enhance the MLLM's perception of video details; (2) \\textit{Visual Question Answering (VQA)}, to deepen the MLLM's understanding of issue definitions and annotation guidelines; (3) \\textit{Chain-of-Thought (CoT)}, to enhance the MLLM's reasoning capability. Experimental results show that our pretraining approach significantly improves the MLLM's performance in both zero-shot and supervised fine-tuning (SFT) settings. In addition, our pretrained model demonstrates strong generalization capabilities to emergent, previously unseen issues.", "date": "2025-09-25", "categories": []}
{"id": "2511.06402", "title": "SugarTextNet: A Transformer-Based Framework for Detecting Sugar Dating-Related Content on Social Media with Context-Aware Focal Loss", "abstract": "Sugar dating-related content has rapidly proliferated on mainstream social media platforms, giving rise to serious societal and regulatory concerns, including commercialization of intimate relationships and the normalization of transactional relationships.~Detecting such content is highly challenging due to the prevalence of subtle euphemisms, ambiguous linguistic cues, and extreme class imbalance in real-world data.~In this work, we present SugarTextNet, a novel transformer-based framework specifically designed to identify sugar dating-related posts on social media.~SugarTextNet integrates a pretrained transformer encoder, an attention-based cue extractor, and a contextual phrase encoder to capture both salient and nuanced features in user-generated text.~To address class imbalance and enhance minority-class detection, we introduce Context-Aware Focal Loss, a tailored loss function that combines focal loss scaling with contextual weighting.~We evaluate SugarTextNet on a newly curated, manually annotated dataset of 3,067 Chinese social media posts from Sina Weibo, demonstrating that our approach substantially outperforms traditional machine learning models, deep learning baselines, and large language models across multiple metrics.~Comprehensive ablation studies confirm the indispensable role of each component.~Our findings highlight the importance of domain-specific, context-aware modeling for sensitive content detection, and provide a robust solution for content moderation in complex, real-world scenarios.", "date": "2025-11-09", "categories": []}
{"id": "2503.07237", "title": "LLM-C3MOD: A Human-LLM Collaborative System for Cross-Cultural Hate Speech Moderation", "abstract": "Content moderation is a global challenge, yet major tech platforms prioritize high-resource languages, leaving low-resource languages with scarce native moderators. Since effective moderation depends on understanding contextual cues, this imbalance increases the risk of improper moderation due to non-native moderators' limited cultural understanding. Through a user study, we identify that non-native moderators struggle with interpreting culturally-specific knowledge, sentiment, and internet culture in the hate speech moderation. To assist them, we present LLM-C3MOD, a human-LLM collaborative pipeline with three steps: (1) RAG-enhanced cultural context annotations; (2) initial LLM-based moderation; and (3) targeted human moderation for cases lacking LLM consensus. Evaluated on a Korean hate speech dataset with Indonesian and German participants, our system achieves 78% accuracy (surpassing GPT-4o's 71% baseline), while reducing human workload by 83.6%. Notably, human moderators excel at nuanced contents where LLMs struggle. Our findings suggest that non-native moderators, when properly supported by LLMs, can effectively contribute to cross-cultural hate speech moderation.", "date": "2025-03-10", "categories": []}
{"id": "2511.05532", "title": "Beyond One-Size-Fits-All: Personalized Harmful Content Detection with In-Context Learning", "abstract": "The proliferation of harmful online content--e.g., toxicity, spam, and negative sentiment--demands robust and adaptable moderation systems. However, prevailing moderation systems are centralized and task-specific, offering limited transparency and neglecting diverse user preferences--an approach ill-suited for privacy-sensitive or decentralized environments. We propose a novel framework that leverages in-context learning (ICL) with foundation models to unify the detection of toxicity, spam, and negative sentiment across binary, multi-class, and multi-label settings. Crucially, our approach enables lightweight personalization, allowing users to easily block new categories, unblock existing ones, or extend detection to semantic variations through simple prompt-based interventions--all without model retraining. Extensive experiments on public benchmarks (TextDetox, UCI SMS, SST2) and a new, annotated Mastodon dataset reveal that: (i) foundation models achieve strong cross-task generalization, often matching or surpassing task-specific fine-tuned models; (ii) effective personalization is achievable with as few as one user-provided example or definition; and (iii) augmenting prompts with label definitions or rationales significantly enhances robustness to noisy, real-world data. Our work demonstrates a definitive shift beyond one-size-fits-all moderation, establishing ICL as a practical, privacy-preserving, and highly adaptable pathway for the next generation of user-centric content safety systems. To foster reproducibility and facilitate future research, we publicly release our code on GitHub and the annotated Mastodon dataset on Hugging Face.", "date": "2025-10-29", "categories": []}
{"id": "2511.06051", "title": "Efficient Hate Speech Detection: A Three-Layer LoRA-Tuned BERTweet Framework", "abstract": "This paper addresses the critical challenge of developing computationally efficient hate speech detection systems that maintain competitive performance while being practical for real-time deployment. We propose a novel three-layer framework that combines rule-based pre-filtering with a parameter-efficient LoRA-tuned BERTweet model and continuous learning capabilities. Our approach achieves 0.85 macro F1 score - representing 94% of the performance of state-of-the-art large language models like SafePhi (Phi-4 based) while using a base model that is 100x smaller (134M vs 14B parameters). Compared to traditional BERT-based approaches with similar computational requirements, our method demonstrates superior performance through strategic dataset unification and optimized fine-tuning. The system requires only 1.87M trainable parameters (1.37% of full fine-tuning) and trains in approximately 2 hours on a single T4 GPU, making robust hate speech detection accessible in resource-constrained environments while maintaining competitive accuracy for real-world deployment.", "date": "2025-11-08", "categories": []}
{"id": "2511.07065", "title": "Aligning Attention with Human Rationales for Self-Explaining Hate Speech Detection", "abstract": "The opaque nature of deep learning models presents significant challenges for the ethical deployment of hate speech detection systems. To address this limitation, we introduce Supervised Rational Attention (SRA), a framework that explicitly aligns model attention with human rationales, improving both interpretability and fairness in hate speech classification. SRA integrates a supervised attention mechanism into transformer-based classifiers, optimizing a joint objective that combines standard classification loss with an alignment loss term that minimizes the discrepancy between attention weights and human-annotated rationales. We evaluated SRA on hate speech benchmarks in English (HateXplain) and Portuguese (HateBRXplain) with rationale annotations. Empirically, SRA achieves 2.4x better explainability compared to current baselines, and produces token-level explanations that are more faithful and human-aligned. In terms of fairness, SRA achieves competitive fairness across all measures, with second-best performance in detecting toxic posts targeting identity groups, while maintaining comparable results on other metrics. These findings demonstrate that incorporating human rationales into attention mechanisms can enhance interpretability and faithfulness without compromising fairness.", "date": "2025-11-10", "categories": []}
{"id": "2511.07304", "title": "Retriv at BLP-2025 Task 1: A Transformer Ensemble and Multi-Task Learning Approach for Bangla Hate Speech Identification", "abstract": "This paper addresses the problem of Bangla hate speech identification, a socially impactful yet linguistically challenging task. As part of the \"Bangla Multi-task Hate Speech Identification\" shared task at the BLP Workshop, IJCNLP-AACL 2025, our team \"Retriv\" participated in all three subtasks: (1A) hate type classification, (1B) target group identification, and (1C) joint detection of type, severity, and target. For subtasks 1A and 1B, we employed a soft-voting ensemble of transformer models (BanglaBERT, MuRIL, IndicBERTv2). For subtask 1C, we trained three multitask variants and aggregated their predictions through a weighted voting ensemble. Our systems achieved micro-f1 scores of 72.75% (1A) and 72.69% (1B), and a weighted micro-f1 score of 72.62% (1C). On the shared task leaderboard, these corresponded to 9th, 10th, and 7th positions, respectively. These results highlight the promise of transformer ensembles and weighted multitask frameworks for advancing Bangla hate speech detection in low-resource contexts. We made experimental scripts publicly available for the community.", "date": "2025-11-10", "categories": []}
{"id": "2504.17902", "title": "TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection", "abstract": "Social media memes are a challenging domain for hate detection because they intertwine visual and textual cues into culturally nuanced messages. To tackle these challenges, we introduce TRACE, a hierarchical multimodal framework that leverages visually grounded context augmentation, along with a novel caption-scoring network to emphasize hate-relevant content, and parameter-efficient fine-tuning of CLIP's text encoder. Our experiments demonstrate that selectively fine-tuning deeper text encoder layers significantly enhances performance compared to simpler projection-layer fine-tuning methods. Specifically, our framework achieves state-of-the-art accuracy (0.807) and F1-score (0.806) on the widely-used Hateful Memes dataset, matching the performance of considerably larger models while maintaining efficiency. Moreover, it achieves superior generalization on the MultiOFF offensive meme dataset (F1-score 0.673), highlighting robustness across meme categories. Additional analyses confirm that robust visual grounding and nuanced text representations significantly reduce errors caused by benign confounders. We publicly release our code to facilitate future research.", "date": "2025-04-24", "categories": []}
{"id": "2509.18293", "title": "Evaluating Large Language Models for Detecting Antisemitism", "abstract": "Detecting hateful content is a challenging and important problem. Automated tools, like machine-learning models, can help, but they require continuous training to adapt to the ever-changing landscape of social media. In this work, we evaluate eight open-source LLMs' capability to detect antisemitic content, specifically leveraging in-context definition. We also study how LLMs understand and explain their decisions given a moderation policy as a guideline. First, we explore various prompting techniques and design a new CoT-like prompt, Guided-CoT, and find that injecting domain-specific thoughts increases performance and utility. Guided-CoT handles the in-context policy well, improving performance and utility by reducing refusals across all evaluated models, regardless of decoding configuration, model size, or reasoning capability. Notably, Llama 3.1 70B outperforms fine-tuned GPT-3.5. Additionally, we examine LLM errors and introduce metrics to quantify semantic divergence in model-generated rationales, revealing notable differences and paradoxical behaviors among LLMs. Our experiments highlight the differences observed across LLMs' utility, explainability, and reliability. Code and resources available at: https://github.com/idramalab/quantify-llm-explanations", "date": "2025-09-22", "categories": []}
{"id": "2510.27038", "title": "Dataset Creation and Baseline Models for Sexism Detection in Hausa", "abstract": "Sexism reinforces gender inequality and social exclusion by perpetuating stereotypes, bias, and discriminatory norms. Noting how online platforms enable various forms of sexism to thrive, there is a growing need for effective sexism detection and mitigation strategies. While computational approaches to sexism detection are widespread in high-resource languages, progress remains limited in low-resource languages where limited linguistic resources and cultural differences affect how sexism is expressed and perceived. This study introduces the first Hausa sexism detection dataset, developed through community engagement, qualitative coding, and data augmentation. For cultural nuances and linguistic representation, we conducted a two-stage user study (n=66) involving native speakers to explore how sexism is defined and articulated in everyday discourse. We further experiment with both traditional machine learning classifiers and pre-trained multilingual language models and evaluating the effectiveness few-shot learning in detecting sexism in Hausa. Our findings highlight challenges in capturing cultural nuance, particularly with clarification-seeking and idiomatic expressions, and reveal a tendency for many false positives in such cases.", "date": "2025-10-30", "categories": []}
{"id": "2505.12116", "title": "A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings", "abstract": "Content moderation research has recently made significant advances, but remains limited in serving the majority of the world's languages due to the lack of resources, leaving millions of vulnerable users to online hostility. This work presents a large-scale human-annotated multi-task benchmark dataset for abusive language detection in Tigrinya social media with joint annotations for three tasks: abusiveness, sentiment, and topic classification. The dataset comprises 13,717 YouTube comments annotated by nine native speakers, collected from 7,373 videos with a total of over 1.2 billion views across 51 channels. We developed an iterative term clustering approach for effective data selection. Recognizing that around 64% of Tigrinya social media content uses Romanized transliterations rather than native Ge'ez script, our dataset accommodates both writing systems to reflect actual language use. We establish strong baselines across the tasks in the benchmark, while leaving significant challenges for future contributions. Our experiments demonstrate that small fine-tuned models outperform prompted frontier large language models (LLMs) in the low-resource setting, achieving 86.67% F1 in abusiveness detection (7+ points over best LLM), and maintain stronger performance in all other tasks. The benchmark is made public to promote research on online safety.", "date": "2025-05-17", "categories": []}
{"id": "2505.14483", "title": "MoMoE: Mixture of Moderation Experts Framework for AI-Assisted Online Governance", "abstract": "Large language models (LLMs) have shown great potential in flagging harmful content in online communities. Yet, existing approaches for moderation require a separate model for every community and are opaque in their decision-making, limiting real-world adoption. We introduce Mixture of Moderation Experts (MoMoE), a modular, cross-community framework that adds post-hoc explanations to scalable content moderation. MoMoE orchestrates four operators -- Allocate, Predict, Aggregate, Explain -- and is instantiated as seven community-specialized experts (MoMoE-Community) and five norm-violation experts (MoMoE-NormVio). On 30 unseen subreddits, the best variants obtain Micro-F1 scores of 0.72 and 0.67, respectively, matching or surpassing strong fine-tuned baselines while consistently producing concise and reliable explanations. Although community-specialized experts deliver the highest peak accuracy, norm-violation experts provide steadier performance across domains. These findings show that MoMoE yields scalable, transparent moderation without needing per-community fine-tuning. More broadly, they suggest that lightweight, explainable expert ensembles can guide future NLP and HCI research on trustworthy human-AI governance of online communities.", "date": "2025-05-20", "categories": []}
{"id": "2510.19882", "title": "Quantifying Feature Importance for Online Content Moderation", "abstract": "Accurately estimating how users respond to moderation interventions is paramount for developing effective and user-centred moderation strategies. However, this requires a clear understanding of which user characteristics are associated with different behavioural responses, which is the goal of this work. We investigate the informativeness of 753 socio-behavioural, linguistic, relational, and psychological features, in predicting the behavioural changes of 16.8K users affected by a major moderation intervention on Reddit. To reach this goal, we frame the problem in terms of \"quantification\", a task well-suited to estimating shifts in aggregate user behaviour. We then apply a greedy feature selection strategy with the double goal of (i) identifying the features that are most predictive of changes in user activity, toxicity, and participation diversity, and (ii) estimating their importance. Our results allow identifying a small set of features that are consistently informative across all tasks, and determining that many others are either task-specific or of limited utility altogether. We also find that predictive performance varies according to the task, with changes in activity and toxicity being easier to estimate than changes in diversity. Overall, our results pave the way for the development of accurate systems that predict user reactions to moderation interventions. Furthermore, our findings highlight the complexity of post-moderation user behaviour, and indicate that effective moderation should be tailored not only to user traits but also to the specific objective of the intervention.", "date": "2025-10-22", "categories": []}
{"id": "2510.17289", "title": "Addressing Antisocial Behavior in Multi-Party Dialogs Through Multimodal Representation Learning", "abstract": "Antisocial behavior (ASB) on social media -- including hate speech, harassment, and cyberbullying -- poses growing risks to platform safety and societal well-being. Prior research has focused largely on networks such as X and Reddit, while \\textit{multi-party conversational settings} remain underexplored due to limited data. To address this gap, we use \\textit{CyberAgressionAdo-Large}, a French open-access dataset simulating ASB in multi-party conversations, and evaluate three tasks: \\textit{abuse detection}, \\textit{bullying behavior analysis}, and \\textit{bullying peer-group identification}. We benchmark six text-based and eight graph-based \\textit{representation-learning methods}, analyzing lexical cues, interactional dynamics, and their multimodal fusion. Results show that multimodal models outperform unimodal baselines. The late fusion model \\texttt{mBERT + WD-SGCN} achieves the best overall results, with top performance on abuse detection (0.718) and competitive scores on peer-group identification (0.286) and bullying analysis (0.606). Error analysis highlights its effectiveness in handling nuanced ASB phenomena such as implicit aggression, role transitions, and context-dependent hostility.", "date": "2025-10-20", "categories": []}
{"id": "2502.11246", "title": "MemeSense: An Adaptive In-Context Framework for Social Commonsense Driven Meme Moderation", "abstract": "Online memes are a powerful yet challenging medium for content moderation, often masking harmful intent behind humor, irony, or cultural symbolism. Conventional moderation systems \"especially those relying on explicit text\" frequently fail to recognize such subtle or implicit harm. We introduce MemeSense, an adaptive framework designed to generate socially grounded interventions for harmful memes by combining visual and textual understanding with curated, semantically aligned examples enriched with commonsense cues. This enables the model to detect nuanced complexed threats like misogyny, stereotyping, or vulgarity \"even in memes lacking overt language\". Across multiple benchmark datasets, MemeSense outperforms state-of-the-art methods, achieving up to 35% higher semantic similarity and 9% improvement in BERTScore for non-textual memes, and notable gains for text-rich memes as well. These results highlight MemeSense as a promising step toward safer, more context-aware AI systems for real-world content moderation. Code and data available at: https://github.com/sayantan11995/MemeSense", "date": "2025-02-16", "categories": []}
{"id": "2510.15685", "title": "Leveraging LLMs for Context-Aware Implicit Textual and Multimodal Hate Speech Detection", "abstract": "This research introduces a novel approach to textual and multimodal Hate Speech Detection (HSD), using Large Language Models (LLMs) as dynamic knowledge bases to generate background context and incorporate it into the input of HSD classifiers. Two context generation strategies are examined: one focused on named entities and the other on full-text prompting. Four methods of incorporating context into the classifier input are compared: text concatenation, embedding concatenation, a hierarchical transformer-based fusion, and LLM-driven text enhancement. Experiments are conducted on the textual Latent Hatred dataset of implicit hate speech and applied in a multimodal setting on the MAMI dataset of misogynous memes. Results suggest that both the contextual information and the method by which it is incorporated are key, with gains of up to 3 and 6 F1 points on textual and multimodal setups respectively, from a zero-context baseline to the highest-performing system, based on embedding concatenation.", "date": "2025-10-17", "categories": []}
{"id": "2510.13837", "title": "Seeing Hate Differently: Hate Subspace Modeling for Culture-Aware Hate Speech Detection", "abstract": "Hate speech detection has been extensively studied, yet existing methods often overlook a real-world complexity: training labels are biased, and interpretations of what is considered hate vary across individuals with different cultural backgrounds. We first analyze these challenges, including data sparsity, cultural entanglement, and ambiguous labeling. To address them, we propose a culture-aware framework that constructs individuals' hate subspaces. To alleviate data sparsity, we model combinations of cultural attributes. For cultural entanglement and ambiguous labels, we use label propagation to capture distinctive features of each combination. Finally, individual hate subspaces, which in turn can further enhance classification performance. Experiments show our method outperforms state-of-the-art by 1.05\\% on average across all metrics.", "date": "2025-10-11", "categories": []}
{"id": "2510.13259", "title": "Hypernetworks for Perspectivist Adaptation", "abstract": "The task of perspective-aware classification introduces a bottleneck in terms of parametric efficiency that did not get enough recognition in existing studies. In this article, we aim to address this issue by applying an existing architecture, the hypernetwork+adapters combination, to perspectivist classification. Ultimately, we arrive at a solution that can compete with specialized models in adopting user perspectives on hate speech and toxicity detection, while also making use of considerably fewer parameters. Our solution is architecture-agnostic and can be applied to a wide range of base models out of the box.", "date": "2025-10-15", "categories": []}
{"id": "2510.12316", "title": "Beating Harmful Stereotypes Through Facts: RAG-based Counter-speech Generation", "abstract": "Counter-speech generation is at the core of many expert activities, such as fact-checking and hate speech, to counter harmful content. Yet, existing work treats counter-speech generation as pure text generation task, mainly based on Large Language Models or NGO experts. These approaches show severe drawbacks due to the limited reliability and coherence in the generated countering text, and in scalability, respectively. To close this gap, we introduce a novel framework to model counter-speech generation as knowledge-wise text generation process. Our framework integrates advanced Retrieval-Augmented Generation (RAG) pipelines to ensure the generation of trustworthy counter-speech for 8 main target groups identified in the hate speech literature, including women, people of colour, persons with disabilities, migrants, Muslims, Jews, LGBT persons, and other. We built a knowledge base over the United Nations Digital Library, EUR-Lex and the EU Agency for Fundamental Rights, comprising a total of 32,792 texts. We use the MultiTarget-CONAN dataset to empirically assess the quality of the generated counter-speech, both through standard metrics (i.e., JudgeLM) and a human evaluation. Results show that our framework outperforms standard LLM baselines and competitive approach, on both assessments. The resulting framework and the knowledge base pave the way for studying trustworthy and sound counter-speech generation, in hate speech and beyond.", "date": "2025-10-14", "categories": []}
{"id": "2410.20490", "title": "Who Speaks Matters: Analysing the Influence of the Speaker's Ethnicity on Hate Classification", "abstract": "Large Language Models (LLMs) offer a lucrative promise for scalable content moderation, including hate speech detection. However, they are also known to be brittle and biased against marginalised communities and dialects. This requires their applications to high-stakes tasks like hate speech detection to be critically scrutinized. In this work, we investigate the robustness of hate speech classification using LLMs particularly when explicit and implicit markers of the speaker's ethnicity are injected into the input. For explicit markers, we inject a phrase that mentions the speaker's linguistic identity. For the implicit markers, we inject dialectal features. By analysing how frequently model outputs flip in the presence of these markers, we reveal varying degrees of brittleness across 3 LLMs and 1 LM and 5 linguistic identities. We find that the presence of implicit dialect markers in inputs causes model outputs to flip more than the presence of explicit markers. Further, the percentage of flips varies across ethnicities. Finally, we find that larger models are more robust. Our findings indicate the need for exercising caution in deploying LLMs for high-stakes tasks like hate speech detection.", "date": "2024-10-27", "categories": []}
{"id": "2506.11919", "title": "Effectiveness of Counter-Speech against Abusive Content: A Multidimensional Annotation and Classification Study", "abstract": "Counter-speech (CS) is a key strategy for mitigating online Hate Speech (HS), yet defining the criteria to assess its effectiveness remains an open challenge. We propose a novel computational framework for CS effectiveness classification, grounded in linguistics, communication and argumentation concepts. Our framework defines six core dimensions - Clarity, Evidence, Emotional Appeal, Rebuttal, Audience Adaptation, and Fairness - which we use to annotate 4,214 CS instances from two benchmark datasets, resulting in a novel linguistic resource released to the community. In addition, we propose two classification strategies, multi-task and dependency-based, achieving strong results (0.94 and 0.96 average F1 respectively on both expert- and user-written CS), outperforming standard baselines, and revealing strong interdependence among dimensions.", "date": "2025-06-13", "categories": []}
{"id": "2510.09935", "title": "Unpacking Hateful Memes: Presupposed Context and False Claims", "abstract": "While memes are often humorous, they are frequently used to disseminate hate, causing serious harm to individuals and society. Current approaches to hateful meme detection mainly rely on pre-trained language models. However, less focus has been dedicated to \\textit{what make a meme hateful}. Drawing on insights from philosophy and psychology, we argue that hateful memes are characterized by two essential features: a \\textbf{presupposed context} and the expression of \\textbf{false claims}. To capture presupposed context, we develop \\textbf{PCM} for modeling contextual information across modalities. To detect false claims, we introduce the \\textbf{FACT} module, which integrates external knowledge and harnesses cross-modal reference graphs. By combining PCM and FACT, we introduce \\textbf{\\textsf{SHIELD}}, a hateful meme detection framework designed to capture the fundamental nature of hate. Extensive experiments show that SHIELD outperforms state-of-the-art methods across datasets and metrics, while demonstrating versatility on other tasks, such as fake news detection.", "date": "2025-10-11", "categories": []}
{"id": "2510.11167", "title": "Bridging Gaps in Hate Speech Detection: Meta-Collections and Benchmarks for Low-Resource Iberian Languages", "abstract": "Hate speech poses a serious threat to social cohesion and individual well-being, particularly on social media, where it spreads rapidly. While research on hate speech detection has progressed, it remains largely focused on English, resulting in limited resources and benchmarks for low-resource languages. Moreover, many of these languages have multiple linguistic varieties, a factor often overlooked in current approaches. At the same time, large language models require substantial amounts of data to perform reliably, a requirement that low-resource languages often cannot meet. In this work, we address these gaps by compiling a meta-collection of hate speech datasets for European Spanish, standardised with unified labels and metadata. This collection is based on a systematic analysis and integration of existing resources, aiming to bridge the data gap and support more consistent and scalable hate speech detection. We extended this collection by translating it into European Portuguese and into a Galician standard that is more convergent with Spanish and another Galician variant that is more convergent with Portuguese, creating aligned multilingual corpora. Using these resources, we establish new benchmarks for hate speech detection in Iberian languages. We evaluate state-of-the-art large language models in zero-shot, few-shot, and fine-tuning settings, providing baseline results for future research. Moreover, we perform a cross-lingual analysis with our target languages. Our findings underscore the importance of multilingual and variety-aware approaches in hate speech detection and offer a foundation for improved benchmarking in underrepresented European languages.", "date": "2025-10-13", "categories": []}
{"id": "2509.21787", "title": "DeHate: A Stable Diffusion-based Multimodal Approach to Mitigate Hate Speech in Images", "abstract": "The rise in harmful online content not only distorts public discourse but also poses significant challenges to maintaining a healthy digital environment. In response to this, we introduce a multimodal dataset uniquely crafted for identifying hate in digital content. Central to our methodology is the innovative application of watermarked, stability-enhanced, stable diffusion techniques combined with the Digital Attention Analysis Module (DAAM). This combination is instrumental in pinpointing the hateful elements within images, thereby generating detailed hate attention maps, which are used to blur these regions from the image, thereby removing the hateful sections of the image. We release this data set as a part of the dehate shared task. This paper also describes the details of the shared task. Furthermore, we present DeHater, a vision-language model designed for multimodal dehatification tasks. Our approach sets a new standard in AI-driven image hate detection given textual prompts, contributing to the development of more ethical AI applications in social media.", "date": "2025-09-26", "categories": []}
{"id": "2510.03683", "title": "Fine-Tuning Large Language Models with QLoRA for Offensive Language Detection in Roman Urdu-English Code-Mixed Text", "abstract": "The use of derogatory terms in languages that employ code mixing, such as Roman Urdu, presents challenges for Natural Language Processing systems due to unstated grammar, inconsistent spelling, and a scarcity of labeled data. In this work, we propose a QLoRA based fine tuning framework to improve offensive language detection in Roman Urdu-English text. We translated the Roman Urdu-English code mixed dataset into English using Google Translate to leverage English LLMs, while acknowledging that this translation reduces direct engagement with code mixing features. Our focus is on classification performance using English translated low resource inputs. We fine tuned several transformers and large language models, including Meta LLaMA 3 8B, Mistral 7B v0.1, LLaMA 2 7B, ModernBERT, and RoBERTa, with QLoRA for memory efficient adaptation. Models were trained and evaluated on a manually annotated Roman Urdu dataset for offensive vs non offensive content. Of all tested models, the highest F1 score of 91.45 was attained by Meta LLaMA 3 8B, followed by Mistral 7B at 89.66, surpassing traditional transformer baselines. These results demonstrate the efficacy of QLoRA in fine tuning high performing models for low resource environments such as code mixed offensive language detection, and confirm the potential of LLMs for this task. This work advances a scalable approach to Roman Urdu moderation and paves the way for future multilingual offensive detection systems based on LLMs.", "date": "2025-10-04", "categories": []}
{"id": "2510.02351", "title": "Language, Culture, and Ideology: Personalizing Offensiveness Detection in Political Tweets with Reasoning LLMs", "abstract": "We explore how large language models (LLMs) assess offensiveness in political discourse when prompted to adopt specific political and cultural perspectives. Using a multilingual subset of the MD-Agreement dataset centered on tweets from the 2020 US elections, we evaluate several recent LLMs - including DeepSeek-R1, o4-mini, GPT-4.1-mini, Qwen3, Gemma, and Mistral - tasked with judging tweets as offensive or non-offensive from the viewpoints of varied political personas (far-right, conservative, centrist, progressive) across English, Polish, and Russian contexts. Our results show that larger models with explicit reasoning abilities (e.g., DeepSeek-R1, o4-mini) are more consistent and sensitive to ideological and cultural variation, while smaller models often fail to capture subtle distinctions. We find that reasoning capabilities significantly improve both the personalization and interpretability of offensiveness judgments, suggesting that such mechanisms are key to adapting LLMs for nuanced sociopolitical text classification across languages and ideologies.", "date": "2025-09-27", "categories": []}
{"id": "2510.01995", "title": "LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target", "abstract": "Online social media platforms are central to everyday communication and information seeking. While these platforms serve positive purposes, they also provide fertile ground for the spread of hate speech, offensive language, and bullying content targeting individuals, organizations, and communities. Such content undermines safety, participation, and equity online. Reliable detection systems are therefore needed, especially for low-resource languages where moderation tools are limited. In Bangla, prior work has contributed resources and models, but most are single-task (e.g., binary hate/offense) with limited coverage of multi-facet signals (type, severity, target). We address these gaps by introducing the first multi-task Bangla hate-speech dataset, BanglaMultiHate, one of the largest manually annotated corpus to date. Building on this resource, we conduct a comprehensive, controlled comparison spanning classical baselines, monolingual pretrained models, and LLMs under zero-shot prompting and LoRA fine-tuning. Our experiments assess LLM adaptability in a low-resource setting and reveal a consistent trend: although LoRA-tuned LLMs are competitive with BanglaBERT, culturally and linguistically grounded pretraining remains critical for robust performance. Together, our dataset and findings establish a stronger benchmark for developing culturally aligned moderation tools in low-resource contexts. For reproducibility, we will release the dataset and all related scripts.", "date": "2025-10-02", "categories": []}
{"id": "2401.09333", "title": "Machines Do See Color: A Guideline to Classify Different Forms of Racist Discourse in Large Corpora", "abstract": "Current methods to identify and classify racist language in text rely on small-n qualitative approaches or large-n approaches focusing exclusively on overt forms of racist discourse. This article provides a step-by-step generalizable guideline to identify and classify different forms of racist discourse in large corpora. In our approach, we start by conceptualizing racism and its different manifestations. We then contextualize these racist manifestations to the time and place of interest, which allows researchers to identify their discursive form. Finally, we apply XLM-RoBERTa (XLM-R), a cross-lingual model for supervised text classification with a cutting-edge contextual understanding of text. We show that XLM-R and XLM-R-Racismo, our pretrained model, outperform other state-of-the-art approaches in classifying racism in large corpora. We illustrate our approach using a corpus of tweets relating to the Ecuadorian ind\\'igena community between 2018 and 2021.", "date": "2024-01-17", "categories": []}
{"id": "2507.15339", "title": "LionGuard 2: Building Lightweight, Data-Efficient & Localised Multilingual Content Moderators", "abstract": "Modern moderation systems increasingly support multiple languages, but often fail to address localisation and low-resource variants - creating safety gaps in real-world deployments. Small models offer a potential alternative to large LLMs, yet still demand considerable data and compute. We present LionGuard 2, a lightweight, multilingual moderation classifier tailored to the Singapore context, supporting English, Chinese, Malay, and partial Tamil. Built on pre-trained OpenAI embeddings and a multi-head ordinal classifier, LionGuard 2 outperforms several commercial and open-source systems across 17 benchmarks, including both Singapore-specific and public English datasets. The system is actively deployed within the Singapore Government, demonstrating practical efficacy at scale. Our findings show that high-quality local data and robust multilingual embeddings can achieve strong moderation performance, without fine-tuning large models. We release our model weights and part of our training data to support future work on LLM safety.", "date": "2025-07-21", "categories": []}
{"id": "2509.24215", "title": "Metamorphic Testing for Audio Content Moderation Software", "abstract": "The rapid growth of audio-centric platforms and applications such as WhatsApp and Twitter has transformed the way people communicate and share audio content in modern society. However, these platforms are increasingly misused to disseminate harmful audio content, such as hate speech, deceptive advertisements, and explicit material, which can have significant negative consequences (e.g., detrimental effects on mental health). In response, researchers and practitioners have been actively developing and deploying audio content moderation tools to tackle this issue. Despite these efforts, malicious actors can bypass moderation systems by making subtle alterations to audio content, such as modifying pitch or inserting noise. Moreover, the effectiveness of modern audio moderation tools against such adversarial inputs remains insufficiently studied. To address these challenges, we propose MTAM, a Metamorphic Testing framework for Audio content Moderation software. Specifically, we conduct a pilot study on 2000 audio clips and define 14 metamorphic relations across two perturbation categories: Audio Features-Based and Heuristic perturbations. MTAM applies these metamorphic relations to toxic audio content to generate test cases that remain harmful while being more likely to evade detection. In our evaluation, we employ MTAM to test five commercial textual content moderation software and an academic model against three kinds of toxic content. The results show that MTAM achieves up to 38.6%, 18.3%, 35.1%, 16.7%, and 51.1% error finding rates (EFR) when testing commercial moderation software provided by Gladia, Assembly AI, Baidu, Nextdata, and Tencent, respectively, and it obtains up to 45.7% EFR when testing the state-of-the-art algorithms from the academy.", "date": "2025-09-29", "categories": []}
{"id": "2509.18562", "title": "CPCLDETECTOR: Knowledge Enhancement and Alignment Selection for Chinese Patronizing and Condescending Language Detection", "abstract": "Chinese Patronizing and Condescending Language (CPCL) is an implicitly discriminatory toxic speech targeting vulnerable groups on Chinese video platforms. The existing dataset lacks user comments, which are a direct reflection of video content. This undermines the model's understanding of video content and results in the failure to detect some CPLC videos. To make up for this loss, this research reconstructs a new dataset PCLMMPLUS that includes 103k comment entries and expands the dataset size. We also propose the CPCLDetector model with alignment selection and knowledge-enhanced comment content modules. Extensive experiments show the proposed CPCLDetector outperforms the SOTA on PCLMM and achieves higher performance on PCLMMPLUS . CPLC videos are detected more accurately, supporting content governance and protecting vulnerable groups. Code and dataset are available at https://github.com/jiaxunyang256/PCLD.", "date": "2025-09-23", "categories": []}
{"id": "2505.19528", "title": "AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection", "abstract": "Implicit hate speech detection is challenging due to its subtlety and reliance on contextual interpretation rather than explicit offensive words. Current approaches rely on contrastive learning, which are shown to be effective on distinguishing hate and non-hate sentences. Humans, however, detect implicit hate speech by first identifying specific targets within the text and subsequently interpreting how these target relate to their surrounding context. Motivated by this reasoning process, we propose AmpleHate, a novel approach designed to mirror human inference for implicit hate detection. AmpleHate identifies explicit target using a pretrained Named Entity Recognition model and capture implicit target information via [CLS] tokens. It computes attention-based relationships between explicit, implicit targets and sentence context and then, directly injects these relational vectors into the final sentence representation. This amplifies the critical signals of target-context relations for determining implicit hate. Experiments demonstrate that AmpleHate achieves state-of-the-art performance, outperforming contrastive learning baselines by an average of 82.14% and achieve faster convergence. Qualitative analyses further reveal that attention patterns produced by AmpleHate closely align with human judgement, underscoring its interpretability and robustness. Our code is publicly available at: https://github.com/leeyejin1231/AmpleHate.", "date": "2025-05-26", "categories": []}
{"id": "2509.14264", "title": "Defining, Understanding, and Detecting Online Toxicity: Challenges and Machine Learning Approaches", "abstract": "Online toxic content has grown into a pervasive phenomenon, intensifying during times of crisis, elections, and social unrest. A significant amount of research has been focused on detecting or analyzing toxic content using machine-learning approaches. The proliferation of toxic content across digital platforms has spurred extensive research into automated detection mechanisms, primarily driven by advances in machine learning and natural language processing. Overall, the present study represents the synthesis of 140 publications on different types of toxic content on digital platforms. We present a comprehensive overview of the datasets used in previous studies focusing on definitions, data sources, challenges, and machine learning approaches employed in detecting online toxicity, such as hate speech, offensive language, and harmful discourse. The dataset encompasses content in 32 languages, covering topics such as elections, spontaneous events, and crises. We examine the possibility of using existing cross-platform data to improve the performance of classification models. We present the recommendations and guidelines for new research on online toxic consent and the use of content moderation for mitigation. Finally, we present some practical guidelines to mitigate toxic content from online platforms.", "date": "2025-09-14", "categories": []}
{"id": "2509.14266", "title": "Efficient Hate Speech Detection: Evaluating 38 Models from Traditional Methods to Transformers", "abstract": "The proliferation of hate speech on social media necessitates automated detection systems that balance accuracy with computational efficiency. This study evaluates 38 model configurations in detecting hate speech across datasets ranging from 6.5K to 451K samples. We analyze transformer architectures (e.g., BERT, RoBERTa, Distil-BERT), deep neural networks (e.g., CNN, LSTM, GRU, Hierarchical Attention Networks), and traditional machine learning methods (e.g., SVM, CatBoost, Random Forest). Our results show that transformers, particularly RoBERTa, consistently achieve superior performance with accuracy and F1-scores exceeding 90%. Among deep learning approaches, Hierarchical Attention Networks yield the best results, while traditional methods like CatBoost and SVM remain competitive, achieving F1-scores above 88% with significantly lower computational costs. Additionally, our analysis highlights the importance of dataset characteristics, with balanced, moderately sized unprocessed datasets outperforming larger, preprocessed datasets. These findings offer valuable insights for developing efficient and effective hate speech detection systems.", "date": "2025-09-14", "categories": []}
{"id": "2509.14712", "title": "From Ground Trust to Truth: Disparities in Offensive Language Judgments on Contemporary Korean Political Discourse", "abstract": "Although offensive language continually evolves over time, even recent studies using LLMs have predominantly relied on outdated datasets and rarely evaluated the generalization ability on unseen texts. In this study, we constructed a large-scale dataset of contemporary political discourse and employed three refined judgments in the absence of ground truth. Each judgment reflects a representative offensive language detection method and is carefully designed for optimal conditions. We identified distinct patterns for each judgment and demonstrated tendencies of label agreement using a leave-one-out strategy. By establishing pseudo-labels as ground trust for quantitative performance assessment, we observed that a strategically designed single prompting achieves comparable performance to more resource-intensive methods. This suggests a feasible approach applicable in real-world settings with inherent constraints.", "date": "2025-09-18", "categories": []}
{"id": "2509.13515", "title": "Multimodal Hate Detection Using Dual-Stream Graph Neural Networks", "abstract": "Hateful videos present serious risks to online safety and real-world well-being, necessitating effective detection methods. Although multimodal classification approaches integrating information from several modalities outperform unimodal ones, they typically neglect that even minimal hateful content defines a video's category. Specifically, they generally treat all content uniformly, instead of emphasizing the hateful components. Additionally, existing multimodal methods cannot systematically capture structured information in videos, limiting the effectiveness of multimodal fusion. To address these limitations, we propose a novel multimodal dual-stream graph neural network model. It constructs an instance graph by separating the given video into several instances to extract instance-level features. Then, a complementary weight graph assigns importance weights to these features, highlighting hateful instances. Importance weights and instance features are combined to generate video labels. Our model employs a graph-based framework to systematically model structured relationships within and across modalities. Extensive experiments on public datasets show that our model is state-of-the-art in hateful video classification and has strong explainability. Code is available: https://github.com/Multimodal-Intelligence-Lab-MIL/MultiHateGNN.", "date": "2025-09-16", "categories": []}
{"id": "2509.07768", "title": "Are LLMs Enough for Hyperpartisan, Fake, Polarized and Harmful Content Detection? Evaluating In-Context Learning vs. Fine-Tuning", "abstract": "The spread of fake news, polarizing, politically biased, and harmful content on online platforms has been a serious concern. With large language models becoming a promising approach, however, no study has properly benchmarked their performance across different models, usage methods, and languages. This study presents a comprehensive overview of different Large Language Models adaptation paradigms for the detection of hyperpartisan and fake news, harmful tweets, and political bias. Our experiments spanned 10 datasets and 5 different languages (English, Spanish, Portuguese, Arabic and Bulgarian), covering both binary and multiclass classification scenarios. We tested different strategies ranging from parameter efficient Fine-Tuning of language models to a variety of different In-Context Learning strategies and prompts. These included zero-shot prompts, codebooks, few-shot (with both randomly-selected and diversely-selected examples using Determinantal Point Processes), and Chain-of-Thought. We discovered that In-Context Learning often underperforms when compared to Fine-Tuning a model. This main finding highlights the importance of Fine-Tuning even smaller models on task-specific settings even when compared to the largest models evaluated in an In-Context Learning setup - in our case LlaMA3.1-8b-Instruct, Mistral-Nemo-Instruct-2407 and Qwen2.5-7B-Instruct.", "date": "2025-09-09", "categories": []}
{"id": "2508.05557", "title": "MV-Debate: Multi-view Agent Debate with Dynamic Reflection Gating for Multimodal Harmful Content Detection in Social Media", "abstract": "Social media has evolved into a complex multimodal environment where text, images, and other signals interact to shape nuanced meanings, often concealing harmful intent. Identifying such intent, whether sarcasm, hate speech, or misinformation, remains challenging due to cross-modal contradictions, rapid cultural shifts, and subtle pragmatic cues. To address these challenges, we propose MV-Debate, a multi-view agent debate framework with dynamic reflection gating for unified multimodal harmful content detection. MV-Debate assembles four complementary debate agents, a surface analyst, a deep reasoner, a modality contrast, and a social contextualist, to analyze content from diverse interpretive perspectives. Through iterative debate and reflection, the agents refine responses under a reflection-gain criterion, ensuring both accuracy and efficiency. Experiments on three benchmark datasets demonstrate that MV-Debate significantly outperforms strong single-model and existing multi-agent debate baselines. This work highlights the promise of multi-agent debate in advancing reliable social intent detection in safety-critical online contexts.", "date": "2025-08-07", "categories": []}
{"id": "2509.02926", "title": "Decoding the Rule Book: Extracting Hidden Moderation Criteria from Reddit Communities", "abstract": "Effective content moderation systems require explicit classification criteria, yet online communities like subreddits often operate with diverse, implicit standards. This work introduces a novel approach to identify and extract these implicit criteria from historical moderation data using an interpretable architecture. We represent moderation criteria as score tables of lexical expressions associated with content removal, enabling systematic comparison across different communities. Our experiments demonstrate that these extracted lexical patterns effectively replicate the performance of neural moderation models while providing transparent insights into decision-making processes. The resulting criteria matrix reveals significant variations in how seemingly shared norms are actually enforced, uncovering previously undocumented moderation patterns including community-specific tolerances for language, features for topical restrictions, and underlying subcategories of the toxic speech classification.", "date": "2025-09-03", "categories": []}
{"id": "2509.00696", "title": "Queuing for Civility: Regulating Emotions and Reducing Toxicity in Digital Discourse", "abstract": "The pervasiveness of online toxicity, including hate speech and trolling, disrupts digital interactions and online well-being. Previous research has mainly focused on post-hoc moderation, overlooking the real-time emotional dynamics of online conversations and the impact of users' emotions on others. This paper presents a graph-based framework to identify the need for emotion regulation within online conversations. This framework promotes self-reflection to manage emotional responses and encourage responsible behaviour in real time. Additionally, a comment queuing mechanism is proposed to address intentional trolls who exploit emotions to inflame conversations. This mechanism introduces a delay in publishing comments, giving users time to self-regulate before further engaging in the conversation and helping maintain emotional balance. Analysis of social media data from Twitter and Reddit demonstrates that the graph-based framework reduced toxicity by 12%, while the comment queuing mechanism decreased the spread of anger by 15%, with only 4% of comments being temporarily held on average. These findings indicate that combining real-time emotion regulation with delayed moderation can significantly improve well-being in online environments.", "date": "2025-08-31", "categories": []}
{"id": "2509.01379", "title": "WATCHED: A Web AI Agent Tool for Combating Hate Speech by Expanding Data", "abstract": "Online harms are a growing problem in digital spaces, putting user safety at risk and reducing trust in social media platforms. One of the most persistent forms of harm is hate speech. To address this, we need tools that combine the speed and scale of automated systems with the judgment and insight of human moderators. These tools should not only find harmful content but also explain their decisions clearly, helping to build trust and understanding. In this paper, we present WATCHED, a chatbot designed to support content moderators in tackling hate speech. The chatbot is built as an Artificial Intelligence Agent system that uses Large Language Models along with several specialised tools. It compares new posts with real examples of hate speech and neutral content, uses a BERT-based classifier to help flag harmful messages, looks up slang and informal language using sources like Urban Dictionary, generates chain-of-thought reasoning, and checks platform guidelines to explain and support its decisions. This combination allows the chatbot not only to detect hate speech but to explain why content is considered harmful, grounded in both precedent and policy. Experimental results show that our proposed method surpasses existing state-of-the-art methods, reaching a macro F1 score of 0.91. Designed for moderators, safety teams, and researchers, the tool helps reduce online harms by supporting collaboration between AI and human oversight.", "date": "2025-09-01", "categories": []}
{"id": "2411.04090", "title": "A Collaborative Content Moderation Framework for Toxicity Detection based on Conformalized Estimates of Annotation Disagreement", "abstract": "Content moderation typically combines the efforts of human moderators and machine learning models. However, these systems often rely on data where significant disagreement occurs during moderation, reflecting the subjective nature of toxicity perception. Rather than dismissing this disagreement as noise, we interpret it as a valuable signal that highlights the inherent ambiguity of the content,an insight missed when only the majority label is considered. In this work, we introduce a novel content moderation framework that emphasizes the importance of capturing annotation disagreement. Our approach uses multitask learning, where toxicity classification serves as the primary task and annotation disagreement is addressed as an auxiliary task. Additionally, we leverage uncertainty estimation techniques, specifically Conformal Prediction, to account for both the ambiguity in comment annotations and the model's inherent uncertainty in predicting toxicity and disagreement.The framework also allows moderators to adjust thresholds for annotation disagreement, offering flexibility in determining when ambiguity should trigger a review. We demonstrate that our joint approach enhances model performance, calibration, and uncertainty estimation, while offering greater parameter efficiency and improving the review process in comparison to single-task methods.", "date": "2024-11-06", "categories": []}
{"id": "2501.12640", "title": "Toxicity Begets Toxicity: Unraveling Conversational Chains in Political Podcasts", "abstract": "Tackling toxic behavior in digital communication continues to be a pressing concern for both academics and industry professionals. While significant research has explored toxicity on platforms like social networks and discussion boards, podcasts despite their rapid rise in popularity remain relatively understudied in this context. This work seeks to fill that gap by curating a dataset of political podcast transcripts and analyzing them with a focus on conversational structure. Specifically, we investigate how toxicity surfaces and intensifies through sequences of replies within these dialogues, shedding light on the organic patterns by which harmful language can escalate across conversational turns. Warning: Contains potentially abusive/toxic contents.", "date": "2025-01-22", "categories": []}
{"id": "2508.20546", "title": "MM-HSD: Multi-Modal Hate Speech Detection in Videos", "abstract": "While hate speech detection (HSD) has been extensively studied in text, existing multi-modal approaches remain limited, particularly in videos. As modalities are not always individually informative, simple fusion methods fail to fully capture inter-modal dependencies. Moreover, previous work often omits relevant modalities such as on-screen text and audio, which may contain subtle hateful content and thus provide essential cues, both individually and in combination with others. In this paper, we present MM-HSD, a multi-modal model for HSD in videos that integrates video frames, audio, and text derived from speech transcripts and from frames (i.e.~on-screen text) together with features extracted by Cross-Modal Attention (CMA). We are the first to use CMA as an early feature extractor for HSD in videos, to systematically compare query/key configurations, and to evaluate the interactions between different modalities in the CMA block. Our approach leads to improved performance when on-screen text is used as a query and the rest of the modalities serve as a key. Experiments on the HateMM dataset show that MM-HSD outperforms state-of-the-art methods on M-F1 score (0.874), using concatenation of transcript, audio, video, on-screen text, and CMA for feature extraction on raw embeddings of the modalities. The code is available at https://github.com/idiap/mm-hsd", "date": "2025-08-28", "categories": []}
{"id": "2508.20750", "title": "Specializing General-purpose LLM Embeddings for Implicit Hate Speech Detection across Datasets", "abstract": "Implicit hate speech (IHS) is indirect language that conveys prejudice or hatred through subtle cues, sarcasm or coded terminology. IHS is challenging to detect as it does not include explicit derogatory or inflammatory words. To address this challenge, task-specific pipelines can be complemented with external knowledge or additional information such as context, emotions and sentiment data. In this paper, we show that, by solely fine-tuning recent general-purpose embedding models based on large language models (LLMs), such as Stella, Jasper, NV-Embed and E5, we achieve state-of-the-art performance. Experiments on multiple IHS datasets show up to 1.10 percentage points improvements for in-dataset, and up to 20.35 percentage points improvements in cross-dataset evaluation, in terms of F1-macro score.", "date": "2025-08-28", "categories": []}
{"id": "2406.04876", "title": "HateDebias: On the Diversity and Variability of Hate Speech Debiasing", "abstract": "Hate speech frequently appears on social media platforms and urgently needs to be effectively controlled. Alleviating the bias caused by hate speech can help resolve various ethical issues. Although existing research has constructed several datasets for hate speech detection, these datasets seldom consider the diversity and variability of bias, making them far from real-world scenarios. To fill this gap, we propose a benchmark HateDebias to analyze the fairness of models under dynamically evolving environments. Specifically, to meet the diversity of biases, we collect hate speech data with different types of biases from real-world scenarios. To further simulate the variability in the real-world scenarios(i.e., the changing of bias attributes in datasets), we construct a dataset to follow the continuous learning setting and evaluate the detection accuracy of models on the HateDebias, where performance degradation indicates a significant bias toward a specific attribute. To provide a potential direction, we further propose a continual debiasing framework tailored to dynamic bias in real-world scenarios, integrating memory replay and bias information regularization to ensure the fairness of the model. Experiment results on the HateDebias benchmark reveal that our methods achieve improved performance in mitigating dynamic biases in real-world scenarios, highlighting the practicality in real-world applications.", "date": "2024-06-07", "categories": []}
{"id": "2508.15810", "title": "Detecting Hope, Hate, and Emotion in Arabic Textual Speech and Multi-modal Memes Using Large Language Models", "abstract": "The rise of social media and online communication platforms has led to the spread of Arabic textual posts and memes as a key form of digital expression. While these contents can be humorous and informative, they are also increasingly being used to spread offensive language and hate speech. Consequently, there is a growing demand for precise analysis of content in Arabic text and memes. This paper explores the potential of large language models to effectively identify hope, hate speech, offensive language, and emotional expressions within such content. We evaluate the performance of base LLMs, fine-tuned LLMs, and pre-trained embedding models. The evaluation is conducted using a dataset of Arabic textual speech and memes proposed in the ArabicNLP MAHED 2025 challenge. The results underscore the capacity of LLMs such as GPT-4o-mini, fine-tuned with Arabic textual speech, and Gemini Flash 2.5, fine-tuned with Arabic memes, to deliver the superior performance. They achieve up to 72.1%, 57.8%, and 79.6% macro F1 scores for tasks 1, 2, and 3, respectively, and secure first place overall in the Mahed 2025 challenge. The proposed solutions offer a more nuanced understanding of both text and memes for accurate and efficient Arabic content moderation systems.", "date": "2025-08-15", "categories": []}
{"id": "2508.16555", "title": "Transfer Learning via Lexical Relatedness: A Sarcasm and Hate Speech Case Study", "abstract": "Detecting hate speech in non-direct forms, such as irony, sarcasm, and innuendos, remains a persistent challenge for social networks. Although sarcasm and hate speech are regarded as distinct expressions, our work explores whether integrating sarcasm as a pre-training step improves implicit hate speech detection and, by extension, explicit hate speech detection. Incorporating samples from ETHOS, Sarcasm on Reddit, and Implicit Hate Corpus, we devised two training strategies to compare the effectiveness of sarcasm pre-training on a CNN+LSTM and BERT+BiLSTM model. The first strategy is a single-step training approach, where a model trained only on sarcasm is then tested on hate speech. The second strategy uses sequential transfer learning to fine-tune models for sarcasm, implicit hate, and explicit hate. Our results show that sarcasm pre-training improved the BERT+BiLSTM's recall by 9.7%, AUC by 7.8%, and F1-score by 6% on ETHOS. On the Implicit Hate Corpus, precision increased by 7.8% when tested only on implicit samples. By incorporating sarcasm into the training process, we show that models can more effectively detect both implicit and explicit hate.", "date": "2025-08-22", "categories": []}
{"id": "2502.15860", "title": "Synthetic vs. Gold: The Role of LLM Generated Labels and Data in Cyberbullying Detection", "abstract": "Cyberbullying (CB) presents a pressing threat, especially to children, underscoring the urgent need for robust detection systems to ensure online safety. While large-scale datasets on online abuse exist, there remains a significant gap in labeled data that specifically reflects the language and communication styles used by children. The acquisition of such data from vulnerable populations, such as children, is challenging due to ethical, legal and technical barriers. Moreover, the creation of these datasets relies heavily on human annotation, which not only strains resources but also raises significant concerns due to annotators exposure to harmful content. In this paper, we address these challenges by leveraging Large Language Models (LLMs) to generate synthetic data and labels. Our experiments demonstrate that synthetic data enables BERT-based CB classifiers to achieve performance close to that of those trained on fully authentic datasets (75.8% vs. 81.5% accuracy). Additionally, LLMs can effectively label authentic yet unlabeled data, allowing BERT classifiers to attain a comparable performance level (79.1% vs. 81.5% accuracy). These results highlight the potential of LLMs as a scalable, ethical, and cost-effective solution for generating data for CB detection.", "date": "2025-02-21", "categories": []}
{"id": "2504.11169", "title": "MuSeD: A Multimodal Spanish Dataset for Sexism Detection in Social Media Videos", "abstract": "Sexism is generally defined as prejudice and discrimination based on sex or gender, affecting every sector of society, from social institutions to relationships and individual behavior. Social media platforms amplify the impact of sexism by conveying discriminatory content not only through text but also across multiple modalities, highlighting the critical need for a multimodal approach to the analysis of sexism online. With the rise of social media platforms where users share short videos, sexism is increasingly spreading through video content. Automatically detecting sexism in videos is a challenging task, as it requires analyzing the combination of verbal, audio, and visual elements to identify sexist content. In this study, (1) we introduce MuSeD, a new Multimodal Spanish dataset for Sexism Detection consisting of $\\approx$ 11 hours of videos extracted from TikTok and BitChute; (2) we propose an innovative annotation framework for analyzing the contributions of textual, vocal, and visual modalities to the classification of content as either sexist or non-sexist; and (3) we evaluate a range of large language models (LLMs) and multimodal LLMs on the task of sexism detection. We find that visual information plays a key role in labeling sexist content for both humans and models. Models effectively detect explicit sexism; however, they struggle with implicit cases, such as stereotypes, instances where annotators also show low agreement. This highlights the inherent difficulty of the task, as identifying implicit sexism depends on the social and cultural context.", "date": "2025-04-15", "categories": []}
{"id": "2508.12828", "title": "Context Matters: Incorporating Target Awareness in Conversational Abusive Language Detection", "abstract": "Abusive language detection has become an increasingly important task as a means to tackle this type of harmful content in social media. There has been a substantial body of research developing models for determining if a social media post is abusive or not; however, this research has primarily focused on exploiting social media posts individually, overlooking additional context that can be derived from surrounding posts. In this study, we look at conversational exchanges, where a user replies to an earlier post by another user (the parent tweet). We ask: does leveraging context from the parent tweet help determine if a reply post is abusive or not, and what are the features that contribute the most? We study a range of content-based and account-based features derived from the context, and compare this to the more widely studied approach of only looking at the features from the reply tweet. For a more generalizable study, we test four different classification models on a dataset made of conversational exchanges (parent-reply tweet pairs) with replies labeled as abusive or not. Our experiments show that incorporating contextual features leads to substantial improvements compared to the use of features derived from the reply tweet only, confirming the importance of leveraging context. We observe that, among the features under study, it is especially the content-based features (what is being posted) that contribute to the classification performance rather than account-based features (who is posting it). While using content-based features, it is best to combine a range of different features to ensure improved performance over being more selective and using fewer features. Our study provides insights into the development of contextualized abusive language detection models in realistic settings involving conversations.", "date": "2025-08-18", "categories": []}
{"id": "2508.01712", "title": "HateClipSeg: A Segment-Level Annotated Dataset for Fine-Grained Hate Video Detection", "abstract": "Detecting hate speech in videos remains challenging due to the complexity of multimodal content and the lack of fine-grained annotations in existing datasets. We present HateClipSeg, a large-scale multimodal dataset with both video-level and segment-level annotations, comprising over 11,714 segments labeled as Normal or across five Offensive categories: Hateful, Insulting, Sexual, Violence, Self-Harm, along with explicit target victim labels. Our three-stage annotation process yields high inter-annotator agreement (Krippendorff's alpha = 0.817). We propose three tasks to benchmark performance: (1) Trimmed Hateful Video Classification, (2) Temporal Hateful Video Localization, and (3) Online Hateful Video Classification. Results highlight substantial gaps in current models, emphasizing the need for more sophisticated multimodal and temporally aware approaches. The HateClipSeg dataset are publicly available at https://github.com/Social-AI-Studio/HateClipSeg.git.", "date": "2025-08-03", "categories": []}
{"id": "2508.06570", "title": "ImpliHateVid: A Benchmark Dataset and Two-stage Contrastive Learning Framework for Implicit Hate Speech Detection in Videos", "abstract": "The existing research has primarily focused on text and image-based hate speech detection, video-based approaches remain underexplored. In this work, we introduce a novel dataset, ImpliHateVid, specifically curated for implicit hate speech detection in videos. ImpliHateVid consists of 2,009 videos comprising 509 implicit hate videos, 500 explicit hate videos, and 1,000 non-hate videos, making it one of the first large-scale video datasets dedicated to implicit hate detection. We also propose a novel two-stage contrastive learning framework for hate speech detection in videos. In the first stage, we train modality-specific encoders for audio, text, and image using contrastive loss by concatenating features from the three encoders. In the second stage, we train cross-encoders using contrastive learning to refine multimodal representations. Additionally, we incorporate sentiment, emotion, and caption-based features to enhance implicit hate detection. We evaluate our method on two datasets, ImpliHateVid for implicit hate speech detection and another dataset for general hate speech detection in videos, HateMM dataset, demonstrating the effectiveness of the proposed multimodal contrastive learning for hateful content detection in videos and the significance of our dataset.", "date": "2025-08-07", "categories": []}
{"id": "2508.11434", "title": "Online Anti-sexist Speech: Identifying Resistance to Gender Bias in Political Discourse", "abstract": "Anti-sexist speech, i.e., public expressions that challenge or resist gendered abuse and sexism, plays a vital role in shaping democratic debate online. Yet automated content moderation systems, increasingly powered by large language models (LLMs), may struggle to distinguish such resistance from the sexism it opposes. This study examines how five LLMs classify sexist, anti-sexist, and neutral political tweets from the UK, focusing on high-salience trigger events involving female Members of Parliament in the year 2022. Our analysis show that models frequently misclassify anti-sexist speech as harmful, particularly during politically charged events where rhetorical styles of harm and resistance converge. These errors risk silencing those who challenge sexism, with disproportionate consequences for marginalised voices. We argue that moderation design must move beyond binary harmful/not-harmful schemas, integrate human-in-the-loop review during sensitive events, and explicitly include counter-speech in training data. By linking feminist scholarship, event-based analysis, and model evaluation, this work highlights the sociotechnical challenges of safeguarding resistance speech in digital political spaces.", "date": "2025-08-15", "categories": []}
{"id": "2508.08274", "title": "Distilling Knowledge from Large Language Models: A Concept Bottleneck Model for Hate and Counter Speech Recognition", "abstract": "The rapid increase in hate speech on social media has exposed an unprecedented impact on society, making automated methods for detecting such content important. Unlike prior black-box models, we propose a novel transparent method for automated hate and counter speech recognition, i.e., \"Speech Concept Bottleneck Model\" (SCBM), using adjectives as human-interpretable bottleneck concepts. SCBM leverages large language models (LLMs) to map input texts to an abstract adjective-based representation, which is then sent to a light-weight classifier for downstream tasks. Across five benchmark datasets spanning multiple languages and platforms (e.g., Twitter, Reddit, YouTube), SCBM achieves an average macro-F1 score of 0.69 which outperforms the most recently reported results from the literature on four out of five datasets. Aside from high recognition accuracy, SCBM provides a high level of both local and global interpretability. Furthermore, fusing our adjective-based concept representation with transformer embeddings, leads to a 1.8% performance increase on average across all datasets, showing that the proposed representation captures complementary information. Our results demonstrate that adjective-based concept representations can serve as compact, interpretable, and effective encodings for hate and counter speech recognition. With adapted adjectives, our method can also be applied to other NLP tasks.", "date": "2025-07-30", "categories": []}
{"id": "2508.04913", "title": "Advancing Hate Speech Detection with Transformers: Insights from the MetaHate", "abstract": "Hate speech is a widespread and harmful form of online discourse, encompassing slurs and defamatory posts that can have serious social, psychological, and sometimes physical impacts on targeted individuals and communities. As social media platforms such as X (formerly Twitter), Facebook, Instagram, Reddit, and others continue to facilitate widespread communication, they also become breeding grounds for hate speech, which has increasingly been linked to real-world hate crimes. Addressing this issue requires the development of robust automated methods to detect hate speech in diverse social media environments. Deep learning approaches, such as vanilla recurrent neural networks (RNNs), long short-term memory (LSTM), and convolutional neural networks (CNNs), have achieved good results, but are often limited by issues such as long-term dependencies and inefficient parallelization. This study represents the comprehensive exploration of transformer-based models for hate speech detection using the MetaHate dataset--a meta-collection of 36 datasets with 1.2 million social media samples. We evaluate multiple state-of-the-art transformer models, including BERT, RoBERTa, GPT-2, and ELECTRA, with fine-tuned ELECTRA achieving the highest performance (F1 score: 0.8980). We also analyze classification errors, revealing challenges with sarcasm, coded language, and label noise.", "date": "2025-08-06", "categories": []}
{"id": "2508.05527", "title": "AI vs. Human Moderators: A Comparative Evaluation of Multimodal LLMs in Content Moderation for Brand Safety", "abstract": "As the volume of video content online grows exponentially, the demand for moderation of unsafe videos has surpassed human capabilities, posing both operational and mental health challenges. While recent studies demonstrated the merits of Multimodal Large Language Models (MLLMs) in various video understanding tasks, their application to multimodal content moderation, a domain that requires nuanced understanding of both visual and textual cues, remains relatively underexplored. In this work, we benchmark the capabilities of MLLMs in brand safety classification, a critical subset of content moderation for safe-guarding advertising integrity. To this end, we introduce a novel, multimodal and multilingual dataset, meticulously labeled by professional reviewers in a multitude of risk categories. Through a detailed comparative analysis, we demonstrate the effectiveness of MLLMs such as Gemini, GPT, and Llama in multimodal brand safety, and evaluate their accuracy and cost efficiency compared to professional human reviewers. Furthermore, we present an in-depth discussion shedding light on limitations of MLLMs and failure cases. We are releasing our dataset alongside this paper to facilitate future research on effective and responsible brand safety and content moderation.", "date": "2025-08-07", "categories": []}
{"id": "2410.16520", "title": "AUTALIC: A Dataset for Anti-AUTistic Ableist Language In Context", "abstract": "As our understanding of autism and ableism continues to increase, so does our understanding of ableist language towards autistic people. Such language poses a significant challenge in NLP research due to its subtle and context-dependent nature. Yet, detecting anti-autistic ableist language remains underexplored, with existing NLP tools often failing to capture its nuanced expressions. We present AUTALIC, the first benchmark dataset dedicated to the detection of anti-autistic ableist language in context, addressing a significant gap in the field. The dataset comprises 2,400 autism-related sentences collected from Reddit, accompanied by surrounding context, and is annotated by trained experts with backgrounds in neurodiversity. Our comprehensive evaluation reveals that current language models, including state-of-the-art LLMs, struggle to reliably identify anti-autistic ableism and align with human judgments, underscoring their limitations in this domain. We publicly release AUTALIC along with the individual annotations which serve as a valuable resource to researchers working on ableism, neurodiversity, and also studying disagreements in annotation tasks. This dataset serves as a crucial step towards developing more inclusive and context-aware NLP systems that better reflect diverse perspectives.", "date": "2024-10-21", "categories": []}
{"id": "2508.03732", "title": "What is Beneath Misogyny: Misogynous Memes Classification and Explanation", "abstract": "Memes are popular in the modern world and are distributed primarily for entertainment. However, harmful ideologies such as misogyny can be propagated through innocent-looking memes. The detection and understanding of why a meme is misogynous is a research challenge due to its multimodal nature (image and text) and its nuanced manifestations across different societal contexts. We introduce a novel multimodal approach, \\textit{namely}, \\textit{\\textbf{MM-Misogyny}} to detect, categorize, and explain misogynistic content in memes. \\textit{\\textbf{MM-Misogyny}} processes text and image modalities separately and unifies them into a multimodal context through a cross-attention mechanism. The resulting multimodal context is then easily processed for labeling, categorization, and explanation via a classifier and Large Language Model (LLM). The evaluation of the proposed model is performed on a newly curated dataset (\\textit{\\textbf{W}hat's \\textbf{B}eneath \\textbf{M}isogynous \\textbf{S}tereotyping (WBMS)}) created by collecting misogynous memes from cyberspace and categorizing them into four categories, \\textit{namely}, Kitchen, Leadership, Working, and Shopping. The model not only detects and classifies misogyny, but also provides a granular understanding of how misogyny operates in domains of life. The results demonstrate the superiority of our approach compared to existing methods. The code and dataset are available at \\href{https://github.com/kushalkanwarNS/WhatisBeneathMisogyny/tree/main}{https://github.com/Misogyny}.", "date": "2025-07-30", "categories": []}
{"id": "2508.04638", "title": "Can NLP Tackle Hate Speech in the Real World? Stakeholder-Informed Feedback and Survey on Counterspeech", "abstract": "Counterspeech, i.e. the practice of responding to online hate speech, has gained traction in NLP as a promising intervention. While early work emphasised collaboration with non-governmental organisation stakeholders, recent research trends have shifted toward automated pipelines that reuse a small set of legacy datasets, often without input from affected communities. This paper presents a systematic review of 74 NLP studies on counterspeech, analysing the extent to which stakeholder participation influences dataset creation, model development, and evaluation. To complement this analysis, we conducted a participatory case study with five NGOs specialising in online Gender-Based Violence (oGBV), identifying stakeholder-informed practices for counterspeech generation. Our findings reveal a growing disconnect between current NLP research and the needs of communities most impacted by toxic online content. We conclude with concrete recommendations for re-centring stakeholder expertise in counterspeech research.", "date": "2025-08-06", "categories": []}
{"id": "2501.13836", "title": "Think Outside the Data: Colonial Biases and Systemic Issues in Automated Moderation Pipelines for Low-Resource Languages", "abstract": "Most social media users come from the Global South, where harmful content usually appears in local languages. Yet, AI-driven moderation systems struggle with low-resource languages spoken in these regions. Through semi-structured interviews with 22 AI experts working on harmful content detection in four low-resource languages: Tamil (South Asia), Swahili (East Africa), Maghrebi Arabic (North Africa), and Quechua (South America)--we examine systemic issues in building automated moderation tools for these languages. Our findings reveal that beyond data scarcity, socio-political factors such as tech companies' monopoly on user data and lack of investment in moderation for low-profit Global South markets exacerbate historic inequities. Even if more data were available, the English-centric and data-intensive design of language models and preprocessing techniques overlooks the need to design for morphologically complex, linguistically diverse, and code-mixed languages. We argue these limitations are not just technical gaps caused by \"data scarcity\" but reflect structural inequities, rooted in colonial suppression of non-Western languages. We discuss multi-stakeholder approaches to strengthen local research capacity, democratize data access, and support language-aware solutions to improve automated moderation for low-resource languages.", "date": "2025-01-23", "categories": []}
{"id": "2508.02094", "title": "\"Harmless to You, Hurtful to Me!\": Investigating the Detection of Toxic Languages Grounded in the Perspective of Youth", "abstract": "Risk perception is subjective, and youth's understanding of toxic content differs from that of adults. Although previous research has conducted extensive studies on toxicity detection in social media, the investigation of youth's unique toxicity, i.e., languages perceived as nontoxic by adults but toxic as youth, is ignored. To address this gap, we aim to explore: 1) What are the features of ``youth-toxicity'' languages in social media (RQ1); 2) Can existing toxicity detection techniques accurately detect these languages (RQ2). For these questions, we took Chinese youth as the research target, constructed the first Chinese ``youth-toxicity'' dataset, and then conducted extensive analysis. Our results suggest that youth's perception of these is associated with several contextual factors, like the source of an utterance and text-related features. Incorporating these meta information into current toxicity detection methods significantly improves accuracy overall. Finally, we propose several insights into future research on youth-centered toxicity detection.", "date": "2025-08-04", "categories": []}
{"id": "2508.00760", "title": "MMBERT: Scaled Mixture-of-Experts Multimodal BERT for Robust Chinese Hate Speech Detection under Cloaking Perturbations", "abstract": "Hate speech detection on Chinese social networks presents distinct challenges, particularly due to the widespread use of cloaking techniques designed to evade conventional text-based detection systems. Although large language models (LLMs) have recently improved hate speech detection capabilities, the majority of existing work has concentrated on English datasets, with limited attention given to multimodal strategies in the Chinese context. In this study, we propose MMBERT, a novel BERT-based multimodal framework that integrates textual, speech, and visual modalities through a Mixture-of-Experts (MoE) architecture. To address the instability associated with directly integrating MoE into BERT-based models, we develop a progressive three-stage training paradigm. MMBERT incorporates modality-specific experts, a shared self-attention mechanism, and a router-based expert allocation strategy to enhance robustness against adversarial perturbations. Empirical results in several Chinese hate speech datasets show that MMBERT significantly surpasses fine-tuned BERT-based encoder models, fine-tuned LLMs, and LLMs utilizing in-context learning approaches.", "date": "2025-08-01", "categories": []}
{"id": "2507.23661", "title": "Arabic Hate Speech Identification and Masking in Social Media using Deep Learning Models and Pre-trained Models Fine-tuning", "abstract": "Hate speech identification in social media has become an increasingly important issue in recent years. In this research, we address two problems: 1) to detect hate speech in Arabic text, 2) to clean a given text from hate speech. The meaning of cleaning here is replacing each bad word with stars based on the number of letters for each word. Regarding the first problem, we conduct several experiments using deep learning models and transformers to determine the best model in terms of the F1 score. Regarding second problem, we consider it as a machine translation task, where the input is a sentence containing dirty text and the output is the same sentence with masking the dirty text. The presented methods achieve the best model in hate speech detection with a 92\\% Macro F1 score and 95\\% accuracy. Regarding the text cleaning experiment, the best result in the hate speech masking model reached 0.3 in BLEU score with 1-gram, which is a good result compared with the state of the art machine translation systems.", "date": "2025-07-31", "categories": []}
{"id": "2505.19010", "title": "Co-AttenDWG: Co-Attentive Dimension-Wise Gating and Expert Fusion for Multi-Modal Offensive Content Detection", "abstract": "Multi-modal learning has emerged as a crucial research direction, as integrating textual and visual information can substantially enhance performance in tasks such as classification, retrieval, and scene understanding. Despite advances with large pre-trained models, existing approaches often suffer from insufficient cross-modal interactions and rigid fusion strategies, failing to fully harness the complementary strengths of different modalities. To address these limitations, we propose Co-AttenDWG, co-attention with dimension-wise gating, and expert fusion. Our approach first projects textual and visual features into a shared embedding space, where a dedicated co-attention mechanism enables simultaneous, fine-grained interactions between modalities. This is further strengthened by a dimension-wise gating network, which adaptively modulates feature contributions at the channel level to emphasize salient information. In parallel, dual-path encoders independently refine modality-specific representations, while an additional cross-attention layer aligns the modalities further. The resulting features are aggregated via an expert fusion module that integrates learned gating and self-attention, yielding a robust unified representation. Experimental results on the MIMIC and SemEval Memotion 1.0 datasets show that Co-AttenDWG achieves state-of-the-art performance and superior cross-modal alignment, highlighting its effectiveness for diverse multi-modal applications.", "date": "2025-05-25", "categories": []}
{"id": "2507.21089", "title": "Emotionally Aware Moderation: The Potential of Emotion Monitoring in Shaping Healthier Social Media Conversations", "abstract": "Social media platforms increasingly employ proactive moderation techniques, such as detecting and curbing toxic and uncivil comments, to prevent the spread of harmful content. Despite these efforts, such approaches are often criticized for creating a climate of censorship and failing to address the underlying causes of uncivil behavior. Our work makes both theoretical and practical contributions by proposing and evaluating two types of emotion monitoring dashboards to users' emotional awareness and mitigate hate speech. In a study involving 211 participants, we evaluate the effects of the two mechanisms on user commenting behavior and emotional experiences. The results reveal that these interventions effectively increase users' awareness of their emotional states and reduce hate speech. However, our findings also indicate potential unintended effects, including increased expression of negative emotions (Angry, Fear, and Sad) when discussing sensitive issues. These insights provide a basis for further research on integrating proactive emotion regulation tools into social media platforms to foster healthier digital interactions.", "date": "2025-06-24", "categories": []}
{"id": "2504.21226", "title": "MemeBLIP2: A novel lightweight multimodal system to detect harmful memes", "abstract": "Memes often merge visuals with brief text to share humor or opinions, yet some memes contain harmful messages such as hate speech. In this paper, we introduces MemeBLIP2, a light weight multimodal system that detects harmful memes by combining image and text features effectively. We build on previous studies by adding modules that align image and text representations into a shared space and fuse them for better classification. Using BLIP-2 as the core vision-language model, our system is evaluated on the PrideMM datasets. The results show that MemeBLIP2 can capture subtle cues in both modalities, even in cases with ironic or culturally specific content, thereby improving the detection of harmful material.", "date": "2025-04-29", "categories": []}
{"id": "2507.20061", "title": "Strategic Filtering for Content Moderation: Free Speech or Free of Distortion?", "abstract": "User-generated content (UGC) on social media platforms is vulnerable to incitements and manipulations, necessitating effective regulations. To address these challenges, those platforms often deploy automated content moderators tasked with evaluating the harmfulness of UGC and filtering out content that violates established guidelines. However, such moderation inevitably gives rise to strategic responses from users, who strive to express themselves within the confines of guidelines. Such phenomena call for a careful balance between: 1. ensuring freedom of speech -- by minimizing the restriction of expression; and 2. reducing social distortion -- measured by the total amount of content manipulation. We tackle the problem of optimizing this balance through the lens of mechanism design, aiming at optimizing the trade-off between minimizing social distortion and maximizing free speech. Although determining the optimal trade-off is NP-hard, we propose practical methods to approximate the optimal solution. Additionally, we provide generalization guarantees determining the amount of finite offline data required to approximate the optimal moderator effectively.", "date": "2025-07-26", "categories": []}
{"id": "2507.20503", "title": "Customize Multi-modal RAI Guardrails with Precedent-based predictions", "abstract": "A multi-modal guardrail must effectively filter image content based on user-defined policies, identifying material that may be hateful, reinforce harmful stereotypes, contain explicit material, or spread misinformation. Deploying such guardrails in real-world applications, however, poses significant challenges. Users often require varied and highly customizable policies and typically cannot provide abundant examples for each custom policy. Consequently, an ideal guardrail should be scalable to the multiple policies and adaptable to evolving user standards with minimal retraining. Existing fine-tuning methods typically condition predictions on pre-defined policies, restricting their generalizability to new policies or necessitating extensive retraining to adapt. Conversely, training-free methods struggle with limited context lengths, making it difficult to incorporate all the policies comprehensively. To overcome these limitations, we propose to condition model's judgment on \"precedents\", which are the reasoning processes of prior data points similar to the given input. By leveraging precedents instead of fixed policies, our approach greatly enhances the flexibility and adaptability of the guardrail. In this paper, we introduce a critique-revise mechanism for collecting high-quality precedents and two strategies that utilize precedents for robust prediction. Experimental results demonstrate that our approach outperforms previous methods across both few-shot and full-dataset scenarios and exhibits superior generalization to novel policies.", "date": "2025-07-28", "categories": []}
{"id": "2507.20528", "title": "Dialogues of Dissent: Thematic and Rhetorical Dimensions of Hate and Counter-Hate Speech in Social Media Conversations", "abstract": "We introduce a novel multi-labeled scheme for joint annotation of hate and counter-hate speech in social media conversations, categorizing hate and counter-hate messages into thematic and rhetorical dimensions. The thematic categories outline different discursive aspects of each type of speech, while the rhetorical dimension captures how hate and counter messages are communicated, drawing on Aristotle's Logos, Ethos and Pathos. We annotate a sample of 92 conversations, consisting of 720 tweets, and conduct statistical analyses, incorporating public metrics, to explore patterns of interaction between the thematic and rhetorical dimensions within and between hate and counter-hate speech. Our findings provide insights into the spread of hate messages on social media, the strategies used to counter them, and their potential impact on online behavior.", "date": "2025-07-28", "categories": []}
{"id": "2409.13725", "title": "Identity-related Speech Suppression in Generative AI Content Moderation", "abstract": "Automated content moderation has long been used to help identify and filter undesired user-generated content online. But such systems have a history of incorrectly flagging content by and about marginalized identities for removal. Generative AI systems now use such filters to keep undesired generated content from being created by or shown to users. While a lot of focus has been given to making sure such systems do not produce undesired outcomes, considerably less attention has been paid to making sure appropriate text can be generated. From classrooms to Hollywood, as generative AI is increasingly used for creative or expressive text generation, whose stories will these technologies allow to be told, and whose will they suppress?\n  In this paper, we define and introduce measures of speech suppression, focusing on speech related to different identity groups incorrectly filtered by a range of content moderation APIs. Using both short-form, user-generated datasets traditional in content moderation and longer generative AI-focused data, including two datasets we introduce in this work, we create a benchmark for measurement of speech suppression for nine identity groups. Across one traditional and four generative AI-focused automated content moderation services tested, we find that identity-related speech is more likely to be incorrectly suppressed than other speech. We find that reasons for incorrect flagging behavior vary by identity based on stereotypes and text associations, with, e.g., disability-related content more likely to be flagged for self-harm or health-related reasons while non-Christian content is more likely to be flagged as violent or hateful. As generative AI systems are increasingly used for creative work, we urge further attention to how this may impact the creation of identity-related content.", "date": "2024-09-09", "categories": []}
{"id": "2507.14679", "title": "GCC-Spam: Spam Detection via GAN, Contrastive Learning, and Character Similarity Networks", "abstract": "The exponential growth of spam text on the Internet necessitates robust detection mechanisms to mitigate risks such as information leakage and social instability. This work addresses two principal challenges: adversarial strategies employed by spammers and the scarcity of labeled data. We propose a novel spam-text detection framework GCC-Spam, which integrates three core innovations. First, a character similarity network captures orthographic and phonetic features to counter character-obfuscation attacks and furthermore produces sentence embeddings for downstream classification. Second, contrastive learning enhances discriminability by optimizing the latent-space distance between spam and normal texts. Third, a Generative Adversarial Network (GAN) generates realistic pseudo-spam samples to alleviate data scarcity while improving model robustness and classification accuracy. Extensive experiments on real-world datasets demonstrate that our model outperforms baseline approaches, achieving higher detection rates with significantly fewer labeled examples.", "date": "2025-07-19", "categories": []}
{"id": "2507.18580", "title": "System Report for CCL25-Eval Task 10: SRAG-MAV for Fine-Grained Chinese Hate Speech Recognition", "abstract": "This paper presents our system for CCL25-Eval Task 10, addressing Fine-Grained Chinese Hate Speech Recognition (FGCHSR). We propose a novel SRAG-MAV framework that synergistically integrates task reformulation(TR), Self-Retrieval-Augmented Generation (SRAG), and Multi-Round Accumulative Voting (MAV). Our method reformulates the quadruplet extraction task into triplet extraction, uses dynamic retrieval from the training set to create contextual prompts, and applies multi-round inference with voting to improve output stability and performance. Our system, based on the Qwen2.5-7B model, achieves a Hard Score of 26.66, a Soft Score of 48.35, and an Average Score of 37.505 on the STATE ToxiCN dataset, significantly outperforming baselines such as GPT-4o (Average Score 15.63) and fine-tuned Qwen2.5-7B (Average Score 35.365). The code is available at https://github.com/king-wang123/CCL25-SRAG-MAV.", "date": "2025-07-24", "categories": []}
{"id": "2507.17204", "title": "Filter-And-Refine: A MLLM Based Cascade System for Industrial-Scale Video Content Moderation", "abstract": "Effective content moderation is essential for video platforms to safeguard user experience and uphold community standards. While traditional video classification models effectively handle well-defined moderation tasks, they struggle with complicated scenarios such as implicit harmful content and contextual ambiguity. Multimodal large language models (MLLMs) offer a promising solution to these limitations with their superior cross-modal reasoning and contextual understanding. However, two key challenges hinder their industrial adoption. First, the high computational cost of MLLMs makes full-scale deployment impractical. Second, adapting generative models for discriminative classification remains an open research problem. In this paper, we first introduce an efficient method to transform a generative MLLM into a multimodal classifier using minimal discriminative training data. To enable industry-scale deployment, we then propose a router-ranking cascade system that integrates MLLMs with a lightweight router model. Offline experiments demonstrate that our MLLM-based approach improves F1 score by 66.50% over traditional classifiers while requiring only 2% of the fine-tuning data. Online evaluations show that our system increases automatic content moderation volume by 41%, while the cascading deployment reduces computational cost to only 1.5% of direct full-scale deployment.", "date": "2025-07-23", "categories": []}
{"id": "2507.16183", "title": "BIDWESH: A Bangla Regional Based Hate Speech Detection Dataset", "abstract": "Hate speech on digital platforms has become a growing concern globally, especially in linguistically diverse countries like Bangladesh, where regional dialects play a major role in everyday communication. Despite progress in hate speech detection for standard Bangla, Existing datasets and systems fail to address the informal and culturally rich expressions found in dialects such as Barishal, Noakhali, and Chittagong. This oversight results in limited detection capability and biased moderation, leaving large sections of harmful content unaccounted for. To address this gap, this study introduces BIDWESH, the first multi-dialectal Bangla hate speech dataset, constructed by translating and annotating 9,183 instances from the BD-SHS corpus into three major regional dialects. Each entry was manually verified and labeled for hate presence, type (slander, gender, religion, call to violence), and target group (individual, male, female, group), ensuring linguistic and contextual accuracy. The resulting dataset provides a linguistically rich, balanced, and inclusive resource for advancing hate speech detection in Bangla. BIDWESH lays the groundwork for the development of dialect-sensitive NLP tools and contributes significantly to equitable and context-aware content moderation in low-resource language settings.", "date": "2025-07-22", "categories": []}
{"id": "2303.09823", "title": "Transformers and Ensemble methods: A solution for Hate Speech Detection in Arabic languages", "abstract": "This paper describes our participation in the shared task of hate speech detection, which is one of the subtasks of the CERIST NLP Challenge 2022. Our experiments evaluate the performance of six transformer models and their combination using 2 ensemble approaches. The best results on the training set, in a five-fold cross validation scenario, were obtained by using the ensemble approach based on the majority vote. The evaluation of this approach on the test set resulted in an F1-score of 0.60 and an Accuracy of 0.86.", "date": "2023-03-17", "categories": []}
{"id": "2507.10996", "title": "Mario at EXIST 2025: A Simple Gateway to Effective Multilingual Sexism Detection", "abstract": "This paper presents our approach to EXIST 2025 Task 1, addressing text-based sexism detection in English and Spanish tweets through hierarchical Low-Rank Adaptation (LoRA) of Llama 3.1 8B. Our method introduces conditional adapter routing that explicitly models label dependencies across three hierarchically structured subtasks: binary sexism identification, source intention detection, and multilabel sexism categorization. Unlike conventional LoRA applications that target only attention layers, we apply adaptation to all linear transformations, enhancing the model's capacity to capture task-specific patterns. In contrast to complex data processing and ensemble approaches, we show that straightforward parameter-efficient fine-tuning achieves strong performance. We train separate LoRA adapters (rank=16, QLoRA 4-bit) for each subtask using unified multilingual training that leverages Llama 3.1's native bilingual capabilities. The method requires minimal preprocessing and uses standard supervised learning. Our multilingual training strategy eliminates the need for separate language-specific models, achieving 1.7-2.4\\% F1 improvements through cross-lingual transfer. With only 1.67\\% trainable parameters compared to full fine-tuning, our approach reduces training time by 75\\% and model storage by 98\\%, while achieving competitive performance across all subtasks (ICM-Hard: 0.6774 for binary classification, 0.4991 for intention detection, 0.6519 for multilabel categorization).", "date": "2025-07-15", "categories": []}
{"id": "2507.11292", "title": "Fine-Grained Chinese Hate Speech Understanding: Span-Level Resources, Coded Term Lexicon, and Enhanced Detection Frameworks", "abstract": "The proliferation of hate speech has inflicted significant societal harm, with its intensity and directionality closely tied to specific targets and arguments. In recent years, numerous machine learning-based methods have been developed to detect hateful comments on online platforms automatically. However, research on Chinese hate speech detection lags behind, and interpretability studies face two major challenges: first, the scarcity of span-level fine-grained annotated datasets limits models' deep semantic understanding of hate speech; second, insufficient research on identifying and interpreting coded hate speech restricts model explainability in complex real-world scenarios. To address these, we make the following contributions: (1) We introduce the Span-level Target-Aware Toxicity Extraction dataset (STATE ToxiCN), the first span-level Chinese hate speech dataset, and evaluate the hate semantic understanding of existing models using it. (2) We conduct the first comprehensive study on Chinese coded hate terms, LLMs' ability to interpret hate semantics. (3) We propose a method to integrate an annotated lexicon into models, significantly enhancing hate speech detection performance. Our work provides valuable resources and insights to advance the interpretability of Chinese hate speech detection research.", "date": "2025-07-15", "categories": []}
{"id": "2507.10177", "title": "Abusive text transformation using LLMs", "abstract": "Although Large Language Models (LLMs) have demonstrated significant advancements in natural language processing tasks, their effectiveness in the classification and transformation of abusive text into non-abusive versions remains an area for exploration. In this study, we aim to use LLMs to transform abusive text (tweets and reviews) featuring hate speech and swear words into non-abusive text, while retaining the intent of the text. We evaluate the performance of two state-of-the-art LLMs, such as Gemini, GPT-4o, DeekSeek and Groq, on their ability to identify abusive text. We them to transform and obtain a text that is clean from abusive and inappropriate content but maintains a similar level of sentiment and semantics, i.e. the transformed text needs to maintain its message. Afterwards, we evaluate the raw and transformed datasets with sentiment analysis and semantic analysis. Our results show Groq provides vastly different results when compared with other LLMs. We have identified similarities between GPT-4o and DeepSeek-V3.", "date": "2025-07-14", "categories": []}
{"id": "2507.10468", "title": "From BERT to Qwen: Hate Detection across architectures", "abstract": "Online platforms struggle to curb hate speech without over-censoring legitimate discourse. Early bidirectional transformer encoders made big strides, but the arrival of ultra-large autoregressive LLMs promises deeper context-awareness. Whether this extra scale actually improves practical hate-speech detection on real-world text remains unverified. Our study puts this question to the test by benchmarking both model families, classic encoders and next-generation LLMs, on curated corpora of online interactions for hate-speech detection (Hate or No Hate).", "date": "2025-07-14", "categories": []}
{"id": "2402.13818", "title": "Beyond Hate Speech: NLP's Challenges and Opportunities in Uncovering Dehumanizing Language", "abstract": "Dehumanization, i.e., denying human qualities to individuals or groups, is a particularly harmful form of hate speech that can normalize violence against marginalized communities. Despite advances in NLP for detecting general hate speech, approaches to identifying dehumanizing language remain limited due to scarce annotated data and the subtle nature of such expressions. In this work, we systematically evaluate four state-of-the-art large language models (LLMs) - Claude, GPT, Mistral, and Qwen - for dehumanization detection. Our results show that only one model-Claude-achieves strong performance (over 80% F1) under an optimized configuration, while others, despite their capabilities, perform only moderately. Performance drops further when distinguishing dehumanization from related hate types such as derogation. We also identify systematic disparities across target groups: models tend to over-predict dehumanization for some identities (e.g., Gay men), while under-identifying it for others (e.g., Refugees). These findings motivate the need for systematic, group-level evaluation when applying pretrained language models to dehumanization detection tasks.", "date": "2024-02-21", "categories": []}
{"id": "2507.07640", "title": "Lost in Pronunciation: Detecting Chinese Offensive Language Disguised by Phonetic Cloaking Replacement", "abstract": "Phonetic Cloaking Replacement (PCR), defined as the deliberate use of homophonic or near-homophonic variants to hide toxic intent, has become a major obstacle to Chinese content moderation. While this problem is well-recognized, existing evaluations predominantly rely on rule-based, synthetic perturbations that ignore the creativity of real users. We organize PCR into a four-way surface-form taxonomy and compile \\ours, a dataset of 500 naturally occurring, phonetically cloaked offensive posts gathered from the RedNote platform. Benchmarking state-of-the-art LLMs on this dataset exposes a serious weakness: the best model reaches only an F1-score of 0.672, and zero-shot chain-of-thought prompting pushes performance even lower. Guided by error analysis, we revisit a Pinyin-based prompting strategy that earlier studies judged ineffective and show that it recovers much of the lost accuracy. This study offers the first comprehensive taxonomy of Chinese PCR, a realistic benchmark that reveals current detectors' limits, and a lightweight mitigation technique that advances research on robust toxicity detection.", "date": "2025-07-10", "categories": []}
{"id": "2507.05455", "title": "ModelCitizens: Representing Community Voices in Online Safety", "abstract": "Automatic toxic language detection is critical for creating safe, inclusive online spaces. However, it is a highly subjective task, with perceptions of toxic language shaped by community norms and lived experience. Existing toxicity detection models are typically trained on annotations that collapse diverse annotator perspectives into a single ground truth, erasing important context-specific notions of toxicity such as reclaimed language. To address this, we introduce MODELCITIZENS, a dataset of 6.8K social media posts and 40K toxicity annotations across diverse identity groups. To capture the role of conversational context on toxicity, typical of social media posts, we augment MODELCITIZENS posts with LLM-generated conversational scenarios. State-of-the-art toxicity detection tools (e.g. OpenAI Moderation API, GPT-o4-mini) underperform on MODELCITIZENS, with further degradation on context-augmented posts. Finally, we release LLAMACITIZEN-8B and GEMMACITIZEN-12B, LLaMA- and Gemma-based models finetuned on MODELCITIZENS, which outperform GPT-o4-mini by 5.5% on in-distribution evaluations. Our findings highlight the importance of community-informed annotation and modeling for inclusive content moderation. The data, models and code are available at https://github.com/asuvarna31/modelcitizens.", "date": "2025-07-07", "categories": []}
{"id": "2507.06908", "title": "MIND: A Multi-agent Framework for Zero-shot Harmful Meme Detection", "abstract": "The rapid expansion of memes on social media has highlighted the urgent need for effective approaches to detect harmful content. However, traditional data-driven approaches struggle to detect new memes due to their evolving nature and the lack of up-to-date annotated data. To address this issue, we propose MIND, a multi-agent framework for zero-shot harmful meme detection that does not rely on annotated data. MIND implements three key strategies: 1) We retrieve similar memes from an unannotated reference set to provide contextual information. 2) We propose a bi-directional insight derivation mechanism to extract a comprehensive understanding of similar memes. 3) We then employ a multi-agent debate mechanism to ensure robust decision-making through reasoned arbitration. Extensive experiments on three meme datasets demonstrate that our proposed framework not only outperforms existing zero-shot approaches but also shows strong generalization across different model architectures and parameter scales, providing a scalable solution for harmful meme detection. The code is available at https://github.com/destroy-lonely/MIND.", "date": "2025-07-09", "categories": []}
{"id": "2507.05271", "title": "An Adaptive Supervised Contrastive Learning Framework for Implicit Sexism Detection in Digital Social Networks", "abstract": "The global reach of social media has amplified the spread of hateful content, including implicit sexism, which is often overlooked by conventional detection methods. In this work, we introduce an Adaptive Supervised Contrastive lEarning framework for implicit sexism detectioN (ASCEND). A key innovation of our method is the incorporation of threshold-based contrastive learning: by computing cosine similarities between embeddings, we selectively treat only those sample pairs as positive if their similarity exceeds a learnable threshold. This mechanism refines the embedding space by robustly pulling together representations of semantically similar texts while pushing apart dissimilar ones, thus reducing false positives and negatives. The final classification is achieved by jointly optimizing a contrastive loss with a cross-entropy loss. Textual features are enhanced through a word-level attention module. Additionally, we employ sentiment, emotion, and toxicity features. Evaluations on the EXIST2021 and MLSC datasets demonstrate that ASCEND significantly outperforms existing methods, with average Macro F1 improvements of 9.86%, 29.63%, and 32.51% across multiple tasks, highlighting its efficacy in capturing the subtle cues of implicit sexist language.", "date": "2025-07-03", "categories": []}
{"id": "2503.23088", "title": "UNITYAI-GUARD: Pioneering Toxicity Detection Across Low-Resource Indian Languages", "abstract": "This work introduces UnityAI-Guard, a framework for binary toxicity classification targeting low-resource Indian languages. While existing systems predominantly cater to high-resource languages, UnityAI-Guard addresses this critical gap by developing state-of-the-art models for identifying toxic content across diverse Brahmic/Indic scripts. Our approach achieves an impressive average F1-score of 84.23% across seven languages, leveraging a dataset of 567k training instances and 30k manually verified test instances. By advancing multilingual content moderation for linguistically diverse regions, UnityAI-Guard also provides public API access to foster broader adoption and application.", "date": "2025-03-29", "categories": []}
{"id": "2507.04350", "title": "HatePRISM: Policies, Platforms, and Research Integration. Advancing NLP for Hate Speech Proactive Mitigation", "abstract": "Despite regulations imposed by nations and social media platforms, e.g. (Government of India, 2021; European Parliament and Council of the European Union, 2022), inter alia, hateful content persists as a significant challenge. Existing approaches primarily rely on reactive measures such as blocking or suspending offensive messages, with emerging strategies focusing on proactive measurements like detoxification and counterspeech. In our work, which we call HatePRISM, we conduct a comprehensive examination of hate speech regulations and strategies from three perspectives: country regulations, social platform policies, and NLP research datasets. Our findings reveal significant inconsistencies in hate speech definitions and moderation practices across jurisdictions and platforms, alongside a lack of alignment with research efforts. Based on these insights, we suggest ideas and research direction for further exploration of a unified framework for automated hate speech moderation incorporating diverse strategies.", "date": "2025-07-06", "categories": []}
{"id": "2501.05443", "title": "A survey of textual cyber abuse detection using cutting-edge language models and large language models", "abstract": "The success of social media platforms has facilitated the emergence of various forms of online abuse within digital communities. This abuse manifests in multiple ways, including hate speech, cyberbullying, emotional abuse, grooming, and sexting. In this paper, we present a comprehensive analysis of the different forms of abuse prevalent in social media, with a particular focus on how emerging technologies, such as Language Models (LMs) and Large Language Models (LLMs), are reshaping both the detection and generation of abusive content within these networks. We delve into the mechanisms through which social media abuse is perpetuated, exploring the psychological and social impact. Additionally, we examine the dual role of advanced language models-highlighting their potential to enhance automated detection systems for abusive behavior while also acknowledging their capacity to generate harmful content. This paper aims to contribute to the ongoing discourse on online safety and ethics, offering insights into the evolving landscape of cyberabuse and the technological innovations that both mitigate and exacerbate it.", "date": "2025-01-09", "categories": []}
{"id": "2507.01066", "title": "Embedding-based Retrieval in Multimodal Content Moderation", "abstract": "Video understanding plays a fundamental role for content moderation on short video platforms, enabling the detection of inappropriate content. While classification remains the dominant approach for content moderation, it often struggles in scenarios requiring rapid and cost-efficient responses, such as trend adaptation and urgent escalations. To address this issue, we introduce an Embedding-Based Retrieval (EBR) method designed to complement traditional classification approaches. We first leverage a Supervised Contrastive Learning (SCL) framework to train a suite of foundation embedding models, including both single-modal and multi-modal architectures. Our models demonstrate superior performance over established contrastive learning methods such as CLIP and MoCo. Building on these embedding models, we design and implement the embedding-based retrieval system that integrates embedding generation and video retrieval to enable efficient and effective trend handling. Comprehensive offline experiments on 25 diverse emerging trends show that EBR improves ROC-AUC from 0.85 to 0.99 and PR-AUC from 0.35 to 0.95. Further online experiments reveal that EBR increases action rates by 10.32% and reduces operational costs by over 80%, while also enhancing interpretability and flexibility compared to classification-based solutions.", "date": "2025-06-30", "categories": []}
{"id": "2506.23930", "title": "Leveraging the Potential of Prompt Engineering for Hate Speech Detection in Low-Resource Languages", "abstract": "The rapid expansion of social media leads to a marked increase in hate speech, which threatens personal lives and results in numerous hate crimes. Detecting hate speech presents several challenges: diverse dialects, frequent code-mixing, and the prevalence of misspelled words in user-generated content on social media platforms. Recent progress in hate speech detection is typically concentrated on high-resource languages. However, low-resource languages still face significant challenges due to the lack of large-scale, high-quality datasets. This paper investigates how we can overcome this limitation via prompt engineering on large language models (LLMs) focusing on low-resource Bengali language. We investigate six prompting strategies - zero-shot prompting, refusal suppression, flattering the classifier, multi-shot prompting, role prompting, and finally our innovative metaphor prompting to detect hate speech effectively in low-resource languages. We pioneer the metaphor prompting to circumvent the built-in safety mechanisms of LLMs that marks a significant departure from existing jailbreaking methods. We investigate all six different prompting strategies on the Llama2-7B model and compare the results extensively with three pre-trained word embeddings - GloVe, Word2Vec, and FastText for three different deep learning models - multilayer perceptron (MLP), convolutional neural network (CNN), and bidirectional gated recurrent unit (BiGRU). To prove the effectiveness of our metaphor prompting in the low-resource Bengali language, we also evaluate it in another low-resource language - Hindi, and two high-resource languages - English and German. The performance of all prompting techniques is evaluated using the F1 score, and environmental impact factor (IF), which measures CO$_2$ emissions, electricity usage, and computational time.", "date": "2025-06-30", "categories": []}
{"id": "2410.11230", "title": "\"Is Hate Lost in Translation?\": Evaluation of Multilingual LGBTQIA+ Hate Speech Detection", "abstract": "This paper explores the challenges of detecting LGBTQIA+ hate speech of large language models across multiple languages, including English, Italian, Chinese and (code-switched) English-Tamil, examining the impact of machine translation and whether the nuances of hate speech are preserved across translation. We examine the hate speech detection ability of zero-shot and fine-tuned GPT. Our findings indicate that: (1) English has the highest performance and the code-switching scenario of English-Tamil being the lowest, (2) fine-tuning improves performance consistently across languages whilst translation yields mixed results. Through simple experimentation with original text and machine-translated text for hate speech detection along with a qualitative error analysis, this paper sheds light on the socio-cultural nuances and complexities of languages that may not be captured by automatic translation.", "date": "2024-10-15", "categories": []}
{"id": "2506.21795", "title": "Offensive Language Detection on Social Media Using XLNet", "abstract": "The widespread use of text-based communication on social media-through chats, comments, and microblogs-has improved user interaction but has also led to an increase in offensive content, including hate speech, racism, and other forms of abuse. Due to the enormous volume of user-generated content, manual moderation is impractical, which creates a need for automated systems that can detect offensive language. Deep learning models, particularly those using transfer learning, have demonstrated significant success in understanding natural language through large-scale pretraining. In this study, we propose an automatic offensive language detection model based on XLNet, a generalized autoregressive pretraining method, and compare its performance with BERT (Bidirectional Encoder Representations from Transformers), which is a widely used baseline in natural language processing (NLP). Both models are evaluated using the Offensive Language Identification Dataset (OLID), a benchmark Twitter dataset that includes hierarchical annotations. Our experimental results show that XLNet outperforms BERT in detecting offensive content and in categorizing the types of offenses, while BERT performs slightly better in identifying the targets of the offenses. Additionally, we find that oversampling and undersampling strategies are effective in addressing class imbalance and improving classification performance. These findings highlight the potential of transfer learning and XLNet-based architectures to create robust systems for detecting offensive language on social media platforms.", "date": "2025-06-26", "categories": []}
{"id": "2411.19832", "title": "Sensitive Content Classification in Social Media: A Holistic Resource and Evaluation", "abstract": "The detection of sensitive content in large datasets is crucial for ensuring that shared and analysed data is free from harmful material. However, current moderation tools, such as external APIs, suffer from limitations in customisation, accuracy across diverse sensitive categories, and privacy concerns. Additionally, existing datasets and open-source models focus predominantly on toxic language, leaving gaps in detecting other sensitive categories such as substance abuse or self-harm. In this paper, we put forward a unified dataset tailored for social media content moderation across six sensitive categories: conflictual language, profanity, sexually explicit material, drug-related content, self-harm, and spam. By collecting and annotating data with consistent retrieval strategies and guidelines, we address the shortcomings of previous focalised research. Our analysis demonstrates that fine-tuning large language models (LLMs) on this novel dataset yields significant improvements in detection performance compared to open off-the-shelf models such as LLaMA, and even proprietary OpenAI models, which underperform by 10-15% overall. This limitation is even more pronounced on popular moderation APIs, which cannot be easily tailored to specific sensitive content categories, among others.", "date": "2024-11-29", "categories": []}
{"id": "2506.19603", "title": "Social Hatred: Efficient Multimodal Detection of Hatemongers", "abstract": "Automatic detection of online hate speech serves as a crucial step in the detoxification of the online discourse. Moreover, accurate classification can promote a better understanding of the proliferation of hate as a social phenomenon. While most prior work focus on the detection of hateful utterances, we argue that focusing on the user level is as important, albeit challenging. In this paper we consider a multimodal aggregative approach for the detection of hate-mongers, taking into account the potentially hateful texts, user activity, and the user network. Evaluating our method on three unique datasets X (Twitter), Gab, and Parler we show that processing a user's texts in her social context significantly improves the detection of hate mongers, compared to previously used text and graph-based methods. We offer comprehensive set of results obtained in different experimental settings as well as qualitative analysis of illustrative cases. Our method can be used to improve the classification of coded messages, dog-whistling, and racial gas-lighting, as well as to inform intervention measures. Moreover, we demonstrate that our multimodal approach performs well across very different content platforms and over large datasets and networks.", "date": "2025-06-24", "categories": []}
{"id": "2506.19831", "title": "How Effectively Can BERT Models Interpret Context and Detect Bengali Communal Violent Text?", "abstract": "The spread of cyber hatred has led to communal violence, fueling aggression and conflicts between various religious, ethnic, and social groups, posing a significant threat to social harmony. Despite its critical importance, the classification of communal violent text remains an underexplored area in existing research. This study aims to enhance the accuracy of detecting text that incites communal violence, focusing specifically on Bengali textual data sourced from social media platforms. We introduce a fine-tuned BanglaBERT model tailored for this task, achieving a macro F1 score of 0.60. To address the issue of data imbalance, our dataset was expanded by adding 1,794 instances, which facilitated the development and evaluation of a fine-tuned ensemble model. This ensemble model demonstrated an improved performance, achieving a macro F1 score of 0.63, thus highlighting its effectiveness in this domain. In addition to quantitative performance metrics, qualitative analysis revealed instances where the models struggled with context understanding, leading to occasional misclassifications, even when predictions were made with high confidence. Through analyzing the cosine similarity between words, we identified certain limitations in the pre-trained BanglaBERT models, particularly in their ability to distinguish between closely related communal and non-communal terms. To further interpret the model's decisions, we applied LIME, which helped to uncover specific areas where the model struggled in understanding context, contributing to errors in classification. These findings highlight the promise of NLP and interpretability tools in reducing online communal violence. Our work contributes to the growing body of research in communal violence detection and offers a foundation for future studies aiming to refine these techniques for better accuracy and societal impact.", "date": "2025-06-24", "categories": []}
{"id": "2409.17130", "title": "Assessing the Level of Toxicity Against Distinct Groups in Bangla Social Media Comments: A Comprehensive Investigation", "abstract": "Social media platforms have a vital role in the modern world, serving as conduits for communication, the exchange of ideas, and the establishment of networks. However, the misuse of these platforms through toxic comments, which can range from offensive remarks to hate speech, is a concerning issue. This study focuses on identifying toxic comments in the Bengali language targeting three specific groups: transgender people, indigenous people, and migrant people, from multiple social media sources. The study delves into the intricate process of identifying and categorizing toxic language while considering the varying degrees of toxicity: high, medium, and low. The methodology involves creating a dataset, manual annotation, and employing pre-trained transformer models like Bangla-BERT, bangla-bert-base, distil-BERT, and Bert-base-multilingual-cased for classification. Diverse assessment metrics such as accuracy, recall, precision, and F1-score are employed to evaluate the model's effectiveness. The experimental findings reveal that Bangla-BERT surpasses alternative models, achieving an F1-score of 0.8903. This research exposes the complexity of toxicity in Bangla social media dialogues, revealing its differing impacts on diverse demographic groups.", "date": "2024-09-25", "categories": []}
{"id": "2506.18576", "title": "A Modular Taxonomy for Hate Speech Definitions and Its Impact on Zero-Shot LLM Classification Performance", "abstract": "Detecting harmful content is a crucial task in the landscape of NLP applications for Social Good, with hate speech being one of its most dangerous forms. But what do we mean by hate speech, how can we define it, and how does prompting different definitions of hate speech affect model performance? The contribution of this work is twofold. At the theoretical level, we address the ambiguity surrounding hate speech by collecting and analyzing existing definitions from the literature. We organize these definitions into a taxonomy of 14 Conceptual Elements-building blocks that capture different aspects of hate speech definitions, such as references to the target of hate (individual or groups) or of the potential consequences of it. At the experimental level, we employ the collection of definitions in a systematic zero-shot evaluation of three LLMs, on three hate speech datasets representing different types of data (synthetic, human-in-the-loop, and real-world). We find that choosing different definitions, i.e., definitions with a different degree of specificity in terms of encoded elements, impacts model performance, but this effect is not consistent across all architectures.", "date": "2025-06-23", "categories": []}
{"id": "2506.16066", "title": "Cyberbullying Detection in Hinglish Text Using MURIL and Explainable AI", "abstract": "The growth of digital communication platforms has led to increased cyberbullying incidents worldwide, creating a need for automated detection systems to protect users. The rise of code-mixed Hindi-English (Hinglish) communication on digital platforms poses challenges for existing cyberbullying detection systems, which were designed primarily for monolingual text. This paper presents a framework for cyberbullying detection in Hinglish text using the Multilingual Representations for Indian Languages (MURIL) architecture to address limitations in current approaches. Evaluation across six benchmark datasets -- Bohra \\textit{et al.}, BullyExplain, BullySentemo, Kumar \\textit{et al.}, HASOC 2021, and Mendeley Indo-HateSpeech -- shows that the MURIL-based approach outperforms existing multilingual models including RoBERTa and IndicBERT, with improvements of 1.36 to 13.07 percentage points and accuracies of 86.97\\% on Bohra, 84.62\\% on BullyExplain, 86.03\\% on BullySentemo, 75.41\\% on Kumar datasets, 83.92\\% on HASOC 2021, and 94.63\\% on Mendeley dataset. The framework includes explainability features through attribution analysis and cross-linguistic pattern recognition. Ablation studies show that selective layer freezing, appropriate classification head design, and specialized preprocessing for code-mixed content improve detection performance, while failure analysis identifies challenges including context-dependent interpretation, cultural understanding, and cross-linguistic sarcasm detection, providing directions for future research in multilingual cyberbullying detection.", "date": "2025-06-19", "categories": []}
{"id": "2506.16190", "title": "Web(er) of Hate: A Survey on How Hate Speech Is Typed", "abstract": "The curation of hate speech datasets involves complex design decisions that balance competing priorities. This paper critically examines these methodological choices in a diverse range of datasets, highlighting common themes and practices, and their implications for dataset reliability. Drawing on Max Weber's notion of ideal types, we argue for a reflexive approach in dataset creation, urging researchers to acknowledge their own value judgments during dataset construction, fostering transparency and methodological rigour.", "date": "2025-06-19", "categories": []}
{"id": "2506.16476", "title": "Towards Generalizable Generic Harmful Speech Datasets for Implicit Hate Speech Detection", "abstract": "Implicit hate speech has recently emerged as a critical challenge for social media platforms. While much of the research has traditionally focused on harmful speech in general, the need for generalizable techniques to detect veiled and subtle forms of hate has become increasingly pressing. Based on lexicon analysis, we hypothesize that implicit hate speech is already present in publicly available harmful speech datasets but may not have been explicitly recognized or labeled by annotators. Additionally, crowdsourced datasets are prone to mislabeling due to the complexity of the task and often influenced by annotators' subjective interpretations. In this paper, we propose an approach to address the detection of implicit hate speech and enhance generalizability across diverse datasets by leveraging existing harmful speech datasets. Our method comprises three key components: influential sample identification, reannotation, and augmentation using Llama-3 70B and GPT-4o. Experimental results demonstrate the effectiveness of our approach in improving implicit hate detection, achieving a +12.9-point F1 score improvement compared to the baseline.", "date": "2025-06-19", "categories": []}
{"id": "2506.16575", "title": "Advancing Harmful Content Detection in Organizational Research: Integrating Large Language Models with Elo Rating System", "abstract": "Large language models (LLMs) offer promising opportunities for organizational research. However, their built-in moderation systems can create problems when researchers try to analyze harmful content, often refusing to follow certain instructions or producing overly cautious responses that undermine validity of the results. This is particularly problematic when analyzing organizational conflicts such as microaggressions or hate speech. This paper introduces an Elo rating-based method that significantly improves LLM performance for harmful content analysis In two datasets, one focused on microaggression detection and the other on hate speech, we find that our method outperforms traditional LLM prompting techniques and conventional machine learning models on key measures such as accuracy, precision, and F1 scores. Advantages include better reliability when analyzing harmful content, fewer false positives, and greater scalability for large-scale datasets. This approach supports organizational applications, including detecting workplace harassment, assessing toxic communication, and fostering safer and more inclusive work environments.", "date": "2025-06-19", "categories": []}
{"id": "2506.12148", "title": "Hatevolution: What Static Benchmarks Don't Tell Us", "abstract": "Language changes over time, including in the hate speech domain, which evolves quickly following social dynamics and cultural shifts. While NLP research has investigated the impact of language evolution on model training and has proposed several solutions for it, its impact on model benchmarking remains under-explored. Yet, hate speech benchmarks play a crucial role to ensure model safety. In this paper, we empirically evaluate the robustness of 20 language models across two evolving hate speech experiments, and we show the temporal misalignment between static and time-sensitive evaluations. Our findings call for time-sensitive linguistic benchmarks in order to correctly and reliably evaluate language models in the hate speech domain.", "date": "2025-06-13", "categories": []}
{"id": "2506.12744", "title": "Rethinking Hate Speech Detection on Social Media: Can LLMs Replace Traditional Models?", "abstract": "Hate speech detection across contemporary social media presents unique challenges due to linguistic diversity and the informal nature of online discourse. These challenges are further amplified in settings involving code-mixing, transliteration, and culturally nuanced expressions. While fine-tuned transformer models, such as BERT, have become standard for this task, we argue that recent large language models (LLMs) not only surpass them but also redefine the landscape of hate speech detection more broadly. To support this claim, we introduce IndoHateMix, a diverse, high-quality dataset capturing Hindi-English code-mixing and transliteration in the Indian context, providing a realistic benchmark to evaluate model robustness in complex multilingual scenarios where existing NLP methods often struggle. Our extensive experiments show that cutting-edge LLMs (such as LLaMA-3.1) consistently outperform task-specific BERT-based models, even when fine-tuned on significantly less data. With their superior generalization and adaptability, LLMs offer a transformative approach to mitigating online hate in diverse environments. This raises the question of whether future works should prioritize developing specialized models or focus on curating richer and more varied datasets to further enhance the effectiveness of LLMs.", "date": "2025-06-15", "categories": []}
{"id": "2506.09992", "title": "Large Language Models for Toxic Language Detection in Low-Resource Balkan Languages", "abstract": "Online toxic language causes real harm, especially in regions with limited moderation tools. In this study, we evaluate how large language models handle toxic comments in Serbian, Croatian, and Bosnian, languages with limited labeled data. We built and manually labeled a dataset of 4,500 YouTube and TikTok comments drawn from videos across diverse categories, including music, politics, sports, modeling, influencer content, discussions of sexism, and general topics. Four models (GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus) were tested in two modes: zero-shot and context-augmented. We measured precision, recall, F1 score, accuracy and false positive rates. Including a short context snippet raised recall by about 0.12 on average and improved F1 score by up to 0.10, though it sometimes increased false positives. The best balance came from Gemini in context-augmented mode, reaching an F1 score of 0.82 and accuracy of 0.82, while zero-shot GPT-4.1 led on precision and had the lowest false alarms. We show how adding minimal context can improve toxic language detection in low-resource settings and suggest practical strategies such as improved prompt design and threshold calibration. These results show that prompt design alone can yield meaningful gains in toxicity detection for underserved Balkan language communities.", "date": "2025-06-11", "categories": []}
{"id": "2406.19349", "title": "IndoToxic2024: A Demographically-Enriched Dataset of Hate Speech and Toxicity Types for Indonesian Language", "abstract": "Hate speech poses a significant threat to social harmony. Over the past two years, Indonesia has seen a ten-fold increase in the online hate speech ratio, underscoring the urgent need for effective detection mechanisms. However, progress is hindered by the limited availability of labeled data for Indonesian texts. The condition is even worse for marginalized minorities, such as Shia, LGBTQ, and other ethnic minorities because hate speech is underreported and less understood by detection tools. Furthermore, the lack of accommodation for subjectivity in current datasets compounds this issue. To address this, we introduce IndoToxic2024, a comprehensive Indonesian hate speech and toxicity classification dataset. Comprising 43,692 entries annotated by 19 diverse individuals, the dataset focuses on texts targeting vulnerable groups in Indonesia, specifically during the hottest political event in the country: the presidential election. We establish baselines for seven binary classification tasks, achieving a macro-F1 score of 0.78 with a BERT model (IndoBERTweet) fine-tuned for hate speech classification. Furthermore, we demonstrate how incorporating demographic information can enhance the zero-shot performance of the large language model, gpt-3.5-turbo. However, we also caution that an overemphasis on demographic information can negatively impact the fine-tuned model performance due to data fragmentation.", "date": "2024-06-27", "categories": []}
{"id": "2410.07991", "title": "Human and LLM Biases in Hate Speech Annotations: A Socio-Demographic Analysis of Annotators and Targets", "abstract": "The rise of online platforms exacerbated the spread of hate speech, demanding scalable and effective detection. However, the accuracy of hate speech detection systems heavily relies on human-labeled data, which is inherently susceptible to biases. While previous work has examined the issue, the interplay between the characteristics of the annotator and those of the target of the hate are still unexplored. We fill this gap by leveraging an extensive dataset with rich socio-demographic information of both annotators and targets, uncovering how human biases manifest in relation to the target's attributes. Our analysis surfaces the presence of widespread biases, which we quantitatively describe and characterize based on their intensity and prevalence, revealing marked differences. Furthermore, we compare human biases with those exhibited by persona-based LLMs. Our findings indicate that while persona-based LLMs do exhibit biases, these differ significantly from those of human annotators. Overall, our work offers new and nuanced results on human biases in hate speech annotations, as well as fresh insights into the design of AI-driven hate speech detection systems.", "date": "2024-10-10", "categories": []}
{"id": "2506.10117", "title": "A Manually Annotated Image-Caption Dataset for Detecting Children in the Wild", "abstract": "Platforms and the law regulate digital content depicting minors (defined as individuals under 18 years of age) differently from other types of content. Given the sheer amount of content that needs to be assessed, machine learning-based automation tools are commonly used to detect content depicting minors. To our knowledge, no dataset or benchmark currently exists for detecting these identification methods in a multi-modal environment. To fill this gap, we release the Image-Caption Children in the Wild Dataset (ICCWD), an image-caption dataset aimed at benchmarking tools that detect depictions of minors. Our dataset is richer than previous child image datasets, containing images of children in a variety of contexts, including fictional depictions and partially visible bodies. ICCWD contains 10,000 image-caption pairs manually labeled to indicate the presence or absence of a child in the image. To demonstrate the possible utility of our dataset, we use it to benchmark three different detectors, including a commercial age estimation system applied to images. Our results suggest that child detection is a challenging task, with the best method achieving a 75.3% true positive rate. We hope the release of our dataset will aid in the design of better minor detection methods in a wide range of scenarios.", "date": "2025-06-11", "categories": []}
{"id": "2506.10789", "title": "FASCIST-O-METER: Classifier for Neo-fascist Discourse Online", "abstract": "Neo-fascism is a political and societal ideology that has been having remarkable growth in the last decade in the United States of America (USA), as well as in other Western societies. It poses a grave danger to democracy and the minorities it targets, and it requires active actions against it to avoid escalation. This work presents the first-of-its-kind neo-fascist coding scheme for digital discourse in the USA societal context, overseen by political science researchers. Our work bridges the gap between Natural Language Processing (NLP) and political science against this phenomena. Furthermore, to test the coding scheme, we collect a tremendous amount of activity on the internet from notable neo-fascist groups (the forums of Iron March and Stormfront.org), and the guidelines are applied to a subset of the collected posts. Through crowdsourcing, we annotate a total of a thousand posts that are labeled as neo-fascist or non-neo-fascist. With this labeled data set, we fine-tune and test both Small Language Models (SLMs) and Large Language Models (LLMs), obtaining the very first classification models for neo-fascist discourse. We find that the prevalence of neo-fascist rhetoric in this kind of forum is ever-present, making them a good target for future research. The societal context is a key consideration for neo-fascist speech when conducting NLP research. Finally, the work against this kind of political movement must be pressed upon and continued for the well-being of a democratic society. Disclaimer: This study focuses on detecting neo-fascist content in text, similar to other hate speech analyses, without labeling individuals or organizations.", "date": "2025-06-12", "categories": []}
{"id": "2506.07667", "title": "Silencing Empowerment, Allowing Bigotry: Auditing the Moderation of Hate Speech on Twitch", "abstract": "To meet the demands of content moderation, online platforms have resorted to automated systems. Newer forms of real-time engagement($\\textit{e.g.}$, users commenting on live streams) on platforms like Twitch exert additional pressures on the latency expected of such moderation systems. Despite their prevalence, relatively little is known about the effectiveness of these systems. In this paper, we conduct an audit of Twitch's automated moderation tool ($\\texttt{AutoMod}$) to investigate its effectiveness in flagging hateful content. For our audit, we create streaming accounts to act as siloed test beds, and interface with the live chat using Twitch's APIs to send over $107,000$ comments collated from $4$ datasets. We measure $\\texttt{AutoMod}$'s accuracy in flagging blatantly hateful content containing misogyny, racism, ableism and homophobia. Our experiments reveal that a large fraction of hateful messages, up to $94\\%$ on some datasets, $\\textit{bypass moderation}$. Contextual addition of slurs to these messages results in $100\\%$ removal, revealing $\\texttt{AutoMod}$'s reliance on slurs as a moderation signal. We also find that contrary to Twitch's community guidelines, $\\texttt{AutoMod}$ blocks up to $89.5\\%$ of benign examples that use sensitive words in pedagogical or empowering contexts. Overall, our audit points to large gaps in $\\texttt{AutoMod}$'s capabilities and underscores the importance for such systems to understand context effectively.", "date": "2025-06-09", "categories": []}
{"id": "2506.08147", "title": "Multilingual Hate Speech Detection in Social Media Using Translation-Based Approaches with Large Language Models", "abstract": "Social media platforms are critical spaces for public discourse, shaping opinions and community dynamics, yet their widespread use has amplified harmful content, particularly hate speech, threatening online safety and inclusivity. While hate speech detection has been extensively studied in languages like English and Spanish, Urdu remains underexplored, especially using translation-based approaches. To address this gap, we introduce a trilingual dataset of 10,193 tweets in English (3,834 samples), Urdu (3,197 samples), and Spanish (3,162 samples), collected via keyword filtering, with a balanced distribution of 4,849 Hateful and 5,344 Not-Hateful labels. Our methodology leverages attention layers as a precursor to transformer-based models and large language models (LLMs), enhancing feature extraction for multilingual hate speech detection. For non-transformer models, we use TF-IDF for feature extraction. The dataset is benchmarked using state-of-the-art models, including GPT-3.5 Turbo and Qwen 2.5 72B, alongside traditional machine learning models like SVM and other transformers (e.g., BERT, RoBERTa). Three annotators, following rigorous guidelines, ensured high dataset quality, achieving a Fleiss' Kappa of 0.821. Our approach, integrating attention layers with GPT-3.5 Turbo and Qwen 2.5 72B, achieves strong performance, with macro F1 scores of 0.87 for English (GPT-3.5 Turbo), 0.85 for Spanish (GPT-3.5 Turbo), 0.81 for Urdu (Qwen 2.5 72B), and 0.88 for the joint multilingual model (Qwen 2.5 72B). These results reflect improvements of 8.75% in English (over SVM baseline 0.80), 8.97% in Spanish (over SVM baseline 0.78), 5.19% in Urdu (over SVM baseline 0.77), and 7.32% in the joint multilingual model (over SVM baseline 0.82). Our framework offers a robust solution for multilingual hate speech detection, fostering safer digital communities worldwide.", "date": "2025-06-09", "categories": []}
{"id": "2506.08593", "title": "Hateful Person or Hateful Model? Investigating the Role of Personas in Hate Speech Detection by Large Language Models", "abstract": "Hate speech detection is a socially sensitive and inherently subjective task, with judgments often varying based on personal traits. While prior work has examined how socio-demographic factors influence annotation, the impact of personality traits on Large Language Models (LLMs) remains largely unexplored. In this paper, we present the first comprehensive study on the role of persona prompts in hate speech classification, focusing on MBTI-based traits. A human annotation survey confirms that MBTI dimensions significantly affect labeling behavior. Extending this to LLMs, we prompt four open-source models with MBTI personas and evaluate their outputs across three hate speech datasets. Our analysis uncovers substantial persona-driven variation, including inconsistencies with ground truth, inter-persona disagreement, and logit-level biases. These findings highlight the need to carefully define persona prompts in LLM-based annotation workflows, with implications for fairness and alignment with human values.", "date": "2025-06-10", "categories": []}
{"id": "2410.00775", "title": "Decoding Hate: Exploring Language Models' Reactions to Hate Speech", "abstract": "Hate speech is a harmful form of online expression, often manifesting as derogatory posts. It is a significant risk in digital environments. With the rise of Large Language Models (LLMs), there is concern about their potential to replicate hate speech patterns, given their training on vast amounts of unmoderated internet data. Understanding how LLMs respond to hate speech is crucial for their responsible deployment. However, the behaviour of LLMs towards hate speech has been limited compared. This paper investigates the reactions of seven state-of-the-art LLMs (LLaMA 2, Vicuna, LLaMA 3, Mistral, GPT-3.5, GPT-4, and Gemini Pro) to hate speech. Through qualitative analysis, we aim to reveal the spectrum of responses these models produce, highlighting their capacity to handle hate speech inputs. We also discuss strategies to mitigate hate speech generation by LLMs, particularly through fine-tuning and guideline guardrailing. Finally, we explore the models' responses to hate speech framed in politically correct language.", "date": "2024-10-01", "categories": []}
{"id": "2506.01484", "title": "LLM in the Loop: Creating the ParaDeHate Dataset for Hate Speech Detoxification", "abstract": "Detoxification, the task of rewriting harmful language into non-toxic text, has become increasingly important amid the growing prevalence of toxic content online. However, high-quality parallel datasets for detoxification, especially for hate speech, remain scarce due to the cost and sensitivity of human annotation. In this paper, we propose a novel LLM-in-the-loop pipeline leveraging GPT-4o-mini for automated detoxification. We first replicate the ParaDetox pipeline by replacing human annotators with an LLM and show that the LLM performs comparably to human annotation. Building on this, we construct ParaDeHate, a large-scale parallel dataset specifically for hatespeech detoxification. We release ParaDeHate as a benchmark of over 8K hate/non-hate text pairs and evaluate a wide range of baseline methods. Experimental results show that models such as BART, fine-tuned on ParaDeHate, achieve better performance in style accuracy, content preservation, and fluency, demonstrating the effectiveness of LLM-generated detoxification text as a scalable alternative to human annotation.", "date": "2025-06-02", "categories": []}
{"id": "2506.05635", "title": "IYKYK: Using language models to decode extremist cryptolects", "abstract": "Extremist groups develop complex in-group language, also referred to as cryptolects, to exclude or mislead outsiders. We investigate the ability of current language technologies to detect and interpret the cryptolects of two online extremist platforms. Evaluating eight models across six tasks, our results indicate that general purpose LLMs cannot consistently detect or decode extremist language. However, performance can be significantly improved by domain adaptation and specialised prompting techniques. These results provide important insights to inform the development and deployment of automated moderation technologies. We further develop and release novel labelled and unlabelled datasets, including 19.4M posts from extremist platforms and lexicons validated by human experts.", "date": "2025-06-05", "categories": []}
{"id": "2506.06238", "title": "Explaining Matters: Leveraging Definitions and Semantic Expansion for Sexism Detection", "abstract": "The detection of sexism in online content remains an open problem, as harmful language disproportionately affects women and marginalized groups. While automated systems for sexism detection have been developed, they still face two key challenges: data sparsity and the nuanced nature of sexist language. Even in large, well-curated datasets like the Explainable Detection of Online Sexism (EDOS), severe class imbalance hinders model generalization. Additionally, the overlapping and ambiguous boundaries of fine-grained categories introduce substantial annotator disagreement, reflecting the difficulty of interpreting nuanced expressions of sexism. To address these challenges, we propose two prompt-based data augmentation techniques: Definition-based Data Augmentation (DDA), which leverages category-specific definitions to generate semantically-aligned synthetic examples, and Contextual Semantic Expansion (CSE), which targets systematic model errors by enriching examples with task-specific semantic features. To further improve reliability in fine-grained classification, we introduce an ensemble strategy that resolves prediction ties by aggregating complementary perspectives from multiple language models. Our experimental evaluation on the EDOS dataset demonstrates state-of-the-art performance across all tasks, with notable improvements of macro F1 by 1.5 points for binary classification (Task A) and 4.1 points for fine-grained classification (Task C).", "date": "2025-06-06", "categories": []}
{"id": "2505.22184", "title": "Breaking the Cloak! Unveiling Chinese Cloaked Toxicity with Homophone Graph and Toxic Lexicon", "abstract": "Social media platforms have experienced a significant rise in toxic content, including abusive language and discriminatory remarks, presenting growing challenges for content moderation. Some users evade censorship by deliberately disguising toxic words through homophonic cloak, which necessitates the task of unveiling cloaked toxicity. Existing methods are mostly designed for English texts, while Chinese cloaked toxicity unveiling has not been solved yet. To tackle the issue, we propose C$^2$TU, a novel training-free and prompt-free method for Chinese cloaked toxic content unveiling. It first employs substring matching to identify candidate toxic words based on Chinese homo-graph and toxic lexicon. Then it filters those candidates that are non-toxic and corrects cloaks to be their corresponding toxicities. Specifically, we develop two model variants for filtering, which are based on BERT and LLMs, respectively. For LLMs, we address the auto-regressive limitation in computing word occurrence probability and utilize the full semantic contexts of a text sequence to reveal cloaked toxic words. Extensive experiments demonstrate that C$^2$TU can achieve superior performance on two Chinese toxic datasets. In particular, our method outperforms the best competitor by up to 71% on the F1 score and 35% on accuracy, respectively. Our code and data are available at https://github.com/XDxc-cuber/C2TU-Chinese-cloaked-toxicity-unveiling.", "date": "2025-05-28", "categories": []}
{"id": "2506.04693", "title": "Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification", "abstract": "The internet has become a hotspot for hate speech (HS), threatening societal harmony and individual well-being. While automatic detection methods perform well in identifying explicit hate speech (ex-HS), they struggle with more subtle forms, such as implicit hate speech (im-HS). We tackle this problem by introducing a new taxonomy for im-HS detection, defining six encoding strategies named codetypes. We present two methods for integrating codetypes into im-HS detection: 1) prompting large language models (LLMs) directly to classify sentences based on generated responses, and 2) using LLMs as encoders with codetypes embedded during the encoding process. Experiments show that the use of codetypes improves im-HS detection in both Chinese and English datasets, validating the effectiveness of our approach across different languages.", "date": "2025-06-05", "categories": []}
{"id": "2506.03312", "title": "Cross-Platform Violence Detection on Social Media: A Dataset and Analysis", "abstract": "Violent threats remain a significant problem across social media platforms. Useful, high-quality data facilitates research into the understanding and detection of malicious content, including violence. In this paper, we introduce a cross-platform dataset of 30,000 posts hand-coded for violent threats and sub-types of violence, including political and sexual violence. To evaluate the signal present in this dataset, we perform a machine learning analysis with an existing dataset of violent comments from YouTube. We find that, despite originating from different platforms and using different coding criteria, we achieve high classification accuracy both by training on one dataset and testing on the other, and in a merged dataset condition. These results have implications for content-classification strategies and for understanding violent content across social media.", "date": "2025-06-03", "categories": []}
{"id": "2506.03458", "title": "Culture Matters in Toxic Language Detection in Persian", "abstract": "Toxic language detection is crucial for creating safer online environments and limiting the spread of harmful content. While toxic language detection has been under-explored in Persian, the current work compares different methods for this task, including fine-tuning, data enrichment, zero-shot and few-shot learning, and cross-lingual transfer learning. What is especially compelling is the impact of cultural context on transfer learning for this task: We show that the language of a country with cultural similarities to Persian yields better results in transfer learning. Conversely, the improvement is lower when the language comes from a culturally distinct country. Warning: This paper contains examples of toxic language that may disturb some readers. These examples are included for the purpose of research on toxic detection.", "date": "2025-06-03", "categories": []}
{"id": "2506.03916", "title": "Compositional Generalisation for Explainable Hate Speech Detection", "abstract": "Hate speech detection is key to online content moderation, but current models struggle to generalise beyond their training data. This has been linked to dataset biases and the use of sentence-level labels, which fail to teach models the underlying structure of hate speech. In this work, we show that even when models are trained with more fine-grained, span-level annotations (e.g., \"artists\" is labeled as target and \"are parasites\" as dehumanising comparison), they struggle to disentangle the meaning of these labels from the surrounding context. As a result, combinations of expressions that deviate from those seen during training remain particularly difficult for models to detect. We investigate whether training on a dataset where expressions occur with equal frequency across all contexts can improve generalisation. To this end, we create U-PLEAD, a dataset of ~364,000 synthetic posts, along with a novel compositional generalisation benchmark of ~8,000 manually validated posts. Training on a combination of U-PLEAD and real data improves compositional generalisation while achieving state-of-the-art performance on the human-sourced PLEAD.", "date": "2025-06-04", "categories": []}
{"id": "2506.04043", "title": "Think Like a Person Before Responding: A Multi-Faceted Evaluation of Persona-Guided LLMs for Countering Hate", "abstract": "Automated counter-narratives (CN) offer a promising strategy for mitigating online hate speech, yet concerns about their affective tone, accessibility, and ethical risks remain. We propose a framework for evaluating Large Language Model (LLM)-generated CNs across four dimensions: persona framing, verbosity and readability, affective tone, and ethical robustness. Using GPT-4o-Mini, Cohere's CommandR-7B, and Meta's LLaMA 3.1-70B, we assess three prompting strategies on the MT-Conan and HatEval datasets. Our findings reveal that LLM-generated CNs are often verbose and adapted for people with college-level literacy, limiting their accessibility. While emotionally guided prompts yield more empathetic and readable responses, there remain concerns surrounding safety and effectiveness.", "date": "2025-06-04", "categories": []}
{"id": "2411.15462", "title": "HateDay: Insights from a Global Hate Speech Dataset Representative of a Day on Twitter", "abstract": "To address the global challenge of online hate speech, prior research has developed detection models to flag such content on social media. However, due to systematic biases in evaluation datasets, the real-world effectiveness of these models remains unclear, particularly across geographies. We introduce HateDay, the first global hate speech dataset representative of social media settings, constructed from a random sample of all tweets posted on September 21, 2022 and covering eight languages and four English-speaking countries. Using HateDay, we uncover substantial variation in the prevalence and composition of hate speech across languages and regions. We show that evaluations on academic datasets greatly overestimate real-world detection performance, which we find is very low, especially for non-European languages. Our analysis identifies key drivers of this gap, including models' difficulty to distinguish hate from offensive speech and a mismatch between the target groups emphasized in academic datasets and those most frequently targeted in real-world settings. We argue that poor model performance makes public models ill-suited for automatic hate speech moderation and find that high moderation rates are only achievable with substantial human oversight. Our results underscore the need to evaluate detection systems on data that reflects the complexity and diversity of real-world social media.", "date": "2024-11-23", "categories": []}
{"id": "2412.17533", "title": "Behind Closed Words: Creating and Investigating the forePLay Annotated Dataset for Polish Erotic Discourse", "abstract": "The surge in online content has created an urgent demand for robust detection systems, especially in non-English contexts where current tools demonstrate significant limitations. We present forePLay, a novel Polish language dataset for erotic content detection, featuring over 24k annotated sentences with a multidimensional taxonomy encompassing ambiguity, violence, and social unacceptability dimensions. Our comprehensive evaluation demonstrates that specialized Polish language models achieve superior performance compared to multilingual alternatives, with transformer-based architectures showing particular strength in handling imbalanced categories. The dataset and accompanying analysis establish essential frameworks for developing linguistically-aware content moderation systems, while highlighting critical considerations for extending such capabilities to morphologically complex languages.", "date": "2024-12-23", "categories": []}
{"id": "2506.02753", "title": "Multi-task Learning with Active Learning for Arabic Offensive Speech Detection", "abstract": "The rapid growth of social media has amplified the spread of offensive, violent, and vulgar speech, which poses serious societal and cybersecurity concerns. Detecting such content in Arabic text is particularly complex due to limited labeled data, dialectal variations, and the language's inherent complexity. This paper proposes a novel framework that integrates multi-task learning (MTL) with active learning to enhance offensive speech detection in Arabic social media text. By jointly training on two auxiliary tasks, violent and vulgar speech, the model leverages shared representations to improve the detection accuracy of the offensive speech. Our approach dynamically adjusts task weights during training to balance the contribution of each task and optimize performance. To address the scarcity of labeled data, we employ an active learning strategy through several uncertainty sampling techniques to iteratively select the most informative samples for model training. We also introduce weighted emoji handling to better capture semantic cues. Experimental results on the OSACT2022 dataset show that the proposed framework achieves a state-of-the-art macro F1-score of 85.42%, outperforming existing methods while using significantly fewer fine-tuning samples. The findings of this study highlight the potential of integrating MTL with active learning for efficient and accurate offensive language detection in resource-constrained settings.", "date": "2025-06-03", "categories": []}
{"id": "2506.03009", "title": "Conditioning Large Language Models on Legal Systems? Detecting Punishable Hate Speech", "abstract": "The assessment of legal problems requires the consideration of a specific legal system and its levels of abstraction, from constitutional law to statutory law to case law. The extent to which Large Language Models (LLMs) internalize such legal systems is unknown. In this paper, we propose and investigate different approaches to condition LLMs at different levels of abstraction in legal systems. This paper examines different approaches to conditioning LLMs at multiple levels of abstraction in legal systems to detect potentially punishable hate speech. We focus on the task of classifying whether a specific social media posts falls under the criminal offense of incitement to hatred as prescribed by the German Criminal Code. The results show that there is still a significant performance gap between models and legal experts in the legal assessment of hate speech, regardless of the level of abstraction with which the models were conditioned. Our analysis revealed, that models conditioned on abstract legal knowledge lacked deep task understanding, often contradicting themselves and hallucinating answers, while models using concrete legal knowledge performed reasonably well in identifying relevant target groups, but struggled with classifying target conducts.", "date": "2025-06-03", "categories": []}
{"id": "2410.03448", "title": "\"Cold, Calculated, and Condescending\": How AI Identifies and Explains Ableism Compared to Disabled People", "abstract": "People with disabilities (PwD) regularly encounter ableist hate and microaggressions online. These spaces are generally moderated by machine learning models, but little is known about how effectively AI models identify ableist speech and how well their judgments align with PwD. To investigate this, we curated a first-of-its-kind dataset of 200 social media comments targeted towards PwD, and prompted state-of-the art AI models (i.e., Toxicity Classifiers, LLMs) to score toxicity and ableism for each comment, and explain their reasoning. Then, we recruited 190 participants to similarly rate and explain the harm, and evaluate LLM explanations. Our mixed-methods analysis highlighted a major disconnect: AI underestimated toxicity compared to PwD ratings, while its ableism assessments were sporadic and varied. Although LLMs identified some biases, its explanations were flawed--they lacked nuance, made incorrect assumptions, and appeared judgmental instead of educational. Going forward, we discuss challenges and opportunities in designing moderation systems for ableism, and advocate for the involvement of intersectional disabled perspectives in AI.", "date": "2024-10-04", "categories": []}
{"id": "2410.13281", "title": "BanTH: A Multi-label Hate Speech Detection Dataset for Transliterated Bangla", "abstract": "The proliferation of transliterated texts in digital spaces has emphasized the need for detecting and classifying hate speech in languages beyond English, particularly in low-resource languages. As online discourse can perpetuate discrimination based on target groups, e.g. gender, religion, and origin, multi-label classification of hateful content can help in comprehending hate motivation and enhance content moderation. While previous efforts have focused on monolingual or binary hate classification tasks, no work has yet addressed the challenge of multi-label hate speech classification in transliterated Bangla. We introduce BanTH, the first multi-label transliterated Bangla hate speech dataset comprising 37.3k samples. The samples are sourced from YouTube comments, where each instance is labeled with one or more target groups, reflecting the regional demographic. We establish novel transformer encoder-based baselines by further pre-training on transliterated Bangla corpus. We also propose a novel translation-based LLM prompting strategy for transliterated text. Experiments reveal that our further pre-trained encoders are achieving state-of-the-art performance on the BanTH dataset, while our translation-based prompting outperforms other strategies in the zero-shot setting. The introduction of BanTH not only fills a critical gap in hate speech research for Bangla but also sets the stage for future exploration into code-mixed and multi-label classification challenges in underrepresented languages.", "date": "2024-10-17", "categories": []}
{"id": "2412.15268", "title": "Enhancing LLM-based Hatred and Toxicity Detection with Meta-Toxic Knowledge Graph", "abstract": "The rapid growth of social media platforms has raised significant concerns regarding online content toxicity. When Large Language Models (LLMs) are used for toxicity detection, two key challenges emerge: 1) the absence of domain-specific toxic knowledge leads to false negatives; 2) the excessive sensitivity of LLMs to toxic speech results in false positives, limiting freedom of speech. To address these issues, we propose a novel method called MetaTox, leveraging graph search on a meta-toxic knowledge graph to enhance hatred and toxicity detection. First, we construct a comprehensive meta-toxic knowledge graph by utilizing LLMs to extract toxic information through a three-step pipeline, with toxic benchmark datasets serving as corpora. Second, we query the graph via retrieval and ranking processes to supplement accurate, relevant toxic knowledge. Extensive experiments and in-depth case studies across multiple datasets demonstrate that our MetaTox significantly decreases the false positive rate while boosting overall toxicity detection performance. Our code is available at https://github.com/YiboZhao624/MetaTox.", "date": "2024-12-17", "categories": []}
{"id": "2505.11958", "title": "Counterspeech the ultimate shield! Multi-Conditioned Counterspeech Generation through Attributed Prefix Learning", "abstract": "Counterspeech has proven to be a powerful tool to combat hate speech online. Previous studies have focused on generating counterspeech conditioned only on specific intents (single attributed). However, a holistic approach considering multiple attributes simultaneously can yield more nuanced and effective responses. Here, we introduce HiPPrO, Hierarchical Prefix learning with Preference Optimization, a novel two-stage framework that utilizes the effectiveness of attribute-specific prefix embedding spaces hierarchically optimized during the counterspeech generation process in the first phase. Thereafter, we incorporate both reference and reward-free preference optimization to generate more constructive counterspeech. Furthermore, we extend IntentCONANv2 by annotating all 13,973 counterspeech instances with emotion labels by five annotators. HiPPrO leverages hierarchical prefix optimization to integrate these dual attributes effectively. An extensive evaluation demonstrates that HiPPrO achieves a ~38 % improvement in intent conformity and a ~3 %, ~2 %, ~3 % improvement in Rouge-1, Rouge-2, and Rouge-L, respectively, compared to several baseline models. Human evaluations further substantiate the superiority of our approach, highlighting the enhanced relevance and appropriateness of the generated counterspeech. This work underscores the potential of multi-attribute conditioning in advancing the efficacy of counterspeech generation systems. Our code is available on Github and dataset is open-sourced on Hugging-face.", "date": "2025-05-17", "categories": []}
{"id": "2505.18927", "title": "Moderating Harm: Benchmarking Large Language Models for Cyberbullying Detection in YouTube Comments", "abstract": "As online platforms grow, comment sections increasingly host harassment that undermines user experience and well-being. This study benchmarks three leading large language models, OpenAI GPT-4.1, Google Gemini 1.5 Pro, and Anthropic Claude 3 Opus, on a corpus of 5,080 YouTube comments sampled from high-abuse threads in gaming, lifestyle, food vlog, and music channels. The dataset comprises 1,334 harmful and 3,746 non-harmful messages in English, Arabic, and Indonesian, annotated independently by two reviewers with substantial agreement (Cohen's kappa = 0.83). Using a unified prompt and deterministic settings, GPT-4.1 achieved the best overall balance with an F1 score of 0.863, precision of 0.887, and recall of 0.841. Gemini flagged the highest share of harmful posts (recall = 0.875) but its precision fell to 0.767 due to frequent false positives. Claude delivered the highest precision at 0.920 and the lowest false-positive rate of 0.022, yet its recall dropped to 0.720. Qualitative analysis showed that all three models struggle with sarcasm, coded insults, and mixed-language slang. These results underscore the need for moderation pipelines that combine complementary models, incorporate conversational context, and fine-tune for under-represented languages and implicit abuse. A de-identified version of the dataset and full prompts is publicly released to promote reproducibility and further progress in automated content moderation.", "date": "2025-05-25", "categories": []}
{"id": "2506.00583", "title": "The Hidden Language of Harm: Examining the Role of Emojis in Harmful Online Communication and Content Moderation", "abstract": "Social media platforms have become central to modern communication, yet they also harbor offensive content that challenges platform safety and inclusivity. While prior research has primarily focused on textual indicators of offense, the role of emojis, ubiquitous visual elements in online discourse, remains underexplored. Emojis, despite being rarely offensive in isolation, can acquire harmful meanings through symbolic associations, sarcasm, and contextual misuse. In this work, we systematically examine emoji contributions to offensive Twitter messages, analyzing their distribution across offense categories and how users exploit emoji ambiguity. To address this, we propose an LLM-powered, multi-step moderation pipeline that selectively replaces harmful emojis while preserving the tweet's semantic intent. Human evaluations confirm our approach effectively reduces perceived offensiveness without sacrificing meaning. Our analysis also reveals heterogeneous effects across offense types, offering nuanced insights for online communication and emoji moderation.", "date": "2025-05-31", "categories": []}
{"id": "2506.00636", "title": "ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances", "abstract": "Toxic speech on online platforms is a growing concern, impacting user experience and online safety. While text-based toxicity detection is well-studied, audio-based approaches remain underexplored, especially for low-resource languages like Vietnamese. This paper introduces ViToSA (Vietnamese Toxic Spans Audio), the first dataset for toxic spans detection in Vietnamese speech, comprising 11,000 audio samples (25 hours) with accurate human-annotated transcripts. We propose a pipeline that combines ASR and toxic spans detection for fine-grained identification of toxic content. Our experiments show that fine-tuning ASR models on ViToSA significantly reduces WER when transcribing toxic speech, while the text-based toxic spans detection (TSD) models outperform existing baselines. These findings establish a novel benchmark for Vietnamese audio-based toxic spans detection, paving the way for future research in speech content moderation.", "date": "2025-05-31", "categories": []}
{"id": "2505.24341", "title": "Exploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings", "abstract": "Detecting toxic content using language models is important but challenging. While large language models (LLMs) have demonstrated strong performance in understanding Chinese, recent studies show that simple character substitutions in toxic Chinese text can easily confuse the state-of-the-art (SOTA) LLMs. In this paper, we highlight the multimodal nature of Chinese language as a key challenge for deploying LLMs in toxic Chinese detection. First, we propose a taxonomy of 3 perturbation strategies and 8 specific approaches in toxic Chinese content. Then, we curate a dataset based on this taxonomy, and benchmark 9 SOTA LLMs (from both the US and China) to assess if they can detect perturbed toxic Chinese text. Additionally, we explore cost-effective enhancement solutions like in-context learning (ICL) and supervised fine-tuning (SFT). Our results reveal two important findings. (1) LLMs are less capable of detecting perturbed multimodal Chinese toxic contents. (2) ICL or SFT with a small number of perturbed examples may cause the LLMs \"overcorrect'': misidentify many normal Chinese contents as toxic.", "date": "2025-05-30", "categories": []}
{"id": "2505.24538", "title": "Don't Erase, Inform! Detecting and Contextualizing Harmful Language in Cultural Heritage Collections", "abstract": "Cultural Heritage (CH) data hold invaluable knowledge, reflecting the history, traditions, and identities of societies, and shaping our understanding of the past and present. However, many CH collections contain outdated or offensive descriptions that reflect historical biases. CH Institutions (CHIs) face significant challenges in curating these data due to the vast scale and complexity of the task. To address this, we develop an AI-powered tool that detects offensive terms in CH metadata and provides contextual insights into their historical background and contemporary perception. We leverage a multilingual vocabulary co-created with marginalized communities, researchers, and CH professionals, along with traditional NLP techniques and Large Language Models (LLMs). Available as a standalone web app and integrated with major CH platforms, the tool has processed over 7.9 million records, contextualizing the contentious terms detected in their metadata. Rather than erasing these terms, our approach seeks to inform, making biases visible and providing actionable insights for creating more inclusive and accessible CH collections.", "date": "2025-05-30", "categories": []}
{"id": "2501.13977", "title": "Re-ranking Using Large Language Models for Mitigating Exposure to Harmful Content on Social Media Platforms", "abstract": "Social media platforms utilize Machine Learning (ML) and Artificial Intelligence (AI) powered recommendation algorithms to maximize user engagement, which can result in inadvertent exposure to harmful content. Current moderation efforts, reliant on classifiers trained with extensive human-annotated data, struggle with scalability and adapting to new forms of harm. To address these challenges, we propose a novel re-ranking approach using Large Language Models (LLMs) in zero-shot and few-shot settings. Our method dynamically assesses and re-ranks content sequences, effectively mitigating harmful content exposure without requiring extensive labeled data. Alongside traditional ranking metrics, we also introduce two new metrics to evaluate the effectiveness of re-ranking in reducing exposure to harmful content. Through experiments on three datasets, three models and across three configurations, we demonstrate that our LLM-based approach significantly outperforms existing proprietary moderation approaches, offering a scalable and adaptable solution for harm mitigation.", "date": "2025-01-23", "categories": []}
{"id": "2505.23386", "title": "VModA: An Effective Framework for Adaptive NSFW Image Moderation", "abstract": "Not Safe/Suitable for Work (NSFW) content is rampant on social networks and poses serious harm to citizens, especially minors. Current detection methods mainly rely on deep learning-based image recognition and classification. However, NSFW images are now presented in increasingly sophisticated ways, often using image details and complex semantics to obscure their true nature or attract more views. Although still understandable to humans, these images often evade existing detection methods, posing a significant threat. Further complicating the issue, varying regulations across platforms and regions create additional challenges for effective moderation, leading to detection bias and reduced accuracy. To address this, we propose VModA, a general and effective framework that adapts to diverse moderation rules and handles complex, semantically rich NSFW content across categories. Experimental results show that VModA significantly outperforms existing methods, achieving up to a 54.3% accuracy improvement across NSFW types, including those with complex semantics. Further experiments demonstrate that our method exhibits strong adaptability across categories, scenarios, and base VLMs. We also identified inconsistent and controversial label samples in public NSFW benchmark datasets, re-annotated them, and submitted corrections to the original maintainers. Two datasets have confirmed the updates so far. Additionally, we evaluate VModA in real-world scenarios to demonstrate its practical effectiveness.", "date": "2025-05-29", "categories": []}
{"id": "2410.16665", "title": "SafetyAnalyst: Interpretable, Transparent, and Steerable Safety Moderation for AI Behavior", "abstract": "The ideal AI safety moderation system would be both structurally interpretable (so its decisions can be reliably explained) and steerable (to align to safety standards and reflect a community's values), which current systems fall short on. To address this gap, we present SafetyAnalyst, a novel AI safety moderation framework. Given an AI behavior, SafetyAnalyst uses chain-of-thought reasoning to analyze its potential consequences by creating a structured \"harm-benefit tree,\" which enumerates harmful and beneficial actions and effects the AI behavior may lead to, along with likelihood, severity, and immediacy labels that describe potential impacts on stakeholders. SafetyAnalyst then aggregates all effects into a harmfulness score using 28 fully interpretable weight parameters, which can be aligned to particular safety preferences. We applied this framework to develop an open-source LLM prompt safety classification system, distilled from 18.5 million harm-benefit features generated by frontier LLMs on 19k prompts. On comprehensive benchmarks, we show that SafetyAnalyst (average F1=0.81) outperforms existing moderation systems (average F1$<$0.72) on prompt safety classification, while offering the additional advantages of interpretability, transparency, and steerability.", "date": "2024-10-22", "categories": []}
{"id": "2505.21710", "title": "Assessing and Refining ChatGPT's Performance in Identifying Targeting and Inappropriate Language: A Comparative Study", "abstract": "This study evaluates the effectiveness of ChatGPT, an advanced AI model for natural language processing, in identifying targeting and inappropriate language in online comments. With the increasing challenge of moderating vast volumes of user-generated content on social network sites, the role of AI in content moderation has gained prominence. We compared ChatGPT's performance against crowd-sourced annotations and expert evaluations to assess its accuracy, scope of detection, and consistency. Our findings highlight that ChatGPT performs well in detecting inappropriate content, showing notable improvements in accuracy through iterative refinements, particularly in Version 6. However, its performance in targeting language detection showed variability, with higher false positive rates compared to expert judgments. This study contributes to the field by demonstrating the potential of AI models like ChatGPT to enhance automated content moderation systems while also identifying areas for further improvement. The results underscore the importance of continuous model refinement and contextual understanding to better support automated moderation and mitigate harmful online behavior.", "date": "2025-05-27", "categories": []}
{"id": "2505.20963", "title": "Context-Aware Content Moderation for German Newspaper Comments", "abstract": "The increasing volume of online discussions requires advanced automatic content moderation to maintain responsible discourse. While hate speech detection on social media is well-studied, research on German-language newspaper forums remains limited. Existing studies often neglect platform-specific context, such as user history and article themes. This paper addresses this gap by developing and evaluating binary classification models for automatic content moderation in German newspaper forums, incorporating contextual information. Using LSTM, CNN, and ChatGPT-3.5 Turbo, and leveraging the One Million Posts Corpus from the Austrian newspaper Der Standard, we assess the impact of context-aware models. Results show that CNN and LSTM models benefit from contextual information and perform competitively with state-of-the-art approaches. In contrast, ChatGPT's zero-shot classification does not improve with added context and underperforms.", "date": "2025-05-27", "categories": []}
{"id": "2505.21380", "title": "PHISH in MESH: Korean Adversarial Phonetic Substitution and Phonetic-Semantic Feature Integration Defense", "abstract": "As malicious users increasingly employ phonetic substitution to evade hate speech detection, researchers have investigated such strategies. However, two key challenges remain. First, existing studies have overlooked the Korean language, despite its vulnerability to phonetic perturbations due to its phonographic nature. Second, prior work has primarily focused on constructing datasets rather than developing architectural defenses. To address these challenges, we propose (1) PHonetic-Informed Substitution for Hangul (PHISH) that exploits the phonological characteristics of the Korean writing system, and (2) Mixed Encoding of Semantic-pHonetic features (MESH) that enhances the detector's robustness by incorporating phonetic information at the architectural level. Our experimental results demonstrate the effectiveness of our proposed methods on both perturbed and unperturbed datasets, suggesting that they not only improve detection performance but also reflect realistic adversarial behaviors employed by malicious users.", "date": "2025-05-27", "categories": []}
{"id": "2503.00092", "title": "EdgeAIGuard: Agentic LLMs for Minor Protection in Digital Spaces", "abstract": "Social media has become integral to minors' daily lives and is used for various purposes, such as making friends, exploring shared interests, and engaging in educational activities. However, the increase in screen time has also led to heightened challenges, including cyberbullying, online grooming, and exploitations posed by malicious actors. Traditional content moderation techniques have proven ineffective against exploiters' evolving tactics. To address these growing challenges, we propose the EdgeAIGuard content moderation approach that is designed to protect minors from online grooming and various forms of digital exploitation. The proposed method comprises a multi-agent architecture deployed strategically at the network edge to enable rapid detection with low latency and prevent harmful content targeting minors. The experimental results show the proposed method is significantly more effective than the existing approaches.", "date": "2025-02-28", "categories": []}
{"id": "2505.06149", "title": "Can Prompting LLMs Unlock Hate Speech Detection across Languages? A Zero-shot and Few-shot Study", "abstract": "Despite growing interest in automated hate speech detection, most existing approaches overlook the linguistic diversity of online content. Multilingual instruction-tuned large language models such as LLaMA, Aya, Qwen, and BloomZ offer promising capabilities across languages, but their effectiveness in identifying hate speech through zero-shot and few-shot prompting remains underexplored. This work evaluates LLM prompting-based detection across eight non-English languages, utilizing several prompting techniques and comparing them to fine-tuned encoder models. We show that while zero-shot and few-shot prompting lag behind fine-tuned encoder models on most of the real-world evaluation sets, they achieve better generalization on functional tests for hate speech detection. Our study also reveals that prompt design plays a critical role, with each language often requiring customized prompting techniques to maximize performance.", "date": "2025-05-09", "categories": []}
{"id": "2505.14272", "title": "Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval with Limited Labeled Data", "abstract": "Considering the importance of detecting hateful language, labeled hate speech data is expensive and time-consuming to collect, particularly for low-resource languages. Prior work has demonstrated the effectiveness of cross-lingual transfer learning and data augmentation in improving performance on tasks with limited labeled data. To develop an efficient and scalable cross-lingual transfer learning approach, we leverage nearest-neighbor retrieval to augment minimal labeled data in the target language, thereby enhancing detection performance. Specifically, we assume access to a small set of labeled training instances in the target language and use these to retrieve the most relevant labeled examples from a large multilingual hate speech detection pool. We evaluate our approach on eight languages and demonstrate that it consistently outperforms models trained solely on the target language data. Furthermore, in most cases, our method surpasses the current state-of-the-art. Notably, our approach is highly data-efficient, retrieving as small as 200 instances in some cases while maintaining superior performance. Moreover, it is scalable, as the retrieval pool can be easily expanded, and the method can be readily adapted to new languages and tasks. We also apply maximum marginal relevance to mitigate redundancy and filter out highly similar retrieved instances, resulting in improvements in some languages.", "date": "2025-05-20", "categories": []}
{"id": "2505.20067", "title": "Community Moderation and the New Epistemology of Fact Checking on Social Media", "abstract": "Social media platforms have traditionally relied on internal moderation teams and partnerships with independent fact-checking organizations to identify and flag misleading content. Recently, however, platforms including X (formerly Twitter) and Meta have shifted towards community-driven content moderation by launching their own versions of crowd-sourced fact-checking -- Community Notes. If effectively scaled and governed, such crowd-checking initiatives have the potential to combat misinformation with increased scale and speed as successfully as community-driven efforts once did with spam. Nevertheless, general content moderation, especially for misinformation, is inherently more complex. Public perceptions of truth are often shaped by personal biases, political leanings, and cultural contexts, complicating consensus on what constitutes misleading content. This suggests that community efforts, while valuable, cannot replace the indispensable role of professional fact-checkers. Here we systemically examine the current approaches to misinformation detection across major platforms, explore the emerging role of community-driven moderation, and critically evaluate both the promises and challenges of crowd-checking at scale.", "date": "2025-05-26", "categories": []}
{"id": "2505.11969", "title": "An Annotated Corpus of Arabic Tweets for Hate Speech Analysis", "abstract": "Identifying hate speech content in the Arabic language is challenging due to the rich quality of dialectal variations. This study introduces a multilabel hate speech dataset in the Arabic language. We have collected 10000 Arabic tweets and annotated each tweet, whether it contains offensive content or not. If a text contains offensive content, we further classify it into different hate speech targets such as religion, gender, politics, ethnicity, origin, and others. A text can contain either single or multiple targets. Multiple annotators are involved in the data annotation task. We calculated the inter-annotator agreement, which was reported to be 0.86 for offensive content and 0.71 for multiple hate speech targets. Finally, we evaluated the data annotation task by employing a different transformers-based model in which AraBERTv2 outperformed with a micro-F1 score of 0.7865 and an accuracy of 0.786.", "date": "2025-05-17", "categories": []}
{"id": "2505.17207", "title": "Content Moderation in TV Search: Balancing Policy Compliance, Relevance, and User Experience", "abstract": "Millions of people rely on search functionality to find and explore content on entertainment platforms. Modern search systems use a combination of candidate generation and ranking approaches, with advanced methods leveraging deep learning and LLM-based techniques to retrieve, generate, and categorize search results. Despite these advancements, search algorithms can still surface inappropriate or irrelevant content due to factors like model unpredictability, metadata errors, or overlooked design flaws. Such issues can misalign with product goals and user expectations, potentially harming user trust and business outcomes. In this work, we introduce an additional monitoring layer using Large Language Models (LLMs) to enhance content moderation. This additional layer flags content if the user did not intend to search for it. This approach serves as a baseline for product quality assurance, with collected feedback used to refine the initial retrieval mechanisms of the search model, ensuring a safer and more reliable user experience.", "date": "2025-05-22", "categories": []}
{"id": "2411.01844", "title": "DeMod: A Holistic Tool with Explainable Detection and Personalized Modification for Toxicity Censorship", "abstract": "Although there have been automated approaches and tools supporting toxicity censorship for social posts, most of them focus on detection. Toxicity censorship is a complex process, wherein detection is just an initial task and a user can have further needs such as rationale understanding and content modification. For this problem, we conduct a needfinding study to investigate people's diverse needs in toxicity censorship and then build a ChatGPT-based censorship tool named DeMod accordingly. DeMod is equipped with the features of explainable Detection and personalized Modification, providing fine-grained detection results, detailed explanations, and personalized modification suggestions. We also implemented the tool and recruited 35 Weibo users for evaluation. The results suggest DeMod's multiple strengths like the richness of functionality, the accuracy of censorship, and ease of use. Based on the findings, we further propose several insights into the design of content censorship systems.", "date": "2024-11-04", "categories": []}
{"id": "2505.16263", "title": "All You Need is \"Leet\": Evading Hate-speech Detection AI", "abstract": "Social media and online forums are increasingly becoming popular. Unfortunately, these platforms are being used for spreading hate speech. In this paper, we design black-box techniques to protect users from hate-speech on online platforms by generating perturbations that can fool state of the art deep learning based hate speech detection models thereby decreasing their efficiency. We also ensure a minimal change in the original meaning of hate-speech. Our best perturbation attack is successfully able to evade hate-speech detection for 86.8 % of hateful text.", "date": "2025-05-22", "categories": []}
{"id": "2505.16847", "title": "Understanding and Analyzing Inappropriately Targeting Language in Online Discourse: A Comparative Annotation Study", "abstract": "This paper introduces a method for detecting inappropriately targeting language in online conversations by integrating crowd and expert annotations with ChatGPT. We focus on English conversation threads from Reddit, examining comments that target individuals or groups. Our approach involves a comprehensive annotation framework that labels a diverse data set for various target categories and specific target words within the conversational context. We perform a comparative analysis of annotations from human experts, crowd annotators, and ChatGPT, revealing strengths and limitations of each method in recognizing both explicit hate speech and subtler discriminatory language. Our findings highlight the significant role of contextual factors in identifying hate speech and uncover new categories of targeting, such as social belief and body image. We also address the challenges and subjective judgments involved in annotation and the limitations of ChatGPT in grasping nuanced language. This study provides insights for improving automated content moderation strategies to enhance online safety and inclusivity.", "date": "2025-05-22", "categories": []}
{"id": "2411.07417", "title": "Untangling Hate Speech Definitions: A Semantic Componential Analysis Across Cultures and Domains", "abstract": "Hate speech relies heavily on cultural influences, leading to varying individual interpretations. For that reason, we propose a Semantic Componential Analysis (SCA) framework for a cross-cultural and cross-domain analysis of hate speech definitions. We create the first dataset of hate speech definitions encompassing 493 definitions from more than 100 cultures, drawn from five key domains: online dictionaries, academic research, Wikipedia, legal texts, and online platforms. By decomposing these definitions into semantic components, our analysis reveals significant variation across definitions, yet many domains borrow definitions from one another without taking into account the target culture. We conduct zero-shot model experiments using our proposed dataset, employing three popular open-sourced LLMs to understand the impact of different definitions on hate speech detection. Our findings indicate that LLMs are sensitive to definitions: responses for hate speech detection change according to the complexity of definitions used in the prompt.", "date": "2024-11-11", "categories": []}
{"id": "2505.15297", "title": "Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites", "abstract": "Detoxifying offensive language while preserving the speaker's original intent is a challenging yet critical goal for improving the quality of online interactions. Although large language models (LLMs) show promise in rewriting toxic content, they often default to overly polite rewrites, distorting the emotional tone and communicative intent. This problem is especially acute in Chinese, where toxicity often arises implicitly through emojis, homophones, or discourse context. We present ToxiRewriteCN, the first Chinese detoxification dataset explicitly designed to preserve sentiment polarity. The dataset comprises 1,556 carefully annotated triplets, each containing a toxic sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans. It covers five real-world scenarios: standard expressions, emoji-induced and homophonic toxicity, as well as single-turn and multi-turn dialogues. We evaluate 17 LLMs, including commercial and open-source models with variant architectures, across four dimensions: detoxification accuracy, fluency, content preservation, and sentiment polarity. Results show that while commercial and MoE models perform best overall, all models struggle to balance safety with emotional fidelity in more subtle or context-heavy settings such as emoji, homophone, and dialogue-based inputs. We release ToxiRewriteCN to support future research on controllable, sentiment-aware detoxification for Chinese.", "date": "2025-05-21", "categories": []}
{"id": "2505.15773", "title": "ToxicTone: A Mandarin Audio Dataset Annotated for Toxicity and Toxic Utterance Tonality", "abstract": "Despite extensive research on toxic speech detection in text, a critical gap remains in handling spoken Mandarin audio. The lack of annotated datasets that capture the unique prosodic cues and culturally specific expressions in Mandarin leaves spoken toxicity underexplored. To address this, we introduce ToxicTone -- the largest public dataset of its kind -- featuring detailed annotations that distinguish both forms of toxicity (e.g., profanity, bullying) and sources of toxicity (e.g., anger, sarcasm, dismissiveness). Our data, sourced from diverse real-world audio and organized into 13 topical categories, mirrors authentic communication scenarios. We also propose a multimodal detection framework that integrates acoustic, linguistic, and emotional features using state-of-the-art speech and emotion encoders. Extensive experiments show our approach outperforms text-only and baseline models, underscoring the essential role of speech-specific cues in revealing hidden toxic expressions.", "date": "2025-05-21", "categories": []}
{"id": "2501.15451", "title": "STATE ToxiCN: A Benchmark for Span-level Target-Aware Toxicity Extraction in Chinese Hate Speech Detection", "abstract": "The proliferation of hate speech has caused significant harm to society. The intensity and directionality of hate are closely tied to the target and argument it is associated with. However, research on hate speech detection in Chinese has lagged behind, and existing datasets lack span-level fine-grained annotations. Furthermore, the lack of research on Chinese hateful slang poses a significant challenge. In this paper, we provide a solution for fine-grained detection of Chinese hate speech. First, we construct a dataset containing Target-Argument-Hateful-Group quadruples (STATE ToxiCN), which is the first span-level Chinese hate speech dataset. Secondly, we evaluate the span-level hate speech detection performance of existing models using STATE ToxiCN. Finally, we conduct the first study on Chinese hateful slang and evaluate the ability of LLMs to detect such expressions. Our work contributes valuable resources and insights to advance span-level hate speech detection in Chinese.", "date": "2025-01-26", "categories": []}
{"id": "2404.17874", "title": "From Languages to Geographies: Towards Evaluating Cultural Bias in Hate Speech Datasets", "abstract": "Perceptions of hate can vary greatly across cultural contexts. Hate speech (HS) datasets, however, have traditionally been developed by language. This hides potential cultural biases, as one language may be spoken in different countries home to different cultures. In this work, we evaluate cultural bias in HS datasets by leveraging two interrelated cultural proxies: language and geography. We conduct a systematic survey of HS datasets in eight languages and confirm past findings on their English-language bias, but also show that this bias has been steadily decreasing in the past few years. For three geographically-widespread languages -- English, Arabic and Spanish -- we then leverage geographical metadata from tweets to approximate geo-cultural contexts by pairing language and country information. We find that HS datasets for these languages exhibit a strong geo-cultural bias, largely overrepresenting a handful of countries (e.g., US and UK for English) relative to their prominence in both the broader social media population and the general population speaking these languages. Based on these findings, we formulate recommendations for the creation of future HS datasets.", "date": "2024-04-27", "categories": []}
{"id": "2502.06207", "title": "Is LLM an Overconfident Judge? Unveiling the Capabilities of LLMs in Detecting Offensive Language with Annotation Disagreement", "abstract": "Large Language Models (LLMs) have become essential for offensive language detection, yet their ability to handle annotation disagreement remains underexplored. Disagreement samples, which arise from subjective interpretations, pose a unique challenge due to their ambiguous nature. Understanding how LLMs process these cases, particularly their confidence levels, can offer insight into their alignment with human annotators. This study systematically evaluates the performance of multiple LLMs in detecting offensive language at varying levels of annotation agreement. We analyze binary classification accuracy, examine the relationship between model confidence and human disagreement, and explore how disagreement samples influence model decision-making during few-shot learning and instruction fine-tuning. Our findings reveal that LLMs struggle with low-agreement samples, often exhibiting overconfidence in these ambiguous cases. However, utilizing disagreement samples in training improves both detection accuracy and model alignment with human judgment. These insights provide a foundation for enhancing LLM-based offensive language detection in real-world moderation tasks.", "date": "2025-02-10", "categories": []}
{"id": "2505.13251", "title": "Stronger Together: Unleashing the Social Impact of Hate Speech Research", "abstract": "The advent of the internet has been both a blessing and a curse for once marginalised communities. When used well, the internet can be used to connect and establish communities crossing different intersections; however, it can also be used as a tool to alienate people and communities as well as perpetuate hate, misinformation, and disinformation especially on social media platforms. We propose steering hate speech research and researchers away from pre-existing computational solutions and consider social methods to inform social solutions to address this social problem. In a similar way linguistics research can inform language planning policy, linguists should apply what we know about language and society to mitigate some of the emergent risks and dangers of anti-social behaviour in digital spaces. We argue linguists and NLP researchers can play a principle role in unleashing the social impact potential of linguistics research working alongside communities, advocates, activists, and policymakers to enable equitable digital inclusion and to close the digital divide.", "date": "2025-05-19", "categories": []}
{"id": "2409.20204", "title": "Divided by discipline? A systematic literature review on the quantification of online sexism and misogyny using a semi-automated approach", "abstract": "Several computational tools have been developed to detect and identify sexism, misogyny, and gender-based hate speech, particularly on online platforms. These tools draw on insights from both social science and computer science. Given the increasing concern over gender-based discrimination in digital spaces, the contested definitions and measurements of sexism, and the rise of interdisciplinary efforts to understand its online manifestations, a systematic literature review is essential for capturing the current state and trajectory of this evolving field. In this review, we make four key contributions: (1) we synthesize the literature into five core themes: definitions of sexism and misogyny, disciplinary divergences, automated detection methods, associated challenges, and design-based interventions; (2) we adopt an interdisciplinary lens, bridging theoretical and methodological divides across disciplines; (3) we highlight critical gaps, including the need for intersectional approaches, the under-representation of non-Western languages and perspectives, and the limited focus on proactive design strategies beyond text classification; and (4) we offer a methodological contribution by applying a rigorous semi-automated systematic review process guided by PRISMA, establishing a replicable standard for future work in this domain. Our findings reveal a clear disciplinary divide in how sexism and misogyny are conceptualized and measured. Through an evidence-based synthesis, we examine how existing studies have attempted to bridge this gap through interdisciplinary collaboration. Drawing on both social science theories and computational modeling practices, we assess the strengths and limitations of current methodologies. Finally, we outline key challenges and future directions for advancing research on the detection and mitigation of online sexism and misogyny.", "date": "2024-09-30", "categories": []}
{"id": "2411.08135", "title": "On the Role of Speech Data in Reducing Toxicity Detection Bias", "abstract": "Text toxicity detection systems exhibit significant biases, producing disproportionate rates of false positives on samples mentioning demographic groups. But what about toxicity detection in speech? To investigate the extent to which text-based biases are mitigated by speech-based systems, we produce a set of high-quality group annotations for the multilingual MuTox dataset, and then leverage these annotations to systematically compare speech- and text-based toxicity classifiers. Our findings indicate that access to speech data during inference supports reduced bias against group mentions, particularly for ambiguous and disagreement-inducing samples. Our results also suggest that improving classifiers, rather than transcription pipelines, is more helpful for reducing group bias. We publicly release our annotations and provide recommendations for future toxicity dataset construction.", "date": "2024-11-12", "categories": []}
{"id": "2505.10588", "title": "Understanding Gen Alpha Digital Language: Evaluation of LLM Safety Systems for Content Moderation", "abstract": "This research offers a unique evaluation of how AI systems interpret the digital language of Generation Alpha (Gen Alpha, born 2010-2024). As the first cohort raised alongside AI, Gen Alpha faces new forms of online risk due to immersive digital engagement and a growing mismatch between their evolving communication and existing safety tools. Their distinct language, shaped by gaming, memes, and AI-driven trends, often conceals harmful interactions from both human moderators and automated systems. We assess four leading AI models (GPT-4, Claude, Gemini, and Llama 3) on their ability to detect masked harassment and manipulation within Gen Alpha discourse. Using a dataset of 100 recent expressions from gaming platforms, social media, and video content, the study reveals critical comprehension failures with direct implications for online safety. This work contributes: (1) a first-of-its-kind dataset capturing Gen Alpha expressions; (2) a framework to improve AI moderation systems for youth protection; (3) a multi-perspective evaluation including AI systems, human moderators, and parents, with direct input from Gen Alpha co-researchers; and (4) an analysis of how linguistic divergence increases youth vulnerability. Findings highlight the urgent need to redesign safety systems attuned to youth communication, especially given Gen Alpha reluctance to seek help when adults fail to understand their digital world. This study combines the insight of a Gen Alpha researcher with systematic academic analysis to address critical digital safety challenges.", "date": "2025-05-14", "categories": []}
{"id": "2505.03746", "title": "Promoting Security and Trust on Social Networks: Explainable\n  Cyberbullying Detection Using Large Language Models in a Stream-Based Machine\n  Learning Framework", "abstract": "  Social media platforms enable instant and ubiquitous connectivity and are\nessential to social interaction and communication in our technological society.\nApart from its advantages, these platforms have given rise to negative\nbehaviors in the online community, the so-called cyberbullying. Despite the\nmany works involving generative Artificial Intelligence (AI) in the literature\nlately, there remain opportunities to study its performance apart from\nzero/few-shot learning strategies. Accordingly, we propose an innovative and\nreal-time solution for cyberbullying detection that leverages stream-based\nMachine Learning (ML) models able to process the incoming samples incrementally\nand Large Language Models (LLMS) for feature engineering to address the\nevolving nature of abusive and hate speech online. An explainability dashboard\nis provided to promote the system's trustworthiness, reliability, and\naccountability. Results on experimental data report promising performance close\nto 90 % in all evaluation metrics and surpassing those obtained by competing\nworks in the literature. Ultimately, our proposal contributes to the safety of\nonline communities by timely detecting abusive behavior to prevent long-lasting\nharassment and reduce the negative consequences in society.\n", "date": "2025-04-07", "categories": []}
{"id": "2401.06526", "title": "MetaHate: A Dataset for Unifying Efforts on Hate Speech Detection", "abstract": "  Hate speech represents a pervasive and detrimental form of online discourse,\noften manifested through an array of slurs, from hateful tweets to defamatory\nposts. As such speech proliferates, it connects people globally and poses\nsignificant social, psychological, and occasionally physical threats to\ntargeted individuals and communities. Current computational linguistic\napproaches for tackling this phenomenon rely on labelled social media datasets\nfor training. For unifying efforts, our study advances in the critical need for\na comprehensive meta-collection, advocating for an extensive dataset to help\ncounteract this problem effectively. We scrutinized over 60 datasets,\nselectively integrating those pertinent into MetaHate. This paper offers a\ndetailed examination of existing collections, highlighting their strengths and\nlimitations. Our findings contribute to a deeper understanding of the existing\ndatasets, paving the way for training more robust and adaptable models. These\nenhanced models are essential for effectively combating the dynamic and complex\nnature of hate speech in the digital realm.\n", "date": "2024-01-12", "categories": []}
{"id": "2412.13698", "title": "Towards Efficient and Explainable Hate Speech Detection via Model\n  Distillation", "abstract": "  Automatic detection of hate and abusive language is essential to combat its\nonline spread. Moreover, recognising and explaining hate speech serves to\neducate people about its negative effects. However, most current detection\nmodels operate as black boxes, lacking interpretability and explainability. In\nthis context, Large Language Models (LLMs) have proven effective for hate\nspeech detection and to promote interpretability. Nevertheless, they are\ncomputationally costly to run. In this work, we propose distilling big language\nmodels by using Chain-of-Thought to extract explanations that support the hate\nspeech classification task. Having small language models for these tasks will\ncontribute to their use in operational settings. In this paper, we demonstrate\nthat distilled models deliver explanations of the same quality as larger models\nwhile surpassing them in classification performance. This dual capability,\nclassifying and explaining, advances hate speech detection making it more\naffordable, understandable and actionable.\n", "date": "2024-12-18", "categories": []}
{"id": "2505.02252", "title": "Personalisation or Prejudice? Addressing Geographic Bias in Hate Speech\n  Detection using Debias Tuning in Large Language Models", "abstract": "  Commercial Large Language Models (LLMs) have recently incorporated memory\nfeatures to deliver personalised responses. This memory retains details such as\nuser demographics and individual characteristics, allowing LLMs to adjust their\nbehaviour based on personal information. However, the impact of integrating\npersonalised information into the context has not been thoroughly assessed,\nleading to questions about its influence on LLM behaviour. Personalisation can\nbe challenging, particularly with sensitive topics. In this paper, we examine\nvarious state-of-the-art LLMs to understand their behaviour in different\npersonalisation scenarios, specifically focusing on hate speech. We prompt the\nmodels to assume country-specific personas and use different languages for hate\nspeech detection. Our findings reveal that context personalisation\nsignificantly influences LLMs' responses in this sensitive area. To mitigate\nthese unwanted biases, we fine-tune the LLMs by penalising inconsistent hate\nspeech classifications made with and without country or language-specific\ncontext. The refined models demonstrate improved performance in both\npersonalised contexts and when no context is provided.\n", "date": "2025-05-04", "categories": []}
{"id": "2505.00150", "title": "Detecting and Mitigating Hateful Content in Multimodal Memes with\n  Vision-Language Models", "abstract": "  The rapid evolution of social media has provided enhanced communication\nchannels for individuals to create online content, enabling them to express\ntheir thoughts and opinions. Multimodal memes, often utilized for playful or\nhumorous expressions with visual and textual elements, are sometimes misused to\ndisseminate hate speech against individuals or groups. While the detection of\nhateful memes is well-researched, developing effective methods to transform\nhateful content in memes remains a significant challenge. Leveraging the\npowerful generation and reasoning capabilities of Vision-Language Models\n(VLMs), we address the tasks of detecting and mitigating hateful content. This\npaper presents two key contributions: first, a definition-guided prompting\ntechnique for detecting hateful memes, and second, a unified framework for\nmitigating hateful content in memes, named UnHateMeme, which works by replacing\nhateful textual and/or visual components. With our definition-guided prompts,\nVLMs achieve impressive performance on hateful memes detection task.\nFurthermore, our UnHateMeme framework, integrated with VLMs, demonstrates a\nstrong capability to convert hateful memes into non-hateful forms that meet\nhuman-level criteria for hate speech and maintain multimodal coherence between\nimage and text. Through empirical experiments, we show the effectiveness of\nstate-of-the-art pretrained VLMs such as LLaVA, Gemini and GPT-4o on the\nproposed tasks, providing a comprehensive analysis of their respective\nstrengths and limitations for these tasks. This paper aims to shed light on\nimportant applications of VLMs for ensuring safe and respectful online\nenvironments.\n", "date": "2025-04-30", "categories": []}
{"id": "2310.18964", "title": "LLMs and Finetuning: Benchmarking cross-domain performance for hate\n  speech detection", "abstract": "  In the evolving landscape of online communication, hate speech detection\nremains a formidable challenge, further compounded by the diversity of digital\nplatforms. This study investigates the effectiveness and adaptability of\npre-trained and fine-tuned Large Language Models (LLMs) in identifying hate\nspeech, to address two central questions: (1) To what extent does the model\nperformance depend on the fine-tuning and training parameters?, (2) To what\nextent do models generalize to cross-domain hate speech detection? and (3) What\nare the specific features of the datasets or models that influence the\ngeneralization potential? The experiment shows that LLMs offer a huge advantage\nover the state-of-the-art even without pretraining. Ordinary least squares\nanalyses suggest that the advantage of training with fine-grained hate speech\nlabels is washed away with the increase in dataset size. While our research\ndemonstrates the potential of large language models (LLMs) for hate speech\ndetection, several limitations remain, particularly regarding the validity and\nthe reproducibility of the results. We conclude with an exhaustive discussion\nof the challenges we faced in our experimentation and offer recommended best\npractices for future scholars designing benchmarking experiments of this kind.\n", "date": "2023-10-29", "categories": []}
{"id": "2408.05794", "title": "HateSieve: A Contrastive Learning Framework for Detecting and Segmenting\n  Hateful Content in Multimodal Memes", "abstract": "  Amidst the rise of Large Multimodal Models (LMMs) and their widespread\napplication in generating and interpreting complex content, the risk of\npropagating biased and harmful memes remains significant. Current safety\nmeasures often fail to detect subtly integrated hateful content within\n``Confounder Memes''. To address this, we introduce \\textsc{HateSieve}, a new\nframework designed to enhance the detection and segmentation of hateful\nelements in memes. \\textsc{HateSieve} features a novel Contrastive Meme\nGenerator that creates semantically paired memes, a customized triplet dataset\nfor contrastive learning, and an Image-Text Alignment module that produces\ncontext-aware embeddings for accurate meme segmentation. Empirical experiments\non the Hateful Meme Dataset show that \\textsc{HateSieve} not only surpasses\nexisting LMMs in performance with fewer trainable parameters but also offers a\nrobust mechanism for precisely identifying and isolating hateful content.\n\\textcolor{red}{Caution: Contains academic discussions of hate speech; viewer\ndiscretion advised.}\n", "date": "2024-08-11", "categories": []}
{"id": "2504.21026", "title": "Creating and Evaluating Code-Mixed Nepali-English and Telugu-English\n  Datasets for Abusive Language Detection Using Traditional and Deep Learning\n  Models", "abstract": "  With the growing presence of multilingual users on social media, detecting\nabusive language in code-mixed text has become increasingly challenging.\nCode-mixed communication, where users seamlessly switch between English and\ntheir native languages, poses difficulties for traditional abuse detection\nmodels, as offensive content may be context-dependent or obscured by linguistic\nblending. While abusive language detection has been extensively explored for\nhigh-resource languages like English and Hindi, low-resource languages such as\nTelugu and Nepali remain underrepresented, leaving gaps in effective\nmoderation. In this study, we introduce a novel, manually annotated dataset of\n2 thousand Telugu-English and 5 Nepali-English code-mixed comments, categorized\nas abusive and non-abusive, collected from various social media platforms. The\ndataset undergoes rigorous preprocessing before being evaluated across multiple\nMachine Learning (ML), Deep Learning (DL), and Large Language Models (LLMs). We\nexperimented with models including Logistic Regression, Random Forest, Support\nVector Machines (SVM), Neural Networks (NN), LSTM, CNN, and LLMs, optimizing\ntheir performance through hyperparameter tuning, and evaluate it using 10-fold\ncross-validation and statistical significance testing (t-test). Our findings\nprovide key insights into the challenges of detecting abusive language in\ncode-mixed settings and offer a comparative analysis of computational\napproaches. This study contributes to advancing NLP for low-resource languages\nby establishing benchmarks for abusive language detection in Telugu-English and\nNepali-English code-mixed text. The dataset and insights can aid in the\ndevelopment of more robust moderation strategies for multilingual social media\nenvironments.\n", "date": "2025-04-23", "categories": []}
{"id": "2303.10430", "title": "NoisyHate: Mining Online Human-Written Perturbations for Realistic\n  Robustness Benchmarking of Content Moderation Models", "abstract": "  Online texts with toxic content are a clear threat to the users on social\nmedia in particular and society in general. Although many platforms have\nadopted various measures (e.g., machine learning-based hate-speech detection\nsystems) to diminish their effect, toxic content writers have also attempted to\nevade such measures by using cleverly modified toxic words, so-called\nhuman-written text perturbations. Therefore, to help build automatic detection\ntools to recognize those perturbations, prior methods have developed\nsophisticated techniques to generate diverse adversarial samples. However, we\nnote that these ``algorithms\"-generated perturbations do not necessarily\ncapture all the traits of ``human\"-written perturbations. Therefore, in this\npaper, we introduce a novel, high-quality dataset of human-written\nperturbations, named as NoisyHate, that was created from real-life\nperturbations that are both written and verified by human-in-the-loop. We show\nthat perturbations in NoisyHate have different characteristics than prior\nalgorithm-generated toxic datasets show, and thus can be in particular useful\nto help develop better toxic speech detection solutions. We thoroughly validate\nNoisyHate against state-of-the-art language models, such as BERT and RoBERTa,\nand black box APIs, such as Perspective API, on two tasks, such as perturbation\nnormalization and understanding.\n", "date": "2023-03-18", "categories": []}
{"id": "2504.16723", "title": "Detecting and Understanding Hateful Contents in Memes Through Captioning\n  and Visual Question-Answering", "abstract": "  Memes are widely used for humor and cultural commentary, but they are\nincreasingly exploited to spread hateful content. Due to their multimodal\nnature, hateful memes often evade traditional text-only or image-only detection\nsystems, particularly when they employ subtle or coded references. To address\nthese challenges, we propose a multimodal hate detection framework that\nintegrates key components: OCR to extract embedded text, captioning to describe\nvisual content neutrally, sub-label classification for granular categorization\nof hateful content, RAG for contextually relevant retrieval, and VQA for\niterative analysis of symbolic and contextual cues. This enables the framework\nto uncover latent signals that simpler pipelines fail to detect. Experimental\nresults on the Facebook Hateful Memes dataset reveal that the proposed\nframework exceeds the performance of unimodal and conventional multimodal\nmodels in both accuracy and AUC-ROC.\n", "date": "2025-04-23", "categories": []}
{"id": "2504.15548", "title": "LLM-based Semantic Augmentation for Harmful Content Detection", "abstract": "  Recent advances in large language models (LLMs) have demonstrated strong\nperformance on simple text classification tasks, frequently under zero-shot\nsettings. However, their efficacy declines when tackling complex social media\nchallenges such as propaganda detection, hateful meme classification, and\ntoxicity identification. Much of the existing work has focused on using LLMs to\ngenerate synthetic training data, overlooking the potential of LLM-based text\npreprocessing and semantic augmentation. In this paper, we introduce an\napproach that prompts LLMs to clean noisy text and provide context-rich\nexplanations, thereby enhancing training sets without substantial increases in\ndata volume. We systematically evaluate on the SemEval 2024 multi-label\nPersuasive Meme dataset and further validate on the Google Jigsaw toxic\ncomments and Facebook hateful memes datasets to assess generalizability. Our\nresults reveal that zero-shot LLM classification underperforms on these\nhigh-context tasks compared to supervised models. In contrast, integrating\nLLM-based semantic augmentation yields performance on par with approaches that\nrely on human-annotated data, at a fraction of the cost. These findings\nunderscore the importance of strategically incorporating LLMs into machine\nlearning (ML) pipeline for social media classification tasks, offering broad\nimplications for combating harmful content online.\n", "date": "2025-04-22", "categories": []}
{"id": "2504.15987", "title": "Few-shot Hate Speech Detection Based on the MindSpore Framework", "abstract": "  The proliferation of hate speech on social media poses a significant threat\nto online communities, requiring effective detection systems. While deep\nlearning models have shown promise, their performance often deteriorates in\nfew-shot or low-resource settings due to reliance on large annotated corpora.\nTo address this, we propose MS-FSLHate, a prompt-enhanced neural framework for\nfew-shot hate speech detection implemented on the MindSpore deep learning\nplatform. The model integrates learnable prompt embeddings, a CNN-BiLSTM\nbackbone with attention pooling, and synonym-based adversarial data\naugmentation to improve generalization. Experimental results on two benchmark\ndatasets-HateXplain and HSOL-demonstrate that our approach outperforms\ncompetitive baselines in precision, recall, and F1-score. Additionally, the\nframework shows high efficiency and scalability, suggesting its suitability for\ndeployment in resource-constrained environments. These findings highlight the\npotential of combining prompt-based learning with adversarial augmentation for\nrobust and adaptable hate speech detection in few-shot scenarios.\n", "date": "2025-04-22", "categories": []}
{"id": "2504.14904", "title": "VLM as Policy: Common-Law Content Moderation Framework for Short Video\n  Platform", "abstract": "  Exponentially growing short video platforms (SVPs) face significant\nchallenges in moderating content detrimental to users' mental health,\nparticularly for minors. The dissemination of such content on SVPs can lead to\ncatastrophic societal consequences. Although substantial efforts have been\ndedicated to moderating such content, existing methods suffer from critical\nlimitations: (1) Manual review is prone to human bias and incurs high\noperational costs. (2) Automated methods, though efficient, lack nuanced\ncontent understanding, resulting in lower accuracy. (3) Industrial moderation\nregulations struggle to adapt to rapidly evolving trends due to long update\ncycles. In this paper, we annotate the first SVP content moderation benchmark\nwith authentic user/reviewer feedback to fill the absence of benchmark in this\nfield. Then we evaluate various methods on the benchmark to verify the\nexistence of the aforementioned limitations. We further propose our common-law\ncontent moderation framework named KuaiMod to address these challenges. KuaiMod\nconsists of three components: training data construction, offline adaptation,\nand online deployment & refinement. Leveraging large vision language model\n(VLM) and Chain-of-Thought (CoT) reasoning, KuaiMod adequately models video\ntoxicity based on sparse user feedback and fosters dynamic moderation policy\nwith rapid update speed and high accuracy. Offline experiments and large-scale\nonline A/B test demonstrates the superiority of KuaiMod: KuaiMod achieves the\nbest moderation performance on our benchmark. The deployment of KuaiMod reduces\nthe user reporting rate by 20% and its application in video recommendation\nincreases both Daily Active User (DAU) and APP Usage Time (AUT) on several\nKuaishou scenarios. We have open-sourced our benchmark at\nhttps://kuaimod.github.io.\n", "date": "2025-04-21", "categories": []}
{"id": "2504.12082", "title": "Selective Demonstration Retrieval for Improved Implicit Hate Speech\n  Detection", "abstract": "  Hate speech detection is a crucial area of research in natural language\nprocessing, essential for ensuring online community safety. However, detecting\nimplicit hate speech, where harmful intent is conveyed in subtle or indirect\nways, remains a major challenge. Unlike explicit hate speech, implicit\nexpressions often depend on context, cultural subtleties, and hidden biases,\nmaking them more challenging to identify consistently. Additionally, the\ninterpretation of such speech is influenced by external knowledge and\ndemographic biases, resulting in varied detection results across different\nlanguage models. Furthermore, Large Language Models often show heightened\nsensitivity to toxic language and references to vulnerable groups, which can\nlead to misclassifications. This over-sensitivity results in false positives\n(incorrectly identifying harmless statements as hateful) and false negatives\n(failing to detect genuinely harmful content). Addressing these issues requires\nmethods that not only improve detection precision but also reduce model biases\nand enhance robustness. To address these challenges, we propose a novel method,\nwhich utilizes in-context learning without requiring model fine-tuning. By\nadaptively retrieving demonstrations that focus on similar groups or those with\nthe highest similarity scores, our approach enhances contextual comprehension.\nExperimental results show that our method outperforms current state-of-the-art\ntechniques. Implementation details and code are available at TBD.\n", "date": "2025-04-16", "categories": []}
{"id": "2409.14740", "title": "ToxiCraft: A Novel Framework for Synthetic Generation of Harmful\n  Information", "abstract": "  In different NLP tasks, detecting harmful content is crucial for online\nenvironments, especially with the growing influence of social media. However,\nprevious research has two main issues: 1) a lack of data in low-resource\nsettings, and 2) inconsistent definitions and criteria for judging harmful\ncontent, requiring classification models to be robust to spurious features and\ndiverse. We propose Toxicraft, a novel framework for synthesizing datasets of\nharmful information to address these weaknesses. With only a small amount of\nseed data, our framework can generate a wide variety of synthetic, yet\nremarkably realistic, examples of toxic information. Experimentation across\nvarious datasets showcases a notable enhancement in detection model robustness\nand adaptability, surpassing or close to the gold labels.\n", "date": "2024-09-23", "categories": []}
{"id": "2503.10650", "title": "AI Enabled User-Specific Cyberbullying Severity Detection with\n  Explainability", "abstract": "  The rise of social media has significantly increased the prevalence of\ncyberbullying (CB), posing serious risks to both mental and physical\nwell-being. Effective detection systems are essential for mitigating its\nimpact. While several machine learning (ML) models have been developed, few\nincorporate victims' psychological, demographic, and behavioral factors\nalongside bullying comments to assess severity. In this study, we propose an AI\nmodel intregrating user-specific attributes, including psychological factors\n(self-esteem, anxiety, depression), online behavior (internet usage,\ndisciplinary history), and demographic attributes (race, gender, ethnicity),\nalong with social media comments. Additionally, we introduce a re-labeling\ntechnique that categorizes social media comments into three severity levels:\nNot Bullying, Mild Bullying, and Severe Bullying, considering user-specific\nfactors.Our LSTM model is trained using 146 features, incorporating emotional,\ntopical, and word2vec representations of social media comments as well as\nuser-level attributes and it outperforms existing baseline models, achieving\nthe highest accuracy of 98\\% and an F1-score of 0.97. To identify key factors\ninfluencing the severity of cyberbullying, we employ explainable AI techniques\n(SHAP and LIME) to interpret the model's decision-making process. Our findings\nreveal that, beyond hate comments, victims belonging to specific racial and\ngender groups are more frequently targeted and exhibit higher incidences of\ndepression, disciplinary issues, and low self-esteem. Additionally, individuals\nwith a prior history of bullying are at a greater risk of becoming victims of\ncyberbullying.\n", "date": "2025-03-04", "categories": []}
{"id": "2403.15449", "title": "Hatred Stems from Ignorance! Distillation of the Persuasion Modes in\n  Countering Conversational Hate Speech", "abstract": "  Examining the factors that the counterspeech uses are at the core of\nunderstanding the optimal methods for confronting hate speech online. Various\nstudies have assessed the emotional base factors used in counter speech, such\nas emotional empathy, offensiveness, and hostility. To better understand the\ncounterspeech used in conversations, this study distills persuasion modes into\nreason, emotion, and credibility and evaluates their use in two types of\nconversation interactions: closed (multi-turn) and open (single-turn)\nconcerning racism, sexism, and religious bigotry. The evaluation covers the\ndistinct behaviors seen with human-sourced as opposed to machine-generated\ncounterspeech. It also assesses the interplay between the stance taken and the\nmode of persuasion seen in the counterspeech.\n  Notably, we observe nuanced differences in the counterspeech persuasion modes\nused in open and closed interactions, especially in terms of the topic, with a\ngeneral tendency to use reason as a persuasion mode to express the counterpoint\nto hate comments. The machine-generated counterspeech tends to exhibit an\nemotional persuasion mode, while human counters lean toward reason.\nFurthermore, our study shows that reason tends to obtain more supportive\nreplies than other persuasion modes. The findings highlight the potential for\nincorporating persuasion modes into studies about countering hate speech, as\nthey can serve as an optimal means of explainability and pave the way for the\nfurther adoption of the reply's stance and the role it plays in assessing what\ncomprises the optimal counterspeech.\n", "date": "2024-03-18", "categories": []}
{"id": "2504.09914", "title": "Improving Multimodal Hateful Meme Detection Exploiting LMM-Generated\n  Knowledge", "abstract": "  Memes have become a dominant form of communication in social media in recent\nyears. Memes are typically humorous and harmless, however there are also memes\nthat promote hate speech, being in this way harmful to individuals and groups\nbased on their identity. Therefore, detecting hateful content in memes has\nemerged as a task of critical importance. The need for understanding the\ncomplex interactions of images and their embedded text renders the hateful meme\ndetection a challenging multimodal task. In this paper we propose to address\nthe aforementioned task leveraging knowledge encoded in powerful Large\nMultimodal Models (LMM). Specifically, we propose to exploit LMMs in a two-fold\nmanner. First, by extracting knowledge oriented to the hateful meme detection\ntask in order to build strong meme representations. Specifically, generic\nsemantic descriptions and emotions that the images along with their embedded\ntexts elicit are extracted, which are then used to train a simple\nclassification head for hateful meme detection. Second, by developing a novel\nhard mining approach introducing directly LMM-encoded knowledge to the training\nprocess, providing further improvements. We perform extensive experiments on\ntwo datasets that validate the effectiveness of the proposed method, achieving\nstate-of-the-art performance. Our code and trained models are publicly\navailable at: https://github.com/IDT-ITI/LMM-CLIP-meme.\n", "date": "2025-04-14", "categories": []}
{"id": "2412.04942", "title": "A Federated Approach to Few-Shot Hate Speech Detection for Marginalized\n  Communities", "abstract": "  Hate speech online remains an understudied issue for marginalized\ncommunities, particularly in the Global South, which includes developing\nsocieties with increasing internet penetration. In this paper, we aim to\nprovide marginalized communities in societies where the dominant language is\nlow-resource with a privacy-preserving tool to protect themselves from online\nhate speech by filtering offensive content in their native languages. Our\ncontributions are twofold: 1) we release REACT (REsponsive hate speech datasets\nAcross ConTexts), a collection of high-quality, culture-specific hate speech\ndetection datasets comprising multiple target groups and low-resource\nlanguages, curated by experienced data collectors; 2) we propose a few-shot\nhate speech detection approach based on federated learning (FL), a\nprivacy-preserving method for collaboratively training a central model that\nexhibits robustness when tackling different target groups and languages. By\nkeeping training local to user devices, we ensure data privacy while leveraging\nthe collective learning benefits of FL. Furthermore, we explore personalized\nclient models tailored to specific target groups and evaluate their\nperformance. Our findings indicate the overall effectiveness of FL across\ndifferent target groups, and point to personalization as a promising direction.\n", "date": "2024-12-06", "categories": []}
{"id": "2504.08408", "title": "BOISHOMMO: Holistic Approach for Bangla Hate Speech", "abstract": "  One of the most alarming issues in digital society is hate speech (HS) on\nsocial media. The severity is so high that researchers across the globe are\ncaptivated by this domain. A notable amount of work has been conducted to\naddress the identification and alarm system. However, a noticeable gap exists,\nespecially for low-resource languages. Comprehensive datasets are the main\nproblem among the constrained resource languages, such as Bangla.\nInterestingly, hate speech or any particular speech has no single\ndimensionality. Similarly, the hate component can simultaneously have multiple\nabusive attributes, which seems to be missed in the existing datasets. Thus, a\nmulti-label Bangla hate speech dataset named BOISHOMMO has been compiled and\nevaluated in this work. That includes categories of HS across race, gender,\nreligion, politics, and more. With over two thousand annotated examples,\nBOISHOMMO provides a nuanced understanding of hate speech in Bangla and\nhighlights the complexities of processing non-Latin scripts. Apart from\nevaluating with multiple algorithmic approaches, it also highlights the\ncomplexities of processing Bangla text and assesses model performance. This\nunique multi-label approach enriches future hate speech detection and analysis\nstudies for low-resource languages by providing a more nuanced, diverse\ndataset.\n", "date": "2025-04-11", "categories": []}
{"id": "2208.10598", "title": "Generalizing Hate Speech Detection Using Multi-Task Learning: A Case\n  Study of Political Public Figures", "abstract": "  Automatic identification of hateful and abusive content is vital in combating\nthe spread of harmful online content and its damaging effects. Most existing\nworks evaluate models by examining the generalization error on train-test\nsplits on hate speech datasets. These datasets often differ in their\ndefinitions and labeling criteria, leading to poor generalization performance\nwhen predicting across new domains and datasets. This work proposes a new\nMulti-task Learning (MTL) pipeline that trains simultaneously across multiple\nhate speech datasets to construct a more encompassing classification model.\nUsing a dataset-level leave-one-out evaluation (designating a dataset for\ntesting and jointly training on all others), we trial the MTL detection on new,\npreviously unseen datasets. Our results consistently outperform a large sample\nof existing work. We show strong results when examining the generalization\nerror in train-test splits and substantial improvements when predicting on\npreviously unseen datasets. Furthermore, we assemble a novel dataset, dubbed\nPubFigs, focusing on the problematic speech of American Public Political\nFigures. We crowdsource-label using Amazon MTurk more than $20,000$ tweets and\nmachine-label problematic speech in all the $305,235$ tweets in PubFigs. We\nfind that the abusive and hate tweeting mainly originates from right-leaning\nfigures and relates to six topics, including Islam, women, ethnicity, and\nimmigrants. We show that MTL builds embeddings that can simultaneously separate\nabusive from hate speech, and identify its topics.\n", "date": "2022-08-22", "categories": []}
{"id": "2504.02178", "title": "Subasa - Adapting Language Models for Low-resourced Offensive Language\n  Detection in Sinhala", "abstract": "  Accurate detection of offensive language is essential for a number of\napplications related to social media safety. There is a sharp contrast in\nperformance in this task between low and high-resource languages. In this\npaper, we adapt fine-tuning strategies that have not been previously explored\nfor Sinhala in the downstream task of offensive language detection. Using this\napproach, we introduce four models: \"Subasa-XLM-R\", which incorporates an\nintermediate Pre-Finetuning step using Masked Rationale Prediction. Two\nvariants of \"Subasa-Llama\" and \"Subasa-Mistral\", are fine-tuned versions of\nLlama (3.2) and Mistral (v0.3), respectively, with a task-specific strategy. We\nevaluate our models on the SOLD benchmark dataset for Sinhala offensive\nlanguage detection. All our models outperform existing baselines. Subasa-XLM-R\nachieves the highest Macro F1 score (0.84) surpassing state-of-the-art large\nlanguage models like GPT-4o when evaluated on the same SOLD benchmark dataset\nunder zero-shot settings. The models and code are publicly available.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.02863", "title": "GS_DravidianLangTech@2025: Women Targeted Abusive Texts Detection on\n  Social Media", "abstract": "  The increasing misuse of social media has become a concern; however,\ntechnological solutions are being developed to moderate its content\neffectively. This paper focuses on detecting abusive texts targeting women on\nsocial media platforms. Abusive speech refers to communication intended to harm\nor incite hatred against vulnerable individuals or groups. Specifically, this\nstudy aims to identify abusive language directed toward women. To achieve this,\nwe utilized logistic regression and BERT as base models to train datasets\nsourced from DravidianLangTech@2025 for Tamil and Malayalam languages. The\nmodels were evaluated on test datasets, resulting in a 0.729 macro F1 score for\nBERT and 0.6279 for logistic regression in Tamil and Malayalam, respectively.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.02310", "title": "Improving Harmful Text Detection with Joint Retrieval and External\n  Knowledge", "abstract": "  Harmful text detection has become a crucial task in the development and\ndeployment of large language models, especially as AI-generated content\ncontinues to expand across digital platforms. This study proposes a joint\nretrieval framework that integrates pre-trained language models with knowledge\ngraphs to improve the accuracy and robustness of harmful text detection.\nExperimental results demonstrate that the joint retrieval approach\nsignificantly outperforms single-model baselines, particularly in low-resource\ntraining scenarios and multilingual environments. The proposed method\neffectively captures nuanced harmful content by leveraging external contextual\ninformation, addressing the limitations of traditional detection models. Future\nresearch should focus on optimizing computational efficiency, enhancing model\ninterpretability, and expanding multimodal detection capabilities to better\ntackle evolving harmful content patterns. This work contributes to the\nadvancement of AI safety, ensuring more trustworthy and reliable content\nmoderation systems.\n", "date": "2025-04-03", "categories": []}
{"id": "2504.01902", "title": "Graphically Speaking: Unmasking Abuse in Social Media with Conversation\n  Insights", "abstract": "  Detecting abusive language in social media conversations poses significant\nchallenges, as identifying abusiveness often depends on the conversational\ncontext, characterized by the content and topology of preceding comments.\nTraditional Abusive Language Detection (ALD) models often overlook this\ncontext, which can lead to unreliable performance metrics. Recent Natural\nLanguage Processing (NLP) methods that integrate conversational context often\ndepend on limited and simplified representations, and report inconsistent\nresults. In this paper, we propose a novel approach that utilize graph neural\nnetworks (GNNs) to model social media conversations as graphs, where nodes\nrepresent comments, and edges capture reply structures. We systematically\ninvestigate various graph representations and context windows to identify the\noptimal configuration for ALD. Our GNN model outperform both context-agnostic\nbaselines and linear context-aware methods, achieving significant improvements\nin F1 scores. These findings demonstrate the critical role of structured\nconversational context and establish GNNs as a robust framework for advancing\ncontext-aware abusive language detection.\n", "date": "2025-04-02", "categories": []}
{"id": "2211.05995", "title": "In-game Toxic Language Detection: Shared Task and Attention Residuals", "abstract": "  In-game toxic language becomes the hot potato in the gaming industry and\ncommunity. There have been several online game toxicity analysis frameworks and\nmodels proposed. However, it is still challenging to detect toxicity due to the\nnature of in-game chat, which has extremely short length. In this paper, we\ndescribe how the in-game toxic language shared task has been established using\nthe real-world in-game chat data. In addition, we propose and introduce the\nmodel/framework for toxic language token tagging (slot filling) from the\nin-game chat. The relevant code is publicly available on GitHub:\nhttps://github.com/Yuanzhe-Jia/In-Game-Toxic-Detection\n", "date": "2022-11-11", "categories": []}
{"id": "2410.15314", "title": "KTCR: Improving Implicit Hate Detection with Knowledge Transfer driven\n  Concept Refinement", "abstract": "  The constant shifts in social and political contexts, driven by emerging\nsocial movements and political events, lead to new forms of hate content and\npreviously unrecognized hate patterns that machine learning models may not have\ncaptured. Some recent literature proposes data augmentation-based techniques to\nenrich existing hate datasets by incorporating samples that reveal new implicit\nhate patterns. This approach aims to improve the model's performance on\nout-of-domain implicit hate instances. It is observed, that further addition of\nmore samples for augmentation results in the decrease of the performance of the\nmodel. In this work, we propose a Knowledge Transfer-driven Concept Refinement\nmethod that distills and refines the concepts related to implicit hate samples\nthrough novel prototype alignment and concept losses, alongside data\naugmentation based on concept activation vectors. Experiments with several\npublicly available datasets show that incorporating additional implicit samples\nreflecting new hate patterns through concept refinement enhances the model's\nperformance, surpassing baseline results while maintaining cross-dataset\ngeneralization capabilities.\n", "date": "2024-10-20", "categories": []}
{"id": "2504.00045", "title": "Measuring Online Hate on 4chan using Pre-trained Deep Learning Models", "abstract": "  Online hate speech can harmfully impact individuals and groups, specifically\non non-moderated platforms such as 4chan where users can post anonymous\ncontent. This work focuses on analysing and measuring the prevalence of online\nhate on 4chan's politically incorrect board (/pol/) using state-of-the-art\nNatural Language Processing (NLP) models, specifically transformer-based models\nsuch as RoBERTa and Detoxify. By leveraging these advanced models, we provide\nan in-depth analysis of hate speech dynamics and quantify the extent of online\nhate non-moderated platforms. The study advances understanding through\nmulti-class classification of hate speech (racism, sexism, religion, etc.),\nwhile also incorporating the classification of toxic content (e.g., identity\nattacks and threats) and a further topic modelling analysis. The results show\nthat 11.20% of this dataset is identified as containing hate in different\ncategories. These evaluations show that online hate is manifested in various\nforms, confirming the complicated and volatile nature of detection in the wild.\n", "date": "2025-03-30", "categories": []}
{"id": "2411.09214", "title": "HateGPT: Unleashing GPT-3.5 Turbo to Combat Hate Speech on X", "abstract": "  The widespread use of social media platforms like Twitter and Facebook has\nenabled people of all ages to share their thoughts and experiences, leading to\nan immense accumulation of user-generated content. However, alongside the\nbenefits, these platforms also face the challenge of managing hate speech and\noffensive content, which can undermine rational discourse and threaten\ndemocratic values. As a result, there is a growing need for automated methods\nto detect and mitigate such content, especially given the complexity of\nconversations that may require contextual analysis across multiple languages,\nincluding code-mixed languages like Hinglish, German-English, and Bangla. We\nparticipated in the English task where we have to classify English tweets into\ntwo categories namely Hate and Offensive and Non Hate-Offensive. In this work,\nwe experiment with state-of-the-art large language models like GPT-3.5 Turbo\nvia prompting to classify tweets into Hate and Offensive or Non Hate-Offensive.\nIn this study, we evaluate the performance of a classification model using\nMacro-F1 scores across three distinct runs. The Macro-F1 score, which balances\nprecision and recall across all classes, is used as the primary metric for\nmodel evaluation. The scores obtained are 0.756 for run 1, 0.751 for run 2, and\n0.754 for run 3, indicating a high level of performance with minimal variance\namong the runs. The results suggest that the model consistently performs well\nin terms of precision and recall, with run 1 showing the highest performance.\nThese findings highlight the robustness and reliability of the model across\ndifferent runs.\n", "date": "2024-11-14", "categories": []}
{"id": "2402.12198", "title": "Exploring the Limits of Zero Shot Vision Language Models for Hate Meme\n  Detection: The Vulnerabilities and their Interpretations", "abstract": "  There is a rapid increase in the use of multimedia content in current social\nmedia platforms. One of the highly popular forms of such multimedia content are\nmemes. While memes have been primarily invented to promote funny and buoyant\ndiscussions, malevolent users exploit memes to target individuals or vulnerable\ncommunities, making it imperative to identify and address such instances of\nhateful memes. Thus social media platforms are in dire need for active\nmoderation of such harmful content. While manual moderation is extremely\ndifficult due to the scale of such content, automatic moderation is challenged\nby the need of good quality annotated data to train hate meme detection\nalgorithms. This makes a perfect pretext for exploring the power of modern day\nvision language models (VLMs) that have exhibited outstanding performance\nacross various tasks. In this paper we study the effectiveness of VLMs in\nhandling intricate tasks such as hate meme detection in a completely zero-shot\nsetting so that there is no dependency on annotated data for the task. We\nperform thorough prompt engineering and query state-of-the-art VLMs using\nvarious prompt types to detect hateful/harmful memes. We further interpret the\nmisclassification cases using a novel superpixel based occlusion method.\nFinally we show that these misclassifications can be neatly arranged into a\ntypology of error classes the knowledge of which should enable the design of\nbetter safety guardrails in future.\n", "date": "2024-02-19", "categories": []}
{"id": "2503.14926", "title": "Covering Cracks in Content Moderation: Delexicalized Distant Supervision\n  for Illicit Drug Jargon Detection", "abstract": "  In light of rising drug-related concerns and the increasing role of social\nmedia, sales and discussions of illicit drugs have become commonplace online.\nSocial media platforms hosting user-generated content must therefore perform\ncontent moderation, which is a difficult task due to the vast amount of jargon\nused in drug discussions. Previous works on drug jargon detection were limited\nto extracting a list of terms, but these approaches have fundamental problems\nin practical application. First, they are trivially evaded using word\nsubstitutions. Second, they cannot distinguish whether euphemistic terms such\nas \"pot\" or \"crack\" are being used as drugs or in their benign meanings. We\nargue that drug content moderation should be done using contexts rather than\nrelying on a banlist. However, manually annotated datasets for training such a\ntask are not only expensive but also prone to becoming obsolete. We present\nJEDIS, a framework for detecting illicit drug jargon terms by analyzing their\ncontexts. JEDIS utilizes a novel approach that combines distant supervision and\ndelexicalization, which allows JEDIS to be trained without human-labeled data\nwhile being robust to new terms and euphemisms. Experiments on two manually\nannotated datasets show JEDIS significantly outperforms state-of-the-art\nword-based baselines in terms of F1-score and detection coverage in drug jargon\ndetection. We also conduct qualitative analysis that demonstrates JEDIS is\nrobust against pitfalls faced by existing approaches.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.07862", "title": "cantnlp@DravidianLangTech2025: A Bag-of-Sounds Approach to Multimodal\n  Hate Speech Detection", "abstract": "  This paper presents the systems and results for the Multimodal Social Media\nData Analysis in Dravidian Languages (MSMDA-DL) shared task at the Fifth\nWorkshop on Speech, Vision, and Language Technologies for Dravidian Languages\n(DravidianLangTech-2025). We took a `bag-of-sounds' approach by training our\nhate speech detection system on the speech (audio) data using transformed Mel\nspectrogram measures. While our candidate model performed poorly on the test\nset, our approach offered promising results during training and development for\nMalayalam and Tamil. With sufficient and well-balanced training data, our\nresults show that it is feasible to use both text and speech (audio) data in\nthe development of multimodal hate speech detection systems.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.10648", "title": "Hate Speech and Sentiment of YouTube Video Comments From Public and\n  Private Sources Covering the Israel-Palestine Conflict", "abstract": "  This study explores the prevalence of hate speech (HS) and sentiment in\nYouTube video comments concerning the Israel-Palestine conflict by analyzing\ncontent from both public and private news sources. The research involved\nannotating 4983 comments for HS and sentiments (neutral, pro-Israel, and\npro-Palestine). Subsequently, machine learning (ML) models were developed,\ndemonstrating robust predictive capabilities with area under the receiver\noperating characteristic (AUROC) scores ranging from 0.83 to 0.90. These models\nwere applied to the extracted comment sections of YouTube videos from public\nand private sources, uncovering a higher incidence of HS in public sources\n(40.4%) compared to private sources (31.6%). Sentiment analysis revealed a\npredominantly neutral stance in both source types, with more pronounced\nsentiments towards Israel and Palestine observed in public sources. This\ninvestigation highlights the dynamic nature of online discourse surrounding the\nIsrael-Palestine conflict and underscores the potential of moderating content\nin a politically charged environment.\n", "date": "2025-03-03", "categories": []}
{"id": "2503.09341", "title": "An Evaluation of LLMs for Detecting Harmful Computing Terms", "abstract": "  Detecting harmful and non-inclusive terminology in technical contexts is\ncritical for fostering inclusive environments in computing. This study explores\nthe impact of model architecture on harmful language detection by evaluating a\ncurated database of technical terms, each paired with specific use cases. We\ntested a range of encoder, decoder, and encoder-decoder language models,\nincluding BERT-base-uncased, RoBERTa large-mnli, Gemini Flash 1.5 and 2.0,\nGPT-4, Claude AI Sonnet 3.5, T5-large, and BART-large-mnli. Each model was\npresented with a standardized prompt to identify harmful and non-inclusive\nlanguage across 64 terms. Results reveal that decoder models, particularly\nGemini Flash 2.0 and Claude AI, excel in nuanced contextual analysis, while\nencoder models like BERT exhibit strong pattern recognition but struggle with\nclassification certainty. We discuss the implications of these findings for\nimproving automated detection tools and highlight model-specific strengths and\nlimitations in fostering inclusive communication in technical domains.\n", "date": "2025-03-12", "categories": []}
{"id": "2502.18101", "title": "Detecting Offensive Memes with Social Biases in Singapore Context Using\n  Multimodal Large Language Models", "abstract": "  Traditional online content moderation systems struggle to classify modern\nmultimodal means of communication, such as memes, a highly nuanced and\ninformation-dense medium. This task is especially hard in a culturally diverse\nsociety like Singapore, where low-resource languages are used and extensive\nknowledge on local context is needed to interpret online content. We curate a\nlarge collection of 112K memes labeled by GPT-4V for fine-tuning a VLM to\nclassify offensive memes in Singapore context. We show the effectiveness of\nfine-tuned VLMs on our dataset, and propose a pipeline containing OCR,\ntranslation and a 7-billion parameter-class VLM. Our solutions reach 80.62%\naccuracy and 0.8192 AUROC on a held-out test set, and can greatly aid human in\nmoderating online contents. The dataset, code, and model weights have been\nopen-sourced at https://github.com/aliencaocao/vlm-for-memes-aisg.\n", "date": "2025-02-25", "categories": []}
{"id": "2503.06534", "title": "SafeSpeech: A Comprehensive and Interactive Tool for Analysing Sexist\n  and Abusive Language in Conversations", "abstract": "  Detecting toxic language including sexism, harassment and abusive behaviour,\nremains a critical challenge, particularly in its subtle and context-dependent\nforms. Existing approaches largely focus on isolated message-level\nclassification, overlooking toxicity that emerges across conversational\ncontexts. To promote and enable future research in this direction, we introduce\nSafeSpeech, a comprehensive platform for toxic content detection and analysis\nthat bridges message-level and conversation-level insights. The platform\nintegrates fine-tuned classifiers and large language models (LLMs) to enable\nmulti-granularity detection, toxic-aware conversation summarization, and\npersona profiling. SafeSpeech also incorporates explainability mechanisms, such\nas perplexity gain analysis, to highlight the linguistic elements driving\npredictions. Evaluations on benchmark datasets, including EDOS, OffensEval, and\nHatEval, demonstrate the reproduction of state-of-the-art performance across\nmultiple tasks, including fine-grained sexism detection.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.06550", "title": "BingoGuard: LLM Content Moderation Tools with Risk Levels", "abstract": "  Malicious content generated by large language models (LLMs) can pose varying\ndegrees of harm. Although existing LLM-based moderators can detect harmful\ncontent, they struggle to assess risk levels and may miss lower-risk outputs.\nAccurate risk assessment allows platforms with different safety thresholds to\ntailor content filtering and rejection. In this paper, we introduce per-topic\nseverity rubrics for 11 harmful topics and build BingoGuard, an LLM-based\nmoderation system designed to predict both binary safety labels and severity\nlevels. To address the lack of annotations on levels of severity, we propose a\nscalable generate-then-filter framework that first generates responses across\ndifferent severity levels and then filters out low-quality responses. Using\nthis framework, we create BingoGuardTrain, a training dataset with 54,897\nexamples covering a variety of topics, response severity, styles, and\nBingoGuardTest, a test set with 988 examples explicitly labeled based on our\nseverity rubrics that enables fine-grained analysis on model behaviors on\ndifferent severity levels. Our BingoGuard-8B, trained on BingoGuardTrain,\nachieves the state-of-the-art performance on several moderation benchmarks,\nincluding WildGuardTest and HarmBench, as well as BingoGuardTest, outperforming\nbest public models, WildGuard, by 4.3\\%. Our analysis demonstrates that\nincorporating severity levels into training significantly enhances detection\nperformance and enables the model to effectively gauge the severity of harmful\nresponses.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.05357", "title": "Improving Hate Speech Classification with Cross-Taxonomy Dataset\n  Integration", "abstract": "  Algorithmic hate speech detection faces significant challenges due to the\ndiverse definitions and datasets used in research and practice. Social media\nplatforms, legal frameworks, and institutions each apply distinct yet\noverlapping definitions, complicating classification efforts. This study\naddresses these challenges by demonstrating that existing datasets and\ntaxonomies can be integrated into a unified model, enhancing prediction\nperformance and reducing reliance on multiple specialized classifiers. The work\nintroduces a universal taxonomy and a hate speech classifier capable of\ndetecting a wide range of definitions within a single framework. Our approach\nis validated by combining two widely used but differently annotated datasets,\nshowing improved classification performance on an independent test set. This\nwork highlights the potential of dataset and taxonomy integration in advancing\nhate speech detection, increasing efficiency, and ensuring broader\napplicability across contexts.\n", "date": "2025-03-07", "categories": []}
{"id": "2503.04279", "title": "Dual-Class Prompt Generation: Enhancing Indonesian Gender-Based Hate\n  Speech Detection through Data Augmentation", "abstract": "  Detecting gender-based hate speech in Indonesian social media remains\nchallenging due to limited labeled datasets. While binary hate speech\nclassification has advanced, a more granular category like gender-targeted hate\nspeech is understudied because of class imbalance issues. This paper addresses\nthis gap by comparing three data augmentation techniques for Indonesian\ngender-based hate speech detection. We evaluate backtranslation, single-class\nprompt generation (using only hate speech examples), and our proposed\ndual-class prompt generation (using both hate speech and non-hate speech\nexamples). Experiments show all augmentation methods improve classification\nperformance, with our dual-class approach achieving the best results (88.5%\naccuracy, 88.1% F1-score using Random Forest). Semantic similarity analysis\nreveals dual-class prompt generation produces the most novel content, while\nT-SNE visualizations confirm these samples occupy distinct feature space\nregions while maintaining class characteristics. Our findings suggest that\nincorporating examples from both classes helps language models generate more\ndiverse yet representative samples, effectively addressing limited data\nchallenges in specialized hate speech detection.\n", "date": "2025-03-06", "categories": []}
{"id": "2503.00417", "title": "A Multi-Labeled Dataset for Indonesian Discourse: Examining Toxicity,\n  Polarization, and Demographics Information", "abstract": "  Polarization is defined as divisive opinions held by two or more groups on\nsubstantive issues. As the world's third-largest democracy, Indonesia faces\ngrowing concerns about the interplay between political polarization and online\ntoxicity, which is often directed at vulnerable minority groups. Despite the\nimportance of this issue, previous NLP research has not fully explored the\nrelationship between toxicity and polarization. To bridge this gap, we present\na novel multi-label Indonesian dataset that incorporates toxicity,\npolarization, and annotator demographic information. Benchmarking this dataset\nusing BERT-base models and large language models (LLMs) shows that polarization\ninformation enhances toxicity classification, and vice versa. Furthermore,\nproviding demographic information significantly improves the performance of\npolarization classification.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.01623", "title": "Lost in Moderation: How Commercial Content Moderation APIs Over- and\n  Under-Moderate Group-Targeted Hate Speech and Linguistic Variations", "abstract": "  Commercial content moderation APIs are marketed as scalable solutions to\ncombat online hate speech. However, the reliance on these APIs risks both\nsilencing legitimate speech, called over-moderation, and failing to protect\nonline platforms from harmful speech, known as under-moderation. To assess such\nrisks, this paper introduces a framework for auditing black-box NLP systems.\nUsing the framework, we systematically evaluate five widely used commercial\ncontent moderation APIs. Analyzing five million queries based on four datasets,\nwe find that APIs frequently rely on group identity terms, such as ``black'',\nto predict hate speech. While OpenAI's and Amazon's services perform slightly\nbetter, all providers under-moderate implicit hate speech, which uses codified\nmessages, especially against LGBTQIA+ individuals. Simultaneously, they\nover-moderate counter-speech, reclaimed slurs and content related to Black,\nLGBTQIA+, Jewish, and Muslim people. We recommend that API providers offer\nbetter guidance on API implementation and threshold setting and more\ntransparency on their APIs' limitations.\n  Warning: This paper contains offensive and hateful terms and concepts. We\nhave chosen to reproduce these terms for reasons of transparency.\n", "date": "2025-03-03", "categories": []}
{"id": "2502.07138", "title": "Towards a Robust Framework for Multimodal Hate Detection: A Study on\n  Video vs. Image-based Content", "abstract": "  Social media platforms enable the propagation of hateful content across\ndifferent modalities such as textual, auditory, and visual, necessitating\neffective detection methods. While recent approaches have shown promise in\nhandling individual modalities, their effectiveness across different modality\ncombinations remains unexplored. This paper presents a systematic analysis of\nfusion-based approaches for multimodal hate detection, focusing on their\nperformance across video and image-based content. Our comprehensive evaluation\nreveals significant modality-specific limitations: while simple embedding\nfusion achieves state-of-the-art performance on video content (HateMM dataset)\nwith a 9.9% points F1-score improvement, it struggles with complex image-text\nrelationships in memes (Hateful Memes dataset). Through detailed ablation\nstudies and error analysis, we demonstrate how current fusion approaches fail\nto capture nuanced cross-modal interactions, particularly in cases involving\nbenign confounders. Our findings provide crucial insights for developing more\nrobust hate detection systems and highlight the need for modality-specific\narchitectural considerations. The code is available at\nhttps://github.com/gak97/Video-vs-Meme-Hate.\n", "date": "2025-02-11", "categories": []}
{"id": "2502.19612", "title": "Evaluation of Hate Speech Detection Using Large Language Models and\n  Geographical Contextualization", "abstract": "  The proliferation of hate speech on social media is one of the serious issues\nthat is bringing huge impacts to society: an escalation of violence,\ndiscrimination, and social fragmentation. The problem of detecting hate speech\nis intrinsically multifaceted due to cultural, linguistic, and contextual\ncomplexities and adversarial manipulations. In this study, we systematically\ninvestigate the performance of LLMs on detecting hate speech across\nmultilingual datasets and diverse geographic contexts. Our work presents a new\nevaluation framework in three dimensions: binary classification of hate speech,\ngeography-aware contextual detection, and robustness to adversarially generated\ntext. Using a dataset of 1,000 comments from five diverse regions, we evaluate\nthree state-of-the-art LLMs: Llama2 (13b), Codellama (7b), and DeepSeekCoder\n(6.7b). Codellama had the best binary classification recall with 70.6% and an\nF1-score of 52.18%, whereas DeepSeekCoder had the best performance in\ngeographic sensitivity, correctly detecting 63 out of 265 locations. The tests\nfor adversarial robustness also showed significant weaknesses; Llama2\nmisclassified 62.5% of manipulated samples. These results bring to light the\ntrade-offs between accuracy, contextual understanding, and robustness in the\ncurrent versions of LLMs. This work has thus set the stage for developing\ncontextually aware, multilingual hate speech detection systems by underlining\nkey strengths and limitations, therefore offering actionable insights for\nfuture research and real-world applications.\n", "date": "2025-02-26", "categories": []}
{"id": "2502.00385", "title": "The Impact of Persona-based Political Perspectives on Hateful Content\n  Detection", "abstract": "  While pretraining language models with politically diverse content has been\nshown to improve downstream task fairness, such approaches require significant\ncomputational resources often inaccessible to many researchers and\norganizations. Recent work has established that persona-based prompting can\nintroduce political diversity in model outputs without additional training.\nHowever, it remains unclear whether such prompting strategies can achieve\nresults comparable to political pretraining for downstream tasks. We\ninvestigate this question using persona-based prompting strategies in\nmultimodal hate-speech detection tasks, specifically focusing on hate speech in\nmemes. Our analysis reveals that when mapping personas onto a political compass\nand measuring persona agreement, inherent political positioning has\nsurprisingly little correlation with classification decisions. Notably, this\nlack of correlation persists even when personas are explicitly injected with\nstronger ideological descriptors. Our findings suggest that while LLMs can\nexhibit political biases in their responses to direct political questions,\nthese biases may have less impact on practical classification tasks than\npreviously assumed. This raises important questions about the necessity of\ncomputationally expensive political pretraining for achieving fair performance\nin downstream tasks.\n", "date": "2025-02-01", "categories": []}
{"id": "2502.18695", "title": "Policy-as-Prompt: Rethinking Content Moderation in the Age of Large\n  Language Models", "abstract": "  Content moderation plays a critical role in shaping safe and inclusive online\nenvironments, balancing platform standards, user expectations, and regulatory\nframeworks. Traditionally, this process involves operationalising policies into\nguidelines, which are then used by downstream human moderators for enforcement,\nor to further annotate datasets for training machine learning moderation\nmodels. However, recent advancements in large language models (LLMs) are\ntransforming this landscape. These models can now interpret policies directly\nas textual inputs, eliminating the need for extensive data curation. This\napproach offers unprecedented flexibility, as moderation can be dynamically\nadjusted through natural language interactions. This paradigm shift raises\nimportant questions about how policies are operationalised and the implications\nfor content moderation practices. In this paper, we formalise the emerging\npolicy-as-prompt framework and identify five key challenges across four\ndomains: Technical Implementation (1. translating policy to prompts, 2.\nsensitivity to prompt structure and formatting), Sociotechnical (3. the risk of\ntechnological determinism in policy formation), Organisational (4. evolving\nroles between policy and machine learning teams), and Governance (5. model\ngovernance and accountability). Through analysing these challenges across\ntechnical, sociotechnical, organisational, and governance dimensions, we\ndiscuss potential mitigation approaches. This research provides actionable\ninsights for practitioners and lays the groundwork for future exploration of\nscalable and adaptive content moderation systems in digital ecosystems.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.10921", "title": "Evolving Hate Speech Online: An Adaptive Framework for Detection and\n  Mitigation", "abstract": "  The proliferation of social media platforms has led to an increase in the\nspread of hate speech, particularly targeting vulnerable communities.\nUnfortunately, existing methods for automatically identifying and blocking\ntoxic language rely on pre-constructed lexicons, making them reactive rather\nthan adaptive. As such, these approaches become less effective over time,\nespecially when new communities are targeted with slurs not included in the\noriginal datasets. To address this issue, we present an adaptive approach that\nuses word embeddings to update lexicons and develop a hybrid model that adjusts\nto emerging slurs and new linguistic patterns. This approach can effectively\ndetect toxic language, including intentional spelling mistakes employed by\naggressors to avoid detection. Our hybrid model, which combines BERT with\nlexicon-based techniques, achieves an accuracy of 95% for most state-of-the-art\ndatasets. Our work has significant implications for creating safer online\nenvironments by improving the detection of toxic content and proactively\nupdating the lexicon. Content Warning: This paper contains examples of hate\nspeech that may be triggering.\n", "date": "2025-02-15", "categories": []}
{"id": "2502.15155", "title": "Extreme Speech Classification in the Era of LLMs: Exploring Open-Source\n  and Proprietary Models", "abstract": "  In recent years, widespread internet adoption and the growth in userbase of\nvarious social media platforms have led to an increase in the proliferation of\nextreme speech online. While traditional language models have demonstrated\nproficiency in distinguishing between neutral text and non-neutral text (i.e.\nextreme speech), categorizing the diverse types of extreme speech presents\nsignificant challenges. The task of extreme speech classification is\nparticularly nuanced, as it requires a deep understanding of socio-cultural\ncontexts to accurately interpret the intent of the language used by the\nspeaker. Even human annotators often disagree on the appropriate classification\nof such content, emphasizing the complex and subjective nature of this task.\nThe use of human moderators also presents a scaling issue, necessitating the\nneed for automated systems for extreme speech classification. The recent launch\nof ChatGPT has drawn global attention to the potential applications of Large\nLanguage Models (LLMs) across a diverse variety of tasks. Trained on vast and\ndiverse corpora, and demonstrating the ability to effectively capture and\nencode contextual information, LLMs emerge as highly promising tools for\ntackling this specific task of extreme speech classification. In this paper, we\nleverage the Indian subset of the extreme speech dataset from Maronikolakis et\nal. (2022) to develop an effective classification framework using LLMs. We\nevaluate open-source Llama models against closed-source OpenAI models, finding\nthat while pre-trained LLMs show moderate efficacy, fine-tuning with\ndomain-specific data significantly enhances performance, highlighting their\nadaptability to linguistic and contextual nuances. Although GPT-based models\noutperform Llama models in zero-shot settings, the performance gap disappears\nafter fine-tuning.\n", "date": "2025-02-21", "categories": []}
{"id": "2411.03888", "title": "Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech\n  Detection with Vision-Language Models", "abstract": "  Warning: this paper contains content that may be offensive or upsetting\n  Hate speech moderation on global platforms poses unique challenges due to the\nmultimodal and multilingual nature of content, along with the varying cultural\nperceptions. How well do current vision-language models (VLMs) navigate these\nnuances? To investigate this, we create the first multimodal and multilingual\nparallel hate speech dataset, annotated by a multicultural set of annotators,\ncalled Multi3Hate. It contains 300 parallel meme samples across 5 languages:\nEnglish, German, Spanish, Hindi, and Mandarin. We demonstrate that cultural\nbackground significantly affects multimodal hate speech annotation in our\ndataset. The average pairwise agreement among countries is just 74%,\nsignificantly lower than that of randomly selected annotator groups. Our\nqualitative analysis indicates that the lowest pairwise label agreement-only\n67% between the USA and India-can be attributed to cultural factors. We then\nconduct experiments with 5 large VLMs in a zero-shot setting, finding that\nthese models align more closely with annotations from the US than with those\nfrom other cultures, even when the memes and prompts are presented in the\ndominant language of the other culture. Code and dataset are available at\nhttps://github.com/MinhDucBui/Multi3Hate.\n", "date": "2024-11-06", "categories": []}
{"id": "2412.12072", "title": "Making FETCH! Happen: Finding Emergent Dog Whistles Through Common\n  Habitats", "abstract": "  WARNING: This paper contains content that maybe upsetting or offensive to\nsome readers. Dog whistles are coded expressions with dual meanings: one\nintended for the general public (outgroup) and another that conveys a specific\nmessage to an intended audience (ingroup). Often, these expressions are used to\nconvey controversial political opinions while maintaining plausible deniability\nand slip by content moderation filters. Identification of dog whistles relies\non curated lexicons, which have trouble keeping up to date. We introduce\nFETCH!, a task for finding novel dog whistles in massive social media corpora.\nWe find that state-of-the-art systems fail to achieve meaningful results across\nthree distinct social media case studies. We present EarShot, a strong baseline\nsystem that combines the strengths of vector databases and Large Language\nModels (LLMs) to efficiently and effectively identify new dog whistles.\n", "date": "2024-12-16", "categories": []}
{"id": "2502.10632", "title": "Code-Mixed Telugu-English Hate Speech Detection", "abstract": "  Hate speech detection in low-resource languages like Telugu is a growing\nchallenge in NLP. This study investigates transformer-based models, including\nTeluguHateBERT, HateBERT, DeBERTa, Muril, IndicBERT, Roberta, and\nHindi-Abusive-MuRIL, for classifying hate speech in Telugu. We fine-tune these\nmodels using Low-Rank Adaptation (LoRA) to optimize efficiency and performance.\nAdditionally, we explore a multilingual approach by translating Telugu text\ninto English using Google Translate to assess its impact on classification\naccuracy.\n  Our experiments reveal that most models show improved performance after\ntranslation, with DeBERTa and Hindi-Abusive-MuRIL achieving higher accuracy and\nF1 scores compared to training directly on Telugu text. Notably,\nHindi-Abusive-MuRIL outperforms all other models in both the original Telugu\ndataset and the translated dataset, demonstrating its robustness across\ndifferent linguistic settings. This suggests that translation enables models to\nleverage richer linguistic features available in English, leading to improved\nclassification performance. The results indicate that multilingual processing\ncan be an effective approach for hate speech detection in low-resource\nlanguages. These findings demonstrate that transformer models, when fine-tuned\nappropriately, can significantly improve hate speech detection in Telugu,\npaving the way for more robust multilingual NLP applications.\n", "date": "2025-02-15", "categories": []}
{"id": "2502.11073", "title": "Demystifying Hateful Content: Leveraging Large Multimodal Models for\n  Hateful Meme Detection with Explainable Decisions", "abstract": "  Hateful meme detection presents a significant challenge as a multimodal task\ndue to the complexity of interpreting implicit hate messages and contextual\ncues within memes. Previous approaches have fine-tuned pre-trained\nvision-language models (PT-VLMs), leveraging the knowledge they gained during\npre-training and their attention mechanisms to understand meme content.\nHowever, the reliance of these models on implicit knowledge and complex\nattention mechanisms renders their decisions difficult to explain, which is\ncrucial for building trust in meme classification. In this paper, we introduce\nIntMeme, a novel framework that leverages Large Multimodal Models (LMMs) for\nhateful meme classification with explainable decisions. IntMeme addresses the\ndual challenges of improving both accuracy and explainability in meme\nmoderation. The framework uses LMMs to generate human-like, interpretive\nanalyses of memes, providing deeper insights into multimodal content and\ncontext. Additionally, it uses independent encoding modules for both memes and\ntheir interpretations, which are then combined to enhance classification\nperformance. Our approach addresses the opacity and misclassification issues\nassociated with PT-VLMs, optimizing the use of LMMs for hateful meme detection.\nWe demonstrate the effectiveness of IntMeme through comprehensive experiments\nacross three datasets, showcasing its superiority over state-of-the-art models.\n", "date": "2025-02-16", "categories": []}
{"id": "2501.16235", "title": "Echoes of Discord: Forecasting Hater Reactions to Counterspeech", "abstract": "  Hate speech (HS) erodes the inclusiveness of online users and propagates\nnegativity and division. Counterspeech has been recognized as a way to mitigate\nthe harmful consequences. While some research has investigated the impact of\nuser-generated counterspeech on social media platforms, few have examined and\nmodeled haters' reactions toward counterspeech, despite the immediate\nalteration of haters' attitudes being an important aspect of counterspeech.\nThis study fills the gap by analyzing the impact of counterspeech from the\nhater's perspective, focusing on whether the counterspeech leads the hater to\nreenter the conversation and if the reentry is hateful. We compile the Reddit\nEchoes of Hate dataset (ReEco), which consists of triple-turn conversations\nfeaturing haters' reactions, to assess the impact of counterspeech. To predict\nhaters' behaviors, we employ two strategies: a two-stage reaction predictor and\na three-way classifier. The linguistic analysis sheds insights on the language\nof counterspeech to hate eliciting different haters' reactions. Experimental\nresults demonstrate that the 3-way classification model outperforms the\ntwo-stage reaction predictor, which first predicts reentry and then determines\nthe reentry type. We conclude the study with an assessment showing the most\ncommon errors identified by the best-performing model.\n", "date": "2025-01-27", "categories": []}
{"id": "2502.07637", "title": "BiaSWE: An Expert Annotated Dataset for Misogyny Detection in Swedish", "abstract": "  In this study, we introduce the process for creating BiaSWE, an\nexpert-annotated dataset tailored for misogyny detection in the Swedish\nlanguage. To address the cultural and linguistic specificity of misogyny in\nSwedish, we collaborated with experts from the social sciences and humanities.\nOur interdisciplinary team developed a rigorous annotation process,\nincorporating both domain knowledge and language expertise, to capture the\nnuances of misogyny in a Swedish context. This methodology ensures that the\ndataset is not only culturally relevant but also aligned with broader efforts\nin bias detection for low-resource languages. The dataset, along with the\nannotation guidelines, is publicly available for further research.\n", "date": "2025-02-11", "categories": []}
{"id": "2404.19252", "title": "ViTHSD: Exploiting Hatred by Targets for Hate Speech Detection on\n  Vietnamese Social Media Texts", "abstract": "  The growth of social networks makes toxic content spread rapidly. Hate speech\ndetection is a task to help decrease the number of harmful comments. With the\ndiversity in the hate speech created by users, it is necessary to interpret the\nhate speech besides detecting it. Hence, we propose a methodology to construct\na system for targeted hate speech detection from online streaming texts from\nsocial media. We first introduce the ViTHSD - a targeted hate speech detection\ndataset for Vietnamese Social Media Texts. The dataset contains 10K comments,\neach comment is labeled to specific targets with three levels: clean,\noffensive, and hate. There are 5 targets in the dataset, and each target is\nlabeled with the corresponding level manually by humans with strict annotation\nguidelines. The inter-annotator agreement obtained from the dataset is 0.45 by\nCohen's Kappa index, which is indicated as a moderate level. Then, we construct\na baseline for this task by combining the Bi-GRU-LSTM-CNN with the pre-trained\nlanguage model to leverage the power of text representation of BERTology.\nFinally, we suggest a methodology to integrate the baseline model for targeted\nhate speech detection into the online streaming system for practical\napplication in preventing hateful and offensive content on social media.\n", "date": "2024-04-30", "categories": []}
{"id": "2312.10269", "title": "The DSA Transparency Database: Auditing Self-reported Moderation Actions\n  by Social Media", "abstract": "  Since September 2023, the Digital Services Act (DSA) obliges large online\nplatforms to submit detailed data on each moderation action they take within\nthe European Union (EU) to the DSA Transparency Database. From its inception,\nthis centralized database has sparked scholarly interest as an unprecedented\nand potentially unique trove of data on real-world online moderation. Here, we\nthoroughly analyze all 353.12M records submitted by the eight largest social\nmedia platforms in the EU during the first 100 days of the database.\nSpecifically, we conduct a platform-wise comparative study of their: volume of\nmoderation actions, grounds for decision, types of applied restrictions, types\nof moderated content, timeliness in undertaking and submitting moderation\nactions, and use of automation. Furthermore, we systematically cross-check the\ncontents of the database with the platforms' own transparency reports. Our\nanalyses reveal that (i) the platforms adhered only in part to the philosophy\nand structure of the database, (ii) the structure of the database is partially\ninadequate for the platforms' reporting needs, (iii) the platforms exhibited\nsubstantial differences in their moderation actions, (iv) a remarkable fraction\nof the database data is inconsistent, (v) the platform X (formerly Twitter)\npresents the most inconsistencies. Our findings have far-reaching implications\nfor policymakers and scholars across diverse disciplines. They offer guidance\nfor future regulations that cater to the reporting needs of online platforms in\ngeneral, but also highlight opportunities to improve and refine the database\nitself.\n", "date": "2023-12-16", "categories": []}
{"id": "2501.16750", "title": "HateBench: Benchmarking Hate Speech Detectors on LLM-Generated Content\n  and Hate Campaigns", "abstract": "  Large Language Models (LLMs) have raised increasing concerns about their\nmisuse in generating hate speech. Among all the efforts to address this issue,\nhate speech detectors play a crucial role. However, the effectiveness of\ndifferent detectors against LLM-generated hate speech remains largely unknown.\nIn this paper, we propose HateBench, a framework for benchmarking hate speech\ndetectors on LLM-generated hate speech. We first construct a hate speech\ndataset of 7,838 samples generated by six widely-used LLMs covering 34 identity\ngroups, with meticulous annotations by three labelers. We then assess the\neffectiveness of eight representative hate speech detectors on the\nLLM-generated dataset. Our results show that while detectors are generally\neffective in identifying LLM-generated hate speech, their performance degrades\nwith newer versions of LLMs. We also reveal the potential of LLM-driven hate\ncampaigns, a new threat that LLMs bring to the field of hate speech detection.\nBy leveraging advanced techniques like adversarial attacks and model stealing\nattacks, the adversary can intentionally evade the detector and automate hate\ncampaigns online. The most potent adversarial attack achieves an attack success\nrate of 0.966, and its attack efficiency can be further improved by\n$13-21\\times$ through model stealing attacks with acceptable attack\nperformance. We hope our study can serve as a call to action for the research\ncommunity and platform moderators to fortify defenses against these emerging\nthreats.\n", "date": "2025-01-28", "categories": []}
{"id": "2501.16925", "title": "Detecting harassment and defamation in cyberbullying with\n  emotion-adaptive training", "abstract": "  Existing research on detecting cyberbullying incidents on social media has\nprimarily concentrated on harassment and is typically approached as a binary\nclassification task. However, cyberbullying encompasses various forms, such as\ndenigration and harassment, which celebrities frequently face. Furthermore,\nsuitable training data for these diverse forms of cyberbullying remains scarce.\nIn this study, we first develop a celebrity cyberbullying dataset that\nencompasses two distinct types of incidents: harassment and defamation. We\ninvestigate various types of transformer-based models, namely masked (RoBERTa,\nBert and DistilBert), replacing(Electra), autoregressive (XLnet),\nmasked&permuted (Mpnet), text-text (T5) and large language models (Llama2 and\nLlama3) under low source settings. We find that they perform competitively on\nexplicit harassment binary detection. However, their performance is\nsubstantially lower on harassment and denigration multi-classification tasks.\nTherefore, we propose an emotion-adaptive training framework (EAT) that helps\ntransfer knowledge from the domain of emotion detection to the domain of\ncyberbullying detection to help detect indirect cyberbullying events. EAT\nconsistently improves the average macro F1, precision and recall by 20% in\ncyberbullying detection tasks across nine transformer-based models under\nlow-resource settings. Our claims are supported by intuitive theoretical\ninsights and extensive experiments.\n", "date": "2025-01-28", "categories": []}
{"id": "2501.15042", "title": "SCCD: A Session-based Dataset for Chinese Cyberbullying Detection", "abstract": "  The rampant spread of cyberbullying content poses a growing threat to\nsocietal well-being. However, research on cyberbullying detection in Chinese\nremains underdeveloped, primarily due to the lack of comprehensive and reliable\ndatasets. Notably, no existing Chinese dataset is specifically tailored for\ncyberbullying detection. Moreover, while comments play a crucial role within\nsessions, current session-based datasets often lack detailed, fine-grained\nannotations at the comment level. To address these limitations, we present a\nnovel Chinese cyber-bullying dataset, termed SCCD, which consists of 677\nsession-level samples sourced from a major social media platform Weibo.\nMoreover, each comment within the sessions is annotated with fine-grained\nlabels rather than conventional binary class labels. Empirically, we evaluate\nthe performance of various baseline methods on SCCD, highlighting the\nchallenges for effective Chinese cyberbullying detection.\n", "date": "2025-01-25", "categories": []}
{"id": "2501.15438", "title": "Cross-Modal Transfer from Memes to Videos: Addressing Data Scarcity in\n  Hateful Video Detection", "abstract": "  Detecting hate speech in online content is essential to ensuring safer\ndigital spaces. While significant progress has been made in text and meme\nmodalities, video-based hate speech detection remains under-explored, hindered\nby a lack of annotated datasets and the high cost of video annotation. This gap\nis particularly problematic given the growing reliance on large models, which\ndemand substantial amounts of training data. To address this challenge, we\nleverage meme datasets as both a substitution and an augmentation strategy for\ntraining hateful video detection models. Our approach introduces a\nhuman-assisted reannotation pipeline to align meme dataset labels with video\ndatasets, ensuring consistency with minimal labeling effort. Using two\nstate-of-the-art vision-language models, we demonstrate that meme data can\nsubstitute for video data in resource-scarce scenarios and augment video\ndatasets to achieve further performance gains. Our results consistently\noutperform state-of-the-art benchmarks, showcasing the potential of cross-modal\ntransfer learning for advancing hateful video detection. Dataset and code are\navailable at https://github.com/Social-AI-Studio/CrossModalTransferLearning.\n", "date": "2025-01-26", "categories": []}
{"id": "2501.13976", "title": "Towards Safer Social Media Platforms: Scalable and Performant Few-Shot\n  Harmful Content Moderation Using Large Language Models", "abstract": "  The prevalence of harmful content on social media platforms poses significant\nrisks to users and society, necessitating more effective and scalable content\nmoderation strategies. Current approaches rely on human moderators, supervised\nclassifiers, and large volumes of training data, and often struggle with\nscalability, subjectivity, and the dynamic nature of harmful content (e.g.,\nviolent content, dangerous challenge trends, etc.). To bridge these gaps, we\nutilize Large Language Models (LLMs) to undertake few-shot dynamic content\nmoderation via in-context learning. Through extensive experiments on multiple\nLLMs, we demonstrate that our few-shot approaches can outperform existing\nproprietary baselines (Perspective and OpenAI Moderation) as well as prior\nstate-of-the-art few-shot learning methods, in identifying harm. We also\nincorporate visual information (video thumbnails) and assess if different\nmultimodal techniques improve model performance. Our results underscore the\nsignificant benefits of employing LLM based methods for scalable and dynamic\nharmful content moderation online.\n", "date": "2025-01-23", "categories": []}
{"id": "2406.14835", "title": "ToVo: Toxicity Taxonomy via Voting", "abstract": "  Existing toxic detection models face significant limitations, such as lack of\ntransparency, customization, and reproducibility. These challenges stem from\nthe closed-source nature of their training data and the paucity of explanations\nfor their evaluation mechanism. To address these issues, we propose a dataset\ncreation mechanism that integrates voting and chain-of-thought processes,\nproducing a high-quality open-source dataset for toxic content detection. Our\nmethodology ensures diverse classification metrics for each sample and includes\nboth classification scores and explanatory reasoning for the classifications.\n  We utilize the dataset created through our proposed mechanism to train our\nmodel, which is then compared against existing widely-used detectors. Our\napproach not only enhances transparency and customizability but also\nfacilitates better fine-tuning for specific use cases. This work contributes a\nrobust framework for developing toxic content detection models, emphasizing\nopenness and adaptability, thus paving the way for more effective and\nuser-specific content moderation solutions.\n", "date": "2024-06-21", "categories": []}
{"id": "2501.13302", "title": "Watching the AI Watchdogs: A Fairness and Robustness Analysis of AI\n  Safety Moderation Classifiers", "abstract": "  AI Safety Moderation (ASM) classifiers are designed to moderate content on\nsocial media platforms and to serve as guardrails that prevent Large Language\nModels (LLMs) from being fine-tuned on unsafe inputs. Owing to their potential\nfor disparate impact, it is crucial to ensure that these classifiers: (1) do\nnot unfairly classify content belonging to users from minority groups as unsafe\ncompared to those from majority groups and (2) that their behavior remains\nrobust and consistent across similar inputs. In this work, we thus examine the\nfairness and robustness of four widely-used, closed-source ASM classifiers:\nOpenAI Moderation API, Perspective API, Google Cloud Natural Language (GCNL)\nAPI, and Clarifai API. We assess fairness using metrics such as demographic\nparity and conditional statistical parity, comparing their performance against\nASM models and a fair-only baseline. Additionally, we analyze robustness by\ntesting the classifiers' sensitivity to small and natural input perturbations.\nOur findings reveal potential fairness and robustness gaps, highlighting the\nneed to mitigate these issues in future versions of these models.\n", "date": "2025-01-23", "categories": []}
{"id": "2407.21004", "title": "Evolver: Chain-of-Evolution Prompting to Boost Large Multimodal Models\n  for Hateful Meme Detection", "abstract": "  Recent advances show that two-stream approaches have achieved outstanding\nperformance in hateful meme detection. However, hateful memes constantly evolve\nas new memes emerge by fusing progressive cultural ideas, making existing\nmethods obsolete or ineffective. In this work, we explore the potential of\nLarge Multimodal Models (LMMs) for hateful meme detection. To this end, we\npropose Evolver, which incorporates LMMs via Chain-of-Evolution (CoE)\nPrompting, by integrating the evolution attribute and in-context information of\nmemes. Specifically, Evolver simulates the evolving and expressing process of\nmemes and reasons through LMMs in a step-by-step manner. First, an evolutionary\npair mining module retrieves the top-k most similar memes in the external\ncurated meme set with the input meme. Second, an evolutionary information\nextractor is designed to summarize the semantic regularities between the paired\nmemes for prompting. Finally, a contextual relevance amplifier enhances the\nin-context hatefulness information to boost the search for evolutionary\nprocesses. Extensive experiments on public FHM, MAMI, and HarM datasets show\nthat CoE prompting can be incorporated into existing LMMs to improve their\nperformance. More encouragingly, it can serve as an interpretive tool to\npromote the understanding of the evolution of social memes. [Homepage]\n(https://github.com/inFaaa/Evolver)\n", "date": "2024-07-30", "categories": []}
{"id": "2412.18216", "title": "ICM-Assistant: Instruction-tuning Multimodal Large Language Models for\n  Rule-based Explainable Image Content Moderation", "abstract": "  Controversial contents largely inundate the Internet, infringing various\ncultural norms and child protection standards. Traditional Image Content\nModeration (ICM) models fall short in producing precise moderation decisions\nfor diverse standards, while recent multimodal large language models (MLLMs),\nwhen adopted to general rule-based ICM, often produce classification and\nexplanation results that are inconsistent with human moderators. Aiming at\nflexible, explainable, and accurate ICM, we design a novel rule-based dataset\ngeneration pipeline, decomposing concise human-defined rules and leveraging\nwell-designed multi-stage prompts to enrich short explicit image annotations.\nOur ICM-Instruct dataset includes detailed moderation explanation and\nmoderation Q-A pairs. Built upon it, we create our ICM-Assistant model in the\nframework of rule-based ICM, making it readily applicable in real practice. Our\nICM-Assistant model demonstrates exceptional performance and flexibility.\nSpecifically, it significantly outperforms existing approaches on various\nsources, improving both the moderation classification (36.8% on average) and\nmoderation explanation quality (26.6% on average) consistently over existing\nMLLMs. Code/Data is available at https://github.com/zhaoyuzhi/ICM-Assistant.\n", "date": "2024-12-24", "categories": []}
{"id": "2501.09722", "title": "Attention based Bidirectional GRU hybrid model for inappropriate content\n  detection in Urdu language", "abstract": "  With the increased use of the internet and social networks for online\ndiscussions, the spread of toxic and inappropriate content on social networking\nsites has also increased. Several studies have been conducted in different\nlanguages. However, there is less work done for South Asian languages for\ninappropriate content identification using deep learning techniques. In Urdu\nlanguage, the spellings are not unique, and people write different common\nspellings for the same word, while mixing it other languages, like English in\nthe text makes it more challenging, and limited research work is available to\nprocess such language with the finest algorithms. The use of attention layer\nwith a deep learning model can help handling the long-term dependencies and\nincrease its efficiency . To explore the effects of the attention layer, this\nstudy proposes attention-based Bidirectional GRU hybrid model for identifying\ninappropriate content in Urdu Unicode text language. Four different baseline\ndeep learning models; LSTM, Bi-LSTM, GRU, and TCN, are used to compare the\nperformance of the proposed model. The results of these models were compared\nbased on evaluation metrics, dataset size, and impact of the word embedding\nlayer. The pre-trained Urdu word2Vec embeddings were utilized for our case. Our\nproposed model BiGRU-A outperformed all other baseline models by yielding 84\\%\naccuracy without using pre-trained word2Vec layer. From our experiments, we\nhave established that the attention layer improves the model's efficiency, and\npre-trained word2Vec embedding does not work well with an inappropriate content\ndataset.\n", "date": "2025-01-16", "categories": []}
{"id": "2501.08284", "title": "AfriHate: A Multilingual Collection of Hate Speech and Abusive Language\n  Datasets for African Languages", "abstract": "  Hate speech and abusive language are global phenomena that need\nsocio-cultural background knowledge to be understood, identified, and\nmoderated. However, in many regions of the Global South, there have been\nseveral documented occurrences of (1) absence of moderation and (2) censorship\ndue to the reliance on keyword spotting out of context. Further, high-profile\nindividuals have frequently been at the center of the moderation process, while\nlarge and targeted hate speech campaigns against minorities have been\noverlooked. These limitations are mainly due to the lack of high-quality data\nin the local languages and the failure to include local communities in the\ncollection, annotation, and moderation processes. To address this issue, we\npresent AfriHate: a multilingual collection of hate speech and abusive language\ndatasets in 15 African languages. Each instance in AfriHate is annotated by\nnative speakers familiar with the local culture. We report the challenges\nrelated to the construction of the datasets and present various classification\nbaseline results with and without using LLMs. The datasets, individual\nannotations, and hate speech and offensive language lexicons are available on\nhttps://github.com/AfriHate/AfriHate\n", "date": "2025-01-14", "categories": []}
{"id": "2410.15911", "title": "DefVerify: Do Hate Speech Models Reflect Their Dataset's Definition?", "abstract": "  When building a predictive model, it is often difficult to ensure that\napplication-specific requirements are encoded by the model that will eventually\nbe deployed. Consider researchers working on hate speech detection. They will\nhave an idea of what is considered hate speech, but building a model that\nreflects their view accurately requires preserving those ideals throughout the\nworkflow of data set construction and model training. Complications such as\nsampling bias, annotation bias, and model misspecification almost always arise,\npossibly resulting in a gap between the application specification and the\nmodel's actual behavior upon deployment. To address this issue for hate speech\ndetection, we propose DefVerify: a 3-step procedure that (i) encodes a\nuser-specified definition of hate speech, (ii) quantifies to what extent the\nmodel reflects the intended definition, and (iii) tries to identify the point\nof failure in the workflow. We use DefVerify to find gaps between definition\nand model behavior when applied to six popular hate speech benchmark datasets.\n", "date": "2024-10-21", "categories": []}
{"id": "2501.05871", "title": "Collaborative Content Moderation in the Fediverse", "abstract": "  The Fediverse, a group of interconnected servers providing a variety of\ninteroperable services (e.g. micro-blogging in Mastodon) has gained rapid\npopularity. This sudden growth, partly driven by Elon Musk's acquisition of\nTwitter, has created challenges for administrators though. This paper focuses\non one particular challenge: content moderation, e.g. the need to remove spam\nor hate speech. While centralized platforms like Facebook and Twitter rely on\nautomated tools for moderation, their dependence on massive labeled datasets\nand specialized infrastructure renders them impractical for decentralized,\nlow-resource settings like the Fediverse. In this work, we design and evaluate\nFedMod, a collaborative content moderation system based on federated learning.\nOur system enables servers to exchange parameters of partially trained local\ncontent moderation models with similar servers, creating a federated model\nshared among collaborating servers. FedMod demonstrates robust performance on\nthree different content moderation tasks: harmful content detection, bot\ncontent detection, and content warning assignment, achieving average per-server\nmacro-F1 scores of 0.71, 0.73, and 0.58, respectively.\n", "date": "2025-01-10", "categories": []}
{"id": "2406.07287", "title": "Bilingual Sexism Classification: Fine-Tuned XLM-RoBERTa and GPT-3.5\n  Few-Shot Learning", "abstract": "  Sexism in online content is a pervasive issue that necessitates effective\nclassification techniques to mitigate its harmful impact. Online platforms\noften have sexist comments and posts that create a hostile environment,\nespecially for women and minority groups. This content not only spreads harmful\nstereotypes but also causes emotional harm. Reliable methods are essential to\nfind and remove sexist content, making online spaces safer and more welcoming.\nTherefore, the sEXism Identification in Social neTworks (EXIST) challenge\naddresses this issue at CLEF 2024. This study aims to improve sexism\nidentification in bilingual contexts (English and Spanish) by leveraging\nnatural language processing models. The tasks are to determine whether a text\nis sexist and what the source intention behind it is. We fine-tuned the\nXLM-RoBERTa model and separately used GPT-3.5 with few-shot learning prompts to\nclassify sexist content. The XLM-RoBERTa model exhibited robust performance in\nhandling complex linguistic structures, while GPT-3.5's few-shot learning\ncapability allowed for rapid adaptation to new data with minimal labeled\nexamples. Our approach using XLM-RoBERTa achieved 4th place in the soft-soft\nevaluation of Task 1 (sexism identification). For Task 2 (source intention), we\nachieved 2nd place in the soft-soft evaluation.\n", "date": "2024-06-11", "categories": []}
{"id": "2501.00697", "title": "PANDA -- Paired Anti-hate Narratives Dataset from Asia: Using an\n  LLM-as-a-Judge to Create the First Chinese Counterspeech Dataset", "abstract": "  Despite the global prevalence of Modern Standard Chinese language,\ncounterspeech (CS) resources for Chinese remain virtually nonexistent. To\naddress this gap in East Asian counterspeech research we introduce the a corpus\nof Modern Standard Mandarin counterspeech that focuses on combating hate speech\nin Mainland China. This paper proposes a novel approach of generating CS by\nusing an LLM-as-a-Judge, simulated annealing, LLMs zero-shot CN generation and\na round-robin algorithm. This is followed by manual verification for quality\nand contextual relevance. This paper details the methodology for creating\neffective counterspeech in Chinese and other non-Eurocentric languages,\nincluding unique cultural patterns of which groups are maligned and linguistic\npatterns in what kinds of discourse markers are programmatically marked as hate\nspeech (HS). Analysis of the generated corpora, we provide strong evidence for\nthe lack of open-source, properly labeled Chinese hate speech data and the\nlimitations of using an LLM-as-Judge to score possible answers in Chinese.\nMoreover, the present corpus serves as the first East Asian language based CS\ncorpus and provides an essential resource for future research on counterspeech\ngeneration and evaluation.\n", "date": "2025-01-01", "categories": []}
{"id": "2501.00907", "title": "U-GIFT: Uncertainty-Guided Firewall for Toxic Speech in Few-Shot\n  Scenario", "abstract": "  With the widespread use of social media, user-generated content has surged on\nonline platforms. When such content includes hateful, abusive, offensive, or\ncyberbullying behavior, it is classified as toxic speech, posing a significant\nthreat to the online ecosystem's integrity and safety. While manual content\nmoderation is still prevalent, the overwhelming volume of content and the\npsychological strain on human moderators underscore the need for automated\ntoxic speech detection. Previously proposed detection methods often rely on\nlarge annotated datasets; however, acquiring such datasets is both costly and\nchallenging in practice. To address this issue, we propose an\nuncertainty-guided firewall for toxic speech in few-shot scenarios, U-GIFT,\nthat utilizes self-training to enhance detection performance even when labeled\ndata is limited. Specifically, U-GIFT combines active learning with Bayesian\nNeural Networks (BNNs) to automatically identify high-quality samples from\nunlabeled data, prioritizing the selection of pseudo-labels with higher\nconfidence for training based on uncertainty estimates derived from model\npredictions. Extensive experiments demonstrate that U-GIFT significantly\noutperforms competitive baselines in few-shot detection scenarios. In the\n5-shot setting, it achieves a 14.92\\% performance improvement over the basic\nmodel. Importantly, U-GIFT is user-friendly and adaptable to various\npre-trained language models (PLMs). It also exhibits robust performance in\nscenarios with sample imbalance and cross-domain settings, while showcasing\nstrong generalization across various language applications. We believe that\nU-GIFT provides an efficient solution for few-shot toxic speech detection,\noffering substantial support for automated content moderation in cyberspace,\nthereby acting as a firewall to promote advancements in cybersecurity.\n", "date": "2025-01-01", "categories": []}
{"id": "2501.01256", "title": "Digital Guardians: Can GPT-4, Perspective API, and Moderation API\n  reliably detect hate speech in reader comments of German online newspapers?", "abstract": "  In recent years, toxic content and hate speech have become widespread\nphenomena on the internet. Moderators of online newspapers and forums are now\nrequired, partly due to legal regulations, to carefully review and, if\nnecessary, delete reader comments. This is a labor-intensive process. Some\nproviders of large language models already offer solutions for automated hate\nspeech detection or the identification of toxic content. These include GPT-4o\nfrom OpenAI, Jigsaw's (Google) Perspective API, and OpenAI's Moderation API.\nBased on the selected German test dataset HOCON34k, which was specifically\ncreated for developing tools to detect hate speech in reader comments of online\nnewspapers, these solutions are compared with each other and against the\nHOCON34k baseline. The test dataset contains 1,592 annotated text samples. For\nGPT-4o, three different promptings are used, employing a Zero-Shot, One-Shot,\nand Few-Shot approach. The results of the experiments demonstrate that GPT-4o\noutperforms both the Perspective API and the Moderation API, and exceeds the\nHOCON34k baseline by approximately 5 percentage points, as measured by a\ncombined metric of MCC and F2-score.\n", "date": "2025-01-02", "categories": []}
{"id": "2310.05650", "title": "ReZG: Retrieval-Augmented Zero-Shot Counter Narrative Generation for\n  Hate Speech", "abstract": "  The proliferation of hate speech (HS) on social media poses a serious threat\nto societal security. Automatic counter narrative (CN) generation, as an active\nstrategy for HS intervention, has garnered increasing attention in recent\nyears. Existing methods for automatically generating CNs mainly rely on\nre-training or fine-tuning pre-trained language models (PLMs) on human-curated\nCN corpora. Unfortunately, the annotation speed of CN corpora cannot keep up\nwith the growth of HS targets, while generating specific and effective CNs for\nunseen targets remains a significant challenge for the model. To tackle this\nissue, we propose Retrieval-Augmented Zero-shot Generation (ReZG) to generate\nCNs with high-specificity for unseen targets. Specifically, we propose a\nmulti-dimensional hierarchical retrieval method that integrates stance,\nsemantics, and fitness, extending the retrieval metric from single dimension to\nmultiple dimensions suitable for the knowledge that refutes HS. Then, we\nimplement an energy-based constrained decoding mechanism that enables PLMs to\nuse differentiable knowledge preservation, countering, and fluency constraint\nfunctions instead of in-target CNs as control signals for generation, thereby\nachieving zero-shot CN generation. With the above techniques, ReZG can\nintegrate external knowledge flexibly and improve the specificity of CNs.\nExperimental results show that ReZG exhibits stronger generalization\ncapabilities and outperforms strong baselines with significant improvements of\n2.0%+ in the relevance and 4.5%+ in the countering success rate metrics.\n", "date": "2023-10-09", "categories": []}
{"id": "2412.17947", "title": "IITR-CIOL@NLU of Devanagari Script Languages 2025: Multilingual Hate\n  Speech Detection and Target Identification in Devanagari-Scripted Languages", "abstract": "  This work focuses on two subtasks related to hate speech detection and target\nidentification in Devanagari-scripted languages, specifically Hindi, Marathi,\nNepali, Bhojpuri, and Sanskrit. Subtask B involves detecting hate speech in\nonline text, while Subtask C requires identifying the specific targets of hate\nspeech, such as individuals, organizations, or communities. We propose the\nMultilingualRobertaClass model, a deep neural network built on the pretrained\nmultilingual transformer model ia-multilingual-transliterated-roberta,\noptimized for classification tasks in multilingual and transliterated contexts.\nThe model leverages contextualized embeddings to handle linguistic diversity,\nwith a classifier head for binary classification. We received 88.40% accuracy\nin Subtask B and 66.11% accuracy in Subtask C, in the test set.\n", "date": "2024-12-23", "categories": []}
{"id": "2412.19928", "title": "Assessing Text Classification Methods for Cyberbullying Detection on\n  Social Media Platforms", "abstract": "  Cyberbullying significantly contributes to mental health issues in\ncommunities by negatively impacting the psychology of victims. It is a\nprevalent problem on social media platforms, necessitating effective, real-time\ndetection and monitoring systems to identify harmful messages. However, current\ncyberbullying detection systems face challenges related to performance, dataset\nquality, time efficiency, and computational costs. This research aims to\nconduct a comparative study by adapting and evaluating existing text\nclassification techniques within the cyberbullying detection domain. The study\nspecifically evaluates the effectiveness and performance of these techniques in\nidentifying cyberbullying instances on social media platforms. It focuses on\nleveraging and assessing large language models, including BERT, RoBERTa, XLNet,\nDistilBERT, and GPT-2.0, for their suitability in this domain. The results show\nthat BERT strikes a balance between performance, time efficiency, and\ncomputational resources: Accuracy of 95%, Precision of 95%, Recall of 95%, F1\nScore of 95%, Error Rate of 5%, Inference Time of 0.053 seconds, RAM Usage of\n35.28 MB, CPU/GPU Usage of 0.4%, and Energy Consumption of 0.000263 kWh. The\nfindings demonstrate that generative AI models, while powerful, do not\nconsistently outperform fine-tuned models on the tested benchmarks. However,\nstate-of-the-art performance can still be achieved through strategic adaptation\nand fine-tuning of existing models for specific datasets and tasks.\n", "date": "2024-12-27", "categories": []}
{"id": "2412.20541", "title": "SAFE-MEME: Structured Reasoning Framework for Robust Hate Speech\n  Detection in Memes", "abstract": "  Memes act as cryptic tools for sharing sensitive ideas, often requiring\ncontextual knowledge to interpret. This makes moderating multimodal memes\nchallenging, as existing works either lack high-quality datasets on nuanced\nhate categories or rely on low-quality social media visuals. Here, we curate\ntwo novel multimodal hate speech datasets, MHS and MHS-Con, that capture\nfine-grained hateful abstractions in regular and confounding scenarios,\nrespectively. We benchmark these datasets against several competing baselines.\nFurthermore, we introduce SAFE-MEME (Structured reAsoning FramEwork), a novel\nmultimodal Chain-of-Thought-based framework employing Q&A-style reasoning\n(SAFE-MEME-QA) and hierarchical categorization (SAFE-MEME-H) to enable robust\nhate speech detection in memes. SAFE-MEME-QA outperforms existing baselines,\nachieving an average improvement of approximately 5% and 4% on MHS and MHS-Con,\nrespectively. In comparison, SAFE-MEME-H achieves an average improvement of 6%\nin MHS while outperforming only multimodal baselines in MHS-Con. We show that\nfine-tuning a single-layer adapter within SAFE-MEME-H outperforms fully\nfine-tuned models in regular fine-grained hateful meme detection. However, the\nfully fine-tuning approach with a Q&A setup is more effective for handling\nconfounding cases. We also systematically examine the error cases, offering\nvaluable insights into the robustness and limitations of the proposed\nstructured reasoning framework for analyzing hateful memes.\n", "date": "2024-12-29", "categories": []}
{"id": "2412.17131", "title": "LLMsAgainstHate @ NLU of Devanagari Script Languages 2025: Hate Speech\n  Detection and Target Identification in Devanagari Languages via Parameter\n  Efficient Fine-Tuning of LLMs", "abstract": "  The detection of hate speech has become increasingly important in combating\nonline hostility and its real-world consequences. Despite recent advancements,\nthere is limited research addressing hate speech detection in\nDevanagari-scripted languages, where resources and tools are scarce. While\nlarge language models (LLMs) have shown promise in language-related tasks,\ntraditional fine-tuning approaches are often infeasible given the size of the\nmodels. In this paper, we propose a Parameter Efficient Fine tuning (PEFT)\nbased solution for hate speech detection and target identification. We evaluate\nmultiple LLMs on the Devanagari dataset provided by (Thapa et al., 2025), which\ncontains annotated instances in 2 languages - Hindi and Nepali. The results\ndemonstrate the efficacy of our approach in handling Devanagari-scripted\ncontent.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.17825", "title": "Leveraging Sentiment for Offensive Text Classification", "abstract": "  In this paper, we conduct experiment to analyze whether models can classify\noffensive texts better with the help of sentiment. We conduct this experiment\non the SemEval 2019 task 6, OLID, dataset. First, we utilize pre-trained\nlanguage models to predict the sentiment of each instance. Later we pick the\nmodel that achieved the best performance on the OLID test set, and train it on\nthe augmented OLID set to analyze the performance. Results show that utilizing\nsentiment increases the overall performance of the model.\n", "date": "2024-12-09", "categories": []}
{"id": "2412.16215", "title": "Zero-Shot Image Moderation in Google Ads with LLM-Assisted Textual\n  Descriptions and Cross-modal Co-embeddings", "abstract": "  We present a scalable and agile approach for ads image content moderation at\nGoogle, addressing the challenges of moderating massive volumes of ads with\ndiverse content and evolving policies. The proposed method utilizes\nhuman-curated textual descriptions and cross-modal text-image co-embeddings to\nenable zero-shot classification of policy violating ads images, bypassing the\nneed for extensive supervised training data and human labeling. By leveraging\nlarge language models (LLMs) and user expertise, the system generates and\nrefines a comprehensive set of textual descriptions representing policy\nguidelines. During inference, co-embedding similarity between incoming images\nand the textual descriptions serves as a reliable signal for policy violation\ndetection, enabling efficient and adaptable ads content moderation. Evaluation\nresults demonstrate the efficacy of this framework in significantly boosting\nthe detection of policy violating content.\n", "date": "2024-12-18", "categories": []}
{"id": "2412.15453", "title": "Northeastern Uni at Multilingual Counterspeech Generation: Enhancing\n  Counter Speech Generation with LLM Alignment through Direct Preference\n  Optimization", "abstract": "  The automatic generation of counter-speech (CS) is a critical strategy for\naddressing hate speech by providing constructive and informed responses.\nHowever, existing methods often fail to generate high-quality, impactful, and\nscalable CS, particularly across diverse linguistic contexts. In this paper, we\npropose a novel methodology to enhance CS generation by aligning Large Language\nModels (LLMs) using Supervised Fine-Tuning (SFT) and Direct Preference\nOptimization (DPO). Our approach leverages DPO to align LLM outputs with human\npreferences, ensuring contextually appropriate and linguistically adaptable\nresponses. Additionally, we incorporate knowledge grounding to enhance the\nfactual accuracy and relevance of generated CS. Experimental results\ndemonstrate that DPO-aligned models significantly outperform SFT baselines on\nCS benchmarks while scaling effectively to multiple languages. These findings\nhighlight the potential of preference-based alignment techniques to advance CS\ngeneration across varied linguistic settings. The model supervision and\nalignment is done in English and the same model is used for reporting metrics\nacross other languages like Basque, Italian, and Spanish.\n", "date": "2024-12-19", "categories": []}
{"id": "2410.14881", "title": "Class-RAG: Real-Time Content Moderation with Retrieval Augmented\n  Generation", "abstract": "  Robust content moderation classifiers are essential for the safety of\nGenerative AI systems. In this task, differences between safe and unsafe inputs\nare often extremely subtle, making it difficult for classifiers (and indeed,\neven humans) to properly distinguish violating vs. benign samples without\ncontext or explanation. Scaling risk discovery and mitigation through\ncontinuous model fine-tuning is also slow, challenging and costly, preventing\ndevelopers from being able to respond quickly and effectively to emergent\nharms. We propose a Classification approach employing Retrieval-Augmented\nGeneration (Class-RAG). Class-RAG extends the capability of its base LLM\nthrough access to a retrieval library which can be dynamically updated to\nenable semantic hotfixing for immediate, flexible risk mitigation. Compared to\nmodel fine-tuning, Class-RAG demonstrates flexibility and transparency in\ndecision-making, outperforms on classification and is more robust against\nadversarial attack, as evidenced by empirical studies. Our findings also\nsuggest that Class-RAG performance scales with retrieval library size,\nindicating that increasing the library size is a viable and low-cost approach\nto improve content moderation.\n", "date": "2024-10-18", "categories": []}
{"id": "2412.12212", "title": "Finding a Wolf in Sheep's Clothing: Combating Adversarial Text-To-Image\n  Prompts with Text Summarization", "abstract": "  Text-to-image models are vulnerable to the stepwise \"Divide-and-Conquer\nAttack\" (DACA) that utilize a large language model to obfuscate inappropriate\ncontent in prompts by wrapping sensitive text in a benign narrative. To\nmitigate stepwise DACA attacks, we propose a two-layer method involving text\nsummarization followed by binary classification. We assembled the Adversarial\nText-to-Image Prompt (ATTIP) dataset ($N=940$), which contained DACA-obfuscated\nand non-obfuscated prompts. From the ATTIP dataset, we created two summarized\nversions: one generated by a small encoder model and the other by a large\nlanguage model. Then, we used an encoder classifier and a GPT-4o classifier to\nperform content moderation on the summarized and unsummarized prompts. When\ncompared with a classifier that operated over the unsummarized data, our method\nimproved F1 score performance by 31%. Further, the highest recorded F1 score\nachieved (98%) was produced by the encoder classifier on a summarized ATTIP\nvariant. This study indicates that pre-classification text summarization can\ninoculate content detection models against stepwise DACA obfuscations.\n", "date": "2024-12-15", "categories": []}
{"id": "2403.02472", "title": "OffensiveLang: A Community Based Implicit Offensive Language Dataset", "abstract": "  The widespread presence of hateful languages on social media has resulted in\nadverse effects on societal well-being. As a result, addressing this issue with\nhigh priority has become very important. Hate speech or offensive languages\nexist in both explicit and implicit forms, with the latter being more\nchallenging to detect. Current research in this domain encounters several\nchallenges. Firstly, the existing datasets primarily rely on the collection of\ntexts containing explicit offensive keywords, making it challenging to capture\nimplicitly offensive contents that are devoid of these keywords. Secondly,\ncommon methodologies tend to focus solely on textual analysis, neglecting the\nvaluable insights that community information can provide. In this research\npaper, we introduce a novel dataset OffensiveLang, a community based implicit\noffensive language dataset generated by ChatGPT 3.5 containing data for 38\ndifferent target groups. Despite limitations in generating offensive texts\nusing ChatGPT due to ethical constraints, we present a prompt-based approach\nthat effectively generates implicit offensive languages. To ensure data\nquality, we evaluate the dataset with human. Additionally, we employ a\nprompt-based zero-shot method with ChatGPT and compare the detection results\nbetween human annotation and ChatGPT annotation. We utilize existing\nstate-of-the-art models to see how effective they are in detecting such\nlanguages. The dataset is available here:\nhttps://github.com/AmitDasRup123/OffensiveLang\n", "date": "2024-03-04", "categories": []}
{"id": "2412.10991", "title": "Navigating Dialectal Bias and Ethical Complexities in Levantine Arabic\n  Hate Speech Detection", "abstract": "  Social media platforms have become central to global communication, yet they\nalso facilitate the spread of hate speech. For underrepresented dialects like\nLevantine Arabic, detecting hate speech presents unique cultural, ethical, and\nlinguistic challenges. This paper explores the complex sociopolitical and\nlinguistic landscape of Levantine Arabic and critically examines the\nlimitations of current datasets used in hate speech detection. We highlight the\nscarcity of publicly available, diverse datasets and analyze the consequences\nof dialectal bias within existing resources. By emphasizing the need for\nculturally and contextually informed natural language processing (NLP) tools,\nwe advocate for a more nuanced and inclusive approach to hate speech detection\nin the Arab world.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.11691", "title": "Multilingual and Explainable Text Detoxification with Parallel Corpora", "abstract": "  Even with various regulations in place across countries and social media\nplatforms (Government of India, 2021; European Parliament and Council of the\nEuropean Union, 2022, digital abusive speech remains a significant issue. One\npotential approach to address this challenge is automatic text detoxification,\na text style transfer (TST) approach that transforms toxic language into a more\nneutral or non-toxic form. To date, the availability of parallel corpora for\nthe text detoxification task (Logachevavet al., 2022; Atwell et al., 2022;\nDementievavet al., 2024a) has proven to be crucial for state-of-the-art\napproaches. With this work, we extend parallel text detoxification corpus to\nnew languages -- German, Chinese, Arabic, Hindi, and Amharic -- testing in the\nextensive multilingual setup TST baselines. Next, we conduct the first of its\nkind an automated, explainable analysis of the descriptive features of both\ntoxic and non-toxic sentences, diving deeply into the nuances, similarities,\nand differences of toxicity and detoxification across 9 languages. Finally,\nbased on the obtained insights, we experiment with a novel text detoxification\nmethod inspired by the Chain-of-Thoughts reasoning approach, enhancing the\nprompting process through clustering on relevant descriptive attributes.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.01408", "title": "Towards Cross-Lingual Audio Abuse Detection in Low-Resource Settings\n  with Few-Shot Learning", "abstract": "  Online abusive content detection, particularly in low-resource settings and\nwithin the audio modality, remains underexplored. We investigate the potential\nof pre-trained audio representations for detecting abusive language in\nlow-resource languages, in this case, in Indian languages using Few Shot\nLearning (FSL). Leveraging powerful representations from models such as Wav2Vec\nand Whisper, we explore cross-lingual abuse detection using the ADIMA dataset\nwith FSL. Our approach integrates these representations within the\nModel-Agnostic Meta-Learning (MAML) framework to classify abusive language in\n10 languages. We experiment with various shot sizes (50-200) evaluating the\nimpact of limited data on performance. Additionally, a feature visualization\nstudy was conducted to better understand model behaviour. This study highlights\nthe generalization ability of pre-trained models in low-resource scenarios and\noffers valuable insights into detecting abusive language in multilingual\ncontexts.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.08163", "title": "NLPineers@ NLU of Devanagari Script Languages 2025: Hate Speech\n  Detection using Ensembling of BERT-based models", "abstract": "  This paper explores hate speech detection in Devanagari-scripted languages,\nfocusing on Hindi and Nepali, for Subtask B of the CHIPSAL@COLING 2025 Shared\nTask. Using a range of transformer-based models such as XLM-RoBERTa, MURIL, and\nIndicBERT, we examine their effectiveness in navigating the nuanced boundary\nbetween hate speech and free expression. Our best performing model, implemented\nas ensemble of multilingual BERT models achieve Recall of 0.7762 (Rank 3/31 in\nterms of recall) and F1 score of 0.6914 (Rank 17/31). To address class\nimbalance, we used backtranslation for data augmentation, and cosine similarity\nto preserve label consistency after augmentation. This work emphasizes the need\nfor hate speech detection in Devanagari-scripted languages and presents a\nfoundation for further research.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.06144", "title": "Hate Speech According to the Law: An Analysis for Effective Detection", "abstract": "  The issue of hate speech extends beyond the confines of the online realm. It\nis a problem with real-life repercussions, prompting most nations to formulate\nlegal frameworks that classify hate speech as a punishable offence. These legal\nframeworks differ from one country to another, contributing to the big chaos\nthat online platforms have to face when addressing reported instances of hate\nspeech. With the definitions of hate speech falling short in introducing a\nrobust framework, we turn our gaze onto hate speech laws. We consult the\nopinion of legal experts on a hate speech dataset and we experiment by\nemploying various approaches such as pretrained models both on hate speech and\nlegal data, as well as exploiting two large language models (Qwen2-7B-Instruct\nand Meta-Llama-3-70B). Due to the time-consuming nature of data acquisition for\nprosecutable hate speech, we use pseudo-labeling to improve our pretrained\nmodels. This study highlights the importance of amplifying research on\nprosecutable hate speech and provides insights into effective strategies for\ncombating hate speech within the parameters of legal frameworks. Our findings\nshow that legal knowledge in the form of annotations can be useful when\nclassifying prosecutable hate speech, yet more focus should be paid on the\ndifferences between the laws.\n", "date": "2024-12-09", "categories": []}
{"id": "2412.04046", "title": "Hostility Detection in UK Politics: A Dataset on Online Abuse Targeting\n  MPs", "abstract": "  Numerous politicians use social media platforms, particularly X, to engage\nwith their constituents. This interaction allows constituents to pose questions\nand offer feedback but also exposes politicians to a barrage of hostile\nresponses, especially given the anonymity afforded by social media. They are\ntypically targeted in relation to their governmental role, but the comments\nalso tend to attack their personal identity. This can discredit politicians and\nreduce public trust in the government. It can also incite anger and disrespect,\nleading to offline harm and violence. While numerous models exist for detecting\nhostility in general, they lack the specificity required for political\ncontexts. Furthermore, addressing hostility towards politicians demands\ntailored approaches due to the distinct language and issues inherent to each\ncountry (e.g., Brexit for the UK). To bridge this gap, we construct a dataset\nof 3,320 English tweets spanning a two-year period manually annotated for\nhostility towards UK MPs. Our dataset also captures the targeted identity\ncharacteristics (race, gender, religion, none) in hostile tweets. We perform\nlinguistic and topical analyses to delve into the unique content of the UK\npolitical data. Finally, we evaluate the performance of pre-trained language\nmodels and large language models on binary hostility detection and multi-class\ntargeted identity type classification tasks. Our study offers valuable data and\ninsights for future research on the prevalence and nature of politics-related\nhostility specific to the UK.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.01413", "title": "Impromptu Cybercrime Euphemism Detection", "abstract": "  Detecting euphemisms is essential for content security on various social\nmedia platforms, but existing methods designed for detecting euphemisms are\nineffective in impromptu euphemisms. In this work, we make a first attempt to\nan exploration of impromptu euphemism detection and introduce the Impromptu\nCybercrime Euphemisms Detection (ICED) dataset. Moreover, we propose a\ndetection framework tailored to this problem, which employs context\naugmentation modeling and multi-round iterative training. Our detection\nframework mainly consists of a coarse-grained and a fine-grained classification\nmodel. The coarse-grained classification model removes most of the harmless\ncontent in the corpus to be detected. The fine-grained model, impromptu\neuphemisms detector, integrates context augmentation and multi-round iterations\ntraining to better predicts the actual meaning of a masked token. In addition,\nwe leverage ChatGPT to evaluate the mode's capability. Experimental results\ndemonstrate that our approach achieves a remarkable 76-fold improvement\ncompared to the previous state-of-the-art euphemism detector.\n", "date": "2024-12-02", "categories": []}
{"id": "2311.04916", "title": "Explainable Identification of Hate Speech towards Islam using Graph\n  Neural Networks", "abstract": "  Islamophobic language on online platforms fosters intolerance, making\ndetection and elimination crucial for promoting harmony. Traditional hate\nspeech detection models rely on NLP techniques like tokenization,\npart-of-speech tagging, and encoder-decoder models. However, Graph Neural\nNetworks (GNNs), with their ability to utilize relationships between data\npoints, offer more effective detection and greater explainability. In this\nwork, we represent speeches as nodes and connect them with edges based on their\ncontext and similarity to develop the graph. This study introduces a novel\nparadigm using GNNs to identify and explain hate speech towards Islam. Our\nmodel leverages GNNs to understand the context and patterns of hate speech by\nconnecting texts via pretrained NLP-generated word embeddings, achieving\nstate-of-the-art performance and enhancing detection accuracy while providing\nvaluable explanations. This highlights the potential of GNNs in combating\nonline hate speech and fostering a safer, more inclusive online environment.\n", "date": "2023-11-02", "categories": []}
{"id": "2412.00609", "title": "Exploration and Evaluation of Bias in Cyberbullying Detection with\n  Machine Learning", "abstract": "  It is well known that the usefulness of a machine learning model is due to\nits ability to generalize to unseen data. This study uses three popular\ncyberbullying datasets to explore the effects of data, how it's collected, and\nhow it's labeled, on the resulting machine learning models. The bias introduced\nfrom differing definitions of cyberbullying and from data collection is\ndiscussed in detail. An emphasis is made on the impact of dataset expansion\nmethods, which utilize current data points to fetch and label new ones.\nFurthermore, explicit testing is performed to evaluate the ability of a model\nto generalize to unseen datasets through cross-dataset evaluation. As\nhypothesized, the models have a significant drop in the Macro F1 Score, with an\naverage drop of 0.222. As such, this study effectively highlights the\nimportance of dataset curation and cross-dataset testing for creating models\nwith real-world applicability. The experiments and other code can be found at\nhttps://github.com/rootdrew27/cyberbullying-ml.\n", "date": "2024-11-30", "categories": []}
{"id": "2412.00681", "title": "MIMIC: Multimodal Islamophobic Meme Identification and Classification", "abstract": "  Anti-Muslim hate speech has emerged within memes, characterized by\ncontext-dependent and rhetorical messages using text and images that seemingly\nmimic humor but convey Islamophobic sentiments. This work presents a novel\ndataset and proposes a classifier based on the Vision-and-Language Transformer\n(ViLT) specifically tailored to identify anti-Muslim hate within memes by\nintegrating both visual and textual representations. Our model leverages joint\nmodal embeddings between meme images and incorporated text to capture nuanced\nIslamophobic narratives that are unique to meme culture, providing both high\ndetection accuracy and interoperability.\n", "date": "2024-12-01", "categories": []}
{"id": "2411.19017", "title": "A Survey on Automatic Online Hate Speech Detection in Low-Resource\n  Languages", "abstract": "  The expanding influence of social media platforms over the past decade has\nimpacted the way people communicate. The level of obscurity provided by social\nmedia and easy accessibility of the internet has facilitated the spread of hate\nspeech. The terms and expressions related to hate speech gets updated with\nchanging times which poses an obstacle to policy-makers and researchers in case\nof hate speech identification. With growing number of individuals using their\nnative languages to communicate with each other, hate speech in these\nlow-resource languages are also growing. Although, there is awareness about the\nEnglish-related approaches, much attention have not been provided to these\nlow-resource languages due to lack of datasets and online available data. This\narticle provides a detailed survey of hate speech detection in low-resource\nlanguages around the world with details of available datasets, features\nutilized and techniques used. This survey further discusses the prevailing\nsurveys, overlapping concepts related to hate speech, research challenges and\nopportunities.\n", "date": "2024-11-28", "categories": []}
{"id": "2410.10153", "title": "Diagnosing Hate Speech Classification: Where Do Humans and Machines\n  Disagree, and Why?", "abstract": "  This study uses the cosine similarity ratio, embedding regression, and manual\nre-annotation to diagnose hate speech classification. We begin by computing\ncosine similarity ratio on a dataset \"Measuring Hate Speech\" that contains\n135,556 annotated comments on social media. This way, we show a basic use of\ncosine similarity as a description of hate speech content. We then diagnose\nhate speech classification starting from understanding the inconsistency of\nhuman annotation from the dataset. Using embedding regression as a basic\ndiagnostic, we found that female annotators are more sensitive to racial slurs\nthat target the black population. We perform with a more complicated diagnostic\nby training a hate speech classifier using a SoTA pre-trained large language\nmodel, NV-Embed-v2, to convert texts to embeddings and run a logistic\nregression. This classifier achieves a testing accuracy of 94%. In diagnosing\nwhere machines disagree with human annotators, we found that machines make\nfewer mistakes than humans despite the fact that human annotations are treated\nas ground truth in the training set. Machines perform better in correctly\nlabeling long statements of facts, but perform worse in labeling short\ninstances of swear words. We hypothesize that this is due to model alignment -\nwhile curating models at their creation prevents the models from producing\nobvious hate speech, it also reduces the model's ability to detect such\ncontent.\n", "date": "2024-10-14", "categories": []}
{"id": "2411.17123", "title": "Advancing Content Moderation: Evaluating Large Language Models for\n  Detecting Sensitive Content Across Text, Images, and Videos", "abstract": "  The widespread dissemination of hate speech, harassment, harmful and sexual\ncontent, and violence across websites and media platforms presents substantial\nchallenges and provokes widespread concern among different sectors of society.\nGovernments, educators, and parents are often at odds with media platforms\nabout how to regulate, control, and limit the spread of such content.\nTechnologies for detecting and censoring the media contents are a key solution\nto addressing these challenges. Techniques from natural language processing and\ncomputer vision have been used widely to automatically identify and filter out\nsensitive content such as offensive languages, violence, nudity, and addiction\nin both text, images, and videos, enabling platforms to enforce content\npolicies at scale. However, existing methods still have limitations in\nachieving high detection accuracy with fewer false positives and false\nnegatives. Therefore, more sophisticated algorithms for understanding the\ncontext of both text and image may open rooms for improvement in content\ncensorship to build a more efficient censorship system. In this paper, we\nevaluate existing LLM-based content moderation solutions such as OpenAI\nmoderation model and Llama-Guard3 and study their capabilities to detect\nsensitive contents. Additionally, we explore recent LLMs such as GPT, Gemini,\nand Llama in identifying inappropriate contents across media outlets. Various\ntextual and visual datasets like X tweets, Amazon reviews, news articles, human\nphotos, cartoons, sketches, and violence videos have been utilized for\nevaluation and comparison. The results demonstrate that LLMs outperform\ntraditional techniques by achieving higher accuracy and lower false positive\nand false negative rates. This highlights the potential to integrate LLMs into\nwebsites, social media platforms, and video-sharing services for regulatory and\ncontent moderation purposes.\n", "date": "2024-11-26", "categories": []}
{"id": "2406.11109", "title": "Investigating Annotator Bias in Large Language Models for Hate Speech\n  Detection", "abstract": "  Data annotation, the practice of assigning descriptive labels to raw data, is\npivotal in optimizing the performance of machine learning models. However, it\nis a resource-intensive process susceptible to biases introduced by annotators.\nThe emergence of sophisticated Large Language Models (LLMs) presents a unique\nopportunity to modernize and streamline this complex procedure. While existing\nresearch extensively evaluates the efficacy of LLMs, as annotators, this paper\ndelves into the biases present in LLMs when annotating hate speech data. Our\nresearch contributes to understanding biases in four key categories: gender,\nrace, religion, and disability with four LLMs: GPT-3.5, GPT-4o, Llama-3.1 and\nGemma-2. Specifically targeting highly vulnerable groups within these\ncategories, we analyze annotator biases. Furthermore, we conduct a\ncomprehensive examination of potential factors contributing to these biases by\nscrutinizing the annotated data. We introduce our custom hate speech detection\ndataset, HateBiasNet, to conduct this research. Additionally, we perform the\nsame experiments on the ETHOS (Mollas et al. 2022) dataset also for comparative\nanalysis. This paper serves as a crucial resource, guiding researchers and\npractitioners in harnessing the potential of LLMs for data annotation, thereby\nfostering advancements in this critical field.\n", "date": "2024-06-17", "categories": []}
{"id": "2410.22587", "title": "Toxicity of the Commons: Curating Open-Source Pre-Training Data", "abstract": "  Open-source large language models are becoming increasingly available and\npopular among researchers and practitioners. While significant progress has\nbeen made on open-weight models, open training data is a practice yet to be\nadopted by the leading open-weight models creators. At the same time, there\nresearchers are working to make language models safer. We propose a data\ncuration pipeline to reduce harmful outputs by models trained on public domain\ndata. There are unique challenges to working with public domain data, as these\nsources differ from web text in both form and content. Many sources are\nhistorical documents and are the result of Optical Character Recognition (OCR).\nConsequently, current state-of-the-art approaches to toxicity filtering are\noften infeasible or inappropriate for open data models. In this paper, we\nintroduce a new fully open-source pipeline for open-data toxicity filtering.\nOur contributions are threefold. We create a custom training dataset,\nToxicCommons, which is composed of texts which have been classified across five\ndifferent dimensions (racial/origin-based, gender/sex-based, religious,\nability-based discrimination, and violence). We use this dataset to train a\ncustom classifier, Celadon, that can be used to detect toxic content in open\ndata more efficiently at a larger scale. Finally, we describe the balanced\napproach to content filtration that optimizes safety filtering with respect to\nthe filtered data available for training.\n", "date": "2024-10-29", "categories": []}
{"id": "2411.10480", "title": "Hateful Meme Detection through Context-Sensitive Prompting and\n  Fine-Grained Labeling", "abstract": "  The prevalence of multi-modal content on social media complicates automated\nmoderation strategies. This calls for an enhancement in multi-modal\nclassification and a deeper understanding of understated meanings in images and\nmemes. Although previous efforts have aimed at improving model performance\nthrough fine-tuning, few have explored an end-to-end optimization pipeline that\naccounts for modalities, prompting, labeling, and fine-tuning. In this study,\nwe propose an end-to-end conceptual framework for model optimization in complex\ntasks. Experiments support the efficacy of this traditional yet novel\nframework, achieving the highest accuracy and AUROC. Ablation experiments\ndemonstrate that isolated optimizations are not ineffective on their own.\n", "date": "2024-11-13", "categories": []}
{"id": "2311.07879", "title": "Toxicity Detection is NOT all you Need: Measuring the Gaps to Supporting\n  Volunteer Content Moderators", "abstract": "  Extensive efforts in automated approaches for content moderation have been\nfocused on developing models to identify toxic, offensive, and hateful content\nwith the aim of lightening the load for moderators. Yet, it remains uncertain\nwhether improvements on those tasks have truly addressed moderators' needs in\naccomplishing their work. In this paper, we surface gaps between past research\nefforts that have aimed to provide automation for aspects of content moderation\nand the needs of volunteer content moderators, regarding identifying violations\nof various moderation rules. To do so, we conduct a model review on Hugging\nFace to reveal the availability of models to cover various moderation rules and\nguidelines from three exemplar forums. We further put state-of-the-art LLMs to\nthe test, evaluating how well these models perform in flagging violations of\nplatform rules from one particular forum. Finally, we conduct a user survey\nstudy with volunteer moderators to gain insight into their perspectives on\nuseful moderation models. Overall, we observe a non-trivial gap, as missing\ndeveloped models and LLMs exhibit moderate to low performance on a significant\nportion of the rules. Moderators' reports provide guides for future work on\ndeveloping moderation assistant models.\n", "date": "2023-11-14", "categories": []}
{"id": "2411.06855", "title": "A Unified Multi-Task Learning Architecture for Hate Detection Leveraging\n  User-Based Information", "abstract": "  Hate speech, offensive language, aggression, racism, sexism, and other\nabusive language are common phenomena in social media. There is a need for\nArtificial Intelligence(AI)based intervention which can filter hate content at\nscale. Most existing hate speech detection solutions have utilized the features\nby treating each post as an isolated input instance for the classification.\nThis paper addresses this issue by introducing a unique model that improves\nhate speech identification for the English language by utilising intra-user and\ninter-user-based information. The experiment is conducted over single-task\nlearning (STL) and multi-task learning (MTL) paradigms that use deep neural\nnetworks, such as convolutional neural networks (CNN), gated recurrent unit\n(GRU), bidirectional encoder representations from the transformer (BERT), and A\nLite BERT (ALBERT). We use three benchmark datasets and conclude that combining\ncertain user features with textual features gives significant improvements in\nmacro-F1 and weighted-F1.\n", "date": "2024-11-11", "categories": []}
{"id": "2411.05819", "title": "Hierarchical Sentiment Analysis Framework for Hate Speech Detection:\n  Implementing Binary and Multiclass Classification Strategy", "abstract": "  A significant challenge in automating hate speech detection on social media\nis distinguishing hate speech from regular and offensive language. These\nidentify an essential category of content that web filters seek to remove. Only\nautomated methods can manage this volume of daily data. To solve this problem,\nthe community of Natural Language Processing is currently investigating\ndifferent ways of hate speech detection. In addition to those, previous\napproaches (e.g., Convolutional Neural Networks, multi-channel BERT models, and\nlexical detection) have always achieved low precision without carefully\ntreating other related tasks like sentiment analysis and emotion\nclassification. They still like to group all messages with specific words in\nthem as hate speech simply because those terms often appear alongside hateful\nrhetoric. In this research, our paper presented the hate speech text\nclassification system model drawn upon deep learning and machine learning. In\nthis paper, we propose a new multitask model integrated with shared emotional\nrepresentations to detect hate speech across the English language. The\nTransformer-based model we used from Hugging Face and sentiment analysis helped\nus prevent false positives. Conclusion. We conclude that utilizing sentiment\nanalysis and a Transformer-based trained model considerably improves hate\nspeech detection across multiple datasets.\n", "date": "2024-11-03", "categories": []}
{"id": "2411.05854", "title": "Harmful YouTube Video Detection: A Taxonomy of Online Harm and MLLMs as\n  Alternative Annotators", "abstract": "  Short video platforms, such as YouTube, Instagram, or TikTok, are used by\nbillions of users globally. These platforms expose users to harmful content,\nranging from clickbait or physical harms to misinformation or online hate. Yet,\ndetecting harmful videos remains challenging due to an inconsistent\nunderstanding of what constitutes harm and limited resources and mental tolls\ninvolved in human annotation. As such, this study advances measures and methods\nto detect harm in video content. First, we develop a comprehensive taxonomy for\nonline harm on video platforms, categorizing it into six categories:\nInformation, Hate and harassment, Addictive, Clickbait, Sexual, and Physical\nharms. Next, we establish multimodal large language models as reliable\nannotators of harmful videos. We analyze 19,422 YouTube videos using 14 image\nframes, 1 thumbnail, and text metadata, comparing the accuracy of crowdworkers\n(Mturk) and GPT-4-Turbo with domain expert annotations serving as the gold\nstandard. Our results demonstrate that GPT-4-Turbo outperforms crowdworkers in\nboth binary classification (harmful vs. harmless) and multi-label harm\ncategorization tasks. Methodologically, this study extends the application of\nLLMs to multi-label and multi-modal contexts beyond text annotation and binary\nclassification. Practically, our study contributes to online harm mitigation by\nguiding the definitions and identification of harmful content on video\nplatforms.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.06138", "title": "StopHC: A Harmful Content Detection and Mitigation Architecture for\n  Social Media Platforms", "abstract": "  The mental health of social media users has started more and more to be put\nat risk by harmful, hateful, and offensive content. In this paper, we propose\n\\textsc{StopHC}, a harmful content detection and mitigation architecture for\nsocial media platforms. Our aim with \\textsc{StopHC} is to create more secure\nonline environments. Our solution contains two modules, one that employs deep\nneural network architecture for harmful content detection, and one that uses a\nnetwork immunization algorithm to block toxic nodes and stop the spread of\nharmful content. The efficacy of our solution is demonstrated by experiments\nconducted on two real-world datasets.\n", "date": "2024-11-09", "categories": []}
{"id": "2411.06213", "title": "Incorporating Human Explanations for Robust Hate Speech Detection", "abstract": "  Given the black-box nature and complexity of large transformer language\nmodels (LM), concerns about generalizability and robustness present ethical\nimplications for domains such as hate speech (HS) detection. Using the content\nrich Social Bias Frames dataset, containing human-annotated stereotypes,\nintent, and targeted groups, we develop a three stage analysis to evaluate if\nLMs faithfully assess hate speech. First, we observe the need for modeling\ncontextually grounded stereotype intents to capture implicit semantic meaning.\nNext, we design a new task, Stereotype Intent Entailment (SIE), which\nencourages a model to contextually understand stereotype presence. Finally,\nthrough ablation tests and user studies, we find a SIE objective improves\ncontent understanding, but challenges remain in modeling implicit intent.\n", "date": "2024-11-09", "categories": []}
{"id": "2411.06477", "title": "VocalTweets: Investigating Social Media Offensive Language Among\n  Nigerian Musicians", "abstract": "  Musicians frequently use social media to express their opinions, but they\noften convey different messages in their music compared to their posts online.\nSome utilize these platforms to abuse their colleagues, while others use it to\nshow support for political candidates or engage in activism, as seen during the\n#EndSars protest. There are extensive research done on offensive language\ndetection on social media, the usage of offensive language by musicians has\nreceived limited attention. In this study, we introduce VocalTweets, a\ncode-switched and multilingual dataset comprising tweets from 12 prominent\nNigerian musicians, labeled with a binary classification method as Normal or\nOffensive. We trained a model using HuggingFace's base-Twitter-RoBERTa,\nachieving an F1 score of 74.5. Additionally, we conducted cross-corpus\nexperiments with the OLID dataset to evaluate the generalizability of our\ndataset.\n", "date": "2024-11-10", "categories": []}
{"id": "2411.06850", "title": "1-800-SHARED-TASKS @ NLU of Devanagari Script Languages: Detection of\n  Language, Hate Speech, and Targets using LLMs", "abstract": "  This paper presents a detailed system description of our entry for the\nCHiPSAL 2025 shared task, focusing on language detection, hate speech\nidentification, and target detection in Devanagari script languages. We\nexperimented with a combination of large language models and their ensembles,\nincluding MuRIL, IndicBERT, and Gemma-2, and leveraged unique techniques like\nfocal loss to address challenges in the natural understanding of Devanagari\nlanguages, such as multilingual processing and class imbalance. Our approach\nachieved competitive results across all tasks: F1 of 0.9980, 0.7652, and 0.6804\nfor Sub-tasks A, B, and C respectively. This work provides insights into the\neffectiveness of transformer models in tasks with domain-specific and\nlinguistic challenges, as well as areas for potential improvement in future\niterations.\n", "date": "2024-11-11", "categories": []}
{"id": "2411.05214", "title": "STAND-Guard: A Small Task-Adaptive Content Moderation Model", "abstract": "  Content moderation, the process of reviewing and monitoring the safety of\ngenerated content, is important for development of welcoming online platforms\nand responsible large language models. Content moderation contains various\ntasks, each with its unique requirements tailored to specific scenarios.\nTherefore, it is crucial to develop a model that can be easily adapted to novel\nor customized content moderation tasks accurately without extensive model\ntuning. This paper presents STAND-GUARD, a Small Task-Adaptive coNtent\nmoDeration model. The basic motivation is: by performing instruct tuning on\nvarious content moderation tasks, we can unleash the power of small language\nmodels (SLMs) on unseen (out-of-distribution) content moderation tasks. We also\ncarefully study the effects of training tasks and model size on the efficacy of\ncross-task fine-tuning mechanism. Experiments demonstrate STAND-Guard is\ncomparable to GPT-3.5-Turbo across over 40 public datasets, as well as\nproprietary datasets derived from real-world business scenarios. Remarkably,\nSTAND-Guard achieved nearly equivalent results to GPT-4-Turbo on unseen English\nbinary classification tasks\n", "date": "2024-11-07", "categories": []}
{"id": "2411.05383", "title": "Towards Low-Resource Harmful Meme Detection with LMM Agents", "abstract": "  The proliferation of Internet memes in the age of social media necessitates\neffective identification of harmful ones. Due to the dynamic nature of memes,\nexisting data-driven models may struggle in low-resource scenarios where only a\nfew labeled examples are available. In this paper, we propose an agency-driven\nframework for low-resource harmful meme detection, employing both outward and\ninward analysis with few-shot annotated samples. Inspired by the powerful\ncapacity of Large Multimodal Models (LMMs) on multimodal reasoning, we first\nretrieve relative memes with annotations to leverage label information as\nauxiliary signals for the LMM agent. Then, we elicit knowledge-revising\nbehavior within the LMM agent to derive well-generalized insights into meme\nharmfulness. By combining these strategies, our approach enables dialectical\nreasoning over intricate and implicit harm-indicative patterns. Extensive\nexperiments conducted on three meme datasets demonstrate that our proposed\napproach achieves superior performance than state-of-the-art methods on the\nlow-resource harmful meme detection task.\n", "date": "2024-11-08", "categories": []}
{"id": "2312.02401", "title": "Enhancing Content Moderation with Culturally-Aware Models", "abstract": "  Content moderation on a global scale must navigate a complex array of local\ncultural distinctions, which can hinder effective enforcement. While global\npolicies aim for consistency and broad applicability, they often miss the\nsubtleties of regional language interpretation, cultural beliefs, and local\nlegislation. This work introduces a flexible framework that enhances foundation\nlanguage models with cultural knowledge. Our approach involves fine-tuning\nencoder-decoder models on media-diet data to capture cultural nuances, and\napplies a continued training regime to effectively integrate these models into\na content moderation pipeline. We evaluate this framework in a case study of an\nonline podcast platform with content spanning various regions. The results show\nthat our culturally adapted models improve the accuracy of local violation\ndetection and offer explanations that align more closely with regional cultural\nnorms. Our findings reinforce the need for an adaptable content moderation\napproach that remains flexible in response to the diverse cultural landscapes\nit operates in and represents a step towards a more equitable and culturally\nsensitive framework for content moderation, demonstrating what is achievable in\nthis domain.\n", "date": "2023-12-05", "categories": []}
{"id": "2411.02587", "title": "A Big Data-empowered System for Real-time Detection of Regional\n  Discriminatory Comments on Vietnamese Social Media", "abstract": "  Regional discrimination is a persistent social issue in Vietnam. While\nexisting research has explored hate speech in the Vietnamese language, the\nspecific issue of regional discrimination remains under-addressed. Previous\nstudies primarily focused on model development without considering practical\nsystem implementation. In this work, we propose a task called Detection of\nRegional Discriminatory Comments on Vietnamese Social Media, leveraging the\npower of machine learning and transfer learning models. We have built the ViRDC\n(Vietnamese Regional Discrimination Comments) dataset, which contains comments\nfrom social media platforms, providing a valuable resource for further research\nand development. Our approach integrates streaming capabilities to process\nreal-time data from social media networks, ensuring the system's scalability\nand responsiveness. We developed the system on the Apache Spark framework to\nefficiently handle increasing data inputs during streaming. Our system offers a\ncomprehensive solution for the real-time detection of regional discrimination\nin Vietnam.\n", "date": "2024-10-29", "categories": []}
{"id": "2310.05779", "title": "Why Should This Article Be Deleted? Transparent Stance Detection in\n  Multilingual Wikipedia Editor Discussions", "abstract": "  The moderation of content on online platforms is usually non-transparent. On\nWikipedia, however, this discussion is carried out publicly and the editors are\nencouraged to use the content moderation policies as explanations for making\nmoderation decisions. Currently, only a few comments explicitly mention those\npolicies -- 20% of the English ones, but as few as 2% of the German and Turkish\ncomments. To aid in this process of understanding how content is moderated, we\nconstruct a novel multilingual dataset of Wikipedia editor discussions along\nwith their reasoning in three languages. The dataset contains the stances of\nthe editors (keep, delete, merge, comment), along with the stated reason, and a\ncontent moderation policy, for each edit decision. We demonstrate that stance\nand corresponding reason (policy) can be predicted jointly with a high degree\nof accuracy, adding transparency to the decision-making process. We release\nboth our joint prediction models and the multilingual content moderation\ndataset for further research on automated transparent content moderation.\n", "date": "2023-10-09", "categories": []}
{"id": "2410.20170", "title": "Cyberbullying or just Sarcasm? Unmasking Coordinated Networks on Reddit", "abstract": "  With the rapid growth of social media usage, a common trend has emerged where\nusers often make sarcastic comments on posts. While sarcasm can sometimes be\nharmless, it can blur the line with cyberbullying, especially when used in\nnegative or harmful contexts. This growing issue has been exacerbated by the\nanonymity and vast reach of the internet, making cyberbullying a significant\nconcern on platforms like Reddit. Our research focuses on distinguishing\ncyberbullying from sarcasm, particularly where online language nuances make it\ndifficult to discern harmful intent. This study proposes a framework using\nnatural language processing (NLP) and machine learning to differentiate between\nthe two, addressing the limitations of traditional sentiment analysis in\ndetecting nuanced behaviors. By analyzing a custom dataset scraped from Reddit,\nwe achieved a 95.15% accuracy in distinguishing harmful content from sarcasm.\nOur findings also reveal that teenagers and minority groups are particularly\nvulnerable to cyberbullying. Additionally, our research uncovers coordinated\ngraphs of groups involved in cyberbullying, identifying common patterns in\ntheir behavior. This research contributes to improving detection capabilities\nfor safer online communities.\n", "date": "2024-10-26", "categories": []}
{"id": "2410.23448", "title": "Venire: A Machine Learning-Guided Panel Review System for Community\n  Content Moderation", "abstract": "  Research into community content moderation often assumes that moderation\nteams govern with a single, unified voice. However, recent work has found that\nmoderators disagree with one another at modest, but concerning rates. The\nproblem is not the root disagreements themselves. Subjectivity in moderation is\nunavoidable, and there are clear benefits to including diverse perspectives\nwithin a moderation team. Instead, the crux of the issue is that, due to\nresource constraints, moderation decisions end up being made by individual\ndecision-makers. The result is decision-making that is inconsistent, which is\nfrustrating for community members. To address this, we develop Venire, an\nML-backed system for panel review on Reddit. Venire uses a machine learning\nmodel trained on log data to identify the cases where moderators are most\nlikely to disagree. Venire fast-tracks these cases for multi-person review.\nIdeally, Venire allows moderators to surface and resolve disagreements that\nwould have otherwise gone unnoticed. We conduct three studies through which we\ndesign and evaluate Venire: a set of formative interviews with moderators,\ntechnical evaluations on two datasets, and a think-aloud study in which\nmoderators used Venire to make decisions on real moderation cases.\nQuantitatively, we demonstrate that Venire is able to improve decision\nconsistency and surface latent disagreements. Qualitatively, we find that\nVenire helps moderators resolve difficult moderation cases more confidently.\nVenire represents a novel paradigm for human-AI content moderation, and shifts\nthe conversation from replacing human decision-making to supporting it.\n", "date": "2024-10-30", "categories": []}
{"id": "2311.08110", "title": "Improving Hateful Meme Detection through Retrieval-Guided Contrastive\n  Learning", "abstract": "  Hateful memes have emerged as a significant concern on the Internet.\nDetecting hateful memes requires the system to jointly understand the visual\nand textual modalities. Our investigation reveals that the embedding space of\nexisting CLIP-based systems lacks sensitivity to subtle differences in memes\nthat are vital for correct hatefulness classification. We propose constructing\na hatefulness-aware embedding space through retrieval-guided contrastive\ntraining. Our approach achieves state-of-the-art performance on the\nHatefulMemes dataset with an AUROC of 87.0, outperforming much larger\nfine-tuned large multimodal models. We demonstrate a retrieval-based hateful\nmemes detection system, which is capable of identifying hatefulness based on\ndata unseen in training. This allows developers to update the hateful memes\ndetection system by simply adding new examples without retraining, a desirable\nfeature for real services in the constantly evolving landscape of hateful memes\non the Internet.\n", "date": "2023-11-14", "categories": []}
{"id": "2401.16727", "title": "Recent Advances in Hate Speech Moderation: Multimodality and the Role of\n  Large Models", "abstract": "  In the evolving landscape of online communication, moderating hate speech\n(HS) presents an intricate challenge, compounded by the multimodal nature of\ndigital content. This comprehensive survey delves into the recent strides in HS\nmoderation, spotlighting the burgeoning role of large language models (LLMs)\nand large multimodal models (LMMs). Our exploration begins with a thorough\nanalysis of current literature, revealing the nuanced interplay between\ntextual, visual, and auditory elements in propagating HS. We uncover a notable\ntrend towards integrating these modalities, primarily due to the complexity and\nsubtlety with which HS is disseminated. A significant emphasis is placed on the\nadvances facilitated by LLMs and LMMs, which have begun to redefine the\nboundaries of detection and moderation capabilities. We identify existing gaps\nin research, particularly in the context of underrepresented languages and\ncultures, and the need for solutions to handle low-resource settings. The\nsurvey concludes with a forward-looking perspective, outlining potential\navenues for future research, including the exploration of novel AI\nmethodologies, the ethical governance of AI in moderation, and the development\nof more nuanced, context-aware systems. This comprehensive overview aims to\ncatalyze further research and foster a collaborative effort towards more\nsophisticated, responsible, and human-centric approaches to HS moderation in\nthe digital era. WARNING: This paper contains offensive examples.\n", "date": "2024-01-30", "categories": []}
{"id": "2410.21321", "title": "User-Aware Multilingual Abusive Content Detection in Social Media", "abstract": "  Despite growing efforts to halt distasteful content on social media,\nmultilingualism has added a new dimension to this problem. The scarcity of\nresources makes the challenge even greater when it comes to low-resource\nlanguages. This work focuses on providing a novel method for abusive content\ndetection in multiple low-resource Indic languages. Our observation indicates\nthat a post's tendency to attract abusive comments, as well as features such as\nuser history and social context, significantly aid in the detection of abusive\ncontent. The proposed method first learns social and text context features in\ntwo separate modules. The integrated representation from these modules is\nlearned and used for the final prediction. To evaluate the performance of our\nmethod against different classical and state-of-the-art methods, we have\nperformed extensive experiments on SCIDN and MACI datasets consisting of 1.5M\nand 665K multilingual comments, respectively. Our proposed method outperforms\nstate-of-the-art baseline methods with an average increase of 4.08% and 9.52%\nin F1-scores on SCIDN and MACI datasets, respectively.\n", "date": "2024-10-26", "categories": []}
{"id": "2309.11611", "title": "Hate speech detection in algerian dialect using deep learning", "abstract": "  With the proliferation of hate speech on social networks under different\nformats, such as abusive language, cyberbullying, and violence, etc., people\nhave experienced a significant increase in violence, putting them in\nuncomfortable situations and threats. Plenty of efforts have been dedicated in\nthe last few years to overcome this phenomenon to detect hate speech in\ndifferent structured languages like English, French, Arabic, and others.\nHowever, a reduced number of works deal with Arabic dialects like Tunisian,\nEgyptian, and Gulf, mainly the Algerian ones. To fill in the gap, we propose in\nthis work a complete approach for detecting hate speech on online Algerian\nmessages. Many deep learning architectures have been evaluated on the corpus we\ncreated from some Algerian social networks (Facebook, YouTube, and Twitter).\nThis corpus contains more than 13.5K documents in Algerian dialect written in\nArabic, labeled as hateful or non-hateful. Promising results are obtained,\nwhich show the efficiency of our approach.\n", "date": "2023-09-20", "categories": []}
{"id": "2410.19687", "title": "ProvocationProbe: Instigating Hate Speech Dataset from Twitter", "abstract": "  In the recent years online social media platforms has been flooded with\nhateful remarks such as racism, sexism, homophobia etc. As a result, there have\nbeen many measures taken by various social media platforms to mitigate the\nspread of hate-speech over the internet. One particular concept within the\ndomain of hate speech is instigating hate, which involves provoking hatred\nagainst a particular community, race, colour, gender, religion or ethnicity. In\nthis work, we introduce \\textit{ProvocationProbe} - a dataset designed to\nexplore what distinguishes instigating hate speech from general hate speech.\nFor this study, we collected around twenty thousand tweets from Twitter,\nencompassing a total of nine global controversies. These controversies span\nvarious themes including racism, politics, and religion. In this paper, i) we\npresent an annotated dataset after comprehensive examination of all the\ncontroversies, ii) we also highlight the difference between hate speech and\ninstigating hate speech by identifying distinguishing features, such as\ntargeted identity attacks and reasons for hate.\n", "date": "2024-10-25", "categories": []}
{"id": "2405.20755", "title": "Improving code-mixed hate detection by native sample mixing: A case\n  study for Hindi-English code-mixed scenario", "abstract": "  Hate detection has long been a challenging task for the NLP community. The\ntask becomes complex in a code-mixed environment because the models must\nunderstand the context and the hate expressed through language alteration.\nCompared to the monolingual setup, we see much less work on code-mixed hate as\nlarge-scale annotated hate corpora are unavailable for the study. To overcome\nthis bottleneck, we propose using native language hate samples (native language\nsamples/ native samples hereafter). We hypothesise that in the era of\nmultilingual language models (MLMs), hate in code-mixed settings can be\ndetected by majorly relying on the native language samples. Even though the NLP\nliterature reports the effectiveness of MLMs on hate detection in many\ncross-lingual settings, their extensive evaluation in a code-mixed scenario is\nyet to be done. This paper attempts to fill this gap through rigorous empirical\nexperiments. We considered the Hindi-English code-mixed setup as a case study\nas we have the linguistic expertise for the same. Some of the interesting\nobservations we got are: (i) adding native hate samples in the code-mixed\ntraining set, even in small quantity, improved the performance of MLMs for\ncode-mixed hate detection, (ii) MLMs trained with native samples alone observed\nto be detecting code-mixed hate to a large extent, (iii) the visualisation of\nattention scores revealed that, when native samples were included in training,\nMLMs could better focus on the hate emitting words in the code-mixed context,\nand (iv) finally, when hate is subjective or sarcastic, naively mixing native\nsamples doesn't help much to detect code-mixed hate. We will release the data\nand code repository to reproduce the reported results.\n", "date": "2024-05-31", "categories": []}
{"id": "2408.16857", "title": "Modeling offensive content detection for TikTok", "abstract": "  The advent of social media transformed interpersonal communication and\ninformation consumption processes. This digital landscape accommodates user\nintentions, also resulting in an increase of offensive language and harmful\nbehavior. Concurrently, social media platforms collect vast datasets comprising\nuser-generated content and behavioral information. These datasets are\ninstrumental for platforms deploying machine learning and data-driven\nstrategies, facilitating customer insights and countermeasures against social\nmanipulation mechanisms like disinformation and offensive content.\nNevertheless, the availability of such datasets, along with the application of\nvarious machine learning techniques, to researchers and practitioners, for\nspecific social media platforms regarding particular events, is limited. In\nparticular for TikTok, which offers unique tools for personalized content\ncreation and sharing, the existing body of knowledge would benefit from having\ndiverse comprehensive datasets and associated data analytics solutions on\noffensive content. While efforts from social media platforms, research, and\npractitioner communities are seen on this behalf, such content continues to\nproliferate. This translates to an essential need to make datasets publicly\navailable and build corresponding intelligent solutions. On this behalf, this\nresearch undertakes the collection and analysis of TikTok data containing\noffensive content, building a series of machine learning and deep learning\nmodels for offensive content detection. This is done aiming at answering the\nfollowing research question: \"How to develop a series of computational models\nto detect offensive content on TikTok?\". To this end, a Data Science\nmethodological approach is considered, 120.423 TikTok comments are collected,\nand on a balanced, binary classification approach, F1 score performance results\nof 0.863 is obtained.\n", "date": "2024-08-29", "categories": []}
{"id": "2409.15652", "title": "English offensive text detection using CNN based Bi-GRU model", "abstract": "  Over the years, the number of users of social media has increased\ndrastically. People frequently share their thoughts through social platforms,\nand this leads to an increase in hate content. In this virtual community,\nindividuals share their views, express their feelings, and post photos, videos,\nblogs, and more. Social networking sites like Facebook and Twitter provide\nplatforms to share vast amounts of content with a single click. However, these\nplatforms do not impose restrictions on the uploaded content, which may include\nabusive language and explicit images unsuitable for social media. To resolve\nthis issue, a new idea must be implemented to divide the inappropriate content.\nNumerous studies have been done to automate the process. In this paper, we\npropose a new Bi-GRU-CNN model to classify whether the text is offensive or\nnot. The combination of the Bi-GRU and CNN models outperforms the existing\nmodel.\n", "date": "2024-09-24", "categories": []}
{"id": "2410.12174", "title": "Exploring Large Language Models for Hate Speech Detection in Rioplatense\n  Spanish", "abstract": "  Hate speech detection deals with many language variants, slang, slurs,\nexpression modalities, and cultural nuances. This outlines the importance of\nworking with specific corpora, when addressing hate speech within the scope of\nNatural Language Processing, recently revolutionized by the irruption of Large\nLanguage Models. This work presents a brief analysis of the performance of\nlarge language models in the detection of Hate Speech for Rioplatense Spanish.\nWe performed classification experiments leveraging chain-of-thought reasoning\nwith ChatGPT 3.5, Mixtral, and Aya, comparing their results with those of a\nstate-of-the-art BERT classifier. These experiments outline that, even if large\nlanguage models show a lower precision compared to the fine-tuned BERT\nclassifier and, in some cases, they find hard-to-get slurs or colloquialisms,\nthey still are sensitive to highly nuanced cases (particularly,\nhomophobic/transphobic hate speech). We make our code and models publicly\navailable for future research.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.09220", "title": "M3Hop-CoT: Misogynous Meme Identification with Multimodal Multi-hop\n  Chain-of-Thought", "abstract": "  In recent years, there has been a significant rise in the phenomenon of hate\nagainst women on social media platforms, particularly through the use of\nmisogynous memes. These memes often target women with subtle and obscure cues,\nmaking their detection a challenging task for automated systems. Recently,\nLarge Language Models (LLMs) have shown promising results in reasoning using\nChain-of-Thought (CoT) prompting to generate the intermediate reasoning chains\nas the rationale to facilitate multimodal tasks, but often neglect cultural\ndiversity and key aspects like emotion and contextual knowledge hidden in the\nvisual modalities. To address this gap, we introduce a Multimodal Multi-hop CoT\n(M3Hop-CoT) framework for Misogynous meme identification, combining a\nCLIP-based classifier and a multimodal CoT module with\nentity-object-relationship integration. M3Hop-CoT employs a three-step\nmultimodal prompting principle to induce emotions, target awareness, and\ncontextual knowledge for meme analysis. Our empirical evaluation, including\nboth qualitative and quantitative analysis, validates the efficacy of the\nM3Hop-CoT framework on the SemEval-2022 Task 5 (MAMI task) dataset,\nhighlighting its strong performance in the macro-F1 score. Furthermore, we\nevaluate the model's generalizability by evaluating it on various benchmark\nmeme datasets, offering a thorough insight into the effectiveness of our\napproach across different datasets.\n", "date": "2024-10-11", "categories": []}
{"id": "2410.10332", "title": "Disentangling Hate Across Target Identities", "abstract": "  Hate speech (HS) classifiers do not perform equally well in detecting hateful\nexpressions towards different target identities. They also demonstrate\nsystematic biases in predicted hatefulness scores. Tapping on two recently\nproposed functionality test datasets for HS detection, we quantitatively\nanalyze the impact of different factors on HS prediction. Experiments on\npopular industrial and academic models demonstrate that HS detectors assign a\nhigher hatefulness score merely based on the mention of specific target\nidentities. Besides, models often confuse hatefulness and the polarity of\nemotions. This result is worrisome as the effort to build HS detectors might\nharm the vulnerable identity groups we wish to protect: posts expressing anger\nor disapproval of hate expressions might be flagged as hateful themselves. We\nalso carry out a study inspired by social psychology theory, which reveals that\nthe accuracy of hatefulness prediction correlates strongly with the intensity\nof the stereotype.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.07673", "title": "Multimodal Clickbait Detection by De-confounding Biases Using Causal\n  Representation Inference", "abstract": "  This paper focuses on detecting clickbait posts on the Web. These posts often\nuse eye-catching disinformation in mixed modalities to mislead users to click\nfor profit. That affects the user experience and thus would be blocked by\ncontent provider. To escape detection, malicious creators use tricks to add\nsome irrelevant non-bait content into bait posts, dressing them up as legal to\nfool the detector. This content often has biased relations with non-bait\nlabels, yet traditional detectors tend to make predictions based on simple\nco-occurrence rather than grasping inherent factors that lead to malicious\nbehavior. This spurious bias would easily cause misjudgments. To address this\nproblem, we propose a new debiased method based on causal inference. We first\nemploy a set of features in multiple modalities to characterize the posts.\nConsidering these features are often mixed up with unknown biases, we then\ndisentangle three kinds of latent factors from them, including the invariant\nfactor that indicates intrinsic bait intention; the causal factor which\nreflects deceptive patterns in a certain scenario, and non-causal noise. By\neliminating the noise that causes bias, we can use invariant and causal factors\nto build a robust model with good generalization ability. Experiments on three\npopular datasets show the effectiveness of our approach.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.08053", "title": "A Target-Aware Analysis of Data Augmentation for Hate Speech Detection", "abstract": "  Hate speech is one of the main threats posed by the widespread use of social\nnetworks, despite efforts to limit it. Although attention has been devoted to\nthis issue, the lack of datasets and case studies centered around scarcely\nrepresented phenomena, such as ableism or ageism, can lead to hate speech\ndetection systems that do not perform well on underrepresented identity groups.\nGiven the unpreceded capabilities of LLMs in producing high-quality data, we\ninvestigate the possibility of augmenting existing data with generative\nlanguage models, reducing target imbalance. We experiment with augmenting 1,000\nposts from the Measuring Hate Speech corpus, an English dataset annotated with\ntarget identity information, adding around 30,000 synthetic examples using both\nsimple data augmentation methods and different types of generative models,\ncomparing autoregressive and sequence-to-sequence approaches. We find\ntraditional DA methods to often be preferable to generative models, but the\ncombination of the two tends to lead to the best results. Indeed, for some hate\ncategories such as origin, religion, and disability, hate speech classification\nusing augmented data for training improves by more than 10% F1 over the no\naugmentation baseline. This work contributes to the development of systems for\nhate speech detection that are not only better performing but also fairer and\nmore inclusive towards targets that have been neglected so far.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.05287", "title": "Hate Speech Detection Using Cross-Platform Social Media Data In English\n  and German Language", "abstract": "  Hate speech has grown into a pervasive phenomenon, intensifying during times\nof crisis, elections, and social unrest. Multiple approaches have been\ndeveloped to detect hate speech using artificial intelligence, but a\ngeneralized model is yet unaccomplished. The challenge for hate speech\ndetection as text classification is the cost of obtaining high-quality training\ndata. This study focuses on detecting bilingual hate speech in YouTube comments\nand measuring the impact of using additional data from other platforms in the\nperformance of the classification model. We examine the value of additional\ntraining datasets from cross-platforms for improving the performance of\nclassification models. We also included factors such as content similarity,\ndefinition similarity, and common hate words to measure the impact of datasets\non performance. Our findings show that adding more similar datasets based on\ncontent similarity, hate words, and definitions improves the performance of\nclassification models. The best performance was obtained by combining datasets\nfrom YouTube comments, Twitter, and Gab with an F1-score of 0.74 and 0.68 for\nEnglish and German YouTube comments.\n", "date": "2024-10-02", "categories": []}
{"id": "2410.05600", "title": "Bridging Modalities: Enhancing Cross-Modality Hate Speech Detection with\n  Few-Shot In-Context Learning", "abstract": "  The widespread presence of hate speech on the internet, including formats\nsuch as text-based tweets and vision-language memes, poses a significant\nchallenge to digital platform safety. Recent research has developed detection\nmodels tailored to specific modalities; however, there is a notable gap in\ntransferring detection capabilities across different formats. This study\nconducts extensive experiments using few-shot in-context learning with large\nlanguage models to explore the transferability of hate speech detection between\nmodalities. Our findings demonstrate that text-based hate speech examples can\nsignificantly enhance the classification accuracy of vision-language hate\nspeech. Moreover, text-based demonstrations outperform vision-language\ndemonstrations in few-shot learning settings. These results highlight the\neffectiveness of cross-modality knowledge transfer and offer valuable insights\nfor improving hate speech detection systems.\n", "date": "2024-10-08", "categories": []}
{"id": "2403.11456", "title": "HateCOT: An Explanation-Enhanced Dataset for Generalizable Offensive\n  Speech Detection via Large Language Models", "abstract": "  The widespread use of social media necessitates reliable and efficient\ndetection of offensive content to mitigate harmful effects. Although\nsophisticated models perform well on individual datasets, they often fail to\ngeneralize due to varying definitions and labeling of \"offensive content.\" In\nthis paper, we introduce HateCOT, an English dataset with over 52,000 samples\nfrom diverse sources, featuring explanations generated by GPT-3.5Turbo and\ncurated by humans. We demonstrate that pretraining on HateCOT significantly\nenhances the performance of open-source Large Language Models on three\nbenchmark datasets for offensive content detection in both zero-shot and\nfew-shot settings, despite differences in domain and task. Additionally,\nHateCOT facilitates effective K-shot fine-tuning of LLMs with limited data and\nimproves the quality of their explanations, as confirmed by our human\nevaluation.\n", "date": "2024-03-18", "categories": []}
{"id": "2409.07246", "title": "Propaganda to Hate: A Multimodal Analysis of Arabic Memes with\n  Multi-Agent LLMs", "abstract": "  In the past decade, social media platforms have been used for information\ndissemination and consumption. While a major portion of the content is posted\nto promote citizen journalism and public awareness, some content is posted to\nmislead users. Among different content types such as text, images, and videos,\nmemes (text overlaid on images) are particularly prevalent and can serve as\npowerful vehicles for propaganda, hate, and humor. In the current literature,\nthere have been efforts to individually detect such content in memes. However,\nthe study of their intersection is very limited. In this study, we explore the\nintersection between propaganda and hate in memes using a multi-agent LLM-based\napproach. We extend the propagandistic meme dataset with coarse and\nfine-grained hate labels. Our finding suggests that there is an association\nbetween propaganda and hate in memes. We provide detailed experimental results\nthat can serve as a baseline for future studies. We will make the experimental\nresources publicly available to the community\n(https://github.com/firojalam/propaganda-and-hateful-memes).\n", "date": "2024-09-11", "categories": []}
{"id": "2410.03466", "title": "Is Safer Better? The Impact of Guardrails on the Argumentative Strength\n  of LLMs in Hate Speech Countering", "abstract": "  The potential effectiveness of counterspeech as a hate speech mitigation\nstrategy is attracting increasing interest in the NLG research community,\nparticularly towards the task of automatically producing it. However,\nautomatically generated responses often lack the argumentative richness which\ncharacterises expert-produced counterspeech. In this work, we focus on two\naspects of counterspeech generation to produce more cogent responses. First, by\ninvestigating the tension between helpfulness and harmlessness of LLMs, we test\nwhether the presence of safety guardrails hinders the quality of the\ngenerations. Secondly, we assess whether attacking a specific component of the\nhate speech results in a more effective argumentative strategy to fight online\nhate. By conducting an extensive human and automatic evaluation, we show how\nthe presence of safety guardrails can be detrimental also to a task that\ninherently aims at fostering positive social interactions. Moreover, our\nresults show that attacking a specific component of the hate speech, and in\nparticular its implicit negative stereotype and its hateful parts, leads to\nhigher-quality generations.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.03543", "title": "Re-examining Sexism and Misogyny Classification with Annotator Attitudes", "abstract": "  Gender-Based Violence (GBV) is an increasing problem online, but existing\ndatasets fail to capture the plurality of possible annotator perspectives or\nensure the representation of affected groups. We revisit two important stages\nin the moderation pipeline for GBV: (1) manual data labelling; and (2)\nautomated classification. For (1), we examine two datasets to investigate the\nrelationship between annotator identities and attitudes and the responses they\ngive to two GBV labelling tasks. To this end, we collect demographic and\nattitudinal information from crowd-sourced annotators using three validated\nsurveys from Social Psychology. We find that higher Right Wing Authoritarianism\nscores are associated with a higher propensity to label text as sexist, while\nfor Social Dominance Orientation and Neosexist Attitudes, higher scores are\nassociated with a negative tendency to do so. For (2), we conduct\nclassification experiments using Large Language Models and five prompting\nstrategies, including infusing prompts with annotator information. We find: (i)\nannotator attitudes affect the ability of classifiers to predict their labels;\n(ii) including attitudinal information can boost performance when we use\nwell-structured brief annotator descriptions; and (iii) models struggle to\nreflect the increased complexity and imbalanced classes of the new label sets.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.02378", "title": "Towards Comprehensive Detection of Chinese Harmful Memes", "abstract": "  This paper has been accepted in the NeurIPS 2024 D & B Track. Harmful memes\nhave proliferated on the Chinese Internet, while research on detecting Chinese\nharmful memes significantly lags behind due to the absence of reliable datasets\nand effective detectors. To this end, we focus on the comprehensive detection\nof Chinese harmful memes. We construct ToxiCN MM, the first Chinese harmful\nmeme dataset, which consists of 12,000 samples with fine-grained annotations\nfor various meme types. Additionally, we propose a baseline detector,\nMultimodal Knowledge Enhancement (MKE), incorporating contextual information of\nmeme content generated by the LLM to enhance the understanding of Chinese\nmemes. During the evaluation phase, we conduct extensive quantitative\nexperiments and qualitative analyses on multiple baselines, including LLMs and\nour MKE. The experimental results indicate that detecting Chinese harmful memes\nis challenging for existing models while demonstrating the effectiveness of\nMKE. The resources for this paper are available at\nhttps://github.com/DUT-lujunyu/ToxiCN_MM.\n", "date": "2024-10-03", "categories": []}
{"id": "2410.02657", "title": "Hate Personified: Investigating the role of LLMs in content moderation", "abstract": "  For subjective tasks such as hate detection, where people perceive hate\ndifferently, the Large Language Model's (LLM) ability to represent diverse\ngroups is unclear. By including additional context in prompts, we\ncomprehensively analyze LLM's sensitivity to geographical priming, persona\nattributes, and numerical information to assess how well the needs of various\ngroups are reflected. Our findings on two LLMs, five languages, and six\ndatasets reveal that mimicking persona-based attributes leads to annotation\nvariability. Meanwhile, incorporating geographical signals leads to better\nregional alignment. We also find that the LLMs are sensitive to numerical\nanchors, indicating the ability to leverage community-based flagging efforts\nand exposure to adversaries. Our work provides preliminary guidelines and\nhighlights the nuances of applying LLMs in culturally sensitive cases.\n", "date": "2024-10-03", "categories": []}
{"id": "2409.10341", "title": "Detecting Sexism in German Online Newspaper Comments with Open-Source\n  Text Embeddings (Team GDA, GermEval2024 Shared Task 1: GerMS-Detect, Subtasks\n  1 and 2, Closed Track)", "abstract": "  Sexism in online media comments is a pervasive challenge that often manifests\nsubtly, complicating moderation efforts as interpretations of what constitutes\nsexism can vary among individuals. We study monolingual and multilingual\nopen-source text embeddings to reliably detect sexism and misogyny in\nGerman-language online comments from an Austrian newspaper. We observed\nclassifiers trained on text embeddings to mimic closely the individual\njudgements of human annotators. Our method showed robust performance in the\nGermEval 2024 GerMS-Detect Subtask 1 challenge, achieving an average macro F1\nscore of 0.597 (4th place, as reported on Codabench). It also accurately\npredicted the distribution of human annotations in GerMS-Detect Subtask 2, with\nan average Jensen-Shannon distance of 0.301 (2nd place). The computational\nefficiency of our approach suggests potential for scalable applications across\nvarious languages and linguistic contexts.\n", "date": "2024-09-16", "categories": []}
{"id": "2409.19734", "title": "T2Vs Meet VLMs: A Scalable Multimodal Dataset for Visual Harmfulness\n  Recognition", "abstract": "  To address the risks of encountering inappropriate or harmful content,\nresearchers managed to incorporate several harmful contents datasets with\nmachine learning methods to detect harmful concepts. However, existing harmful\ndatasets are curated by the presence of a narrow range of harmful objects, and\nonly cover real harmful content sources. This hinders the generalizability of\nmethods based on such datasets, potentially leading to misjudgments. Therefore,\nwe propose a comprehensive harmful dataset, Visual Harmful Dataset 11K\n(VHD11K), consisting of 10,000 images and 1,000 videos, crawled from the\nInternet and generated by 4 generative models, across a total of 10 harmful\ncategories covering a full spectrum of harmful concepts with nontrivial\ndefinition. We also propose a novel annotation framework by formulating the\nannotation process as a multi-agent Visual Question Answering (VQA) task,\nhaving 3 different VLMs \"debate\" about whether the given image/video is\nharmful, and incorporating the in-context learning strategy in the debating\nprocess. Therefore, we can ensure that the VLMs consider the context of the\ngiven image/video and both sides of the arguments thoroughly before making\ndecisions, further reducing the likelihood of misjudgments in edge cases.\nEvaluation and experimental results demonstrate that (1) the great alignment\nbetween the annotation from our novel annotation framework and those from\nhuman, ensuring the reliability of VHD11K; (2) our full-spectrum harmful\ndataset successfully identifies the inability of existing harmful content\ndetection methods to detect extensive harmful contents and improves the\nperformance of existing harmfulness recognition methods; (3) VHD11K outperforms\nthe baseline dataset, SMID, as evidenced by the superior improvement in\nharmfulness recognition methods. The complete dataset and code can be found at\nhttps://github.com/nctu-eva-lab/VHD11K.\n", "date": "2024-09-29", "categories": []}
{"id": "2403.17146", "title": "Outcome-Constrained Large Language Models for Countering Hate Speech", "abstract": "  Automatic counterspeech generation methods have been developed to assist\nefforts in combating hate speech. Existing research focuses on generating\ncounterspeech with linguistic attributes such as being polite, informative, and\nintent-driven. However, the real impact of counterspeech in online environments\nis seldom considered. This study aims to develop methods for generating\ncounterspeech constrained by conversation outcomes and evaluate their\neffectiveness. We experiment with large language models (LLMs) to incorporate\ninto the text generation process two desired conversation outcomes: low\nconversation incivility and non-hateful hater reentry. Specifically, we\nexperiment with instruction prompts, LLM finetuning, and LLM reinforcement\nlearning (RL). Evaluation results show that our methods effectively steer the\ngeneration of counterspeech toward the desired outcomes. Our analyses, however,\nshow that there are differences in the quality and style depending on the\nmodel.\n", "date": "2024-03-25", "categories": []}
{"id": "2410.00361", "title": "PclGPT: A Large Language Model for Patronizing and Condescending\n  Language Detection", "abstract": "  Disclaimer: Samples in this paper may be harmful and cause discomfort!\n  Patronizing and condescending language (PCL) is a form of speech directed at\nvulnerable groups. As an essential branch of toxic language, this type of\nlanguage exacerbates conflicts and confrontations among Internet communities\nand detrimentally impacts disadvantaged groups. Traditional pre-trained\nlanguage models (PLMs) perform poorly in detecting PCL due to its implicit\ntoxicity traits like hypocrisy and false sympathy. With the rise of large\nlanguage models (LLMs), we can harness their rich emotional semantics to\nestablish a paradigm for exploring implicit toxicity. In this paper, we\nintroduce PclGPT, a comprehensive LLM benchmark designed specifically for PCL.\nWe collect, annotate, and integrate the Pcl-PT/SFT dataset, and then develop a\nbilingual PclGPT-EN/CN model group through a comprehensive pre-training and\nsupervised fine-tuning staircase process to facilitate implicit toxic\ndetection. Group detection results and fine-grained detection from PclGPT and\nother models reveal significant variations in the degree of bias in PCL towards\ndifferent vulnerable groups, necessitating increased societal attention to\nprotect them.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.00403", "title": "TikGuard: A Deep Learning Transformer-Based Solution for Detecting\n  Unsuitable TikTok Content for Kids", "abstract": "  The rise of short-form videos on platforms like TikTok has brought new\nchallenges in safeguarding young viewers from inappropriate content.\nTraditional moderation methods often fall short in handling the vast and\nrapidly changing landscape of user-generated videos, increasing the risk of\nchildren encountering harmful material. This paper introduces TikGuard, a\ntransformer-based deep learning approach aimed at detecting and flagging\ncontent unsuitable for children on TikTok. By using a specially curated\ndataset, TikHarm, and leveraging advanced video classification techniques,\nTikGuard achieves an accuracy of 86.7%, showing a notable improvement over\nexisting methods in similar contexts. While direct comparisons are limited by\nthe uniqueness of the TikHarm dataset, TikGuard's performance highlights its\npotential in enhancing content moderation, contributing to a safer online\nexperience for minors. This study underscores the effectiveness of transformer\nmodels in video classification and sets a foundation for future research in\nthis area.\n", "date": "2024-10-01", "categories": []}
{"id": "2409.20498", "title": "Enhancing Romanian Offensive Language Detection through Knowledge\n  Distillation, Multi-Task Learning, and Data Augmentation", "abstract": "  This paper highlights the significance of natural language processing (NLP)\nwithin artificial intelligence, underscoring its pivotal role in comprehending\nand modeling human language. Recent advancements in NLP, particularly in\nconversational bots, have garnered substantial attention and adoption among\ndevelopers. This paper explores advanced methodologies for attaining smaller\nand more efficient NLP models. Specifically, we employ three key approaches:\n(1) training a Transformer-based neural network to detect offensive language,\n(2) employing data augmentation and knowledge distillation techniques to\nincrease performance, and (3) incorporating multi-task learning with knowledge\ndistillation and teacher annealing using diverse datasets to enhance\nefficiency. The culmination of these methods has yielded demonstrably improved\noutcomes.\n", "date": "2024-09-30", "categories": []}
{"id": "2303.15110", "title": "Beyond Toxic: Toxicity Detection Datasets are Not Enough for Brand\n  Safety", "abstract": "  The rapid growth in user generated content on social media has resulted in a\nsignificant rise in demand for automated content moderation. Various methods\nand frameworks have been proposed for the tasks of hate speech detection and\ntoxic comment classification. In this work, we combine common datasets to\nextend these tasks to brand safety. Brand safety aims to protect commercial\nbranding by identifying contexts where advertisements should not appear and\ncovers not only toxicity, but also other potentially harmful content. As these\ndatasets contain different label sets, we approach the overall problem as a\nbinary classification task. We demonstrate the need for building brand safety\nspecific datasets via the application of common toxicity detection datasets to\na subset of brand safety and empirically analyze the effects of weighted\nsampling strategies in text classification.\n", "date": "2023-03-27", "categories": []}
{"id": "2409.17156", "title": "An Art-centric perspective on AI-based content moderation of nudity", "abstract": "  At a time when the influence of generative Artificial Intelligence on visual\narts is a highly debated topic, we raise the attention towards a more subtle\nphenomenon: the algorithmic censorship of artistic nudity online. We analyze\nthe performance of three \"Not-Safe-For-Work'' image classifiers on artistic\nnudity, and empirically uncover the existence of a gender and a stylistic bias,\nas well as evident technical limitations, especially when only considering\nvisual information. Hence, we propose a multi-modal zero-shot classification\napproach that improves artistic nudity classification. From our research, we\ndraw several implications that we hope will inform future research on this\ntopic.\n", "date": "2024-09-10", "categories": []}
{"id": "2409.17191", "title": "An Effective, Robust and Fairness-aware Hate Speech Detection Framework", "abstract": "  With the widespread online social networks, hate speeches are spreading\nfaster and causing more damage than ever before. Existing hate speech detection\nmethods have limitations in several aspects, such as handling data\ninsufficiency, estimating model uncertainty, improving robustness against\nmalicious attacks, and handling unintended bias (i.e., fairness). There is an\nurgent need for accurate, robust, and fair hate speech classification in online\nsocial networks. To bridge the gap, we design a data-augmented, fairness\naddressed, and uncertainty estimated novel framework. As parts of the\nframework, we propose Bidirectional Quaternion-Quasi-LSTM layers to balance\neffectiveness and efficiency. To build a generalized model, we combine five\ndatasets collected from three platforms. Experiment results show that our model\noutperforms eight state-of-the-art methods under both no attack scenario and\nvarious attack scenarios, indicating the effectiveness and robustness of our\nmodel. We share our code along with combined dataset for better future research\n", "date": "2024-09-25", "categories": []}
{"id": "2409.17467", "title": "What is the social benefit of hate speech detection research? A\n  Systematic Review", "abstract": "  While NLP research into hate speech detection has grown exponentially in the\nlast three decades, there has been minimal uptake or engagement from policy\nmakers and non-profit organisations. We argue the absence of ethical frameworks\nhave contributed to this rift between current practice and best practice. By\nadopting appropriate ethical frameworks, NLP researchers may enable the social\nimpact potential of hate speech research. This position paper is informed by\nreviewing forty-eight hate speech detection systems associated with\nthirty-seven publications from different venues.\n", "date": "2024-09-26", "categories": []}
{"id": "2409.16673", "title": "SWE2: SubWord Enriched and Significant Word Emphasized Framework for\n  Hate Speech Detection", "abstract": "  Hate speech detection on online social networks has become one of the\nemerging hot topics in recent years. With the broad spread and fast propagation\nspeed across online social networks, hate speech makes significant impacts on\nsociety by increasing prejudice and hurting people. Therefore, there are\naroused attention and concern from both industry and academia. In this paper,\nwe address the hate speech problem and propose a novel hate speech detection\nframework called SWE2, which only relies on the content of messages and\nautomatically identifies hate speech. In particular, our framework exploits\nboth word-level semantic information and sub-word knowledge. It is intuitively\npersuasive and also practically performs well under a situation with/without\ncharacter-level adversarial attack. Experimental results show that our proposed\nmodel achieves 0.975 accuracy and 0.953 macro F1, outperforming 7\nstate-of-the-art baselines under no adversarial attack. Our model robustly and\nsignificantly performed well under extreme adversarial attack (manipulation of\n50% messages), achieving 0.967 accuracy and 0.934 macro F1.\n", "date": "2024-09-25", "categories": []}
{"id": "2409.15623", "title": "Safe Guard: an LLM-agent for Real-time Voice-based Hate Speech Detection\n  in Social Virtual Reality", "abstract": "  In this paper, we present Safe Guard, an LLM-agent for the detection of hate\nspeech in voice-based interactions in social VR (VRChat). Our system leverages\nOpen AI GPT and audio feature extraction for real-time voice interactions. We\ncontribute a system design and evaluation of the system that demonstrates the\ncapability of our approach in detecting hate speech, and reducing false\npositives compared to currently available approaches. Our results indicate the\npotential of LLM-based agents in creating safer virtual environments and set\nthe groundwork for further advancements in LLM-driven moderation approaches.\n", "date": "2024-09-23", "categories": []}
{"id": "2403.10829", "title": "Deciphering Hate: Identifying Hateful Memes and Their Targets", "abstract": "  Internet memes have become a powerful means for individuals to express\nemotions, thoughts, and perspectives on social media. While often considered as\na source of humor and entertainment, memes can also disseminate hateful content\ntargeting individuals or communities. Most existing research focuses on the\nnegative aspects of memes in high-resource languages, overlooking the\ndistinctive challenges associated with low-resource languages like Bengali\n(also known as Bangla). Furthermore, while previous work on Bengali memes has\nfocused on detecting hateful memes, there has been no work on detecting their\ntargeted entities. To bridge this gap and facilitate research in this arena, we\nintroduce a novel multimodal dataset for Bengali, BHM (Bengali Hateful Memes).\nThe dataset consists of 7,148 memes with Bengali as well as code-mixed\ncaptions, tailored for two tasks: (i) detecting hateful memes, and (ii)\ndetecting the social entities they target (i.e., Individual, Organization,\nCommunity, and Society). To solve these tasks, we propose DORA (Dual cO\nattention fRAmework), a multimodal deep neural network that systematically\nextracts the significant modality features from the memes and jointly evaluates\nthem with the modality-specific features to understand the context better. Our\nexperiments show that DORA is generalizable on other low-resource hateful meme\ndatasets and outperforms several state-of-the-art rivaling baselines.\n", "date": "2024-03-16", "categories": []}
{"id": "2409.14464", "title": "AggregHate: An Efficient Aggregative Approach for the Detection of\n  Hatemongers on Social Platforms", "abstract": "  Automatic detection of online hate speech serves as a crucial step in the\ndetoxification of the online discourse. Moreover, accurate classification can\npromote a better understanding of the proliferation of hate as a social\nphenomenon. While most prior work focus on the detection of hateful utterances,\nwe argue that focusing on the user level is as important, albeit challenging.\nIn this paper we consider a multimodal aggregative approach for the detection\nof hate-mongers, taking into account the potentially hateful texts, user\nactivity, and the user network. We evaluate our methods on three unique\ndatasets X (Twitter), Gab, and Parler showing that a processing a user's texts\nin her social context significantly improves the detection of hate mongers,\ncompared to previously used text and graph-based methods. Our method can be\nthen used to improve the classification of coded messages, dog-whistling, and\nracial gas-lighting, as well as inform intervention measures. Moreover, our\napproach is highly efficient even for very large datasets and networks.\n", "date": "2024-09-22", "categories": []}
{"id": "2409.13557", "title": "Trustworthy Hate Speech Detection Through Visual Augmentation", "abstract": "  The surge of hate speech on social media platforms poses a significant\nchallenge, with hate speech detection~(HSD) becoming increasingly critical.\nCurrent HSD methods focus on enriching contextual information to enhance\ndetection performance, but they overlook the inherent uncertainty of hate\nspeech. We propose a novel HSD method, named trustworthy hate speech detection\nmethod through visual augmentation (TrusV-HSD), which enhances semantic\ninformation through integration with diffused visual images and mitigates\nuncertainty with trustworthy loss. TrusV-HSD learns semantic representations by\neffectively extracting trustworthy information through multi-modal connections\nwithout paired data. Our experiments on public HSD datasets demonstrate the\neffectiveness of TrusV-HSD, showing remarkable improvements over conventional\nmethods.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.12658", "title": "Exploring the topics, sentiments and hate speech in the Spanish\n  information environment", "abstract": "  In the digital era, the internet and social media have transformed\ncommunication but have also facilitated the spread of hate speech and\ndisinformation, leading to radicalization, polarization, and toxicity. This is\nespecially concerning for media outlets due to their significant role in\nshaping public discourse. This study examines the topics, sentiments, and hate\nprevalence in 337,807 response messages (website comments and tweets) to news\nfrom five Spanish media outlets (La Vanguardia, ABC, El Pa\\'is, El Mundo, and\n20 Minutos) in January 2021. These public reactions were originally labeled as\ndistinct types of hate by experts following an original procedure, and they are\nnow classified into three sentiment values (negative, neutral, or positive) and\nmain topics. The BERTopic unsupervised framework was used to extract 81 topics,\nmanually named with the help of Large Language Models (LLMs) and grouped into\nnine primary categories.\n  Results show social issues (22.22%), expressions and slang (20.35%), and\npolitical issues (11.80%) as the most discussed. Content is mainly negative\n(62.7%) and neutral (28.57%), with low positivity (8.73%). Toxic narratives\nrelate to conversation expressions, gender, feminism, and COVID-19. Despite low\nlevels of hate speech (3.98%), the study confirms high toxicity in online\nresponses to social and political topics.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.05136", "title": "MHS-STMA: Multimodal Hate Speech Detection via Scalable\n  Transformer-Based Multilevel Attention Framework", "abstract": "  Social media has a significant impact on people's lives. Hate speech on\nsocial media has emerged as one of society's most serious issues in recent\nyears. Text and pictures are two forms of multimodal data that are distributed\nwithin articles. Unimodal analysis has been the primary emphasis of earlier\napproaches. Additionally, when doing multimodal analysis, researchers neglect\nto preserve the distinctive qualities associated with each modality. To address\nthese shortcomings, the present article suggests a scalable architecture for\nmultimodal hate content detection called transformer-based multilevel attention\n(STMA). This architecture consists of three main parts: a combined\nattention-based deep learning mechanism, a vision attention-mechanism encoder,\nand a caption attention-mechanism encoder. To identify hate content, each\ncomponent uses various attention processes and handles multimodal data in a\nunique way. Several studies employing multiple assessment criteria on three\nhate speech datasets such as Hateful memes, MultiOff, and MMHS150K, validate\nthe suggested architecture's efficacy. The outcomes demonstrate that on all\nthree datasets, the suggested strategy performs better than the baseline\napproaches.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.08963", "title": "Safeguarding Decentralized Social Media: LLM Agents for Automating\n  Community Rule Compliance", "abstract": "  Ensuring content compliance with community guidelines is crucial for\nmaintaining healthy online social environments. However, traditional\nhuman-based compliance checking struggles with scaling due to the increasing\nvolume of user-generated content and a limited number of moderators. Recent\nadvancements in Natural Language Understanding demonstrated by Large Language\nModels unlock new opportunities for automated content compliance verification.\nThis work evaluates six AI-agents built on Open-LLMs for automated rule\ncompliance checking in Decentralized Social Networks, a challenging environment\ndue to heterogeneous community scopes and rules. Analyzing over 50,000 posts\nfrom hundreds of Mastodon servers, we find that AI-agents effectively detect\nnon-compliant content, grasp linguistic subtleties, and adapt to diverse\ncommunity contexts. Most agents also show high inter-rater reliability and\nconsistency in score justification and suggestions for compliance. Human-based\nevaluation with domain experts confirmed the agents' reliability and\nusefulness, rendering them promising tools for semi-automated or\nhuman-in-the-loop content moderation systems.\n", "date": "2024-09-13", "categories": []}
{"id": "2409.05134", "title": "Hate Content Detection via Novel Pre-Processing Sequencing and Ensemble\n  Methods", "abstract": "  Social media, particularly Twitter, has seen a significant increase in\nincidents like trolling and hate speech. Thus, identifying hate speech is the\nneed of the hour. This paper introduces a computational framework to curb the\nhate content on the web. Specifically, this study presents an exhaustive study\nof pre-processing approaches by studying the impact of changing the sequence of\ntext pre-processing operations for the identification of hate content. The\nbest-performing pre-processing sequence, when implemented with popular\nclassification approaches like Support Vector Machine, Random Forest, Decision\nTree, Logistic Regression and K-Neighbor provides a considerable boost in\nperformance. Additionally, the best pre-processing sequence is used in\nconjunction with different ensemble methods, such as bagging, boosting and\nstacking to improve the performance further. Three publicly available benchmark\ndatasets (WZ-LS, DT, and FOUNTA), were used to evaluate the proposed approach\nfor hate speech identification. The proposed approach achieves a maximum\naccuracy of 95.14% highlighting the effectiveness of the unique pre-processing\napproach along with an ensemble classifier.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.03939", "title": "Experimentation in Content Moderation using RWKV", "abstract": "  This paper investigates the RWKV model's efficacy in content moderation\nthrough targeted experimentation. We introduce a novel dataset specifically\ndesigned for distillation into smaller models, enhancing content moderation\npractices. This comprehensive dataset encompasses images, videos, sounds, and\ntext data that present societal challenges. Leveraging advanced Large Language\nModels (LLMs), we generated an extensive set of responses -- 558,958 for text\nand 83,625 for images -- to train and refine content moderation systems. Our\ncore experimentation involved fine-tuning the RWKV model, capitalizing on its\nCPU-efficient architecture to address large-scale content moderation tasks. By\nhighlighting the dataset's potential for knowledge distillation, this study not\nonly demonstrates RWKV's capability in improving the accuracy and efficiency of\ncontent moderation systems but also paves the way for developing more compact,\nresource-efficient models in this domain. Datasets and models can be found in\nHuggingFace: https://huggingface.co/modrwkv\n", "date": "2024-09-05", "categories": []}
{"id": "2409.02114", "title": "Tiny-Toxic-Detector: A compact transformer-based model for toxic content\n  detection", "abstract": "  This paper presents Tiny-toxic-detector, a compact transformer-based model\ndesigned for toxic content detection. Despite having only 2.1 million\nparameters, Tiny-toxic-detector achieves competitive performance on benchmark\ndatasets, with 90.97% accuracy on ToxiGen and 86.98% accuracy on the Jigsaw\ndataset, rivaling models over 50 times its size. This efficiency enables\ndeployment in resource-constrained environments, addressing the need for\neffective content moderation tools that balance performance with computational\nefficiency. The model architecture features 4 transformer encoder layers, each\nwith 2 attention heads, an embedding dimension of 64, and a feedforward\ndimension of 128. Trained on both public and private datasets,\nTiny-toxic-detector demonstrates the potential of efficient, task-specific\nmodels for addressing online toxicity. The paper covers the model architecture,\ntraining process, performance benchmarks, and limitations, underscoring its\nsuitability for applications such as social media monitoring and content\nmoderation. By achieving results comparable to much larger models while\nsignificantly reducing computational demands, Tiny-toxic-detector represents\nprogress toward more sustainable and scalable AI-driven content moderation\nsolutions.\n", "date": "2024-08-29", "categories": []}
{"id": "2409.00940", "title": "Large Language Models for Automatic Detection of Sensitive Topics", "abstract": "  Sensitive information detection is crucial in content moderation to maintain\nsafe online communities. Assisting in this traditionally manual process could\nrelieve human moderators from overwhelming and tedious tasks, allowing them to\nfocus solely on flagged content that may pose potential risks. Rapidly\nadvancing large language models (LLMs) are known for their capability to\nunderstand and process natural language and so present a potential solution to\nsupport this process. This study explores the capabilities of five LLMs for\ndetecting sensitive messages in the mental well-being domain within two online\ndatasets and assesses their performance in terms of accuracy, precision,\nrecall, F1 scores, and consistency. Our findings indicate that LLMs have the\npotential to be integrated into the moderation workflow as a convenient and\nprecise detection tool. The best-performing model, GPT-4o, achieved an average\naccuracy of 99.5\\% and an F1-score of 0.99. We discuss the advantages and\npotential challenges of using LLMs in the moderation workflow and suggest that\nfuture research should address the ethical considerations of utilising this\ntechnology.\n", "date": "2024-09-02", "categories": []}
{"id": "2408.16749", "title": "Assessing Large Language Models for Online Extremism Research:\n  Identification, Explanation, and New Knowledge", "abstract": "  The United States has experienced a significant increase in violent\nextremism, prompting the need for automated tools to detect and limit the\nspread of extremist ideology online. This study evaluates the performance of\nBidirectional Encoder Representations from Transformers (BERT) and Generative\nPre-Trained Transformers (GPT) in detecting and classifying online domestic\nextremist posts. We collected social media posts containing \"far-right\" and\n\"far-left\" ideological keywords and manually labeled them as extremist or\nnon-extremist. Extremist posts were further classified into one or more of five\ncontributing elements of extremism based on a working definitional framework.\nThe BERT model's performance was evaluated based on training data size and\nknowledge transfer between categories. We also compared the performance of GPT\n3.5 and GPT 4 models using different prompts: na\\\"ive, layperson-definition,\nrole-playing, and professional-definition. Results showed that the best\nperforming GPT models outperformed the best performing BERT models, with more\ndetailed prompts generally yielding better results. However, overly complex\nprompts may impair performance. Different versions of GPT have unique\nsensitives to what they consider extremist. GPT 3.5 performed better at\nclassifying far-left extremist posts, while GPT 4 performed better at\nclassifying far-right extremist posts. Large language models, represented by\nGPT models, hold significant potential for online extremism classification\ntasks, surpassing traditional BERT models in a zero-shot setting. Future\nresearch should explore human-computer interactions in optimizing GPT models\nfor extremist detection and classification tasks to develop more efficient\n(e.g., quicker, less effort) and effective (e.g., fewer errors or mistakes)\nmethods for identifying extremist content.\n", "date": "2024-08-29", "categories": []}
{"id": "2408.12035", "title": "Let Community Rules Be Reflected in Online Content Moderation", "abstract": "  Content moderation is a widely used strategy to prevent the dissemination of\nirregular information on social media platforms. Despite extensive research on\ndeveloping automated models to support decision-making in content moderation,\nthere remains a notable scarcity of studies that integrate the rules of online\ncommunities into content moderation. This study addresses this gap by proposing\na community rule-based content moderation framework that directly integrates\ncommunity rules into the moderation of user-generated content. Our experiment\nresults with datasets collected from two domains demonstrate the superior\nperformance of models based on the framework to baseline models across all\nevaluation metrics. In particular, incorporating community rules substantially\nenhances model performance in content moderation. The findings of this research\nhave significant research and practical implications for improving the\neffectiveness and generalizability of content moderation models in online\ncommunities.\n", "date": "2024-08-21", "categories": []}
{"id": "2310.18330", "title": "Towards Detecting Contextual Real-Time Toxicity for In-Game Chat", "abstract": "  Real-time toxicity detection in online environments poses a significant\nchallenge, due to the increasing prevalence of social media and gaming\nplatforms. We introduce ToxBuster, a simple and scalable model that reliably\ndetects toxic content in real-time for a line of chat by including chat history\nand metadata. ToxBuster consistently outperforms conventional toxicity models\nacross popular multiplayer games, including Rainbow Six Siege, For Honor, and\nDOTA 2. We conduct an ablation study to assess the importance of each model\ncomponent and explore ToxBuster's transferability across the datasets.\nFurthermore, we showcase ToxBuster's efficacy in post-game moderation,\nsuccessfully flagging 82.1% of chat-reported players at a precision level of\n90.0%. Additionally, we show how an additional 6% of unreported toxic players\ncan be proactively moderated.\n", "date": "2023-10-20", "categories": []}
{"id": "2408.07479", "title": "A Study on Bias Detection and Classification in Natural Language\n  Processing", "abstract": "  Human biases have been shown to influence the performance of models and\nalgorithms in various fields, including Natural Language Processing. While the\nstudy of this phenomenon is garnering focus in recent years, the available\nresources are still relatively scarce, often focusing on different forms or\nmanifestations of biases. The aim of our work is twofold: 1) gather\npublicly-available datasets and determine how to better combine them to\neffectively train models in the task of hate speech detection and\nclassification; 2) analyse the main issues with these datasets, such as\nscarcity, skewed resources, and reliance on non-persistent data. We discuss\nthese issues in tandem with the development of our experiments, in which we\nshow that the combinations of different datasets greatly impact the models'\nperformance.\n", "date": "2024-08-14", "categories": []}
{"id": "2408.03468", "title": "MultiHateClip: A Multilingual Benchmark Dataset for Hateful Video\n  Detection on YouTube and Bilibili", "abstract": "  Hate speech is a pressing issue in modern society, with significant effects\nboth online and offline. Recent research in hate speech detection has primarily\ncentered on text-based media, largely overlooking multimodal content such as\nvideos. Existing studies on hateful video datasets have predominantly focused\non English content within a Western context and have been limited to binary\nlabels (hateful or non-hateful), lacking detailed contextual information. This\nstudy presents MultiHateClip1 , an novel multilingual dataset created through\nhate lexicons and human annotation. It aims to enhance the detection of hateful\nvideos on platforms such as YouTube and Bilibili, including content in both\nEnglish and Chinese languages. Comprising 2,000 videos annotated for\nhatefulness, offensiveness, and normalcy, this dataset provides a\ncross-cultural perspective on gender-based hate speech. Through a detailed\nexamination of human annotation results, we discuss the differences between\nChinese and English hateful videos and underscore the importance of different\nmodalities in hateful and offensive video analysis. Evaluations of\nstate-of-the-art video classification models, such as VLM, GPT-4V and Qwen-VL,\non MultiHateClip highlight the existing challenges in accurately distinguishing\nbetween hateful and offensive content and the urgent need for models that are\nboth multimodally and culturally nuanced. MultiHateClip represents a\nfoundational advance in enhancing hateful video detection by underscoring the\nnecessity of a multimodal and culturally sensitive approach in combating online\nhate speech.\n", "date": "2024-07-28", "categories": []}
{"id": "2408.05655", "title": "WiDe-analysis: Enabling One-click Content Moderation Analysis on\n  Wikipedia's Articles for Deletion", "abstract": "  Content moderation in online platforms is crucial for ensuring activity\ntherein adheres to existing policies, especially as these platforms grow. NLP\nresearch in this area has typically focused on automating some part of it given\nthat it is not feasible to monitor all active discussions effectively. Past\nworks have focused on revealing deletion patterns with like sentiment analysis,\nor on developing platform-specific models such as Wikipedia policy or stance\ndetectors. Unsurprisingly, however, this valuable body of work is rather\nscattered, with little to no agreement with regards to e.g., the deletion\ndiscussions corpora used for training or the number of stance labels. Moreover,\nwhile efforts have been made to connect stance with rationales (e.g., to ground\na deletion decision on the relevant policy), there is little explanability work\nbeyond that. In this paper, we introduce a suite of experiments on Wikipedia\ndeletion discussions and wide-analyis (Wikipedia Deletion Analysis), a Python\npackage aimed at providing one click analysis to content moderation\ndiscussions. We release all assets associated with wide-analysis, including\ndata, models and the Python package, and a HuggingFace space with the goal to\naccelerate research on automating content moderation in Wikipedia and beyond.\n", "date": "2024-08-10", "categories": []}
{"id": "2408.06065", "title": "An Investigation Into Explainable Audio Hate Speech Detection", "abstract": "  Research on hate speech has predominantly revolved around detection and\ninterpretation from textual inputs, leaving verbal content largely unexplored.\nWhile there has been limited exploration into hate speech detection within\nverbal acoustic speech inputs, the aspect of interpretability has been\noverlooked. Therefore, we introduce a new task of explainable audio hate speech\ndetection. Specifically, we aim to identify the precise time intervals,\nreferred to as audio frame-level rationales, which serve as evidence for hate\nspeech classification. Towards this end, we propose two different approaches:\ncascading and End-to-End (E2E). The cascading approach initially converts audio\nto transcripts, identifies hate speech within these transcripts, and\nsubsequently locates the corresponding audio time frames. Conversely, the E2E\napproach processes audio utterances directly, which allows it to pinpoint hate\nspeech within specific time frames. Additionally, due to the lack of\nexplainable audio hate speech datasets that include audio frame-level\nrationales, we curated a synthetic audio dataset to train our models. We\nfurther validated these models on actual human speech utterances and found that\nthe E2E approach outperforms the cascading method in terms of the audio frame\nIntersection over Union (IoU) metric. Furthermore, we observed that including\nframe-level rationales significantly enhances hate speech detection accuracy\nfor the E2E approach.\n  \\textbf{Disclaimer} The reader may encounter content of an offensive or\nhateful nature. However, given the nature of the work, this cannot be avoided.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.03849", "title": "Hate Speech Detection and Classification in Amharic Text with Deep\n  Learning", "abstract": "  Hate speech is a growing problem on social media. It can seriously impact\nsociety, especially in countries like Ethiopia, where it can trigger conflicts\namong diverse ethnic and religious groups. While hate speech detection in\nresource rich languages are progressing, for low resource languages such as\nAmharic are lacking. To address this gap, we develop Amharic hate speech data\nand SBi-LSTM deep learning model that can detect and classify text into four\ncategories of hate speech: racial, religious, gender, and non-hate speech. We\nhave annotated 5k Amharic social media post and comment data into four\ncategories. The data is annotated using a custom annotation tool by a total of\n100 native Amharic speakers. The model achieves a 94.8 F1-score performance.\nFuture improvements will include expanding the dataset and develop state-of-the\nart models.\n  Keywords: Amharic hate speech detection, classification, Amharic dataset,\nDeep Learning, SBi-LSTM\n", "date": "2024-08-07", "categories": []}
{"id": "2407.21772", "title": "ShieldGemma: Generative AI Content Moderation Based on Gemma", "abstract": "  We present ShieldGemma, a comprehensive suite of LLM-based safety content\nmoderation models built upon Gemma2. These models provide robust,\nstate-of-the-art predictions of safety risks across key harm types (sexually\nexplicit, dangerous content, harassment, hate speech) in both user input and\nLLM-generated output. By evaluating on both public and internal benchmarks, we\ndemonstrate superior performance compared to existing models, such as Llama\nGuard (+10.8\\% AU-PRC on public benchmarks) and WildCard (+4.3\\%).\nAdditionally, we present a novel LLM-based data curation pipeline, adaptable to\na variety of safety-related tasks and beyond. We have shown strong\ngeneralization performance for model trained mainly on synthetic data. By\nreleasing ShieldGemma, we provide a valuable resource to the research\ncommunity, advancing LLM safety and enabling the creation of more effective\ncontent moderation solutions for developers.\n", "date": "2024-07-31", "categories": []}
{"id": "2407.20076", "title": "Investigating the Impact of Semi-Supervised Methods with Data\n  Augmentation on Offensive Language Detection in Romanian Language", "abstract": "  Offensive language detection is a crucial task in today's digital landscape,\nwhere online platforms grapple with maintaining a respectful and inclusive\nenvironment. However, building robust offensive language detection models\nrequires large amounts of labeled data, which can be expensive and\ntime-consuming to obtain. Semi-supervised learning offers a feasible solution\nby utilizing labeled and unlabeled data to create more accurate and robust\nmodels. In this paper, we explore a few different semi-supervised methods, as\nwell as data augmentation techniques. Concretely, we implemented eight\nsemi-supervised methods and ran experiments for them using only the available\ndata in the RO-Offense dataset and applying five augmentation techniques before\nfeeding the data to the models. Experimental results demonstrate that some of\nthem benefit more from augmentations than others.\n", "date": "2024-07-29", "categories": []}
{"id": "2407.20729", "title": "Adapting Safe-for-Work Classifier for Malaysian Language Text: Enhancing\n  Alignment in LLM-Ops Framework", "abstract": "  As large language models (LLMs) become increasingly integrated into\noperational workflows (LLM-Ops), there is a pressing need for effective\nguardrails to ensure safe and aligned interactions, including the ability to\ndetect potentially unsafe or inappropriate content across languages. However,\nexisting safe-for-work classifiers are primarily focused on English text. To\naddress this gap for the Malaysian language, we present a novel safe-for-work\ntext classifier tailored specifically for Malaysian language content. By\ncurating and annotating a first-of-its-kind dataset of Malaysian text spanning\nmultiple content categories, we trained a classification model capable of\nidentifying potentially unsafe material using state-of-the-art natural language\nprocessing techniques. This work represents an important step in enabling safer\ninteractions and content filtering to mitigate potential risks and ensure\nresponsible deployment of LLMs. To maximize accessibility and promote further\nresearch towards enhancing alignment in LLM-Ops for the Malaysian context, the\nmodel is publicly released at\nhttps://huggingface.co/malaysia-ai/malaysian-sfw-classifier.\n", "date": "2024-07-30", "categories": []}
{"id": "2407.20910", "title": "Enabling Contextual Soft Moderation on Social Media through Contrastive\n  Textual Deviation", "abstract": "  Automated soft moderation systems are unable to ascertain if a post supports\nor refutes a false claim, resulting in a large number of contextual false\npositives. This limits their effectiveness, for example undermining trust in\nhealth experts by adding warnings to their posts or resorting to vague warnings\ninstead of granular fact-checks, which result in desensitizing users. In this\npaper, we propose to incorporate stance detection into existing automated\nsoft-moderation pipelines, with the goal of ruling out contextual false\npositives and providing more precise recommendations for social media content\nthat should receive warnings. We develop a textual deviation task called\nContrastive Textual Deviation (CTD) and show that it outperforms existing\nstance detection approaches when applied to soft moderation.We then integrate\nCTD into the stateof-the-art system for automated soft moderation Lambretta,\nshowing that our approach can reduce contextual false positives from 20% to\n2.1%, providing another important building block towards deploying reliable\nautomated soft moderation tools on social media.\n", "date": "2024-07-30", "categories": []}
{"id": "2407.20987", "title": "PIXELMOD: Improving Soft Moderation of Visual Misleading Information on\n  Twitter", "abstract": "  Images are a powerful and immediate vehicle to carry misleading or outright\nfalse messages, yet identifying image-based misinformation at scale poses\nunique challenges. In this paper, we present PIXELMOD, a system that leverages\nperceptual hashes, vector databases, and optical character recognition (OCR) to\nefficiently identify images that are candidates to receive soft moderation\nlabels on Twitter. We show that PIXELMOD outperforms existing image similarity\napproaches when applied to soft moderation, with negligible performance\noverhead. We then test PIXELMOD on a dataset of tweets surrounding the 2020 US\nPresidential Election, and find that it is able to identify visually misleading\nimages that are candidates for soft moderation with 0.99% false detection and\n2.06% false negatives.\n", "date": "2024-07-30", "categories": []}
{"id": "2407.18738", "title": "Towards Generalized Offensive Language Identification", "abstract": "  The prevalence of offensive content on the internet, encompassing hate speech\nand cyberbullying, is a pervasive issue worldwide. Consequently, it has\ngarnered significant attention from the machine learning (ML) and natural\nlanguage processing (NLP) communities. As a result, numerous systems have been\ndeveloped to automatically identify potentially harmful content and mitigate\nits impact. These systems can follow two approaches; (1) Use publicly available\nmodels and application endpoints, including prompting large language models\n(LLMs) (2) Annotate datasets and train ML models on them. However, both\napproaches lack an understanding of how generalizable they are. Furthermore,\nthe applicability of these systems is often questioned in off-domain and\npractical environments. This paper empirically evaluates the generalizability\nof offensive language detection models and datasets across a novel generalized\nbenchmark. We answer three research questions on generalizability. Our findings\nwill be useful in creating robust real-world offensive language detection\nsystems.\n", "date": "2024-07-26", "categories": []}
{"id": "2402.11406", "title": "Don't Go To Extremes: Revealing the Excessive Sensitivity and\n  Calibration Limitations of LLMs in Implicit Hate Speech Detection", "abstract": "  The fairness and trustworthiness of Large Language Models (LLMs) are\nreceiving increasing attention. Implicit hate speech, which employs indirect\nlanguage to convey hateful intentions, occupies a significant portion of\npractice. However, the extent to which LLMs effectively address this issue\nremains insufficiently examined. This paper delves into the capability of LLMs\nto detect implicit hate speech (Classification Task) and express confidence in\ntheir responses (Calibration Task). Our evaluation meticulously considers\nvarious prompt patterns and mainstream uncertainty estimation methods. Our\nfindings highlight that LLMs exhibit two extremes: (1) LLMs display excessive\nsensitivity towards groups or topics that may cause fairness issues, resulting\nin misclassifying benign statements as hate speech. (2) LLMs' confidence scores\nfor each method excessively concentrate on a fixed range, remaining unchanged\nregardless of the dataset's complexity. Consequently, the calibration\nperformance is heavily reliant on primary classification accuracy. These\ndiscoveries unveil new limitations of LLMs, underscoring the need for caution\nwhen optimizing models to ensure they do not veer towards extremes. This serves\nas a reminder to carefully consider sensitivity and confidence in the pursuit\nof model fairness.\n", "date": "2024-02-18", "categories": []}
{"id": "2407.15227", "title": "A Community-Centric Perspective for Characterizing and Detecting\n  Anti-Asian Violence-Provoking Speech", "abstract": "  Violence-provoking speech -- speech that implicitly or explicitly promotes\nviolence against the members of the targeted community, contributed to a\nmassive surge in anti-Asian crimes during the pandemic. While previous works\nhave characterized and built tools for detecting other forms of harmful speech,\nlike fear speech and hate speech, our work takes a community-centric approach\nto studying anti-Asian violence-provoking speech. Using data from ~420k Twitter\nposts spanning a 3-year duration (January 1, 2020 to February 1, 2023), we\ndevelop a codebook to characterize anti-Asian violence-provoking speech and\ncollect a community-crowdsourced dataset to facilitate its large-scale\ndetection using state-of-the-art classifiers. We contrast the capabilities of\nnatural language processing classifiers, ranging from BERT-based to LLM-based\nclassifiers, in detecting violence-provoking speech with their capabilities to\ndetect anti-Asian hateful speech. In contrast to prior work that has\ndemonstrated the effectiveness of such classifiers in detecting hateful speech\n($F_1 = 0.89$), our work shows that accurate and reliable detection of\nviolence-provoking speech is a challenging task ($F_1 = 0.69$). We discuss the\nimplications of our findings, particularly the need for proactive interventions\nto support Asian communities during public health crises. The resources related\nto the study are available at\nhttps://claws-lab.github.io/violence-provoking-speech/.\n", "date": "2024-07-21", "categories": []}
{"id": "2407.13942", "title": "Harmful Suicide Content Detection", "abstract": "  Harmful suicide content on the Internet is a significant risk factor inducing\nsuicidal thoughts and behaviors among vulnerable populations. Despite global\nefforts, existing resources are insufficient, specifically in high-risk regions\nlike the Republic of Korea. Current research mainly focuses on understanding\nnegative effects of such content or suicide risk in individuals, rather than on\nautomatically detecting the harmfulness of content. To fill this gap, we\nintroduce a harmful suicide content detection task for classifying online\nsuicide content into five harmfulness levels. We develop a multi-modal\nbenchmark and a task description document in collaboration with medical\nprofessionals, and leverage large language models (LLMs) to explore efficient\nmethods for moderating such content. Our contributions include proposing a\nnovel detection task, a multi-modal Korean benchmark with expert annotations,\nand suggesting strategies using LLMs to detect illegal and harmful content.\nOwing to the potential harm involved, we publicize our implementations and\nbenchmark, incorporating an ethical verification process.\n", "date": "2024-06-03", "categories": []}
{"id": "2407.12836", "title": "OSPC: Artificial VLM Features for Hateful Meme Detection", "abstract": "  The digital revolution and the advent of the world wide web have transformed\nhuman communication, notably through the emergence of memes. While memes are a\npopular and straightforward form of expression, they can also be used to spread\nmisinformation and hate due to their anonymity and ease of use. In response to\nthese challenges, this paper introduces a solution developed by team 'Baseline'\nfor the AI Singapore Online Safety Prize Challenge. Focusing on computational\nefficiency and feature engineering, the solution achieved an AUROC of 0.76 and\nan accuracy of 0.69 on the test dataset. As key features, the solution\nleverages the inherent probabilistic capabilities of large Vision-Language\nModels (VLMs) to generate task-adapted feature encodings from text, and applies\na distilled quantization tailored to the specific cultural nuances present in\nSingapore. This type of processing and fine-tuning can be adapted to various\nvisual and textual understanding and classification tasks, and even applied on\nprivate VLMs such as OpenAI's GPT. Finally it can eliminate the need for\nextensive model training on large GPUs for resource constrained applications,\nalso offering a solution when little or no data is available.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.12018", "title": "Empirical Evaluation of Public HateSpeech Datasets", "abstract": "  Despite the extensive communication benefits offered by social media\nplatforms, numerous challenges must be addressed to ensure user safety. One of\nthe most significant risks faced by users on these platforms is targeted hate\nspeech. Social media platforms are widely utilised for generating datasets\nemployed in training and evaluating machine learning algorithms for hate speech\ndetection. However, existing public datasets exhibit numerous limitations,\nhindering the effective training of these algorithms and leading to inaccurate\nhate speech classification. This study provides a comprehensive empirical\nevaluation of several public datasets commonly used in automated hate speech\nclassification. Through rigorous analysis, we present compelling evidence\nhighlighting the limitations of current hate speech datasets. Additionally, we\nconduct a range of statistical analyses to elucidate the strengths and\nweaknesses inherent in these datasets. This work aims to advance the\ndevelopment of more accurate and reliable machine learning models for hate\nspeech detection by addressing the dataset limitations identified.\n", "date": "2024-06-27", "categories": []}
{"id": "2405.20628", "title": "ToxVidLM: A Multimodal Framework for Toxicity Detection in Code-Mixed\n  Videos", "abstract": "  In an era of rapidly evolving internet technology, the surge in multimodal\ncontent, including videos, has expanded the horizons of online communication.\nHowever, the detection of toxic content in this diverse landscape, particularly\nin low-resource code-mixed languages, remains a critical challenge. While\nsubstantial research has addressed toxic content detection in textual data, the\nrealm of video content, especially in non-English languages, has been\nrelatively underexplored. This paper addresses this research gap by introducing\na benchmark dataset, the first of its kind, consisting of 931 videos with 4021\ncode-mixed Hindi-English utterances collected from YouTube. Each utterance\nwithin this dataset has been meticulously annotated for toxicity, severity, and\nsentiment labels. We have developed an advanced Multimodal Multitask framework\nbuilt for Toxicity detection in Video Content by leveraging Language Models\n(LMs), crafted for the primary objective along with the additional tasks of\nconducting sentiment and severity analysis. ToxVidLM incorporates three key\nmodules - the Encoder module, Cross-Modal Synchronization module, and Multitask\nmodule - crafting a generic multimodal LM customized for intricate video\nclassification tasks. Our experiments reveal that incorporating multiple\nmodalities from the videos substantially enhances the performance of toxic\ncontent detection by achieving an Accuracy and Weighted F1 score of 94.29% and\n94.35%, respectively.\n", "date": "2024-05-31", "categories": []}
{"id": "2407.06309", "title": "Multimodal Chain-of-Thought Reasoning via ChatGPT to Protect Children\n  from Age-Inappropriate Apps", "abstract": "  Mobile applications (Apps) could expose children to inappropriate themes such\nas sexual content, violence, and drug use. Maturity rating offers a quick and\neffective method for potential users, particularly guardians, to assess the\nmaturity levels of apps. Determining accurate maturity ratings for mobile apps\nis essential to protect children's health in today's saturated digital\nmarketplace. Existing approaches to maturity rating are either inaccurate\n(e.g., self-reported rating by developers) or costly (e.g., manual\nexamination). In the literature, there are few text-mining-based approaches to\nmaturity rating. However, each app typically involves multiple modalities,\nnamely app description in the text, and screenshots in the image. In this\npaper, we present a framework for determining app maturity levels that utilize\nmultimodal large language models (MLLMs), specifically ChatGPT-4 Vision.\nPowered by Chain-of-Thought (CoT) reasoning, our framework systematically\nleverages ChatGPT-4 to process multimodal app data (i.e., textual descriptions\nand screenshots) and guide the MLLM model through a step-by-step reasoning\npathway from initial content analysis to final maturity rating determination.\nAs a result, through explicitly incorporating CoT reasoning, our framework\nenables ChatGPT to understand better and apply maturity policies to facilitate\nmaturity rating. Experimental results indicate that the proposed method\noutperforms all baseline models and other fusion strategies.\n", "date": "2024-07-08", "categories": []}
{"id": "2312.15099", "title": "Moderating New Waves of Online Hate with Chain-of-Thought Reasoning in\n  Large Language Models", "abstract": "  Online hate is an escalating problem that negatively impacts the lives of\nInternet users, and is also subject to rapid changes due to evolving events,\nresulting in new waves of online hate that pose a critical threat. Detecting\nand mitigating these new waves present two key challenges: it demands\nreasoning-based complex decision-making to determine the presence of hateful\ncontent, and the limited availability of training samples hinders updating the\ndetection model. To address this critical issue, we present a novel framework\ncalled HATEGUARD for effectively moderating new waves of online hate. HATEGUARD\nemploys a reasoning-based approach that leverages the recently introduced\nchain-of-thought (CoT) prompting technique, harnessing the capabilities of\nlarge language models (LLMs). HATEGUARD further achieves prompt-based zero-shot\ndetection by automatically generating and updating detection prompts with new\nderogatory terms and targets in new wave samples to effectively address new\nwaves of online hate. To demonstrate the effectiveness of our approach, we\ncompile a new dataset consisting of tweets related to three recently witnessed\nnew waves: the 2022 Russian invasion of Ukraine, the 2021 insurrection of the\nUS Capitol, and the COVID-19 pandemic. Our studies reveal crucial longitudinal\npatterns in these new waves concerning the evolution of events and the pressing\nneed for techniques to rapidly update existing moderation tools to counteract\nthem. Comparative evaluations against state-of-the-art tools illustrate the\nsuperiority of our framework, showcasing a substantial 22.22% to 83.33%\nimprovement in detecting the three new waves of online hate. Our work\nhighlights the severe threat posed by the emergence of new waves of online hate\nand represents a paradigm shift in addressing this threat practically.\n", "date": "2023-12-22", "categories": []}
{"id": "2407.02448", "title": "Ensemble of pre-trained language models and data augmentation for hate\n  speech detection from Arabic tweets", "abstract": "  Today, hate speech classification from Arabic tweets has drawn the attention\nof several researchers. Many systems and techniques have been developed to\nresolve this classification task. Nevertheless, two of the major challenges\nfaced in this context are the limited performance and the problem of imbalanced\ndata. In this study, we propose a novel approach that leverages ensemble\nlearning and semi-supervised learning based on previously manually labeled. We\nconducted experiments on a benchmark dataset by classifying Arabic tweets into\n5 distinct classes: non-hate, general hate, racial, religious, or sexism.\nExperimental results show that: (1) ensemble learning based on pre-trained\nlanguage models outperforms existing related works; (2) Our proposed data\naugmentation improves the accuracy results of hate speech detection from Arabic\ntweets and outperforms existing related works. Our main contribution is the\nachievement of encouraging results in Arabic hate speech detection.\n", "date": "2024-07-02", "categories": []}
{"id": "2407.01149", "title": "Sociocultural Considerations in Monitoring Anti-LGBTQ+ Content on Social\n  Media", "abstract": "  The purpose of this paper is to ascertain the influence of sociocultural\nfactors (i.e., social, cultural, and political) in the development of hate\nspeech detection systems. We set out to investigate the suitability of using\nopen-source training data to monitor levels of anti-LGBTQ+ content on social\nmedia across different national-varieties of English. Our findings suggests the\nsocial and cultural alignment of open-source hate speech data sets influences\nthe predicted outputs. Furthermore, the keyword-search approach of anti-LGBTQ+\nslurs in the development of open-source training data encourages detection\nmodels to overfit on slurs; therefore, anti-LGBTQ+ content may go undetected.\nWe recommend combining empirical outputs with qualitative insights to ensure\nthese systems are fit for purpose.\n", "date": "2024-07-01", "categories": []}
{"id": "2403.19836", "title": "Target Span Detection for Implicit Harmful Content", "abstract": "  Identifying the targets of hate speech is a crucial step in grasping the\nnature of such speech and, ultimately, in improving the detection of offensive\nposts on online forums. Much harmful content on online platforms uses implicit\nlanguage especially when targeting vulnerable and protected groups such as\nusing stereotypical characteristics instead of explicit target names, making it\nharder to detect and mitigate the language. In this study, we focus on\nidentifying implied targets of hate speech, essential for recognizing subtler\nhate speech and enhancing the detection of harmful content on digital\nplatforms. We define a new task aimed at identifying the targets even when they\nare not explicitly stated. To address that task, we collect and annotate target\nspans in three prominent implicit hate speech datasets: SBIC, DynaHate, and\nIHC. We call the resulting merged collection Implicit-Target-Span. The\ncollection is achieved using an innovative pooling method with matching scores\nbased on human annotations and Large Language Models (LLMs). Our experiments\nindicate that Implicit-Target-Span provides a challenging test bed for target\nspan detection methods.\n", "date": "2024-03-28", "categories": []}
{"id": "2406.19543", "title": "Demarked: A Strategy for Enhanced Abusive Speech Moderation through\n  Counterspeech, Detoxification, and Message Management", "abstract": "  Despite regulations imposed by nations and social media platforms, such as\nrecent EU regulations targeting digital violence, abusive content persists as a\nsignificant challenge. Existing approaches primarily rely on binary solutions,\nsuch as outright blocking or banning, yet fail to address the complex nature of\nabusive speech. In this work, we propose a more comprehensive approach called\nDemarcation scoring abusive speech based on four aspect -- (i) severity scale;\n(ii) presence of a target; (iii) context scale; (iv) legal scale -- and\nsuggesting more options of actions like detoxification, counter speech\ngeneration, blocking, or, as a final measure, human intervention. Through a\nthorough analysis of abusive speech regulations across diverse jurisdictions,\nplatforms, and research papers we highlight the gap in preventing measures and\nadvocate for tailored proactive steps to combat its multifaceted\nmanifestations. Our work aims to inform future strategies for effectively\naddressing abusive speech online.\n", "date": "2024-06-27", "categories": []}
{"id": "2401.05060", "title": "MuTox: Universal MUltilingual Audio-based TOXicity Dataset and Zero-shot\n  Detector", "abstract": "  Research in toxicity detection in natural language processing for the speech\nmodality (audio-based) is quite limited, particularly for languages other than\nEnglish. To address these limitations and lay the groundwork for truly\nmultilingual audio-based toxicity detection, we introduce MuTox, the first\nhighly multilingual audio-based dataset with toxicity labels. The dataset\ncomprises 20,000 audio utterances for English and Spanish, and 4,000 for the\nother 19 languages. To demonstrate the quality of this dataset, we trained the\nMuTox audio-based toxicity classifier, which enables zero-shot toxicity\ndetection across a wide range of languages. This classifier outperforms\nexisting text-based trainable classifiers by more than 1% AUC, while expanding\nthe language coverage more than tenfold. When compared to a wordlist-based\nclassifier that covers a similar number of languages, MuTox improves precision\nand recall by approximately 2.5 times. This significant improvement underscores\nthe potential of MuTox in advancing the field of audio-based toxicity\ndetection.\n", "date": "2024-01-10", "categories": []}
{"id": "2406.17793", "title": "Deep Learning Approaches for Detecting Adversarial Cyberbullying and\n  Hate Speech in Social Networks", "abstract": "  Cyberbullying is a significant concern intricately linked to technology that\ncan find resolution through technological means. Despite its prevalence,\ntechnology also provides solutions to mitigate cyberbullying. To address\ngrowing concerns regarding the adverse impact of cyberbullying on individuals'\nonline experiences, various online platforms and researchers are actively\nadopting measures to enhance the safety of digital environments. While\nresearchers persist in crafting detection models to counteract or minimize\ncyberbullying, malicious actors are deploying adversarial techniques to\ncircumvent these detection methods. This paper focuses on detecting\ncyberbullying in adversarial attack content within social networking site text\ndata, specifically emphasizing hate speech. Utilizing a deep learning-based\napproach with a correction algorithm, this paper yielded significant results.\nAn LSTM model with a fixed epoch of 100 demonstrated remarkable performance,\nachieving high accuracy, precision, recall, F1-score, and AUC-ROC scores of\n87.57%, 88.73%, 87.57%, 88.15%, and 91% respectively. Additionally, the LSTM\nmodel's performance surpassed that of previous studies.\n", "date": "2024-05-30", "categories": []}
{"id": "2311.09086", "title": "The Uli Dataset: An Exercise in Experience Led Annotation of oGBV", "abstract": "  Online gender based violence has grown concomitantly with adoption of the\ninternet and social media. Its effects are worse in the Global majority where\nmany users use social media in languages other than English. The scale and\nvolume of conversations on the internet has necessitated the need for automated\ndetection of hate speech, and more specifically gendered abuse. There is,\nhowever, a lack of language specific and contextual data to build such\nautomated tools. In this paper we present a dataset on gendered abuse in three\nlanguages- Hindi, Tamil and Indian English. The dataset comprises of tweets\nannotated along three questions pertaining to the experience of gender abuse,\nby experts who identify as women or a member of the LGBTQIA community in South\nAsia. Through this dataset we demonstrate a participatory approach to creating\ndatasets that drive AI systems.\n", "date": "2023-11-15", "categories": []}
{"id": "2403.05268", "title": "Deep Prompt Multi-task Network for Abuse Language Detection", "abstract": "  The detection of abusive language remains a long-standing challenge with the\nextensive use of social networks. The detection task of abusive language\nsuffers from limited accuracy. We argue that the existing detection methods\nutilize the fine-tuning technique of the pre-trained language models (PLMs) to\nhandle downstream tasks. Hence, these methods fail to stimulate the general\nknowledge of the PLMs. To address the problem, we propose a novel Deep Prompt\nMulti-task Network (DPMN) for abuse language detection. Specifically, DPMN\nfirst attempts to design two forms of deep prompt tuning and light prompt\ntuning for the PLMs. The effects of different prompt lengths, tuning\nstrategies, and prompt initialization methods on detecting abusive language are\nstudied. In addition, we propose a Task Head based on Bi-LSTM and FFN, which\ncan be used as a short text classifier. Eventually, DPMN utilizes multi-task\nlearning to improve detection metrics further. The multi-task network has the\nfunction of transferring effective knowledge. The proposed DPMN is evaluated\nagainst eight typical methods on three public datasets: OLID, SOLID, and\nAbuseAnalyzer. The experimental results show that our DPMN outperforms the\nstate-of-the-art methods.\n", "date": "2024-03-08", "categories": []}
{"id": "2403.19260", "title": "NaijaHate: Evaluating Hate Speech Detection on Nigerian Twitter Using\n  Representative Data", "abstract": "  To address the global issue of online hate, hate speech detection (HSD)\nsystems are typically developed on datasets from the United States, thereby\nfailing to generalize to English dialects from the Majority World. Furthermore,\nHSD models are often evaluated on non-representative samples, raising concerns\nabout overestimating model performance in real-world settings. In this work, we\nintroduce NaijaHate, the first dataset annotated for HSD which contains a\nrepresentative sample of Nigerian tweets. We demonstrate that HSD evaluated on\nbiased datasets traditionally used in the literature consistently overestimates\nreal-world performance by at least two-fold. We then propose NaijaXLM-T, a\npretrained model tailored to the Nigerian Twitter context, and establish the\nkey role played by domain-adaptive pretraining and finetuning in maximizing HSD\nperformance. Finally, owing to the modest performance of HSD systems in\nreal-world conditions, we find that content moderators would need to review\nabout ten thousand Nigerian tweets flagged as hateful daily to moderate 60% of\nall hateful content, highlighting the challenges of moderating hate speech at\nscale as social media usage continues to grow globally. Taken together, these\nresults pave the way towards robust HSD systems and a better protection of\nsocial media users from hateful content in low-resource settings.\n", "date": "2024-03-28", "categories": []}
{"id": "2406.00020", "title": "Harmful Speech Detection by Language Models Exhibits Gender-Queer\n  Dialect Bias", "abstract": "  Content moderation on social media platforms shapes the dynamics of online\ndiscourse, influencing whose voices are amplified and whose are suppressed.\nRecent studies have raised concerns about the fairness of content moderation\npractices, particularly for aggressively flagging posts from transgender and\nnon-binary individuals as toxic. In this study, we investigate the presence of\nbias in harmful speech classification of gender-queer dialect online, focusing\nspecifically on the treatment of reclaimed slurs. We introduce a novel dataset,\nQueerReclaimLex, based on 109 curated templates exemplifying non-derogatory\nuses of LGBTQ+ slurs. Dataset instances are scored by gender-queer annotators\nfor potential harm depending on additional context about speaker identity. We\nsystematically evaluate the performance of five off-the-shelf language models\nin assessing the harm of these texts and explore the effectiveness of\nchain-of-thought prompting to teach large language models (LLMs) to leverage\nauthor identity context. We reveal a tendency for these models to inaccurately\nflag texts authored by gender-queer individuals as harmful. Strikingly, across\nall LLMs the performance is poorest for texts that show signs of being written\nby individuals targeted by the featured slur (F1 <= 0.24). We highlight an\nurgent need for fairness and inclusivity in content moderation systems. By\nuncovering these biases, this work aims to inform the development of more\nequitable content moderation practices and contribute to the creation of\ninclusive online spaces for all users.\n", "date": "2024-05-23", "categories": []}
{"id": "2406.15869", "title": "A multitask learning framework for leveraging subjectivity of annotators\n  to identify misogyny", "abstract": "  Identifying misogyny using artificial intelligence is a form of combating\nonline toxicity against women. However, the subjective nature of interpreting\nmisogyny poses a significant challenge to model the phenomenon. In this paper,\nwe propose a multitask learning approach that leverages the subjectivity of\nthis task to enhance the performance of the misogyny identification systems. We\nincorporated diverse perspectives from annotators in our model design,\nconsidering gender and age across six profile groups, and conducted extensive\nexperiments and error analysis using two language models to validate our four\nalternative designs of the multitask learning technique to identify\nmisogynistic content in English tweets. The results demonstrate that\nincorporating various viewpoints enhances the language models' ability to\ninterpret different forms of misogyny. This research advances content\nmoderation and highlights the importance of embracing diverse perspectives to\nbuild effective online moderation systems.\n", "date": "2024-06-22", "categories": []}
{"id": "2406.06840", "title": "Silent Signals, Loud Impact: LLMs for Word-Sense Disambiguation of Coded\n  Dog Whistles", "abstract": "  A dog whistle is a form of coded communication that carries a secondary\nmeaning to specific audiences and is often weaponized for racial and\nsocioeconomic discrimination. Dog whistling historically originated from United\nStates politics, but in recent years has taken root in social media as a means\nof evading hate speech detection systems and maintaining plausible deniability.\nIn this paper, we present an approach for word-sense disambiguation of dog\nwhistles from standard speech using Large Language Models (LLMs), and leverage\nthis technique to create a dataset of 16,550 high-confidence coded examples of\ndog whistles used in formal and informal communication. Silent Signals is the\nlargest dataset of disambiguated dog whistle usage, created for applications in\nhate speech detection, neology, and political science.\n  The dataset can be found at\nhttps://huggingface.co/datasets/SALT-NLP/silent_signals.\n", "date": "2024-06-10", "categories": []}
{"id": "2406.14131", "title": "Detecting sexually explicit content in the context of the child sexual\n  abuse materials (CSAM): end-to-end classifiers and region-based networks", "abstract": "  Child sexual abuse materials (CSAM) pose a significant threat to the safety\nand well-being of children worldwide. Detecting and preventing the distribution\nof such materials is a critical task for law enforcement agencies and\ntechnology companies. As content moderation is often manual, developing an\nautomated detection system can help reduce human reviewers' exposure to\npotentially harmful images and accelerate the process of counteracting. This\nstudy presents methods for classifying sexually explicit content, which plays a\ncrucial role in the automated CSAM detection system. Several approaches are\nexplored to solve the task: an end-to-end classifier, a classifier with person\ndetection and a private body parts detector. All proposed methods are tested on\nthe images obtained from the online tool for reporting illicit content. Due to\nlegal constraints, access to the data is limited, and all algorithms are\nexecuted remotely on the isolated server. The end-to-end classifier yields the\nmost promising results, with an accuracy of 90.17%, after augmenting the\ntraining set with the additional neutral samples and adult pornography. While\ndetection-based methods may not achieve higher accuracy rates and cannot serve\nas a final classifier on their own, their inclusion in the system can be\nbeneficial. Human body-oriented approaches generate results that are easier to\ninterpret, and obtaining more interpretable results is essential when analyzing\nmodels that are trained without direct access to data.\n", "date": "2024-06-20", "categories": []}
{"id": "2406.12223", "title": "ToxiCloakCN: Evaluating Robustness of Offensive Language Detection in\n  Chinese with Cloaking Perturbations", "abstract": "  Detecting hate speech and offensive language is essential for maintaining a\nsafe and respectful digital environment. This study examines the limitations of\nstate-of-the-art large language models (LLMs) in identifying offensive content\nwithin systematically perturbed data, with a focus on Chinese, a language\nparticularly susceptible to such perturbations. We introduce\n\\textsf{ToxiCloakCN}, an enhanced dataset derived from ToxiCN, augmented with\nhomophonic substitutions and emoji transformations, to test the robustness of\nLLMs against these cloaking perturbations. Our findings reveal that existing\nmodels significantly underperform in detecting offensive content when these\nperturbations are applied. We provide an in-depth analysis of how different\ntypes of offensive content are affected by these perturbations and explore the\nalignment between human and model explanations of offensiveness. Our work\nhighlights the urgent need for more advanced techniques in offensive language\ndetection to combat the evolving tactics used to evade detection mechanisms.\n", "date": "2024-06-18", "categories": []}
{"id": "2406.12304", "title": "COT: A Generative Approach for Hate Speech Counter-Narratives via\n  Contrastive Optimal Transport", "abstract": "  Counter-narratives, which are direct responses consisting of non-aggressive\nfact-based arguments, have emerged as a highly effective approach to combat the\nproliferation of hate speech. Previous methodologies have primarily focused on\nfine-tuning and post-editing techniques to ensure the fluency of generated\ncontents, while overlooking the critical aspects of individualization and\nrelevance concerning the specific hatred targets, such as LGBT groups,\nimmigrants, etc. This research paper introduces a novel framework based on\ncontrastive optimal transport, which effectively addresses the challenges of\nmaintaining target interaction and promoting diversification in generating\ncounter-narratives. Firstly, an Optimal Transport Kernel (OTK) module is\nleveraged to incorporate hatred target information in the token\nrepresentations, in which the comparison pairs are extracted between original\nand transported features. Secondly, a self-contrastive learning module is\nemployed to address the issue of model degeneration. This module achieves this\nby generating an anisotropic distribution of token representations. Finally, a\ntarget-oriented search method is integrated as an improved decoding strategy to\nexplicitly promote domain relevance and diversification in the inference\nprocess. This strategy modifies the model's confidence score by considering\nboth token similarity and target relevance. Quantitative and qualitative\nexperiments have been evaluated on two benchmark datasets, which demonstrate\nthat our proposed model significantly outperforms current methods evaluated by\nmetrics from multiple aspects.\n", "date": "2024-06-18", "categories": []}
{"id": "2406.09779", "title": "OSPC: Detecting Harmful Memes with Large Language Model as a Catalyst", "abstract": "  Memes, which rapidly disseminate personal opinions and positions across the\ninternet, also pose significant challenges in propagating social bias and\nprejudice. This study presents a novel approach to detecting harmful memes,\nparticularly within the multicultural and multilingual context of Singapore.\nOur methodology integrates image captioning, Optical Character Recognition\n(OCR), and Large Language Model (LLM) analysis to comprehensively understand\nand classify harmful memes. Utilizing the BLIP model for image captioning,\nPP-OCR and TrOCR for text recognition across multiple languages, and the Qwen\nLLM for nuanced language understanding, our system is capable of identifying\nharmful content in memes created in English, Chinese, Malay, and Tamil. To\nenhance the system's performance, we fine-tuned our approach by leveraging\nadditional data labeled using GPT-4V, aiming to distill the understanding\ncapability of GPT-4V for harmful memes to our system. Our framework achieves\ntop-1 at the public leaderboard of the Online Safety Prize Challenge hosted by\nAI Singapore, with the AUROC as 0.7749 and accuracy as 0.7087, significantly\nahead of the other teams. Notably, our approach outperforms previous\nbenchmarks, with FLAVA achieving an AUROC of 0.5695 and VisualBERT an AUROC of\n0.5561.\n", "date": "2024-06-14", "categories": []}
{"id": "2405.08562", "title": "The Unseen Targets of Hate -- A Systematic Review of Hateful\n  Communication Datasets", "abstract": "  Machine learning (ML)-based content moderation tools are essential to keep\nonline spaces free from hateful communication. Yet, ML tools can only be as\ncapable as the quality of the data they are trained on allows them. While there\nis increasing evidence that they underperform in detecting hateful\ncommunications directed towards specific identities and may discriminate\nagainst them, we know surprisingly little about the provenance of such bias. To\nfill this gap, we present a systematic review of the datasets for the automated\ndetection of hateful communication introduced over the past decade, and unpack\nthe quality of the datasets in terms of the identities that they embody: those\nof the targets of hateful communication that the data curators focused on, as\nwell as those unintentionally included in the datasets. We find, overall, a\nskewed representation of selected target identities and mismatches between the\ntargets that research conceptualizes and ultimately includes in datasets. Yet,\nby contextualizing these findings in the language and location of origin of the\ndatasets, we highlight a positive trend towards the broadening and\ndiversification of this research space.\n", "date": "2024-05-14", "categories": []}
{"id": "2406.07841", "title": "Labeling Comic Mischief Content in Online Videos with a Multimodal\n  Hierarchical-Cross-Attention Model", "abstract": "  We address the challenge of detecting questionable content in online media,\nspecifically the subcategory of comic mischief. This type of content combines\nelements such as violence, adult content, or sarcasm with humor, making it\ndifficult to detect. Employing a multimodal approach is vital to capture the\nsubtle details inherent in comic mischief content. To tackle this problem, we\npropose a novel end-to-end multimodal system for the task of comic mischief\ndetection. As part of this contribution, we release a novel dataset for the\ntargeted task consisting of three modalities: video, text (video captions and\nsubtitles), and audio. We also design a HIerarchical Cross-attention model with\nCAPtions (HICCAP) to capture the intricate relationships among these\nmodalities. The results show that the proposed approach makes a significant\nimprovement over robust baselines and state-of-the-art models for comic\nmischief detection and its type classification. This emphasizes the potential\nof our system to empower users, to make informed decisions about the online\ncontent they choose to see. In addition, we conduct experiments on the UCF101,\nHMDB51, and XD-Violence datasets, comparing our model against other\nstate-of-the-art approaches showcasing the outstanding performance of our\nproposed model in various scenarios.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.07886", "title": "Label-aware Hard Negative Sampling Strategies with Momentum Contrastive\n  Learning for Implicit Hate Speech Detection", "abstract": "  Detecting implicit hate speech that is not directly hateful remains a\nchallenge. Recent research has attempted to detect implicit hate speech by\napplying contrastive learning to pre-trained language models such as BERT and\nRoBERTa, but the proposed models still do not have a significant advantage over\ncross-entropy loss-based learning. We found that contrastive learning based on\nrandomly sampled batch data does not encourage the model to learn hard negative\nsamples. In this work, we propose Label-aware Hard Negative sampling strategies\n(LAHN) that encourage the model to learn detailed features from hard negative\nsamples, instead of naive negative samples in random batch, using\nmomentum-integrated contrastive learning. LAHN outperforms the existing models\nfor implicit hate speech detection both in- and cross-datasets. The code is\navailable at https://github.com/Hanyang-HCC-Lab/LAHN\n", "date": "2024-06-12", "categories": []}
{"id": "2406.08080", "title": "AustroTox: A Dataset for Target-Based Austrian German Offensive Language\n  Detection", "abstract": "  Model interpretability in toxicity detection greatly profits from token-level\nannotations. However, currently such annotations are only available in English.\nWe introduce a dataset annotated for offensive language detection sourced from\na news forum, notable for its incorporation of the Austrian German dialect,\ncomprising 4,562 user comments. In addition to binary offensiveness\nclassification, we identify spans within each comment constituting vulgar\nlanguage or representing targets of offensive statements. We evaluate\nfine-tuned language models as well as large language models in a zero- and\nfew-shot fashion. The results indicate that while fine-tuned models excel in\ndetecting linguistic peculiarities such as vulgar dialect, large language\nmodels demonstrate superior performance in detecting offensiveness in\nAustroTox. We publish the data and code.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.00046", "title": "Hate Speech Detection with Generalizable Target-aware Fairness", "abstract": "  To counter the side effect brought by the proliferation of social media\nplatforms, hate speech detection (HSD) plays a vital role in halting the\ndissemination of toxic online posts at an early stage. However, given the\nubiquitous topical communities on social media, a trained HSD classifier easily\nbecomes biased towards specific targeted groups (e.g., female and black\npeople), where a high rate of false positive/negative results can significantly\nimpair public trust in the fairness of content moderation mechanisms, and\neventually harm the diversity of online society. Although existing\nfairness-aware HSD methods can smooth out some discrepancies across targeted\ngroups, they are mostly specific to a narrow selection of targets that are\nassumed to be known and fixed. This inevitably prevents those methods from\ngeneralizing to real-world use cases where new targeted groups constantly\nemerge over time. To tackle this defect, we propose Generalizable target-aware\nFairness (GetFair), a new method for fairly classifying each post that contains\ndiverse and even unseen targets during inference. To remove the HSD\nclassifier's spurious dependence on target-related features, GetFair trains a\nseries of filter functions in an adversarial pipeline, so as to deceive the\ndiscriminator that recovers the targeted group from filtered post embeddings.\nTo maintain scalability and generalizability, we innovatively parameterize all\nfilter functions via a hypernetwork that is regularized by the semantic\naffinity among targets. Taking a target's pretrained word embedding as input,\nthe hypernetwork generates the weights used by each target-specific filter\non-the-fly without storing dedicated filter parameters. Finally, comparative\nexperiments on two HSD datasets have shown advantageous performance of GetFair\non out-of-sample targets.\n", "date": "2024-05-28", "categories": []}
{"id": "2406.05344", "title": "MemeGuard: An LLM and VLM-based Framework for Advancing Content\n  Moderation via Meme Intervention", "abstract": "  In the digital world, memes present a unique challenge for content moderation\ndue to their potential to spread harmful content. Although detection methods\nhave improved, proactive solutions such as intervention are still limited, with\ncurrent research focusing mostly on text-based content, neglecting the\nwidespread influence of multimodal content like memes. Addressing this gap, we\npresent \\textit{MemeGuard}, a comprehensive framework leveraging Large Language\nModels (LLMs) and Visual Language Models (VLMs) for meme intervention.\n\\textit{MemeGuard} harnesses a specially fine-tuned VLM, \\textit{VLMeme}, for\nmeme interpretation, and a multimodal knowledge selection and ranking mechanism\n(\\textit{MKS}) for distilling relevant knowledge. This knowledge is then\nemployed by a general-purpose LLM to generate contextually appropriate\ninterventions. Another key contribution of this work is the\n\\textit{\\textbf{I}ntervening} \\textit{\\textbf{C}yberbullying in\n\\textbf{M}ultimodal \\textbf{M}emes (ICMM)} dataset, a high-quality, labeled\ndataset featuring toxic memes and their corresponding human-annotated\ninterventions. We leverage \\textit{ICMM} to test \\textit{MemeGuard},\ndemonstrating its proficiency in generating relevant and effective responses to\ntoxic memes.\n", "date": "2024-06-08", "categories": []}
{"id": "2406.04892", "title": "Sexism Detection on a Data Diet", "abstract": "  There is an increase in the proliferation of online hate commensurate with\nthe rise in the usage of social media. In response, there is also a significant\nadvancement in the creation of automated tools aimed at identifying harmful\ntext content using approaches grounded in Natural Language Processing and Deep\nLearning. Although it is known that training Deep Learning models require a\nsubstantial amount of annotated data, recent line of work suggests that models\ntrained on specific subsets of the data still retain performance comparable to\nthe model that was trained on the full dataset. In this work, we show how we\ncan leverage influence scores to estimate the importance of a data point while\ntraining a model and designing a pruning strategy applied to the case of sexism\ndetection. We evaluate the model performance trained on data pruned with\ndifferent pruning strategies on three out-of-domain datasets and find, that in\naccordance with other work a large fraction of instances can be removed without\nsignificant performance drop. However, we also discover that the strategies for\npruning data, previously successful in Natural Language Inference tasks, do not\nreadily apply to the detection of harmful content and instead amplify the\nalready prevalent class imbalance even more, leading in the worst-case to a\ncomplete absence of the hateful class.\n", "date": "2024-06-07", "categories": []}
{"id": "2406.02169", "title": "A multilingual dataset for offensive language and hate speech detection\n  for hausa, yoruba and igbo languages", "abstract": "  The proliferation of online offensive language necessitates the development\nof effective detection mechanisms, especially in multilingual contexts. This\nstudy addresses the challenge by developing and introducing novel datasets for\noffensive language detection in three major Nigerian languages: Hausa, Yoruba,\nand Igbo. We collected data from Twitter and manually annotated it to create\ndatasets for each of the three languages, using native speakers. We used\npre-trained language models to evaluate their efficacy in detecting offensive\nlanguage in our datasets. The best-performing model achieved an accuracy of\n90\\%. To further support research in offensive language detection, we plan to\nmake the dataset and our models publicly available.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.03953", "title": "Tox-BART: Leveraging Toxicity Attributes for Explanation Generation of\n  Implicit Hate Speech", "abstract": "  Employing language models to generate explanations for an incoming implicit\nhate post is an active area of research. The explanation is intended to make\nexplicit the underlying stereotype and aid content moderators. The training\noften combines top-k relevant knowledge graph (KG) tuples to provide world\nknowledge and improve performance on standard metrics. Interestingly, our study\npresents conflicting evidence for the role of the quality of KG tuples in\ngenerating implicit explanations. Consequently, simpler models incorporating\nexternal toxicity signals outperform KG-infused models. Compared to the\nKG-based setup, we observe a comparable performance for SBIC (LatentHatred)\ndatasets with a performance variation of +0.44 (+0.49), +1.83 (-1.56), and\n-4.59 (+0.77) in BLEU, ROUGE-L, and BERTScore. Further human evaluation and\nerror analysis reveal that our proposed setup produces more precise\nexplanations than zero-shot GPT-3.5, highlighting the intricate nature of the\ntask.\n", "date": "2024-06-06", "categories": []}
{"id": "2405.14141", "title": "ViHateT5: Enhancing Hate Speech Detection in Vietnamese With A Unified\n  Text-to-Text Transformer Model", "abstract": "  Recent advancements in hate speech detection (HSD) in Vietnamese have made\nsignificant progress, primarily attributed to the emergence of\ntransformer-based pre-trained language models, particularly those built on the\nBERT architecture. However, the necessity for specialized fine-tuned models has\nresulted in the complexity and fragmentation of developing a multitasking HSD\nsystem. Moreover, most current methodologies focus on fine-tuning general\npre-trained models, primarily trained on formal textual datasets like\nWikipedia, which may not accurately capture human behavior on online platforms.\nIn this research, we introduce ViHateT5, a T5-based model pre-trained on our\nproposed large-scale domain-specific dataset named VOZ-HSD. By harnessing the\npower of a text-to-text architecture, ViHateT5 can tackle multiple tasks using\na unified model and achieve state-of-the-art performance across all standard\nHSD benchmarks in Vietnamese. Our experiments also underscore the significance\nof label distribution in pre-training data on model efficacy. We provide our\nexperimental materials for research purposes, including the VOZ-HSD dataset,\npre-trained checkpoint, the unified HSD-multitask ViHateT5 model, and related\nsource code on GitHub publicly.\n", "date": "2024-05-23", "categories": []}
{"id": "2305.02637", "title": "Towards Weakly-Supervised Hate Speech Classification Across Datasets", "abstract": "  As pointed out by several scholars, current research on hate speech (HS)\nrecognition is characterized by unsystematic data creation strategies and\ndiverging annotation schemata. Subsequently, supervised-learning models tend to\ngeneralize poorly to datasets they were not trained on, and the performance of\nthe models trained on datasets labeled using different HS taxonomies cannot be\ncompared. To ease this problem, we propose applying extremely weak supervision\nthat only relies on the class name rather than on class samples from the\nannotated data. We demonstrate the effectiveness of a state-of-the-art\nweakly-supervised text classification model in various in-dataset and\ncross-dataset settings. Furthermore, we conduct an in-depth quantitative and\nqualitative analysis of the source of poor generalizability of HS\nclassification models.\n", "date": "2023-05-04", "categories": []}
{"id": "2402.15238", "title": "GPT-HateCheck: Can LLMs Write Better Functional Tests for Hate Speech\n  Detection?", "abstract": "  Online hate detection suffers from biases incurred in data sampling,\nannotation, and model pre-training. Therefore, measuring the averaged\nperformance over all examples in held-out test data is inadequate. Instead, we\nmust identify specific model weaknesses and be informed when it is more likely\nto fail. A recent proposal in this direction is HateCheck, a suite for testing\nfine-grained model functionalities on synthesized data generated using\ntemplates of the kind \"You are just a [slur] to me.\" However, despite enabling\nmore detailed diagnostic insights, the HateCheck test cases are often generic\nand have simplistic sentence structures that do not match the real-world data.\nTo address this limitation, we propose GPT-HateCheck, a framework to generate\nmore diverse and realistic functional tests from scratch by instructing large\nlanguage models (LLMs). We employ an additional natural language inference\n(NLI) model to verify the generations. Crowd-sourced annotation demonstrates\nthat the generated test cases are of high quality. Using the new functional\ntests, we can uncover model weaknesses that would be overlooked using the\noriginal HateCheck dataset.\n", "date": "2024-02-23", "categories": []}
{"id": "2405.13011", "title": "Unveiling Social Media Comments with a Novel Named Entity Recognition\n  System for Identity Groups", "abstract": "  While civilized users employ social media to stay informed and discuss daily\noccurrences, haters perceive these platforms as fertile ground for attacking\ngroups and individuals. The prevailing approach to counter this phenomenon\ninvolves detecting such attacks by identifying toxic language. Effective\nplatform measures aim to report haters and block their network access. In this\ncontext, employing hate speech detection methods aids in identifying these\nattacks amidst vast volumes of text, which are impossible for humans to analyze\nmanually. In our study, we expand upon the usual hate speech detection methods,\ntypically based on text classifiers, to develop a Named Entity Recognition\n(NER) System for Identity Groups. To achieve this, we created a dataset that\nallows extending a conventional NER to recognize identity groups. Consequently,\nour tool not only detects whether a sentence contains an attack but also tags\nthe sentence tokens corresponding to the mentioned group. Results indicate that\nthe model performs competitively in identifying groups with an average f1-score\nof 0.75, outperforming in identifying ethnicity attack spans with an f1-score\nof 0.80 compared to other identity groups. Moreover, the tool shows an\noutstanding generalization capability to minority classes concerning sexual\norientation and gender, achieving an f1-score of 0.77 and 0.72, respectively.\nWe tested the utility of our tool in a case study on social media, annotating\nand comparing comments from Facebook related to news mentioning identity\ngroups. The case study reveals differences in the types of attacks recorded,\neffectively detecting named entities related to the categories of the analyzed\nnews articles. Entities are accurately tagged within their categories, with a\nnegligible error rate for inter-category tagging.\n", "date": "2024-05-13", "categories": []}
{"id": "2405.13754", "title": "Grounding Toxicity in Real-World Events across Languages", "abstract": "  Social media conversations frequently suffer from toxicity, creating\nsignificant issues for users, moderators, and entire communities. Events in the\nreal world, like elections or conflicts, can initiate and escalate toxic\nbehavior online. Our study investigates how real-world events influence the\norigin and spread of toxicity in online discussions across various languages\nand regions. We gathered Reddit data comprising 4.5 million comments from 31\nthousand posts in six different languages (Dutch, English, German, Arabic,\nTurkish and Spanish). We target fifteen major social and political world events\nthat occurred between 2020 and 2023. We observe significant variations in\ntoxicity, negative sentiment, and emotion expressions across different events\nand language communities, showing that toxicity is a complex phenomenon in\nwhich many different factors interact and still need to be investigated. We\nwill release the data for further research along with our code.\n", "date": "2024-05-22", "categories": []}
{"id": "2403.16685", "title": "ToXCL: A Unified Framework for Toxic Speech Detection and Explanation", "abstract": "  The proliferation of online toxic speech is a pertinent problem posing\nthreats to demographic groups. While explicit toxic speech contains offensive\nlexical signals, implicit one consists of coded or indirect language.\nTherefore, it is crucial for models not only to detect implicit toxic speech\nbut also to explain its toxicity. This draws a unique need for unified\nframeworks that can effectively detect and explain implicit toxic speech. Prior\nworks mainly formulated the task of toxic speech detection and explanation as a\ntext generation problem. Nonetheless, models trained using this strategy can be\nprone to suffer from the consequent error propagation problem. Moreover, our\nexperiments reveal that the detection results of such models are much lower\nthan those that focus only on the detection task. To bridge these gaps, we\nintroduce ToXCL, a unified framework for the detection and explanation of\nimplicit toxic speech. Our model consists of three modules: a (i) Target Group\nGenerator to generate the targeted demographic group(s) of a given post; an\n(ii) Encoder-Decoder Model in which the encoder focuses on detecting implicit\ntoxic speech and is boosted by a (iii) Teacher Classifier via knowledge\ndistillation, and the decoder generates the necessary explanation. ToXCL\nachieves new state-of-the-art effectiveness, and outperforms baselines\nsignificantly.\n", "date": "2024-03-25", "categories": []}
{"id": "2405.09221", "title": "Bridging the gap in online hate speech detection: a comparative analysis\n  of BERT and traditional models for homophobic content identification on\n  X/Twitter", "abstract": "  Our study addresses a significant gap in online hate speech detection\nresearch by focusing on homophobia, an area often neglected in sentiment\nanalysis research. Utilising advanced sentiment analysis models, particularly\nBERT, and traditional machine learning methods, we developed a nuanced approach\nto identify homophobic content on X/Twitter. This research is pivotal due to\nthe persistent underrepresentation of homophobia in detection models. Our\nfindings reveal that while BERT outperforms traditional methods, the choice of\nvalidation technique can impact model performance. This underscores the\nimportance of contextual understanding in detecting nuanced hate speech. By\nreleasing the largest open-source labelled English dataset for homophobia\ndetection known to us, an analysis of various models' performance and our\nstrongest BERT-based model, we aim to enhance online safety and inclusivity.\nFuture work will extend to broader LGBTQIA+ hate speech detection, addressing\nthe challenges of sourcing diverse datasets. Through this endeavour, we\ncontribute to the larger effort against online hate, advocating for a more\ninclusive digital landscape. Our study not only offers insights into the\neffective detection of homophobic content by improving on previous research\nresults, but it also lays groundwork for future advancements in hate speech\nanalysis.\n", "date": "2024-05-15", "categories": []}
{"id": "2405.06128", "title": "Enhanced Multimodal Content Moderation of Children's Videos using\n  Audiovisual Fusion", "abstract": "  Due to the rise in video content creation targeted towards children, there is\na need for robust content moderation schemes for video hosting platforms. A\nvideo that is visually benign may include audio content that is inappropriate\nfor young children while being impossible to detect with a unimodal content\nmoderation system. Popular video hosting platforms for children such as YouTube\nKids still publish videos which contain audio content that is not conducive to\na child's healthy behavioral and physical development. A robust classification\nof malicious videos requires audio representations in addition to video\nfeatures. However, recent content moderation approaches rarely employ\nmultimodal architectures that explicitly consider non-speech audio cues. To\naddress this, we present an efficient adaptation of CLIP (Contrastive\nLanguage-Image Pre-training) that can leverage contextual audio cues for\nenhanced content moderation. We incorporate 1) the audio modality and 2) prompt\nlearning, while keeping the backbone modules of each modality frozen. We\nconduct our experiments on a multimodal version of the MOB (Malicious or\nBenign) dataset in supervised and few-shot settings.\n", "date": "2024-05-09", "categories": []}
{"id": "2403.12403", "title": "Towards Interpretable Hate Speech Detection using Large Language\n  Model-extracted Rationales", "abstract": "  Although social media platforms are a prominent arena for users to engage in\ninterpersonal discussions and express opinions, the facade and anonymity\noffered by social media may allow users to spew hate speech and offensive\ncontent. Given the massive scale of such platforms, there arises a need to\nautomatically identify and flag instances of hate speech. Although several hate\nspeech detection methods exist, most of these black-box methods are not\ninterpretable or explainable by design. To address the lack of\ninterpretability, in this paper, we propose to use state-of-the-art Large\nLanguage Models (LLMs) to extract features in the form of rationales from the\ninput text, to train a base hate speech classifier, thereby enabling faithful\ninterpretability by design. Our framework effectively combines the textual\nunderstanding capabilities of LLMs and the discriminative power of\nstate-of-the-art hate speech classifiers to make these classifiers faithfully\ninterpretable. Our comprehensive evaluation on a variety of English language\nsocial media hate speech datasets demonstrate: (1) the goodness of the\nLLM-extracted rationales, and (2) the surprising retention of detector\nperformance even after training to ensure interpretability. All code and data\nwill be made available at https://github.com/AmritaBh/shield.\n", "date": "2024-03-19", "categories": []}
{"id": "2405.03794", "title": "Detecting Anti-Semitic Hate Speech using Transformer-based Large\n  Language Models", "abstract": "  Academic researchers and social media entities grappling with the\nidentification of hate speech face significant challenges, primarily due to the\nvast scale of data and the dynamic nature of hate speech. Given the ethical and\npractical limitations of large predictive models like ChatGPT in directly\naddressing such sensitive issues, our research has explored alternative\nadvanced transformer-based and generative AI technologies since 2019.\nSpecifically, we developed a new data labeling technique and established a\nproof of concept targeting anti-Semitic hate speech, utilizing a variety of\ntransformer models such as BERT (arXiv:1810.04805), DistillBERT\n(arXiv:1910.01108), RoBERTa (arXiv:1907.11692), and LLaMA-2 (arXiv:2307.09288),\ncomplemented by the LoRA fine-tuning approach (arXiv:2106.09685). This paper\ndelineates and evaluates the comparative efficacy of these cutting-edge methods\nin tackling the intricacies of hate speech detection, highlighting the need for\nresponsible and carefully managed AI applications within sensitive contexts.\n", "date": "2024-05-06", "categories": []}
{"id": "2305.14081", "title": "How to Solve Few-Shot Abusive Content Detection Using the Data We\n  Actually Have", "abstract": "  Due to the broad range of social media platforms, the requirements of abusive\nlanguage detection systems are varied and ever-changing. Already a large set of\nannotated corpora with different properties and label sets were created, such\nas hate or misogyny detection, but the form and targets of abusive speech are\nconstantly evolving. Since, the annotation of new corpora is expensive, in this\nwork we leverage datasets we already have, covering a wide range of tasks\nrelated to abusive language detection. Our goal is to build models cheaply for\na new target label set and/or language, using only a few training examples of\nthe target domain. We propose a two-step approach: first we train our model in\na multitask fashion. We then carry out few-shot adaptation to the target\nrequirements. Our experiments show that using already existing datasets and\nonly a few-shots of the target task the performance of models improve both\nmonolingually and across languages. Our analysis also shows that our models\nacquire a general understanding of abusive language, since they improve the\nprediction of labels which are present only in the target dataset and can\nbenefit from knowledge about labels which are not directly used for the target\ntask.\n", "date": "2023-05-23", "categories": []}
{"id": "2405.01577", "title": "HateTinyLLM : Hate Speech Detection Using Tiny Large Language Models", "abstract": "  Hate speech encompasses verbal, written, or behavioral communication that\ntargets derogatory or discriminatory language against individuals or groups\nbased on sensitive characteristics. Automated hate speech detection plays a\ncrucial role in curbing its propagation, especially across social media\nplatforms. Various methods, including recent advancements in deep learning,\nhave been devised to address this challenge. In this study, we introduce\nHateTinyLLM, a novel framework based on fine-tuned decoder-only tiny large\nlanguage models (tinyLLMs) for efficient hate speech detection. Our\nexperimental findings demonstrate that the fine-tuned HateTinyLLM outperforms\nthe pretrained mixtral-7b model by a significant margin. We explored various\ntiny LLMs, including PY007/TinyLlama-1.1B-step-50K-105b, Microsoft/phi-2, and\nfacebook/opt-1.3b, and fine-tuned them using LoRA and adapter methods. Our\nobservations indicate that all LoRA-based fine-tuned models achieved over 80\\%\naccuracy.\n", "date": "2024-04-26", "categories": []}
{"id": "2405.01842", "title": "SGHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource\n  Languages of Singapore", "abstract": "  To address the limitations of current hate speech detection models, we\nintroduce \\textsf{SGHateCheck}, a novel framework designed for the linguistic\nand cultural context of Singapore and Southeast Asia. It extends the functional\ntesting approach of HateCheck and MHC, employing large language models for\ntranslation and paraphrasing into Singapore's main languages, and refining\nthese with native annotators. \\textsf{SGHateCheck} reveals critical flaws in\nstate-of-the-art models, highlighting their inadequacy in sensitive content\nmoderation. This work aims to foster the development of more effective hate\nspeech detection tools for diverse linguistic environments, particularly for\nSingapore and Southeast Asia contexts.\n", "date": "2024-05-03", "categories": []}
{"id": "2404.18180", "title": "EkoHate: Abusive Language and Hate Speech Detection for Code-switched\n  Political Discussions on Nigerian Twitter", "abstract": "  Nigerians have a notable online presence and actively discuss political and\ntopical matters. This was particularly evident throughout the 2023 general\nelection, where Twitter was used for campaigning, fact-checking and\nverification, and even positive and negative discourse. However, little or none\nhas been done in the detection of abusive language and hate speech in Nigeria.\nIn this paper, we curated code-switched Twitter data directed at three\nmusketeers of the governorship election on the most populous and economically\nvibrant state in Nigeria; Lagos state, with the view to detect offensive speech\nin political discussions. We developed EkoHate -- an abusive language and hate\nspeech dataset for political discussions between the three candidates and their\nfollowers using a binary (normal vs offensive) and fine-grained four-label\nannotation scheme. We analysed our dataset and provided an empirical evaluation\nof state-of-the-art methods across both supervised and cross-lingual transfer\nlearning settings. In the supervised setting, our evaluation results in both\nbinary and four-label annotation schemes show that we can achieve 95.1 and 70.3\nF1 points respectively. Furthermore, we show that our dataset adequately\ntransfers very well to three publicly available offensive datasets (OLID,\nHateUS2020, and FountaHate), generalizing to political discussions in other\nregions like the US.\n", "date": "2024-04-28", "categories": []}
{"id": "2404.18726", "title": "The Constant in HATE: Analyzing Toxicity in Reddit across Topics and\n  Languages", "abstract": "  Toxic language remains an ongoing challenge on social media platforms,\npresenting significant issues for users and communities. This paper provides a\ncross-topic and cross-lingual analysis of toxicity in Reddit conversations. We\ncollect 1.5 million comment threads from 481 communities in six languages:\nEnglish, German, Spanish, Turkish,Arabic, and Dutch, covering 80 topics such as\nCulture, Politics, and News. We thoroughly analyze how toxicity spikes within\ndifferent communities in relation to specific topics. We observe consistent\npatterns of increased toxicity across languages for certain topics, while also\nnoting significant variations within specific language communities.\n", "date": "2024-04-29", "categories": []}
{"id": "2304.10619", "title": "\"HOT\" ChatGPT: The promise of ChatGPT in detecting and discriminating\n  hateful, offensive, and toxic comments on social media", "abstract": "  Harmful content is pervasive on social media, poisoning online communities\nand negatively impacting participation. A common approach to address this issue\nis to develop detection models that rely on human annotations. However, the\ntasks required to build such models expose annotators to harmful and offensive\ncontent and may require significant time and cost to complete. Generative AI\nmodels have the potential to understand and detect harmful content. To\ninvestigate this potential, we used ChatGPT and compared its performance with\nMTurker annotations for three frequently discussed concepts related to harmful\ncontent: Hateful, Offensive, and Toxic (HOT). We designed five prompts to\ninteract with ChatGPT and conducted four experiments eliciting HOT\nclassifications. Our results show that ChatGPT can achieve an accuracy of\napproximately 80% when compared to MTurker annotations. Specifically, the model\ndisplays a more consistent classification for non-HOT comments than HOT\ncomments compared to human annotations. Our findings also suggest that ChatGPT\nclassifications align with provided HOT definitions, but ChatGPT classifies\n\"hateful\" and \"offensive\" as subsets of \"toxic.\" Moreover, the choice of\nprompts used to interact with ChatGPT impacts its performance. Based on these\nin-sights, our study provides several meaningful implications for employing\nChatGPT to detect HOT content, particularly regarding the reliability and\nconsistency of its performance, its understand-ing and reasoning of the HOT\nconcept, and the impact of prompts on its performance. Overall, our study\nprovides guidance about the potential of using generative AI models to moderate\nlarge volumes of user-generated content on social media.\n", "date": "2023-04-20", "categories": []}
{"id": "2307.16811", "title": "DoDo Learning: DOmain-DemOgraphic Transfer in Language Models for\n  Detecting Abuse Targeted at Public Figures", "abstract": "  Public figures receive a disproportionate amount of abuse on social media,\nimpacting their active participation in public life. Automated systems can\nidentify abuse at scale but labelling training data is expensive, complex and\npotentially harmful. So, it is desirable that systems are efficient and\ngeneralisable, handling both shared and specific aspects of online abuse. We\nexplore the dynamics of cross-group text classification in order to understand\nhow well classifiers trained on one domain or demographic can transfer to\nothers, with a view to building more generalisable abuse classifiers. We\nfine-tune language models to classify tweets targeted at public figures across\nDOmains (sport and politics) and DemOgraphics (women and men) using our novel\nDODO dataset, containing 28,000 labelled entries, split equally across four\ndomain-demographic pairs. We find that (i) small amounts of diverse data are\nhugely beneficial to generalisation and model adaptation; (ii) models transfer\nmore easily across demographics but models trained on cross-domain data are\nmore generalisable; (iii) some groups contribute more to generalisability than\nothers; and (iv) dataset similarity is a signal of transferability.\n", "date": "2023-07-31", "categories": []}
{"id": "2404.11752", "title": "Mapping Violence: Developing an Extensive Framework to Build a Bangla\n  Sectarian Expression Dataset from Social Media Interactions", "abstract": "  Communal violence in online forums has become extremely prevalent in South\nAsia, where many communities of different cultures coexist and share resources.\nThese societies exhibit a phenomenon characterized by strong bonds within their\nown groups and animosity towards others, leading to conflicts that frequently\nescalate into violent confrontations. To address this issue, we have developed\nthe first comprehensive framework for the automatic detection of communal\nviolence markers in online Bangla content accompanying the largest collection\n(13K raw sentences) of social media interactions that fall under the definition\nof four major violence class and their 16 coarse expressions. Our workflow\nintroduces a 7-step expert annotation process incorporating insights from\nsocial scientists, linguists, and psychologists. By presenting data statistics\nand benchmarking performance using this dataset, we have determined that, aside\nfrom the category of Non-communal violence, Religio-communal violence is\nparticularly pervasive in Bangla text. Moreover, we have substantiated the\neffectiveness of fine-tuning language models in identifying violent comments by\nconducting preliminary benchmarking on the state-of-the-art Bangla deep\nlearning model.\n", "date": "2024-04-17", "categories": []}
{"id": "2404.12042", "title": "Exploring Boundaries and Intensities in Offensive and Hate Speech:\n  Unveiling the Complex Spectrum of Social Media Discourse", "abstract": "  The prevalence of digital media and evolving sociopolitical dynamics have\nsignificantly amplified the dissemination of hateful content. Existing studies\nmainly focus on classifying texts into binary categories, often overlooking the\ncontinuous spectrum of offensiveness and hatefulness inherent in the text. In\nthis research, we present an extensive benchmark dataset for Amharic,\ncomprising 8,258 tweets annotated for three distinct tasks: category\nclassification, identification of hate targets, and rating offensiveness and\nhatefulness intensities. Our study highlights that a considerable majority of\ntweets belong to the less offensive and less hate intensity levels,\nunderscoring the need for early interventions by stakeholders. The prevalence\nof ethnic and political hatred targets, with significant overlaps in our\ndataset, emphasizes the complex relationships within Ethiopia's sociopolitical\nlandscape. We build classification and regression models and investigate the\nefficacy of models in handling these tasks. Our results reveal that hate and\noffensive speech can not be addressed by a simplistic binary classification,\ninstead manifesting as variables across a continuous range of values. The\nAfro-XLMR-large model exhibits the best performances achieving F1-scores of\n75.30%, 70.59%, and 29.42% for the category, target, and regression tasks,\nrespectively. The 80.22% correlation coefficient of the Afro-XLMR-large model\nindicates strong alignments.\n", "date": "2024-04-18", "categories": []}
{"id": "2305.09846", "title": "CPL-NoViD: Context-Aware Prompt-based Learning for Norm Violation\n  Detection in Online Communities", "abstract": "  Detecting norm violations in online communities is critical to maintaining\nhealthy and safe spaces for online discussions. Existing machine learning\napproaches often struggle to adapt to the diverse rules and interpretations\nacross different communities due to the inherent challenges of fine-tuning\nmodels for such context-specific tasks. In this paper, we introduce\nContext-aware Prompt-based Learning for Norm Violation Detection (CPL-NoViD), a\nnovel method that employs prompt-based learning to detect norm violations\nacross various types of rules. CPL-NoViD outperforms the baseline by\nincorporating context through natural language prompts and demonstrates\nimproved performance across different rule types. Significantly, it not only\nexcels in cross-rule-type and cross-community norm violation detection but also\nexhibits adaptability in few-shot learning scenarios. Most notably, it\nestablishes a new state-of-the-art in norm violation detection, surpassing\nexisting benchmarks. Our work highlights the potential of prompt-based learning\nfor context-sensitive norm violation detection and paves the way for future\nresearch on more adaptable, context-aware models to better support online\ncommunity moderators.\n", "date": "2023-05-16", "categories": []}
{"id": "2404.03048", "title": "Decentralised Moderation for Interoperable Social Networks: A\n  Conversation-based Approach for Pleroma and the Fediverse", "abstract": "  The recent development of decentralised and interoperable social networks\n(such as the \"fediverse\") creates new challenges for content moderators. This\nis because millions of posts generated on one server can easily \"spread\" to\nanother, even if the recipient server has very different moderation policies.\nAn obvious solution would be to leverage moderation tools to automatically tag\n(and filter) posts that contravene moderation policies, e.g. related to toxic\nspeech. Recent work has exploited the conversational context of a post to\nimprove this automatic tagging, e.g. using the replies to a post to help\nclassify if it contains toxic speech. This has shown particular potential in\nenvironments with large training sets that contain complete conversations.\nThis, however, creates challenges in a decentralised context, as a single\nconversation may be fragmented across multiple servers. Thus, each server only\nhas a partial view of an entire conversation because conversations are often\nfederated across servers in a non-synchronized fashion. To address this, we\npropose a decentralised conversation-aware content moderation approach suitable\nfor the fediverse. Our approach employs a graph deep learning model (GraphNLI)\ntrained locally on each server. The model exploits local data to train a model\nthat combines post and conversational information captured through random walks\nto detect toxicity. We evaluate our approach with data from Pleroma, a major\ndecentralised and interoperable micro-blogging network containing 2 million\nconversations. Our model effectively detects toxicity on larger instances,\nexclusively trained using their local post information (0.8837 macro-F1). Our\napproach has considerable scope to improve moderation in decentralised and\ninteroperable social networks such as Pleroma or Mastodon.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.10857", "title": "D3CODE: Disentangling Disagreements in Data across Cultures on\n  Offensiveness Detection and Evaluation", "abstract": "  While human annotations play a crucial role in language technologies,\nannotator subjectivity has long been overlooked in data collection. Recent\nstudies that have critically examined this issue are often situated in the\nWestern context, and solely document differences across age, gender, or racial\ngroups. As a result, NLP research on subjectivity have overlooked the fact that\nindividuals within demographic groups may hold diverse values, which can\ninfluence their perceptions beyond their group norms. To effectively\nincorporate these considerations into NLP pipelines, we need datasets with\nextensive parallel annotations from various social and cultural groups. In this\npaper we introduce the \\dataset dataset: a large-scale cross-cultural dataset\nof parallel annotations for offensive language in over 4.5K sentences annotated\nby a pool of over 4k annotators, balanced across gender and age, from across 21\ncountries, representing eight geo-cultural regions. The dataset contains\nannotators' moral values captured along six moral foundations: care, equality,\nproportionality, authority, loyalty, and purity. Our analyses reveal\nsubstantial regional variations in annotators' perceptions that are shaped by\nindividual moral values, offering crucial insights for building pluralistic,\nculturally sensitive NLP models.\n", "date": "2024-04-16", "categories": []}
{"id": "2404.11036", "title": "Cross-Platform Hate Speech Detection with Weakly Supervised Causal\n  Disentanglement", "abstract": "  Content moderation faces a challenging task as social media's ability to\nspread hate speech contrasts with its role in promoting global connectivity.\nWith rapidly evolving slang and hate speech, the adaptability of conventional\ndeep learning to the fluid landscape of online dialogue remains limited. In\nresponse, causality inspired disentanglement has shown promise by segregating\nplatform specific peculiarities from universal hate indicators. However, its\ndependency on available ground truth target labels for discerning these nuances\nfaces practical hurdles with the incessant evolution of platforms and the\nmutable nature of hate speech. Using confidence based reweighting and\ncontrastive regularization, this study presents HATE WATCH, a novel framework\nof weakly supervised causal disentanglement that circumvents the need for\nexplicit target labeling and effectively disentangles input features into\ninvariant representations of hate. Empirical validation across platforms two\nwith target labels and two without positions HATE WATCH as a novel method in\ncross platform hate speech detection with superior performance. HATE WATCH\nadvances scalable content moderation techniques towards developing safer online\ncommunities.\n", "date": "2024-04-17", "categories": []}
{"id": "2402.17014", "title": "Z-AGI Labs at ClimateActivism 2024: Stance and Hate Event Detection on\n  Social Media", "abstract": "  In the digital realm, rich data serves as a crucial source of insights into\nthe complexities of social, political, and economic landscapes. Addressing the\ngrowing need for high-quality information on events and the imperative to\ncombat hate speech, this research led to the establishment of the Shared Task\non Climate Activism Stance and Hate Event Detection at CASE 2024. Focused on\nclimate activists contending with hate speech on social media, our study\ncontributes to hate speech identification from tweets. Analyzing three\nsub-tasks - Hate Speech Detection (Sub-task A), Targets of Hate Speech\nIdentification (Sub-task B), and Stance Detection (Sub-task C) - Team Z-AGI\nLabs evaluated various models, including LSTM, Xgboost, and LGBM based on\nTf-Idf. Results unveiled intriguing variations, with Catboost excelling in\nSubtask-B (F1: 0.5604) and Subtask-C (F1: 0.7081), while LGBM emerged as the\ntop-performing model for Subtask-A (F1: 0.8684). This research provides\nvaluable insights into the suitability of classical machine learning models for\nclimate hate speech and stance detection, aiding informed model selection for\nrobust mechanisms.\n", "date": "2024-02-26", "categories": []}
{"id": "2404.05540", "title": "OPSD: an Offensive Persian Social media Dataset and its baseline\n  evaluations", "abstract": "  The proliferation of hate speech and offensive comments on social media has\nbecome increasingly prevalent due to user activities. Such comments can have\ndetrimental effects on individuals' psychological well-being and social\nbehavior. While numerous datasets in the English language exist in this domain,\nfew equivalent resources are available for Persian language. To address this\ngap, this paper introduces two offensive datasets. The first dataset comprises\nannotations provided by domain experts, while the second consists of a large\ncollection of unlabeled data obtained through web crawling for unsupervised\nlearning purposes. To ensure the quality of the former dataset, a meticulous\nthree-stage labeling process was conducted, and kappa measures were computed to\nassess inter-annotator agreement. Furthermore, experiments were performed on\nthe dataset using state-of-the-art language models, both with and without\nemploying masked language modeling techniques, as well as machine learning\nalgorithms, in order to establish the baselines for the dataset using\ncontemporary cutting-edge approaches. The obtained F1-scores for the\nthree-class and two-class versions of the dataset were 76.9% and 89.9% for\nXLM-RoBERTa, respectively.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.03686", "title": "Securing Social Spaces: Harnessing Deep Learning to Eradicate\n  Cyberbullying", "abstract": "  In today's digital world, cyberbullying is a serious problem that can harm\nthe mental and physical health of people who use social media. This paper\nexplains just how serious cyberbullying is and how it really affects\nindi-viduals exposed to it. It also stresses how important it is to find better\nways to detect cyberbullying so that online spaces can be safer. Plus, it talks\nabout how making more accurate tools to spot cyberbullying will be really\nhelpful in the future. Our paper introduces a deep learning-based ap-proach,\nprimarily employing BERT and BiLSTM architectures, to effective-ly address\ncyberbullying. This approach is designed to analyse large vol-umes of posts and\npredict potential instances of cyberbullying in online spaces. Our results\ndemonstrate the superiority of the hateBERT model, an extension of BERT focused\non hate speech detection, among the five mod-els, achieving an accuracy rate of\n89.16%. This research is a significant con-tribution to \"Computational\nIntelligence for Social Transformation,\" prom-ising a safer and more inclusive\ndigital landscape.\n", "date": "2024-04-01", "categories": []}
{"id": "2308.16705", "title": "Exploring Cross-Cultural Differences in English Hate Speech Annotations:\n  From Dataset Construction to Analysis", "abstract": "  Warning: this paper contains content that may be offensive or upsetting.\n  Most hate speech datasets neglect the cultural diversity within a single\nlanguage, resulting in a critical shortcoming in hate speech detection. To\naddress this, we introduce CREHate, a CRoss-cultural English Hate speech\ndataset. To construct CREHate, we follow a two-step procedure: 1) cultural post\ncollection and 2) cross-cultural annotation. We sample posts from the SBIC\ndataset, which predominantly represents North America, and collect posts from\nfour geographically diverse English-speaking countries (Australia, United\nKingdom, Singapore, and South Africa) using culturally hateful keywords we\nretrieve from our survey. Annotations are collected from the four countries\nplus the United States to establish representative labels for each country. Our\nanalysis highlights statistically significant disparities across countries in\nhate speech annotations. Only 56.2% of the posts in CREHate achieve consensus\namong all countries, with the highest pairwise label difference rate of 26%.\nQualitative analysis shows that label disagreement occurs mostly due to\ndifferent interpretations of sarcasm and the personal bias of annotators on\ndivisive topics. Lastly, we evaluate large language models (LLMs) under a\nzero-shot setting and show that current LLMs tend to show higher accuracies on\nAnglosphere country labels in CREHate. Our dataset and codes are available at:\nhttps://github.com/nlee0212/CREHate\n", "date": "2023-08-31", "categories": []}
{"id": "2404.02013", "title": "Breaking the Silence Detecting and Mitigating Gendered Abuse in Hindi,\n  Tamil, and Indian English Online Spaces", "abstract": "  Online gender-based harassment is a widespread issue limiting the free\nexpression and participation of women and marginalized genders in digital\nspaces. Detecting such abusive content can enable platforms to curb this\nmenace. We participated in the Gendered Abuse Detection in Indic Languages\nshared task at ICON2023 that provided datasets of annotated Twitter posts in\nEnglish, Hindi and Tamil for building classifiers to identify gendered abuse.\nOur team CNLP-NITS-PP developed an ensemble approach combining CNN and BiLSTM\nnetworks that can effectively model semantic and sequential patterns in textual\ndata. The CNN captures localized features indicative of abusive language\nthrough its convolution filters applied on embedded input text. To determine\ncontext-based offensiveness, the BiLSTM analyzes this sequence for dependencies\namong words and phrases. Multiple variations were trained using FastText and\nGloVe word embeddings for each language dataset comprising over 7,600\ncrowdsourced annotations across labels for explicit abuse, targeted minority\nattacks and general offences. The validation scores showed strong performance\nacross f1-measures, especially for English 0.84. Our experiments reveal how\ncustomizing embeddings and model hyperparameters can improve detection\ncapability. The proposed architecture ranked 1st in the competition, proving\nits ability to handle real-world noisy text with code-switching. This technique\nhas a promising scope as platforms aim to combat cyber harassment facing Indic\nlanguage internet users. Our Code is at\nhttps://github.com/advaithavetagiri/CNLP-NITS-PP\n", "date": "2024-04-02", "categories": []}
{"id": "2404.02681", "title": "PejorativITy: Disambiguating Pejorative Epithets to Improve Misogyny\n  Detection in Italian Tweets", "abstract": "  Misogyny is often expressed through figurative language. Some neutral words\ncan assume a negative connotation when functioning as pejorative epithets.\nDisambiguating the meaning of such terms might help the detection of misogyny.\nIn order to address such task, we present PejorativITy, a novel corpus of 1,200\nmanually annotated Italian tweets for pejorative language at the word level and\nmisogyny at the sentence level. We evaluate the impact of injecting information\nabout disambiguated words into a model targeting misogyny detection. In\nparticular, we explore two different approaches for injection: concatenation of\npejorative information and substitution of ambiguous words with univocal terms.\nOur experimental results, both on our corpus and on two popular benchmarks on\nItalian tweets, show that both approaches lead to a major classification\nimprovement, indicating that word sense disambiguation is a promising\npreliminary step for misogyny detection. Furthermore, we investigate LLMs'\nunderstanding of pejorative epithets by means of contextual word embeddings\nanalysis and prompting.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.01651", "title": "NLP Systems That Can't Tell Use from Mention Censor Counterspeech, but\n  Teaching the Distinction Helps", "abstract": "  The use of words to convey speaker's intent is traditionally distinguished\nfrom the `mention' of words for quoting what someone said, or pointing out\nproperties of a word. Here we show that computationally modeling this\nuse-mention distinction is crucial for dealing with counterspeech online.\nCounterspeech that refutes problematic content often mentions harmful language\nbut is not harmful itself (e.g., calling a vaccine dangerous is not the same as\nexpressing disapproval of someone for calling vaccines dangerous). We show that\neven recent language models fail at distinguishing use from mention, and that\nthis failure propagates to two key downstream tasks: misinformation and hate\nspeech detection, resulting in censorship of counterspeech. We introduce\nprompting mitigations that teach the use-mention distinction, and show they\nreduce these errors. Our work highlights the importance of the use-mention\ndistinction for NLP and CSS and offers ways to address it.\n", "date": "2024-04-02", "categories": []}
{"id": "2309.06415", "title": "Down the Toxicity Rabbit Hole: A Novel Framework to Bias Audit Large\n  Language Models", "abstract": "  This paper makes three contributions. First, it presents a generalizable,\nnovel framework dubbed \\textit{toxicity rabbit hole} that iteratively elicits\ntoxic content from a wide suite of large language models. Spanning a set of\n1,266 identity groups, we first conduct a bias audit of \\texttt{PaLM 2}\nguardrails presenting key insights. Next, we report generalizability across\nseveral other models. Through the elicited toxic content, we present a broad\nanalysis with a key emphasis on racism, antisemitism, misogyny, Islamophobia,\nhomophobia, and transphobia. Finally, driven by concrete examples, we discuss\npotential ramifications.\n", "date": "2023-09-08", "categories": []}
{"id": "2403.18314", "title": "Chinese Offensive Language Detection:Current Status and Future\n  Directions", "abstract": "  Despite the considerable efforts being made to monitor and regulate\nuser-generated content on social media platforms, the pervasiveness of\noffensive language, such as hate speech or cyberbullying, in the digital space\nremains a significant challenge. Given the importance of maintaining a\ncivilized and respectful online environment, there is an urgent and growing\nneed for automatic systems capable of detecting offensive speech in real time.\nHowever, developing effective systems for processing languages such as Chinese\npresents a significant challenge, owing to the language's complex and nuanced\nnature, which makes it difficult to process automatically. This paper provides\na comprehensive overview of offensive language detection in Chinese, examining\ncurrent benchmarks and approaches and highlighting specific models and tools\nfor addressing the unique challenges of detecting offensive language in this\ncomplex language. The primary objective of this survey is to explore the\nexisting techniques and identify potential avenues for further research that\ncan address the cultural and linguistic complexities of Chinese.\n", "date": "2024-03-27", "categories": []}
{"id": "2404.00303", "title": "A Comprehensive Study on NLP Data Augmentation for Hate Speech\n  Detection: Legacy Methods, BERT, and LLMs", "abstract": "  The surge of interest in data augmentation within the realm of NLP has been\ndriven by the need to address challenges posed by hate speech domains, the\ndynamic nature of social media vocabulary, and the demands for large-scale\nneural networks requiring extensive training data. However, the prevalent use\nof lexical substitution in data augmentation has raised concerns, as it may\ninadvertently alter the intended meaning, thereby impacting the efficacy of\nsupervised machine learning models. In pursuit of suitable data augmentation\nmethods, this study explores both established legacy approaches and\ncontemporary practices such as Large Language Models (LLM), including GPT in\nHate Speech detection. Additionally, we propose an optimized utilization of\nBERT-based encoder models with contextual cosine similarity filtration,\nexposing significant limitations in prior synonym substitution methods. Our\ncomparative analysis encompasses five popular augmentation techniques: WordNet\nand Fast-Text synonym replacement, Back-translation, BERT-mask contextual\naugmentation, and LLM. Our analysis across five benchmarked datasets revealed\nthat while traditional methods like back-translation show low label alteration\nrates (0.3-1.5%), and BERT-based contextual synonym replacement offers sentence\ndiversity but at the cost of higher label alteration rates (over 6%). Our\nproposed BERT-based contextual cosine similarity filtration markedly reduced\nlabel alteration to just 0.05%, demonstrating its efficacy in 0.7% higher F1\nperformance. However, augmenting data with GPT-3 not only avoided overfitting\nwith up to sevenfold data increase but also improved embedding space coverage\nby 15% and classification F1 score by 1.4% over traditional methods, and by\n0.8% over our method.\n", "date": "2024-03-30", "categories": []}
{"id": "2403.20103", "title": "NLP for Counterspeech against Hate: A Survey and How-To Guide", "abstract": "  In recent years, counterspeech has emerged as one of the most promising\nstrategies to fight online hate. These non-escalatory responses tackle online\nabuse while preserving the freedom of speech of the users, and can have a\ntangible impact in reducing online and offline violence. Recently, there has\nbeen growing interest from the Natural Language Processing (NLP) community in\naddressing the challenges of analysing, collecting, classifying, and\nautomatically generating counterspeech, to reduce the huge burden of manually\nproducing it. In particular, researchers have taken different directions in\naddressing these challenges, thus providing a variety of related tasks and\nresources. In this paper, we provide a guide for doing research on\ncounterspeech, by describing - with detailed examples - the steps to undertake,\nand providing best practices that can be learnt from the NLP studies on this\ntopic. Finally, we discuss open challenges and future directions of\ncounterspeech research in NLP.\n", "date": "2024-03-29", "categories": []}
{"id": "2212.00851", "title": "SOLD: Sinhala Offensive Language Dataset", "abstract": "  The widespread of offensive content online, such as hate speech and\ncyber-bullying, is a global phenomenon. This has sparked interest in the\nartificial intelligence (AI) and natural language processing (NLP) communities,\nmotivating the development of various systems trained to detect potentially\nharmful content automatically. These systems require annotated datasets to\ntrain the machine learning (ML) models. However, with a few notable exceptions,\nmost datasets on this topic have dealt with English and a few other\nhigh-resource languages. As a result, the research in offensive language\nidentification has been limited to these languages. This paper addresses this\ngap by tackling offensive language identification in Sinhala, a low-resource\nIndo-Aryan language spoken by over 17 million people in Sri Lanka. We introduce\nthe Sinhala Offensive Language Dataset (SOLD) and present multiple experiments\non this dataset. SOLD is a manually annotated dataset containing 10,000 posts\nfrom Twitter annotated as offensive and not offensive at both sentence-level\nand token-level, improving the explainability of the ML models. SOLD is the\nfirst large publicly available offensive language dataset compiled for Sinhala.\nWe also introduce SemiSOLD, a larger dataset containing more than 145,000\nSinhala tweets, annotated following a semi-supervised approach.\n", "date": "2022-12-01", "categories": []}
{"id": "2403.19559", "title": "Improving Adversarial Data Collection by Supporting Annotators: Lessons\n  from GAHD, a German Hate Speech Dataset", "abstract": "  Hate speech detection models are only as good as the data they are trained\non. Datasets sourced from social media suffer from systematic gaps and biases,\nleading to unreliable models with simplistic decision boundaries. Adversarial\ndatasets, collected by exploiting model weaknesses, promise to fix this\nproblem. However, adversarial data collection can be slow and costly, and\nindividual annotators have limited creativity. In this paper, we introduce\nGAHD, a new German Adversarial Hate speech Dataset comprising ca.\\ 11k\nexamples. During data collection, we explore new strategies for supporting\nannotators, to create more diverse adversarial examples more efficiently and\nprovide a manual analysis of annotator disagreements for each strategy. Our\nexperiments show that the resulting dataset is challenging even for\nstate-of-the-art hate speech detection models, and that training on GAHD\nclearly improves model robustness. Further, we find that mixing multiple\nsupport strategies is most advantageous. We make GAHD publicly available at\nhttps://github.com/jagol/gahd.\n", "date": "2024-03-28", "categories": []}
{"id": "2308.10592", "title": "BAN-PL: a Novel Polish Dataset of Banned Harmful and Offensive Content\n  from Wykop.pl web service", "abstract": "  Since the Internet is flooded with hate, it is one of the main tasks for NLP\nexperts to master automated online content moderation. However, advancements in\nthis field require improved access to publicly available accurate and\nnon-synthetic datasets of social media content. For the Polish language, such\nresources are very limited. In this paper, we address this gap by presenting a\nnew open dataset of offensive social media content for the Polish language. The\ndataset comprises content from Wykop.pl, a popular online service often\nreferred to as the \"Polish Reddit\", reported by users and banned in the\ninternal moderation process. It contains a total of 691,662 posts and comments,\nevenly divided into two categories: \"harmful\" and \"neutral\" (\"non-harmful\").\nThe anonymized subset of the BAN-PL dataset consisting on 24,000 pieces (12,000\nfor each class), along with preprocessing scripts have been made publicly\navailable. Furthermore the paper offers valuable insights into real-life\ncontent moderation processes and delves into an analysis of linguistic features\nand content characteristics of the dataset. Moreover, a comprehensive\nanonymization procedure has been meticulously described and applied. The\nprevalent biases encountered in similar datasets, including post-moderation and\npre-selection biases, are also discussed.\n", "date": "2023-08-21", "categories": []}
{"id": "2403.15467", "title": "Don't be a Fool: Pooling Strategies in Offensive Language Detection from\n  User-Intended Adversarial Attacks", "abstract": "  Offensive language detection is an important task for filtering out abusive\nexpressions and improving online user experiences. However, malicious users\noften attempt to avoid filtering systems through the involvement of textual\nnoises. In this paper, we propose these evasions as user-intended adversarial\nattacks that insert special symbols or leverage the distinctive features of the\nKorean language. Furthermore, we introduce simple yet effective pooling\nstrategies in a layer-wise manner to defend against the proposed attacks,\nfocusing on the preceding layers not just the last layer to capture both\noffensiveness and token embeddings. We demonstrate that these pooling\nstrategies are more robust to performance degradation even when the attack rate\nis increased, without directly training of such patterns. Notably, we found\nthat models pre-trained on clean texts could achieve a comparable performance\nin detecting attacked offensive language, to models pre-trained on noisy texts\nby employing these pooling strategies.\n", "date": "2024-03-20", "categories": []}
{"id": "2307.12418", "title": "HateModerate: Testing Hate Speech Detectors against Content Moderation\n  Policies", "abstract": "  To protect users from massive hateful content, existing works studied\nautomated hate speech detection. Despite the existing efforts, one question\nremains: do automated hate speech detectors conform to social media content\npolicies? A platform's content policies are a checklist of content moderated by\nthe social media platform. Because content moderation rules are often uniquely\ndefined, existing hate speech datasets cannot directly answer this question.\n  This work seeks to answer this question by creating HateModerate, a dataset\nfor testing the behaviors of automated content moderators against content\npolicies. First, we engage 28 annotators and GPT in a six-step annotation\nprocess, resulting in a list of hateful and non-hateful test suites matching\neach of Facebook's 41 hate speech policies. Second, we test the performance of\nstate-of-the-art hate speech detectors against HateModerate, revealing\nsubstantial failures these models have in their conformity to the policies.\nThird, using HateModerate, we augment the training data of a top-downloaded\nhate detector on HuggingFace. We observe significant improvement in the models'\nconformity to content policies while having comparable scores on the original\ntest data. Our dataset and code can be found in the attachment.\n", "date": "2023-07-23", "categories": []}
{"id": "2305.06626", "title": "When the Majority is Wrong: Modeling Annotator Disagreement for\n  Subjective Tasks", "abstract": "  Though majority vote among annotators is typically used for ground truth\nlabels in natural language processing, annotator disagreement in tasks such as\nhate speech detection may reflect differences in opinion across groups, not\nnoise. Thus, a crucial problem in hate speech detection is determining whether\na statement is offensive to the demographic group that it targets, when that\ngroup may constitute a small fraction of the annotator pool. We construct a\nmodel that predicts individual annotator ratings on potentially offensive text\nand combines this information with the predicted target group of the text to\nmodel the opinions of target group members. We show gains across a range of\nmetrics, including raising performance over the baseline by 22% at predicting\nindividual annotators' ratings and by 33% at predicting variance among\nannotators, which provides a metric for model uncertainty downstream. We find\nthat annotator ratings can be predicted using their demographic information and\nopinions on online content, without the need to track identifying annotator IDs\nthat link each annotator to their ratings. We also find that use of\nnon-invasive survey questions on annotators' online experiences helps to\nmaximize privacy and minimize unnecessary collection of demographic information\nwhen predicting annotators' opinions.\n", "date": "2023-05-11", "categories": []}
{"id": "2401.11021", "title": "Analysis and Detection of Multilingual Hate Speech Using Transformer\n  Based Deep Learning", "abstract": "  Hate speech is harmful content that directly attacks or promotes hatred\nagainst members of groups or individuals based on actual or perceived aspects\nof identity, such as racism, religion, or sexual orientation. This can affect\nsocial life on social media platforms as hateful content shared through social\nmedia can harm both individuals and communities. As the prevalence of hate\nspeech increases online, the demand for automated detection as an NLP task is\nincreasing. In this work, the proposed method is using transformer-based model\nto detect hate speech in social media, like twitter, Facebook, WhatsApp,\nInstagram, etc. The proposed model is independent of languages and has been\ntested on Italian, English, German, Bengali. The Gold standard datasets were\ncollected from renowned researcher Zeerak Talat, Sara Tonelli, Melanie Siegel,\nand Rezaul Karim. The success rate of the proposed model for hate speech\ndetection is higher than the existing baseline and state-of-the-art models with\naccuracy in Bengali dataset is 89%, in English: 91%, in German dataset 91% and\nin Italian dataset it is 77%. The proposed algorithm shows substantial\nimprovement to the benchmark method.\n", "date": "2024-01-19", "categories": []}
{"id": "2403.11108", "title": "HarmPot: An Annotation Framework for Evaluating Offline Harm Potential\n  of Social Media Text", "abstract": "  In this paper, we discuss the development of an annotation schema to build\ndatasets for evaluating the offline harm potential of social media texts. We\ndefine \"harm potential\" as the potential for an online public post to cause\nreal-world physical harm (i.e., violence). Understanding that real-world\nviolence is often spurred by a web of triggers, often combining several online\ntactics and pre-existing intersectional fissures in the social milieu, to\nresult in targeted physical violence, we do not focus on any single divisive\naspect (i.e., caste, gender, religion, or other identities of the victim and\nperpetrators) nor do we focus on just hate speech or mis/dis-information.\nRather, our understanding of the intersectional causes of such triggers focuses\nour attempt at measuring the harm potential of online content, irrespective of\nwhether it is hateful or not. In this paper, we discuss the development of a\nframework/annotation schema that allows annotating the data with different\naspects of the text including its socio-political grounding and intent of the\nspeaker (as expressed through mood and modality) that together contribute to it\nbeing a trigger for offline harm. We also give a comparative analysis and\nmapping of our framework with some of the existing frameworks.\n", "date": "2024-03-17", "categories": []}
{"id": "2403.09709", "title": "Exploratory Data Analysis on Code-mixed Misogynistic Comments", "abstract": "  The problems of online hate speech and cyberbullying have significantly\nworsened since the increase in popularity of social media platforms such as\nYouTube and Twitter (X). Natural Language Processing (NLP) techniques have\nproven to provide a great advantage in automatic filtering such toxic content.\nWomen are disproportionately more likely to be victims of online abuse.\nHowever, there appears to be a lack of studies that tackle misogyny detection\nin under-resourced languages. In this short paper, we present a novel dataset\nof YouTube comments in mix-code Hinglish collected from YouTube videos which\nhave been weak labelled as `Misogynistic' and `Non-misogynistic'.\nPre-processing and Exploratory Data Analysis (EDA) techniques have been applied\non the dataset to gain insights on its characteristics. The process has\nprovided a better understanding of the dataset through sentiment scores, word\nclouds, etc.\n", "date": "2024-03-09", "categories": []}
{"id": "2403.10088", "title": "Intent-conditioned and Non-toxic Counterspeech Generation using\n  Multi-Task Instruction Tuning with RLAIF", "abstract": "  Counterspeech, defined as a response to mitigate online hate speech, is\nincreasingly used as a non-censorial solution. Addressing hate speech\neffectively involves dispelling the stereotypes, prejudices, and biases often\nsubtly implied in brief, single-sentence statements or abuses. These implicit\nexpressions challenge language models, especially in seq2seq tasks, as model\nperformance typically excels with longer contexts. Our study introduces CoARL,\na novel framework enhancing counterspeech generation by modeling the pragmatic\nimplications underlying social biases in hateful statements. CoARL's first two\nphases involve sequential multi-instruction tuning, teaching the model to\nunderstand intents, reactions, and harms of offensive statements, and then\nlearning task-specific low-rank adapter weights for generating\nintent-conditioned counterspeech. The final phase uses reinforcement learning\nto fine-tune outputs for effectiveness and non-toxicity. CoARL outperforms\nexisting benchmarks in intent-conditioned counterspeech generation, showing an\naverage improvement of 3 points in intent-conformity and 4 points in\nargument-quality metrics. Extensive human evaluation supports CoARL's efficacy\nin generating superior and more context-appropriate responses compared to\nexisting systems, including prominent LLMs like ChatGPT.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.08035", "title": "Harnessing Artificial Intelligence to Combat Online Hate: Exploring the\n  Challenges and Opportunities of Large Language Models in Hate Speech\n  Detection", "abstract": "  Large language models (LLMs) excel in many diverse applications beyond\nlanguage generation, e.g., translation, summarization, and sentiment analysis.\nOne intriguing application is in text classification. This becomes pertinent in\nthe realm of identifying hateful or toxic speech -- a domain fraught with\nchallenges and ethical dilemmas. In our study, we have two objectives: firstly,\nto offer a literature review revolving around LLMs as classifiers, emphasizing\ntheir role in detecting and classifying hateful or toxic content. Subsequently,\nwe explore the efficacy of several LLMs in classifying hate speech: identifying\nwhich LLMs excel in this task as well as their underlying attributes and\ntraining. Providing insight into the factors that contribute to an LLM\nproficiency (or lack thereof) in discerning hateful content. By combining a\ncomprehensive literature review with an empirical analysis, our paper strives\nto shed light on the capabilities and constraints of LLMs in the crucial domain\nof hate speech detection.\n", "date": "2024-03-12", "categories": []}
{"id": "2309.10561", "title": "A multimodal deep learning architecture for smoking detection with a\n  small data approach", "abstract": "  Introduction: Covert tobacco advertisements often raise regulatory measures.\nThis paper presents that artificial intelligence, particularly deep learning,\nhas great potential for detecting hidden advertising and allows unbiased,\nreproducible, and fair quantification of tobacco-related media content.\nMethods: We propose an integrated text and image processing model based on deep\nlearning, generative methods, and human reinforcement, which can detect smoking\ncases in both textual and visual formats, even with little available training\ndata. Results: Our model can achieve 74\\% accuracy for images and 98\\% for\ntext. Furthermore, our system integrates the possibility of expert intervention\nin the form of human reinforcement. Conclusions: Using the pre-trained\nmultimodal, image, and text processing models available through deep learning\nmakes it possible to detect smoking in different media even with few training\ndata.\n", "date": "2023-09-19", "categories": []}
{"id": "2403.05548", "title": "Monitoring the evolution of antisemitic discourse on extremist social\n  media using BERT", "abstract": "  Racism and intolerance on social media contribute to a toxic online\nenvironment which may spill offline to foster hatred, and eventually lead to\nphysical violence. That is the case with online antisemitism, the specific\ncategory of hatred considered in this study. Tracking antisemitic themes and\ntheir associated terminology over time in online discussions could help monitor\nthe sentiments of their participants and their evolution, and possibly offer\navenues for intervention that may prevent the escalation of hatred. Due to the\nlarge volume and constant evolution of online traffic, monitoring conversations\nmanually is impractical. Instead, we propose an automated method that extracts\nantisemitic themes and terminology from extremist social media over time and\ncaptures their evolution. Since supervised learning would be too limited for\nsuch a task, we created an unsupervised online machine learning approach that\nuses large language models to assess the contextual similarity of posts. The\nmethod clusters similar posts together, dividing, and creating additional\nclusters over time when sub-themes emerge from existing ones or new themes\nappear. The antisemitic terminology used within each theme is extracted from\nthe posts in each cluster. Our experiments show that our methodology\noutperforms existing baselines and demonstrates the kind of themes and\nsub-themes it discovers within antisemitic discourse along with their\nassociated terminology. We believe that our approach will be useful for\nmonitoring the evolution of all kinds of hatred beyond antisemitism on social\nplatforms.\n", "date": "2024-02-06", "categories": []}
{"id": "2403.01270", "title": "A comprehensive cross-language framework for harmful content detection\n  with the aid of sentiment analysis", "abstract": "  In today's digital world, social media plays a significant role in\nfacilitating communication and content sharing. However, the exponential rise\nin user-generated content has led to challenges in maintaining a respectful\nonline environment. In some cases, users have taken advantage of anonymity in\norder to use harmful language, which can negatively affect the user experience\nand pose serious social problems. Recognizing the limitations of manual\nmoderation, automatic detection systems have been developed to tackle this\nproblem. Nevertheless, several obstacles persist, including the absence of a\nuniversal definition for harmful language, inadequate datasets across\nlanguages, the need for detailed annotation guideline, and most importantly, a\ncomprehensive framework. This study aims to address these challenges by\nintroducing, for the first time, a detailed framework adaptable to any\nlanguage. This framework encompasses various aspects of harmful language\ndetection. A key component of the framework is the development of a general and\ndetailed annotation guideline. Additionally, the integration of sentiment\nanalysis represents a novel approach to enhancing harmful language detection.\nAlso, a definition of harmful language based on the review of different related\nconcepts is presented. To demonstrate the effectiveness of the proposed\nframework, its implementation in a challenging low-resource language is\nconducted. We collected a Persian dataset and applied the annotation guideline\nfor harmful detection and sentiment analysis. Next, we present baseline\nexperiments utilizing machine and deep learning methods to set benchmarks.\nResults prove the framework's high performance, achieving an accuracy of 99.4%\nin offensive language detection and 66.2% in sentiment analysis.\n", "date": "2024-03-02", "categories": []}
{"id": "2403.02121", "title": "Leveraging Weakly Annotated Data for Hate Speech Detection in Code-Mixed\n  Hinglish: A Feasibility-Driven Transfer Learning Approach with Large Language\n  Models", "abstract": "  The advent of Large Language Models (LLMs) has advanced the benchmark in\nvarious Natural Language Processing (NLP) tasks. However, large amounts of\nlabelled training data are required to train LLMs. Furthermore, data annotation\nand training are computationally expensive and time-consuming. Zero and\nfew-shot learning have recently emerged as viable options for labelling data\nusing large pre-trained models. Hate speech detection in mix-code low-resource\nlanguages is an active problem area where the use of LLMs has proven\nbeneficial. In this study, we have compiled a dataset of 100 YouTube comments,\nand weakly labelled them for coarse and fine-grained misogyny classification in\nmix-code Hinglish. Weak annotation was applied due to the labor-intensive\nannotation process. Zero-shot learning, one-shot learning, and few-shot\nlearning and prompting approaches have then been applied to assign labels to\nthe comments and compare them to human-assigned labels. Out of all the\napproaches, zero-shot classification using the Bidirectional Auto-Regressive\nTransformers (BART) large model and few-shot prompting using Generative\nPre-trained Transformer- 3 (ChatGPT-3) achieve the best results\n", "date": "2024-03-04", "categories": []}
{"id": "2403.02268", "title": "Subjective $\\textit{Isms}$? On the Danger of Conflating Hate and Offence\n  in Abusive Language Detection", "abstract": "  Natural language processing research has begun to embrace the notion of\nannotator subjectivity, motivated by variations in labelling. This approach\nunderstands each annotator's view as valid, which can be highly suitable for\ntasks that embed subjectivity, e.g., sentiment analysis. However, this\nconstruction may be inappropriate for tasks such as hate speech detection, as\nit affords equal validity to all positions on e.g., sexism or racism. We argue\nthat the conflation of hate and offence can invalidate findings on hate speech,\nand call for future work to be situated in theory, disentangling hate from its\northogonal concept, offence.\n", "date": "2024-03-04", "categories": []}
{"id": "2402.16458", "title": "ID-XCB: Data-independent Debiasing for Fair and Accurate\n  Transformer-based Cyberbullying Detection", "abstract": "  Swear words are a common proxy to collect datasets with cyberbullying\nincidents. Our focus is on measuring and mitigating biases derived from\nspurious associations between swear words and incidents occurring as a result\nof such data collection strategies. After demonstrating and quantifying these\nbiases, we introduce ID-XCB, the first data-independent debiasing technique\nthat combines adversarial training, bias constraints and debias fine-tuning\napproach aimed at alleviating model attention to bias-inducing words without\nimpacting overall model performance. We explore ID-XCB on two popular\nsession-based cyberbullying datasets along with comprehensive ablation and\ngeneralisation studies. We show that ID-XCB learns robust cyberbullying\ndetection capabilities while mitigating biases, outperforming state-of-the-art\ndebiasing methods in both performance and bias mitigation. Our quantitative and\nqualitative analyses demonstrate its generalisability to unseen data.\n", "date": "2024-02-26", "categories": []}
{"id": "2307.09312", "title": "Multi-Modal Discussion Transformer: Integrating Text, Images and Graph\n  Transformers to Detect Hate Speech on Social Media", "abstract": "  We present the Multi-Modal Discussion Transformer (mDT), a novel methodfor\ndetecting hate speech in online social networks such as Reddit discussions. In\ncontrast to traditional comment-only methods, our approach to labelling a\ncomment as hate speech involves a holistic analysis of text and images grounded\nin the discussion context. This is done by leveraging graph transformers to\ncapture the contextual relationships in the discussion surrounding a comment\nand grounding the interwoven fusion layers that combine text and image\nembeddings instead of processing modalities separately. To evaluate our work,\nwe present a new dataset, HatefulDiscussions, comprising complete multi-modal\ndiscussions from multiple online communities on Reddit. We compare the\nperformance of our model to baselines that only process individual comments and\nconduct extensive ablation studies.\n", "date": "2023-07-18", "categories": []}
{"id": "2402.14590", "title": "Scaling Up LLM Reviews for Google Ads Content Moderation", "abstract": "  Large language models (LLMs) are powerful tools for content moderation, but\ntheir inference costs and latency make them prohibitive for casual use on large\ndatasets, such as the Google Ads repository. This study proposes a method for\nscaling up LLM reviews for content moderation in Google Ads. First, we use\nheuristics to select candidates via filtering and duplicate removal, and create\nclusters of ads for which we select one representative ad per cluster. We then\nuse LLMs to review only the representative ads. Finally, we propagate the LLM\ndecisions for the representative ads back to their clusters. This method\nreduces the number of reviews by more than 3 orders of magnitude while\nachieving a 2x recall compared to a baseline non-LLM model. The success of this\napproach is a strong function of the representations used in clustering and\nlabel propagation; we found that cross-modal similarity representations yield\nbetter results than uni-modal representations.\n", "date": "2024-02-07", "categories": []}
{"id": "2402.01967", "title": "MasonPerplexity at Multimodal Hate Speech Event Detection 2024: Hate\n  Speech and Target Detection Using Transformer Ensembles", "abstract": "  The automatic identification of offensive language such as hate speech is\nimportant to keep discussions civil in online communities. Identifying hate\nspeech in multimodal content is a particularly challenging task because\noffensiveness can be manifested in either words or images or a juxtaposition of\nthe two. This paper presents the MasonPerplexity submission for the Shared Task\non Multimodal Hate Speech Event Detection at CASE 2024 at EACL 2024. The task\nis divided into two sub-tasks: sub-task A focuses on the identification of hate\nspeech and sub-task B focuses on the identification of targets in text-embedded\nimages during political events. We use an XLM-roBERTa-large model for sub-task\nA and an ensemble approach combining XLM-roBERTa-base, BERTweet-large, and\nBERT-base for sub-task B. Our approach obtained 0.8347 F1-score in sub-task A\nand 0.6741 F1-score in sub-task B ranking 3rd on both sub-tasks.\n", "date": "2024-02-03", "categories": []}
{"id": "2402.11845", "title": "Modularized Networks for Few-shot Hateful Meme Detection", "abstract": "  In this paper, we address the challenge of detecting hateful memes in the\nlow-resource setting where only a few labeled examples are available. Our\napproach leverages the compositionality of Low-rank adaptation (LoRA), a widely\nused parameter-efficient tuning technique. We commence by fine-tuning large\nlanguage models (LLMs) with LoRA on selected tasks pertinent to hateful meme\ndetection, thereby generating a suite of LoRA modules. These modules are\ncapable of essential reasoning skills for hateful meme detection. We then use\nthe few available annotated samples to train a module composer, which assigns\nweights to the LoRA modules based on their relevance. The model's learnable\nparameters are directly proportional to the number of LoRA modules. This\nmodularized network, underpinned by LLMs and augmented with LoRA modules,\nexhibits enhanced generalization in the context of hateful meme detection. Our\nevaluation spans three datasets designed for hateful meme detection in a\nfew-shot learning context. The proposed method demonstrates superior\nperformance to traditional in-context learning, which is also more\ncomputationally intensive during inference.We then use the few available\nannotated samples to train a module composer, which assigns weights to the LoRA\nmodules based on their relevance. The model's learnable parameters are directly\nproportional to the number of LoRA modules. This modularized network,\nunderpinned by LLMs and augmented with LoRA modules, exhibits enhanced\ngeneralization in the context of hateful meme detection. Our evaluation spans\nthree datasets designed for hateful meme detection in a few-shot learning\ncontext. The proposed method demonstrates superior performance to traditional\nin-context learning, which is also more computationally intensive during\ninference.\n", "date": "2024-02-19", "categories": []}
{"id": "2402.10231", "title": "A Multi-faceted Semi-Synthetic Dataset for Automated Cyberbullying\n  Detection", "abstract": "  In recent years, the rising use of social media has propelled automated\ncyberbullying detection into a prominent research domain. However, challenges\npersist due to the absence of a standardized definition and universally\naccepted datasets. Many researchers now view cyberbullying as a facet of\ncyberaggression, encompassing factors like repetition, peer relationships, and\nharmful intent in addition to online aggression. Acquiring comprehensive data\nreflective of all cyberbullying components from social media networks proves to\nbe a complex task. This paper provides a description of an extensive\nsemi-synthetic cyberbullying dataset that incorporates all of the essential\naspects of cyberbullying, including aggression, repetition, peer relationships,\nand intent to harm. The method of creating the dataset is succinctly outlined,\nand a detailed overview of the publicly accessible dataset is additionally\npresented. This accompanying data article provides an in-depth look at the\ndataset, increasing transparency and enabling replication. It also aids in a\ndeeper understanding of the data, supporting broader research use.\n", "date": "2024-02-09", "categories": []}
{"id": "2401.14681", "title": "MasonTigers@LT-EDI-2024: An Ensemble Approach Towards Detecting\n  Homophobia and Transphobia in Social Media Comments", "abstract": "  In this paper, we describe our approaches and results for Task 2 of the\nLT-EDI 2024 Workshop, aimed at detecting homophobia and/or transphobia across\nten languages. Our methodologies include monolingual transformers and ensemble\nmethods, capitalizing on the strengths of each to enhance the performance of\nthe models. The ensemble models worked well, placing our team, MasonTigers, in\nthe top five for eight of the ten languages, as measured by the macro F1 score.\nOur work emphasizes the efficacy of ensemble methods in multilingual scenarios,\naddressing the complexities of language-specific tasks.\n", "date": "2024-01-26", "categories": []}
{"id": "2402.09738", "title": "Align before Attend: Aligning Visual and Textual Features for Multimodal\n  Hateful Content Detection", "abstract": "  Multimodal hateful content detection is a challenging task that requires\ncomplex reasoning across visual and textual modalities. Therefore, creating a\nmeaningful multimodal representation that effectively captures the interplay\nbetween visual and textual features through intermediate fusion is critical.\nConventional fusion techniques are unable to attend to the modality-specific\nfeatures effectively. Moreover, most studies exclusively concentrated on\nEnglish and overlooked other low-resource languages. This paper proposes a\ncontext-aware attention framework for multimodal hateful content detection and\nassesses it for both English and non-English languages. The proposed approach\nincorporates an attention layer to meaningfully align the visual and textual\nfeatures. This alignment enables selective focus on modality-specific features\nbefore fusing them. We evaluate the proposed approach on two benchmark hateful\nmeme datasets, viz. MUTE (Bengali code-mixed) and MultiOFF (English).\nEvaluation results demonstrate our proposed approach's effectiveness with\nF1-scores of $69.7$% and $70.3$% for the MUTE and MultiOFF datasets. The scores\nshow approximately $2.5$% and $3.2$% performance improvement over the\nstate-of-the-art systems on these datasets. Our implementation is available at\nhttps://github.com/eftekhar-hossain/Bengali-Hateful-Memes.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.08764", "title": "A Dataset for the Detection of Dehumanizing Language", "abstract": "  Dehumanization is a mental process that enables the exclusion and ill\ntreatment of a group of people. In this paper, we present two data sets of\ndehumanizing text, a large, automatically collected corpus and a smaller,\nmanually annotated data set. Both data sets include a combination of political\ndiscourse and dialogue from movie subtitles. Our methods give us a broad and\nvaried amount of dehumanization data to work with, enabling further exploratory\nanalysis and automatic classification of dehumanization patterns. Both data\nsets will be publicly released.\n", "date": "2024-02-13", "categories": []}
{"id": "2402.06549", "title": "Bryndza at ClimateActivism 2024: Stance, Target and Hate Event Detection\n  via Retrieval-Augmented GPT-4 and LLaMA", "abstract": "  This study details our approach for the CASE 2024 Shared Task on Climate\nActivism Stance and Hate Event Detection, focusing on Hate Speech Detection,\nHate Speech Target Identification, and Stance Detection as classification\nchallenges. We explored the capability of Large Language Models (LLMs),\nparticularly GPT-4, in zero- or few-shot settings enhanced by retrieval\naugmentation and re-ranking for Tweet classification. Our goal was to determine\nif LLMs could match or surpass traditional methods in this context.\n  We conducted an ablation study with LLaMA for comparison, and our results\nindicate that our models significantly outperformed the baselines, securing\nsecond place in the Target Detection task. The code for our submission is\navailable at https://github.com/NaiveNeuron/bryndza-case-2024\n", "date": "2024-02-09", "categories": []}
{"id": "2402.05624", "title": "Efficient Models for the Detection of Hate, Abuse and Profanity", "abstract": "  Large Language Models (LLMs) are the cornerstone for many Natural Language\nProcessing (NLP) tasks like sentiment analysis, document classification, named\nentity recognition, question answering, summarization, etc. LLMs are often\ntrained on data which originates from the web. This data is prone to having\ncontent with Hate, Abuse and Profanity (HAP). For a detailed definition of HAP,\nplease refer to the Appendix. Due to the LLMs being exposed to HAP content\nduring training, the models learn it and may then generate hateful or profane\ncontent. For example, when the open-source RoBERTa model (specifically, the\nRoBERTA base model) from the HuggingFace (HF) Transformers library is prompted\nto replace the mask token in `I do not know that Persian people are that MASK`\nit returns the word `stupid` with the highest score. This is unacceptable in\ncivil discourse.The detection of Hate, Abuse and Profanity in text is a vital\ncomponent of creating civil and unbiased LLMs, which is needed not only for\nEnglish, but for all languages. In this article, we briefly describe the\ncreation of HAP detectors and various ways of using them to make models civil\nand acceptable in the output they generate.\n", "date": "2024-02-08", "categories": []}
{"id": "2402.04967", "title": "Text or Image? What is More Important in Cross-Domain Generalization\n  Capabilities of Hate Meme Detection Models?", "abstract": "  This paper delves into the formidable challenge of cross-domain\ngeneralization in multimodal hate meme detection, presenting compelling\nfindings. We provide enough pieces of evidence supporting the hypothesis that\nonly the textual component of hateful memes enables the existing multimodal\nclassifier to generalize across different domains, while the image component\nproves highly sensitive to a specific training dataset. The evidence includes\ndemonstrations showing that hate-text classifiers perform similarly to\nhate-meme classifiers in a zero-shot setting. Simultaneously, the introduction\nof captions generated from images of memes to the hate-meme classifier worsens\nperformance by an average F1 of 0.02. Through blackbox explanations, we\nidentify a substantial contribution of the text modality (average of 83%),\nwhich diminishes with the introduction of meme's image captions (52%).\nAdditionally, our evaluation on a newly created confounder dataset reveals\nhigher performance on text confounders as compared to image confounders with an\naverage $\\Delta$F1 of 0.18.\n", "date": "2024-02-07", "categories": []}
{"id": "2402.04088", "title": "The Use of a Large Language Model for Cyberbullying Detection", "abstract": "  The dominance of social media has added to the channels of bullying for\nperpetrators. Unfortunately, cyberbullying (CB) is the most prevalent\nphenomenon in todays cyber world, and is a severe threat to the mental and\nphysical health of citizens. This opens the need to develop a robust system to\nprevent bullying content from online forums, blogs, and social media platforms\nto manage the impact in our society. Several machine learning (ML) algorithms\nhave been proposed for this purpose. However, their performances are not\nconsistent due to high class imbalance and generalisation issues. In recent\nyears, large language models (LLMs) like BERT and RoBERTa have achieved\nstate-of-the-art (SOTA) results in several natural language processing (NLP)\ntasks. Unfortunately, the LLMs have not been applied extensively for CB\ndetection. In our paper, we explored the use of these models for cyberbullying\n(CB) detection. We have prepared a new dataset (D2) from existing studies\n(Formspring and Twitter). Our experimental results for dataset D1 and D2 showed\nthat RoBERTa outperformed other models.\n", "date": "2024-02-06", "categories": []}
{"id": "2402.01752", "title": "Identifying False Content and Hate Speech in Sinhala YouTube Videos by\n  Analyzing the Audio", "abstract": "  YouTube faces a global crisis with the dissemination of false information and\nhate speech. To counter these issues, YouTube has implemented strict rules\nagainst uploading content that includes false information or promotes hate\nspeech. While numerous studies have been conducted to reduce offensive\nEnglish-language content, there's a significant lack of research on Sinhala\ncontent. This study aims to address the aforementioned gap by proposing a\nsolution to minimize the spread of violence and misinformation in Sinhala\nYouTube videos. The approach involves developing a rating system that assesses\nwhether a video contains false information by comparing the title and\ndescription with the audio content and evaluating whether the video includes\nhate speech. The methodology encompasses several steps, including audio\nextraction using the Pytube library, audio transcription via the fine-tuned\nWhisper model, hate speech detection employing the distilroberta-base model and\na text classification LSTM model, and text summarization through the fine-tuned\nBART-Large- XSUM model. Notably, the Whisper model achieved a 48.99\\% word\nerror rate, while the distilroberta-base model demonstrated an F1 score of\n0.856 and a recall value of 0.861 in comparison to the LSTM model, which\nexhibited signs of overfitting.\n", "date": "2024-01-30", "categories": []}
{"id": "2402.02144", "title": "Probing Critical Learning Dynamics of PLMs for Hate Speech Detection", "abstract": "  Despite the widespread adoption, there is a lack of research into how various\ncritical aspects of pretrained language models (PLMs) affect their performance\nin hate speech detection. Through five research questions, our findings and\nrecommendations lay the groundwork for empirically investigating different\naspects of PLMs' use in hate speech detection. We deep dive into comparing\ndifferent pretrained models, evaluating their seed robustness, finetuning\nsettings, and the impact of pretraining data collection time. Our analysis\nreveals early peaks for downstream tasks during pretraining, the limited\nbenefit of employing a more recent pretraining corpus, and the significance of\nspecific layers during finetuning. We further call into question the use of\ndomain-specific models and highlight the need for dynamic datasets for\nbenchmarking hate speech detection.\n", "date": "2024-02-03", "categories": []}
{"id": "2402.03221", "title": "\"Define Your Terms\" : Enhancing Efficient Offensive Speech\n  Classification with Definition", "abstract": "  The propagation of offensive content through social media channels has\ngarnered attention of the research community. Multiple works have proposed\nvarious semantically related yet subtle distinct categories of offensive\nspeech. In this work, we explore meta-earning approaches to leverage the\ndiversity of offensive speech corpora to enhance their reliable and efficient\ndetection. We propose a joint embedding architecture that incorporates the\ninput's label and definition for classification via Prototypical Network. Our\nmodel achieves at least 75% of the maximal F1-score while using less than 10%\nof the available training data across 4 datasets. Our experimental findings\nalso provide a case study of training strategies valuable to combat resource\nscarcity.\n", "date": "2024-02-05", "categories": []}
{"id": "2312.07228", "title": "Toxic language detection: a systematic review of Arabic datasets", "abstract": "  The detection of toxic language in the Arabic language has emerged as an\nactive area of research in recent years, and reviewing the existing datasets\nemployed for training the developed solutions has become a pressing need. This\npaper offers a comprehensive survey of Arabic datasets focused on online toxic\nlanguage. We systematically gathered a total of 54 available datasets and their\ncorresponding papers and conducted a thorough analysis, considering 18 criteria\nacross four primary dimensions: availability details, content, annotation\nprocess, and reusability. This analysis enabled us to identify existing gaps\nand make recommendations for future research works. For the convenience of the\nresearch community, the list of the analysed datasets is maintained in a GitHub\nrepository (https://github.com/Imene1/Arabic-toxic-language).\n", "date": "2023-12-12", "categories": []}
{"id": "2401.16748", "title": "Detecting Racist Text in Bengali: An Ensemble Deep Learning Framework", "abstract": "  Racism is an alarming phenomenon in our country as well as all over the\nworld. Every day we have come across some racist comments in our daily life and\nvirtual life. Though we can eradicate this racism from virtual life (such as\nSocial Media). In this paper, we have tried to detect those racist comments\nwith NLP and deep learning techniques. We have built a novel dataset in the\nBengali Language. Further, we annotated the dataset and conducted data label\nvalidation. After extensive utilization of deep learning methodologies, we have\nsuccessfully achieved text detection with an impressive accuracy rate of\n87.94\\% using the Ensemble approach. We have applied RNN and LSTM models using\nBERT Embeddings. However, the MCNN-LSTM model performed highest among all those\nmodels. Lastly, the Ensemble approach has been followed to combine all the\nmodel results to increase overall performance.\n", "date": "2024-01-30", "categories": []}
{"id": "2401.15777", "title": "cantnlp@LT-EDI-2024: Automatic Detection of Anti-LGBTQ+ Hate Speech in\n  Under-resourced Languages", "abstract": "  This paper describes our homophobia/transphobia in social media comments\ndetection system developed as part of the shared task at LT-EDI-2024. We took a\ntransformer-based approach to develop our multiclass classification model for\nten language conditions (English, Spanish, Gujarati, Hindi, Kannada, Malayalam,\nMarathi, Tamil, Tulu, and Telugu). We introduced synthetic and organic\ninstances of script-switched language data during domain adaptation to mirror\nthe linguistic realities of social media language as seen in the labelled\ntraining data. Our system ranked second for Gujarati and Telugu with varying\nlevels of performance for other language conditions. The results suggest\nincorporating elements of paralinguistic behaviour such as script-switching may\nimprove the performance of language detection systems especially in the cases\nof under-resourced languages conditions.\n", "date": "2024-01-28", "categories": []}
{"id": "2401.14616", "title": "Alternative Speech: Complementary Method to Counter-Narrative for Better\n  Discourse", "abstract": "  We introduce the concept of \"Alternative Speech\" as a new way to directly\ncombat hate speech and complement the limitations of counter-narrative. An\nalternative speech provides practical alternatives to hate speech in real-world\nscenarios by offering speech-level corrections to speakers while considering\nthe surrounding context and promoting speakers to reform. Further, an\nalternative speech can combat hate speech alongside counter-narratives,\noffering a useful tool to address social issues such as racial discrimination\nand gender inequality. We propose the new concept and provide detailed\nguidelines for constructing the necessary dataset. Through discussion, we\ndemonstrate that combining alternative speech and counter-narrative can be a\nmore effective strategy for combating hate speech by complementing specificity\nand guiding capacity of counter-narrative. This paper presents another\nperspective for dealing with hate speech, offering viable remedies to\ncomplement the constraints of current approaches to mitigating harmful bias.\n", "date": "2024-01-26", "categories": []}
{"id": "2401.10841", "title": "Using LLMs to discover emerging coded antisemitic hate-speech in\n  extremist social media", "abstract": "  Online hate speech proliferation has created a difficult problem for social\nmedia platforms. A particular challenge relates to the use of coded language by\ngroups interested in both creating a sense of belonging for its users and\nevading detection. Coded language evolves quickly and its use varies over time.\nThis paper proposes a methodology for detecting emerging coded hate-laden\nterminology. The methodology is tested in the context of online antisemitic\ndiscourse. The approach considers posts scraped from social media platforms,\noften used by extremist users. The posts are scraped using seed expressions\nrelated to previously known discourse of hatred towards Jews. The method begins\nby identifying the expressions most representative of each post and calculating\ntheir frequency in the whole corpus. It filters out grammatically incoherent\nexpressions as well as previously encountered ones so as to focus on emergent\nwell-formed terminology. This is followed by an assessment of semantic\nsimilarity to known antisemitic terminology using a fine-tuned large language\nmodel, and subsequent filtering out of the expressions that are too distant\nfrom known expressions of hatred. Emergent antisemitic expressions containing\nterms clearly relating to Jewish topics are then removed to return only coded\nexpressions of hatred.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.13298", "title": "Towards Explainable Harmful Meme Detection through Multimodal Debate\n  between Large Language Models", "abstract": "  The age of social media is flooded with Internet memes, necessitating a clear\ngrasp and effective identification of harmful ones. This task presents a\nsignificant challenge due to the implicit meaning embedded in memes, which is\nnot explicitly conveyed through the surface text and image. However, existing\nharmful meme detection methods do not present readable explanations that unveil\nsuch implicit meaning to support their detection decisions. In this paper, we\npropose an explainable approach to detect harmful memes, achieved through\nreasoning over conflicting rationales from both harmless and harmful positions.\nSpecifically, inspired by the powerful capacity of Large Language Models (LLMs)\non text generation and reasoning, we first elicit multimodal debate between\nLLMs to generate the explanations derived from the contradictory arguments.\nThen we propose to fine-tune a small language model as the debate judge for\nharmfulness inference, to facilitate multimodal fusion between the harmfulness\nrationales and the intrinsic multimodal information within memes. In this way,\nour model is empowered to perform dialectical reasoning over intricate and\nimplicit harm-indicative patterns, utilizing multimodal explanations\noriginating from both harmless and harmful arguments. Extensive experiments on\nthree public meme datasets demonstrate that our harmful meme detection approach\nachieves much better performance than state-of-the-art methods and exhibits a\nsuperior capacity for explaining the meme harmfulness of the model predictions.\n", "date": "2024-01-24", "categories": []}
{"id": "2401.10653", "title": "Attentive Fusion: A Transformer-based Approach to Multimodal Hate Speech\n  Detection", "abstract": "  With the recent surge and exponential growth of social media usage,\nscrutinizing social media content for the presence of any hateful content is of\nutmost importance. Researchers have been diligently working since the past\ndecade on distinguishing between content that promotes hatred and content that\ndoes not. Traditionally, the main focus has been on analyzing textual content.\nHowever, recent research attempts have also commenced into the identification\nof audio-based content. Nevertheless, studies have shown that relying solely on\naudio or text-based content may be ineffective, as recent upsurge indicates\nthat individuals often employ sarcasm in their speech and writing. To overcome\nthese challenges, we present an approach to identify whether a speech promotes\nhate or not utilizing both audio and textual representations. Our methodology\nis based on the Transformer framework that incorporates both audio and text\nsampling, accompanied by our very own layer called \"Attentive Fusion\". The\nresults of our study surpassed previous state-of-the-art techniques, achieving\nan impressive macro F1 score of 0.927 on the Test Set.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.07810", "title": "Consolidating Strategies for Countering Hate Speech Using Persuasive\n  Dialogues", "abstract": "  Hateful comments are prevalent on social media platforms. Although tools for\nautomatically detecting, flagging, and blocking such false, offensive, and\nharmful content online have lately matured, such reactive and brute force\nmethods alone provide short-term and superficial remedies while the\nperpetrators persist. With the public availability of large language models\nwhich can generate articulate synthetic and engaging content at scale, there\nare concerns about the rapid growth of dissemination of such malicious content\non the web. There is now a need to focus on deeper, long-term solutions that\ninvolve engaging with the human perpetrator behind the source of the content to\nchange their viewpoint or at least bring down the rhetoric using persuasive\nmeans. To do that, we propose defining and experimenting with controllable\nstrategies for generating counter-arguments to hateful comments in online\nconversations. We experiment with controlling response generation using\nfeatures based on (i) argument structure and reasoning-based Walton argument\nschemes, (ii) counter-argument speech acts, and (iii) human\ncharacteristics-based qualities such as Big-5 personality traits and human\nvalues. Using automatic and human evaluations, we determine the best\ncombination of features that generate fluent, argumentative, and logically\nsound arguments for countering hate. We further share the developed\ncomputational models for automatically annotating text with such features, and\na silver-standard annotated version of an existing hate speech dialog corpora.\n", "date": "2024-01-15", "categories": []}
{"id": "2401.03346", "title": "An Investigation of Large Language Models for Real-World Hate Speech\n  Detection", "abstract": "  Hate speech has emerged as a major problem plaguing our social spaces today.\nWhile there have been significant efforts to address this problem, existing\nmethods are still significantly limited in effectively detecting hate speech\nonline. A major limitation of existing methods is that hate speech detection is\na highly contextual problem, and these methods cannot fully capture the context\nof hate speech to make accurate predictions. Recently, large language models\n(LLMs) have demonstrated state-of-the-art performance in several natural\nlanguage tasks. LLMs have undergone extensive training using vast amounts of\nnatural language data, enabling them to grasp intricate contextual details.\nHence, they could be used as knowledge bases for context-aware hate speech\ndetection. However, a fundamental problem with using LLMs to detect hate speech\nis that there are no studies on effectively prompting LLMs for context-aware\nhate speech detection. In this study, we conduct a large-scale study of hate\nspeech detection, employing five established hate speech datasets. We discover\nthat LLMs not only match but often surpass the performance of current benchmark\nmachine learning models in identifying hate speech. By proposing four diverse\nprompting strategies that optimize the use of LLMs in detecting hate speech.\nOur study reveals that a meticulously crafted reasoning prompt can effectively\ncapture the context of hate speech by fully utilizing the knowledge base in\nLLMs, significantly outperforming existing techniques. Furthermore, although\nLLMs can provide a rich knowledge base for the contextual detection of hate\nspeech, suitable prompting strategies play a crucial role in effectively\nleveraging this knowledge base for efficient detection.\n", "date": "2024-01-07", "categories": []}
{"id": "2401.03677", "title": "Overview of the 2023 ICON Shared Task on Gendered Abuse Detection in\n  Indic Languages", "abstract": "  This paper reports the findings of the ICON 2023 on Gendered Abuse Detection\nin Indic Languages. The shared task deals with the detection of gendered abuse\nin online text. The shared task was conducted as a part of ICON 2023, based on\na novel dataset in Hindi, Tamil and the Indian dialect of English. The\nparticipants were given three subtasks with the train dataset consisting of\napproximately 6500 posts sourced from Twitter. For the test set, approximately\n1200 posts were provided. The shared task received a total of 9 registrations.\nThe best F-1 scores are 0.616 for subtask 1, 0.572 for subtask 2 and, 0.616 and\n0.582 for subtask 3. The paper contains examples of hateful content owing to\nits topic.\n", "date": "2024-01-08", "categories": []}
{"id": "2312.17704", "title": "TuPy-E: detecting hate speech in Brazilian Portuguese social media with\n  a novel dataset and comprehensive analysis of models", "abstract": "  Social media has become integral to human interaction, providing a platform\nfor communication and expression. However, the rise of hate speech on these\nplatforms poses significant risks to individuals and communities. Detecting and\naddressing hate speech is particularly challenging in languages like Portuguese\ndue to its rich vocabulary, complex grammar, and regional variations. To\naddress this, we introduce TuPy-E, the largest annotated Portuguese corpus for\nhate speech detection. TuPy-E leverages an open-source approach, fostering\ncollaboration within the research community. We conduct a detailed analysis\nusing advanced techniques like BERT models, contributing to both academic\nunderstanding and practical applications\n", "date": "2023-12-29", "categories": []}
{"id": "2306.03502", "title": "Russo-Ukrainian War: Prediction and explanation of Twitter suspension", "abstract": "  On 24 February 2022, Russia invaded Ukraine, starting what is now known as\nthe Russo-Ukrainian War, initiating an online discourse on social media.\nTwitter as one of the most popular SNs, with an open and democratic character,\nenables a transparent discussion among its large user base. Unfortunately, this\noften leads to Twitter's policy violations, propaganda, abusive actions, civil\nintegrity violation, and consequently to user accounts' suspension and\ndeletion. This study focuses on the Twitter suspension mechanism and the\nanalysis of shared content and features of the user accounts that may lead to\nthis. Toward this goal, we have obtained a dataset containing 107.7M tweets,\noriginating from 9.8 million users, using Twitter API. We extract the\ncategories of shared content of the suspended accounts and explain their\ncharacteristics, through the extraction of text embeddings in junction with\ncosine similarity clustering. Our results reveal scam campaigns taking\nadvantage of trending topics regarding the Russia-Ukrainian conflict for\nBitcoin and Ethereum fraud, spam, and advertisement campaigns. Additionally, we\napply a machine learning methodology including a SHapley Additive\nexplainability model to understand and explain how user accounts get suspended.\n", "date": "2023-06-06", "categories": []}
{"id": "2312.16755", "title": "Graph Neural Networks for Antisocial Behavior Detection on Twitter", "abstract": "  Social media resurgence of antisocial behavior has exerted a downward spiral\non stereotypical beliefs, and hateful comments towards individuals and social\ngroups, as well as false or distorted news. The advances in graph neural\nnetworks employed on massive quantities of graph-structured data raise high\nhopes for the future of mediating communication on social media platforms. An\napproach based on graph convolutional data was employed to better capture the\ndependencies between the heterogeneous types of data.\n  Utilizing past and present experiences on the topic, we proposed and\nevaluated a graph-based approach for antisocial behavior detection, with\ngeneral applicability that is both language- and context-independent. In this\nresearch, we carried out an experimental validation of our graph-based approach\non several PAN datasets provided as part of their shared tasks, that enable the\ndiscussion of the results obtained by the proposed solution.\n", "date": "2023-12-28", "categories": []}
{"id": "2312.13193", "title": "HCDIR: End-to-end Hate Context Detection, and Intensity Reduction model\n  for online comments", "abstract": "  Warning: This paper contains examples of the language that some people may\nfind offensive.\n  Detecting and reducing hateful, abusive, offensive comments is a critical and\nchallenging task on social media. Moreover, few studies aim to mitigate the\nintensity of hate speech. While studies have shown that context-level semantics\nare crucial for detecting hateful comments, most of this research focuses on\nEnglish due to the ample datasets available. In contrast, low-resource\nlanguages, like Indian languages, remain under-researched because of limited\ndatasets. Contrary to hate speech detection, hate intensity reduction remains\nunexplored in high-resource and low-resource languages. In this paper, we\npropose a novel end-to-end model, HCDIR, for Hate Context Detection, and Hate\nIntensity Reduction in social media posts. First, we fine-tuned several\npre-trained language models to detect hateful comments to ascertain the\nbest-performing hateful comments detection model. Then, we identified the\ncontextual hateful words. Identification of such hateful words is justified\nthrough the state-of-the-art explainable learning model, i.e., Integrated\nGradient (IG). Lastly, the Masked Language Modeling (MLM) model has been\nemployed to capture domain-specific nuances to reduce hate intensity. We masked\nthe 50\\% hateful words of the comments identified as hateful and predicted the\nalternative words for these masked terms to generate convincing sentences. An\noptimal replacement for the original hate comments from the feasible sentences\nis preferred. Extensive experiments have been conducted on several recent\ndatasets using automatic metric-based evaluation (BERTScore) and thorough human\nevaluation. To enhance the faithfulness in human evaluation, we arranged a\ngroup of three human annotators with varied expertise.\n", "date": "2023-12-20", "categories": []}
{"id": "2312.11504", "title": "The performance of multiple language models in identifying offensive\n  language on social media", "abstract": "  Text classification is an important topic in the field of natural language\nprocessing. It has been preliminarily applied in information retrieval, digital\nlibrary, automatic abstracting, text filtering, word semantic discrimination\nand many other fields. The aim of this research is to use a variety of\nalgorithms to test the ability to identify offensive posts and evaluate their\nperformance against a variety of assessment methods. The motivation for this\nproject is to reduce the harm of these languages to human censors by automating\nthe screening of offending posts. The field is a new one, and despite much\ninterest in the past two years, there has been no focus on the object of the\noffence. Through the experiment of this project, it should inspire future\nresearch on identification methods as well as identification content.\n", "date": "2023-12-10", "categories": []}
{"id": "2311.14865", "title": "Improving Cross-Domain Hate Speech Generalizability with Emotion\n  Knowledge", "abstract": "  Reliable automatic hate speech (HS) detection systems must adapt to the\nin-flow of diverse new data to curtail hate speech. However, hate speech\ndetection systems commonly lack generalizability in identifying hate speech\ndissimilar to data used in training, impeding their robustness in real-world\ndeployments. In this work, we propose a hate speech generalization framework\nthat leverages emotion knowledge in a multitask architecture to improve the\ngeneralizability of hate speech detection in a cross-domain setting. We\ninvestigate emotion corpora with varying emotion categorical scopes to\ndetermine the best corpus scope for supplying emotion knowledge to foster\ngeneralized hate speech detection. We further assess the relationship between\nusing pretrained Transformers models adapted for hate speech and its effect on\nour emotion-enriched hate speech generalization model. We perform extensive\nexperiments on six publicly available datasets sourced from different online\ndomains and show that our emotion-enriched HS detection generalization method\ndemonstrates consistent generalization improvement in cross-domain evaluation,\nincreasing generalization performance up to 18.1% and average cross-domain\nperformance up to 8.5%, according to the F1 measure.\n", "date": "2023-11-24", "categories": []}
{"id": "2312.10528", "title": "Cross-Linguistic Offensive Language Detection: BERT-Based Analysis of\n  Bengali, Assamese, & Bodo Conversational Hateful Content from Social Media", "abstract": "  In today's age, social media reigns as the paramount communication platform,\nproviding individuals with the avenue to express their conjectures,\nintellectual propositions, and reflections. Unfortunately, this freedom often\ncomes with a downside as it facilitates the widespread proliferation of hate\nspeech and offensive content, leaving a deleterious impact on our world. Thus,\nit becomes essential to discern and eradicate such offensive material from the\nrealm of social media. This article delves into the comprehensive results and\nkey revelations from the HASOC-2023 offensive language identification result.\nThe primary emphasis is placed on the meticulous detection of hate speech\nwithin the linguistic domains of Bengali, Assamese, and Bodo, forming the\nframework for Task 4: Annihilate Hates. In this work, we used BERT models,\nincluding XML-Roberta, L3-cube, IndicBERT, BenglaBERT, and BanglaHateBERT. The\nresearch outcomes were promising and showed that XML-Roberta-lagre performed\nbetter than monolingual models in most cases. Our team 'TeamBD' achieved rank\n3rd for Task 4 - Assamese, & 5th for Bengali.\n", "date": "2023-12-16", "categories": []}
{"id": "2312.11344", "title": "Muted: Multilingual Targeted Offensive Speech Identification and\n  Visualization", "abstract": "  Offensive language such as hate, abuse, and profanity (HAP) occurs in various\ncontent on the web. While previous work has mostly dealt with sentence level\nannotations, there have been a few recent attempts to identify offensive spans\nas well. We build upon this work and introduce Muted, a system to identify\nmultilingual HAP content by displaying offensive arguments and their targets\nusing heat maps to indicate their intensity. Muted can leverage any\ntransformer-based HAP-classification model and its attention mechanism\nout-of-the-box to identify toxic spans, without further fine-tuning. In\naddition, we use the spaCy library to identify the specific targets and\narguments for the words predicted by the attention heatmaps. We present the\nmodel's performance on identifying offensive spans and their targets in\nexisting datasets and present new annotations on German text. Finally, we\ndemonstrate our proposed visualization tool on multilingual inputs.\n", "date": "2023-12-18", "categories": []}
{"id": "2305.10547", "title": "Rethinking Multimodal Content Moderation from an Asymmetric Angle with\n  Mixed-modality", "abstract": "  There is a rapidly growing need for multimodal content moderation (CM) as\nmore and more content on social media is multimodal in nature. Existing\nunimodal CM systems may fail to catch harmful content that crosses modalities\n(e.g., memes or videos), which may lead to severe consequences. In this paper,\nwe present a novel CM model, Asymmetric Mixed-Modal Moderation (AM3), to target\nmultimodal and unimodal CM tasks. Specifically, to address the asymmetry in\nsemantics between vision and language, AM3 has a novel asymmetric fusion\narchitecture that is designed to not only fuse the common knowledge in both\nmodalities but also to exploit the unique information in each modality. Unlike\nprevious works that focus on representing the two modalities into a similar\nfeature space while overlooking the intrinsic difference between the\ninformation conveyed in multimodality and in unimodality (asymmetry in\nmodalities), we propose a novel cross-modality contrastive loss to learn the\nunique knowledge that only appears in multimodality. This is critical as some\nharmful intent may only be conveyed through the intersection of both\nmodalities. With extensive experiments, we show that AM3 outperforms all\nexisting state-of-the-art methods on both multimodal and unimodal CM\nbenchmarks.\n", "date": "2023-05-17", "categories": []}
{"id": "2312.07831", "title": "Abusive Span Detection for Vietnamese Narrative Texts", "abstract": "  Abuse in its various forms, including physical, psychological, verbal,\nsexual, financial, and cultural, has a negative impact on mental health.\nHowever, there are limited studies on applying natural language processing\n(NLP) in this field in Vietnam. Therefore, we aim to contribute by building a\nhuman-annotated Vietnamese dataset for detecting abusive content in Vietnamese\nnarrative texts. We sourced these texts from VnExpress, Vietnam's popular\nonline newspaper, where readers often share stories containing abusive content.\nIdentifying and categorizing abusive spans in these texts posed significant\nchallenges during dataset creation, but it also motivated our research. We\nexperimented with lightweight baseline models by freezing PhoBERT and\nXLM-RoBERTa and using their hidden states in a BiLSTM to assess the complexity\nof the dataset. According to our experimental results, PhoBERT outperforms\nother models in both labeled and unlabeled abusive span detection tasks. These\nresults indicate that it has the potential for future improvements.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.08303", "title": "Efficient Toxic Content Detection by Bootstrapping and Distilling Large\n  Language Models", "abstract": "  Toxic content detection is crucial for online services to remove\ninappropriate content that violates community standards. To automate the\ndetection process, prior works have proposed varieties of machine learning (ML)\napproaches to train Language Models (LMs) for toxic content detection. However,\nboth their accuracy and transferability across datasets are limited. Recently,\nLarge Language Models (LLMs) have shown promise in toxic content detection due\nto their superior zero-shot and few-shot in-context learning ability as well as\nbroad transferability on ML tasks. However, efficiently designing prompts for\nLLMs remains challenging. Moreover, the high run-time cost of LLMs may hinder\ntheir deployments in production. To address these challenges, in this work, we\npropose BD-LLM, a novel and efficient approach to Bootstrapping and Distilling\nLLMs for toxic content detection. Specifically, we design a novel prompting\nmethod named Decision-Tree-of-Thought (DToT) to bootstrap LLMs' detection\nperformance and extract high-quality rationales. DToT can automatically select\nmore fine-grained context to re-prompt LLMs when their responses lack\nconfidence. Additionally, we use the rationales extracted via DToT to fine-tune\nstudent LMs. Our experimental results on various datasets demonstrate that DToT\ncan improve the accuracy of LLMs by up to 4.6%. Furthermore, student LMs\nfine-tuned with rationales extracted via DToT outperform baselines on all\ndatasets with up to 16.9\\% accuracy improvement, while being more than 60x\nsmaller than conventional LLMs. Finally, we observe that student LMs fine-tuned\nwith rationales exhibit better cross-dataset transferability.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.06861", "title": "Disentangling Perceptions of Offensiveness: Cultural and Moral\n  Correlates", "abstract": "  Perception of offensiveness is inherently subjective, shaped by the lived\nexperiences and socio-cultural values of the perceivers. Recent years have seen\nsubstantial efforts to build AI-based tools that can detect offensive language\nat scale, as a means to moderate social media platforms, and to ensure safety\nof conversational AI technologies such as ChatGPT and Bard. However, existing\napproaches treat this task as a technical endeavor, built on top of data\nannotated for offensiveness by a global crowd workforce without any attention\nto the crowd workers' provenance or the values their perceptions reflect. We\nargue that cultural and psychological factors play a vital role in the\ncognitive processing of offensiveness, which is critical to consider in this\ncontext. We re-frame the task of determining offensiveness as essentially a\nmatter of moral judgment -- deciding the boundaries of ethically wrong vs.\nright language within an implied set of socio-cultural norms. Through a\nlarge-scale cross-cultural study based on 4309 participants from 21 countries\nacross 8 cultural regions, we demonstrate substantial cross-cultural\ndifferences in perceptions of offensiveness. More importantly, we find that\nindividual moral values play a crucial role in shaping these variations: moral\nconcerns about Care and Purity are significant mediating factors driving\ncross-cultural differences. These insights are of crucial importance as we\nbuild AI models for the pluralistic world, where the values they espouse should\naim to respect and account for moral values in diverse geo-cultural contexts.\n", "date": "2023-12-11", "categories": []}
{"id": "2308.02080", "title": "Causality Guided Disentanglement for Cross-Platform Hate Speech\n  Detection", "abstract": "  Social media platforms, despite their value in promoting open discourse, are\noften exploited to spread harmful content. Current deep learning and natural\nlanguage processing models used for detecting this harmful content overly rely\non domain-specific terms affecting their capabilities to adapt to generalizable\nhate speech detection. This is because they tend to focus too narrowly on\nparticular linguistic signals or the use of certain categories of words.\nAnother significant challenge arises when platforms lack high-quality annotated\ndata for training, leading to a need for cross-platform models that can adapt\nto different distribution shifts. Our research introduces a cross-platform hate\nspeech detection model capable of being trained on one platform's data and\ngeneralizing to multiple unseen platforms. To achieve good generalizability\nacross platforms, one way is to disentangle the input representations into\ninvariant and platform-dependent features. We also argue that learning causal\nrelationships, which remain constant across diverse environments, can\nsignificantly aid in understanding invariant representations in hate speech. By\ndisentangling input into platform-dependent features (useful for predicting\nhate targets) and platform-independent features (used to predict the presence\nof hate), we learn invariant representations resistant to distribution shifts.\nThese features are then used to predict hate speech across unseen platforms.\nOur extensive experiments across four platforms highlight our model's enhanced\nefficacy compared to existing state-of-the-art methods in detecting generalized\nhate speech.\n", "date": "2023-08-03", "categories": []}
{"id": "2312.05434", "title": "Beneath the Surface: Unveiling Harmful Memes with Multimodal Reasoning\n  Distilled from Large Language Models", "abstract": "  The age of social media is rife with memes. Understanding and detecting\nharmful memes pose a significant challenge due to their implicit meaning that\nis not explicitly conveyed through the surface text and image. However,\nexisting harmful meme detection approaches only recognize superficial\nharm-indicative signals in an end-to-end classification manner but ignore\nin-depth cognition of the meme text and image. In this paper, we attempt to\ndetect harmful memes based on advanced reasoning over the interplay of\nmultimodal information in memes. Inspired by the success of Large Language\nModels (LLMs) on complex reasoning, we first conduct abductive reasoning with\nLLMs. Then we propose a novel generative framework to learn reasonable thoughts\nfrom LLMs for better multimodal fusion and lightweight fine-tuning, which\nconsists of two training stages: 1) Distill multimodal reasoning knowledge from\nLLMs; and 2) Fine-tune the generative framework to infer harmfulness. Extensive\nexperiments conducted on three meme datasets demonstrate that our proposed\napproach achieves superior performance than state-of-the-art methods on the\nharmful meme detection task.\n", "date": "2023-12-09", "categories": []}
{"id": "2312.05467", "title": "Textual Toxicity in Social Media: Understanding the Bangla Toxic\n  Language Expressed in Facebook Comment", "abstract": "  Social Media is a repository of digital literature including user-generated\ncontent. The users of social media are expressing their opinion with diverse\nmediums such as text, emojis, memes, and also through other visual and textual\nmediums. A major portion of these media elements could be treated as harmful to\nothers and they are known by many words including Cyberbullying and Toxic\nLanguage . The goal of this research paper is to analyze a curated and\nvalue-added dataset of toxic language titled ToxLex_bn . It is an exhaustive\nwordlist that can be used as classifier material to detect toxicity in social\nmedia. The toxic language/script used by the Bengali community as\ncyberbullying, hate speech and moral policing became major trends in social\nmedia culture in Bangladesh and West Bengal. The toxicity became so high that\nthe victims has to post as a counter or release explanation video for the\nhaters. Most cases are pointed to women celebrity and their relation, dress,\nlifestyle are became trolled and toxicity flooded in comments boxes. Not only\ncelebrity bashing but also hates occurred between Hindu Muslims,\nIndia-Bangladesh, Two opponents of 1971 and these are very common for virtual\nconflict in the comment thread. Even many times facebook comment causes sue and\nlegal matters in Bangladesh and thus it requires more study. In this study, a\nBangla toxic language dataset has been analyzed which was inputted by the user\nin Bengali script & language. For this, about 1968 unique bigrams or phrases as\nwordlists have been analyzed which are derived from 2207590 comments. It is\nassumed that this analysis will reinforce the detection of Bangla's toxic\nlanguage used in social media and thus cure this virtual disease.\n", "date": "2023-12-09", "categories": []}
{"id": "2312.05671", "title": "Hate Speech and Offensive Content Detection in Indo-Aryan Languages: A\n  Battle of LSTM and Transformers", "abstract": "  Social media platforms serve as accessible outlets for individuals to express\ntheir thoughts and experiences, resulting in an influx of user-generated data\nspanning all age groups. While these platforms enable free expression, they\nalso present significant challenges, including the proliferation of hate speech\nand offensive content. Such objectionable language disrupts objective discourse\nand can lead to radicalization of debates, ultimately threatening democratic\nvalues. Consequently, organizations have taken steps to monitor and curb\nabusive behavior, necessitating automated methods for identifying suspicious\nposts. This paper contributes to Hate Speech and Offensive Content\nIdentification in English and Indo-Aryan Languages (HASOC) 2023 shared tasks\ntrack. We, team Z-AGI Labs, conduct a comprehensive comparative analysis of\nhate speech classification across five distinct languages: Bengali, Assamese,\nBodo, Sinhala, and Gujarati. Our study encompasses a wide range of pre-trained\nmodels, including Bert variants, XLM-R, and LSTM models, to assess their\nperformance in identifying hate speech across these languages. Results reveal\nintriguing variations in model performance. Notably, Bert Base Multilingual\nCased emerges as a strong performer across languages, achieving an F1 score of\n0.67027 for Bengali and 0.70525 for Assamese. At the same time, it\nsignificantly outperforms other models with an impressive F1 score of 0.83009\nfor Bodo. In Sinhala, XLM-R stands out with an F1 score of 0.83493, whereas for\nGujarati, a custom LSTM-based model outshined with an F1 score of 0.76601. This\nstudy offers valuable insights into the suitability of various pre-trained\nmodels for hate speech detection in multilingual settings. By considering the\nnuances of each, our research contributes to an informed model selection for\nbuilding robust hate speech detection systems.\n", "date": "2023-12-09", "categories": []}
{"id": "2312.03727", "title": "Content-Localization based System for Analyzing Sentiment and Hate\n  Behaviors in Low-Resource Dialectal Arabic: English to Levantine and Gulf", "abstract": "  Even though online social movements can quickly become viral on social media,\nlanguages can be a barrier to timely monitoring and analyzing the underlying\nonline social behaviors (OSB). This is especially true for under-resourced\nlanguages on social media like dialectal Arabic; the primary language used by\nArabs on social media. Therefore, it is crucial to provide solutions to\nefficiently exploit resources from high-resourced languages to solve\nlanguage-dependent OSB analysis in under-resourced languages. This paper\nproposes to localize content of resources in high-resourced languages into\nunder-resourced Arabic dialects. Content localization goes beyond content\ntranslation that converts text from one language to another; content\nlocalization adapts culture, language nuances and regional preferences from one\nlanguage to a specific language/dialect. Automating understanding of the\nnatural and familiar day-to-day expressions in different regions, is the key to\nachieve a wider analysis of OSB especially for smart cities. In this paper, we\nutilize content-localization based neural machine translation to develop\nsentiment and hate classifiers for two low-resourced Arabic dialects: Levantine\nand Gulf. Not only this but we also leverage unsupervised learning to\nfacilitate the analysis of sentiment and hate predictions by inferring hidden\ntopics from the corresponding data and providing coherent interpretations of\nthose topics in their native language/dialects. The experimental evaluations\nand proof-of-concept COVID-19 case study on real data have validated the\neffectiveness of our proposed system in precisely distinguishing sentiments and\naccurately identifying hate content in both Levantine and Gulf Arabic dialects.\nOur findings shed light on the importance of considering the unique nature of\ndialects within the same language and ignoring the dialectal aspect would lead\nto misleading analysis.\n", "date": "2023-11-27", "categories": []}
{"id": "2312.03743", "title": "Easy Data Augmentation in Sentiment Analysis of Cyberbullying", "abstract": "  Instagram, a social media platform, has in the vicinity of 2 billion active\nusers in 2023. The platform allows users to post photos and videos with one\nanother. However, cyberbullying remains a significant problem for about 50% of\nyoung Indonesians. To address this issue, sentiment analysis for comment\nfiltering uses a Support Vector Machine (SVM) and Easy Data Augmentation (EDA).\nEDA will augment the dataset, enabling robust prediction and analysis of\ncyberbullying by introducing more variation. Based on the tests, SVM\ncombination with EDA results in a 2.52% increase in the k-Fold Cross Validation\nscore. Our proposed approach shows an improved accuracy of 92.5%, 2.5% higher\nthan that of the existing state-of-the-art method. To maintain the\nreproducibility and replicability of this research, the source code can be\naccessed at uns.id/eda_svm.\n", "date": "2023-11-29", "categories": []}
{"id": "2312.03936", "title": "The Potential of Vision-Language Models for Content Moderation of\n  Children's Videos", "abstract": "  Natural language supervision has been shown to be effective for zero-shot\nlearning in many computer vision tasks, such as object detection and activity\nrecognition. However, generating informative prompts can be challenging for\nmore subtle tasks, such as video content moderation. This can be difficult, as\nthere are many reasons why a video might be inappropriate, beyond violence and\nobscenity. For example, scammers may attempt to create junk content that is\nsimilar to popular educational videos but with no meaningful information. This\npaper evaluates the performance of several CLIP variations for content\nmoderation of children's cartoons in both the supervised and zero-shot setting.\nWe show that our proposed model (Vanilla CLIP with Projection Layer)\noutperforms previous work conducted on the Malicious or Benign (MOB) benchmark\nfor video content moderation. This paper presents an in depth analysis of how\ncontext-specific language prompts affect content moderation performance. Our\nresults indicate that it is important to include more context in content\nmoderation prompts, particularly for cartoon videos as they are not well\nrepresented in the CLIP training data.\n", "date": "2023-12-06", "categories": []}
{"id": "2311.09428", "title": "Beyond Detection: Unveiling Fairness Vulnerabilities in Abusive Language\n  Models", "abstract": "  This work investigates the potential of undermining both fairness and\ndetection performance in abusive language detection. In a dynamic and complex\ndigital world, it is crucial to investigate the vulnerabilities of these\ndetection models to adversarial fairness attacks to improve their fairness\nrobustness. We propose a simple yet effective framework FABLE that leverages\nbackdoor attacks as they allow targeted control over the fairness and detection\nperformance. FABLE explores three types of trigger designs (i.e., rare,\nartificial, and natural triggers) and novel sampling strategies. Specifically,\nthe adversary can inject triggers into samples in the minority group with the\nfavored outcome (i.e., \"non-abusive\") and flip their labels to the unfavored\noutcome, i.e., \"abusive\". Experiments on benchmark datasets demonstrate the\neffectiveness of FABLE attacking fairness and utility in abusive language\ndetection.\n", "date": "2023-11-15", "categories": []}
{"id": "2312.03379", "title": "A Text-to-Text Model for Multilingual Offensive Language Identification", "abstract": "  The ubiquity of offensive content on social media is a growing cause for\nconcern among companies and government organizations. Recently,\ntransformer-based models such as BERT, XLNET, and XLM-R have achieved\nstate-of-the-art performance in detecting various forms of offensive content\n(e.g. hate speech, cyberbullying, and cyberaggression). However, the majority\nof these models are limited in their capabilities due to their encoder-only\narchitecture, which restricts the number and types of labels in downstream\ntasks. Addressing these limitations, this study presents the first pre-trained\nmodel with encoder-decoder architecture for offensive language identification\nwith text-to-text transformers (T5) trained on two large offensive language\nidentification datasets; SOLID and CCTK. We investigate the effectiveness of\ncombining two datasets and selecting an optimal threshold in semi-supervised\ninstances in SOLID in the T5 retraining step. Our pre-trained T5 model\noutperforms other transformer-based models fine-tuned for offensive language\ndetection, such as fBERT and HateBERT, in multiple English benchmarks.\nFollowing a similar approach, we also train the first multilingual pre-trained\nmodel for offensive language identification using mT5 and evaluate its\nperformance on a set of six different languages (German, Hindi, Korean,\nMarathi, Sinhala, and Spanish). The results demonstrate that this multilingual\nmodel achieves a new state-of-the-art on all the above datasets, showing its\nusefulness in multilingual scenarios. Our proposed T5-based models will be made\nfreely available to the community.\n", "date": "2023-12-06", "categories": []}
{"id": "2308.16549", "title": "Thesis Distillation: Investigating The Impact of Bias in NLP Models on\n  Hate Speech Detection", "abstract": "  This paper is a summary of the work done in my PhD thesis. Where I\ninvestigate the impact of bias in NLP models on the task of hate speech\ndetection from three perspectives: explainability, offensive stereotyping bias,\nand fairness. Then, I discuss the main takeaways from my thesis and how they\ncan benefit the broader NLP community. Finally, I discuss important future\nresearch directions. The findings of my thesis suggest that the bias in NLP\nmodels impacts the task of hate speech detection from all three perspectives.\nAnd that unless we start incorporating social sciences in studying bias in NLP\nmodels, we will not effectively overcome the current limitations of measuring\nand mitigating bias in NLP models.\n", "date": "2023-08-31", "categories": []}
{"id": "2312.01787", "title": "Developing Linguistic Patterns to Mitigate Inherent Human Bias in\n  Offensive Language Detection", "abstract": "  With the proliferation of social media, there has been a sharp increase in\noffensive content, particularly targeting vulnerable groups, exacerbating\nsocial problems such as hatred, racism, and sexism. Detecting offensive\nlanguage use is crucial to prevent offensive language from being widely shared\non social media. However, the accurate detection of irony, implication, and\nvarious forms of hate speech on social media remains a challenge. Natural\nlanguage-based deep learning models require extensive training with large,\ncomprehensive, and labeled datasets. Unfortunately, manually creating such\ndatasets is both costly and error-prone. Additionally, the presence of\nhuman-bias in offensive language datasets is a major concern for deep learning\nmodels. In this paper, we propose a linguistic data augmentation approach to\nreduce bias in labeling processes, which aims to mitigate the influence of\nhuman bias by leveraging the power of machines to improve the accuracy and\nfairness of labeling processes. This approach has the potential to improve\noffensive language classification tasks across multiple languages and reduce\nthe prevalence of offensive content on social media.\n", "date": "2023-12-04", "categories": []}
{"id": "2312.00053", "title": "Anti-Sexism Alert System: Identification of Sexist Comments on Social\n  Media Using AI Techniques", "abstract": "  Social relationships in the digital sphere are becoming more usual and\nfrequent, and they constitute a very important aspect for all of us. {Violent\ninteractions in this sphere are very frequent, and have serious effects on the\nvictims}. Within this global scenario, there is one kind of digital violence\nthat is becoming really worrying: sexism against women. Sexist comments that\nare publicly posted in social media (newspaper comments, social networks,\netc.), usually obtain a lot of attention and become viral, with consequent\ndamage to the persons involved. In this paper, we introduce an anti-sexism\nalert system, based on natural language processing (NLP) and artificial\nintelligence (AI), that analyzes any public post, and decides if it could be\nconsidered a sexist comment or not. Additionally, this system also works on\nanalyzing all the public comments linked to any multimedia content (piece of\nnews, video, tweet, etc.) and decides, using a color-based system similar to\ntraffic lights, if there is sexism in the global set of posts. We have created\na labeled data set in Spanish, since the majority of studies focus on English,\nto train our system, which offers a very good performance after the validation\nexperiments.\n", "date": "2023-11-28", "categories": []}
{"id": "2311.18147", "title": "DisCGen: A Framework for Discourse-Informed Counterspeech Generation", "abstract": "  Counterspeech can be an effective method for battling hateful content on\nsocial media. Automated counterspeech generation can aid in this process.\nGenerated counterspeech, however, can be viable only when grounded in the\ncontext of topic, audience and sensitivity as these factors influence both the\nefficacy and appropriateness. In this work, we propose a novel framework based\non theories of discourse to study the inferential links that connect counter\nspeeches to the hateful comment. Within this framework, we propose: i) a\ntaxonomy of counterspeech derived from discourse frameworks, and ii)\ndiscourse-informed prompting strategies for generating contextually-grounded\ncounterspeech. To construct and validate this framework, we present a process\nfor collecting an in-the-wild dataset of counterspeech from Reddit. Using this\nprocess, we manually annotate a dataset of 3.9k Reddit comment pairs for the\npresence of hatespeech and counterspeech. The positive pairs are annotated for\n10 classes in our proposed taxonomy. We annotate these pairs with paraphrased\ncounterparts to remove offensiveness and first-person references. We show that\nby using our dataset and framework, large language models can generate\ncontextually-grounded counterspeech informed by theories of discourse.\nAccording to our human evaluation, our approaches can act as a safeguard\nagainst critical failures of discourse-agnostic models.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.18778", "title": "Mavericks at BLP-2023 Task 1: Ensemble-based Approach Using Language\n  Models for Violence Inciting Text Detection", "abstract": "  This paper presents our work for the Violence Inciting Text Detection shared\ntask in the First Workshop on Bangla Language Processing. Social media has\naccelerated the propagation of hate and violence-inciting speech in society. It\nis essential to develop efficient mechanisms to detect and curb the propagation\nof such texts. The problem of detecting violence-inciting texts is further\nexacerbated in low-resource settings due to sparse research and less data. The\ndata provided in the shared task consists of texts in the Bangla language,\nwhere each example is classified into one of the three categories defined based\non the types of violence-inciting texts. We try and evaluate several BERT-based\nmodels, and then use an ensemble of the models as our final submission. Our\nsubmission is ranked 10th in the final leaderboard of the shared task with a\nmacro F1 score of 0.737.\n", "date": "2023-11-30", "categories": []}
{"id": "2311.15023", "title": "Offensive Language Identification in Transliterated and Code-Mixed\n  Bangla", "abstract": "  Identifying offensive content in social media is vital for creating safe\nonline communities. Several recent studies have addressed this problem by\ncreating datasets for various languages. In this paper, we explore offensive\nlanguage identification in texts with transliterations and code-mixing,\nlinguistic phenomena common in multilingual societies, and a known challenge\nfor NLP systems. We introduce TB-OLID, a transliterated Bangla offensive\nlanguage dataset containing 5,000 manually annotated comments. We train and\nfine-tune machine learning models on TB-OLID, and we evaluate their results on\nthis dataset. Our results show that English pre-trained transformer-based\nmodels, such as fBERT and HateBERT achieve the best performance on this\ndataset.\n", "date": "2023-11-25", "categories": []}
{"id": "2311.15029", "title": "nlpBDpatriots at BLP-2023 Task 1: A Two-Step Classification for Violence\n  Inciting Text Detection in Bangla", "abstract": "  In this paper, we discuss the nlpBDpatriots entry to the shared task on\nViolence Inciting Text Detection (VITD) organized as part of the first workshop\non Bangla Language Processing (BLP) co-located with EMNLP. The aim of this task\nis to identify and classify the violent threats, that provoke further unlawful\nviolent acts. Our best-performing approach for the task is two-step\nclassification using back translation and multilinguality which ranked 6th out\nof 27 teams with a macro F1 score of 0.74.\n", "date": "2023-11-25", "categories": []}
{"id": "2311.00321", "title": "HARE: Explainable Hate Speech Detection with Step-by-Step Reasoning", "abstract": "  With the proliferation of social media, accurate detection of hate speech has\nbecome critical to ensure safety online. To combat nuanced forms of hate\nspeech, it is important to identify and thoroughly explain hate speech to help\nusers understand its harmful effects. Recent benchmarks have attempted to\ntackle this issue by training generative models on free-text annotations of\nimplications in hateful text. However, we find significant reasoning gaps in\nthe existing annotations schemes, which may hinder the supervision of detection\nmodels. In this paper, we introduce a hate speech detection framework, HARE,\nwhich harnesses the reasoning capabilities of large language models (LLMs) to\nfill these gaps in explanations of hate speech, thus enabling effective\nsupervision of detection models. Experiments on SBIC and Implicit Hate\nbenchmarks show that our method, using model-generated data, consistently\noutperforms baselines, using existing free-text human annotations. Analysis\ndemonstrates that our method enhances the explanation quality of trained models\nand improves generalization to unseen datasets. Our code is available at\nhttps://github.com/joonkeekim/hare-hate-speech.git.\n", "date": "2023-11-01", "categories": []}
{"id": "2311.10899", "title": "Extraction and Summarization of Explicit Video Content using Multi-Modal\n  Deep Learning", "abstract": "  With the increase in video-sharing platforms across the internet, it is\ndifficult for humans to moderate the data for explicit content. Hence, an\nautomated pipeline to scan through video data for explicit content has become\nthe need of the hour. We propose a novel pipeline that uses multi-modal deep\nlearning to first extract the explicit segments of input videos and then\nsummarize their content using text to determine its age appropriateness and age\nrating. We also evaluate our pipeline's effectiveness in the end using standard\nmetrics.\n", "date": "2023-11-17", "categories": []}
{"id": "2311.10236", "title": "Latent Feature-based Data Splits to Improve Generalisation Evaluation: A\n  Hate Speech Detection Case Study", "abstract": "  With the ever-growing presence of social media platforms comes the increased\nspread of harmful content and the need for robust hate speech detection\nsystems. Such systems easily overfit to specific targets and keywords, and\nevaluating them without considering distribution shifts that might occur\nbetween train and test data overestimates their benefit. We challenge hate\nspeech models via new train-test splits of existing datasets that rely on the\nclustering of models' hidden representations. We present two split variants\n(Subset-Sum-Split and Closest-Split) that, when applied to two datasets using\nfour pretrained models, reveal how models catastrophically fail on blind spots\nin the latent space. This result generalises when developing a split with one\nmodel and evaluating it on another. Our analysis suggests that there is no\nclear surface-level property of the data split that correlates with the\ndecreased performance, which underscores that task difficulty is not always\nhumanly interpretable. We recommend incorporating latent feature-based splits\nin model development and release two splits via the GenBench benchmark.\n", "date": "2023-11-16", "categories": []}
{"id": "2311.09834", "title": "Overview of the HASOC Subtrack at FIRE 2023: Identification of Tokens\n  Contributing to Explicit Hate in English by Span Detection", "abstract": "  As hate speech continues to proliferate on the web, it is becoming\nincreasingly important to develop computational methods to mitigate it.\nReactively, using black-box models to identify hateful content can perplex\nusers as to why their posts were automatically flagged as hateful. On the other\nhand, proactive mitigation can be achieved by suggesting rephrasing before a\npost is made public. However, both mitigation techniques require information\nabout which part of a post contains the hateful aspect, i.e., what spans within\na text are responsible for conveying hate. Better detection of such spans can\nsignificantly reduce explicitly hateful content on the web. To further\ncontribute to this research area, we organized HateNorm at HASOC-FIRE 2023,\nfocusing on explicit span detection in English Tweets. A total of 12 teams\nparticipated in the competition, with the highest macro-F1 observed at 0.58.\n", "date": "2023-11-16", "categories": []}
{"id": "2311.09993", "title": "Generative AI for Hate Speech Detection: Evaluation and Findings", "abstract": "  Automatic hate speech detection using deep neural models is hampered by the\nscarcity of labeled datasets, leading to poor generalization. To mitigate this\nproblem, generative AI has been utilized to generate large amounts of synthetic\nhate speech sequences from available labeled examples, leveraging the generated\ndata in finetuning large pre-trained language models (LLMs). In this chapter,\nwe provide a review of relevant methods, experimental setups and evaluation of\nthis approach. In addition to general LLMs, such as BERT, RoBERTa and ALBERT,\nwe apply and evaluate the impact of train set augmentation with generated data\nusing LLMs that have been already adapted for hate detection, including\nRoBERTa-Toxicity, HateBERT, HateXplain, ToxDect, and ToxiGen. An empirical\nstudy corroborates our previous findings, showing that this approach improves\nhate speech generalization, boosting recall performance across data\ndistributions. In addition, we explore and compare the performance of the\nfinetuned LLMs with zero-shot hate detection using a GPT-3.5 model. Our results\ndemonstrate that while better generalization is achieved using the GPT-3.5\nmodel, it achieves mediocre recall and low precision on most datasets. It is an\nopen question whether the sensitivity of models such as GPT-3.5, and onward,\ncan be improved using similar techniques of text generation.\n", "date": "2023-11-16", "categories": []}
{"id": "2310.04313", "title": "KoMultiText: Large-Scale Korean Text Dataset for Classifying Biased\n  Speech in Real-World Online Services", "abstract": "  With the growth of online services, the need for advanced text classification\nalgorithms, such as sentiment analysis and biased text detection, has become\nincreasingly evident. The anonymous nature of online services often leads to\nthe presence of biased and harmful language, posing challenges to maintaining\nthe health of online communities. This phenomenon is especially relevant in\nSouth Korea, where large-scale hate speech detection algorithms have not yet\nbeen broadly explored. In this paper, we introduce \"KoMultiText\", a new\ncomprehensive, large-scale dataset collected from a well-known South Korean SNS\nplatform. Our proposed dataset provides annotations including (1) Preferences,\n(2) Profanities, and (3) Nine types of Bias for the text samples, enabling\nmulti-task learning for simultaneous classification of user-generated texts.\nLeveraging state-of-the-art BERT-based language models, our approach surpasses\nhuman-level accuracy across diverse classification tasks, as measured by\nvarious metrics. Beyond academic contributions, our work can provide practical\nsolutions for real-world hate speech and bias mitigation, contributing directly\nto the improvement of online community health. Our work provides a robust\nfoundation for future research aiming to improve the quality of online\ndiscourse and foster societal well-being. All source codes and datasets are\npublicly accessible at https://github.com/Dasol-Choi/KoMultiText.\n", "date": "2023-10-06", "categories": []}
{"id": "2311.06446", "title": "THOS: A Benchmark Dataset for Targeted Hate and Offensive Speech", "abstract": "  Detecting harmful content on social media, such as Twitter, is made difficult\nby the fact that the seemingly simple yes/no classification conceals a\nsignificant amount of complexity. Unfortunately, while several datasets have\nbeen collected for training classifiers in hate and offensive speech, there is\na scarcity of datasets labeled with a finer granularity of target classes and\nspecific targets. In this paper, we introduce THOS, a dataset of 8.3k tweets\nmanually labeled with fine-grained annotations about the target of the message.\nWe demonstrate that this dataset makes it feasible to train classifiers, based\non Large Language Models, to perform classification at this level of\ngranularity.\n", "date": "2023-11-11", "categories": []}
{"id": "2311.06737", "title": "Detecting and Correcting Hate Speech in Multimodal Memes with Large\n  Visual Language Model", "abstract": "  Recently, large language models (LLMs) have taken the spotlight in natural\nlanguage processing. Further, integrating LLMs with vision enables the users to\nexplore more emergent abilities in multimodality. Visual language models\n(VLMs), such as LLaVA, Flamingo, or GPT-4, have demonstrated impressive\nperformance on various visio-linguistic tasks. Consequently, there are enormous\napplications of large models that could be potentially used on social media\nplatforms. Despite that, there is a lack of related work on detecting or\ncorrecting hateful memes with VLMs. In this work, we study the ability of VLMs\non hateful meme detection and hateful meme correction tasks with zero-shot\nprompting. From our empirical experiments, we show the effectiveness of the\npretrained LLaVA model and discuss its strengths and weaknesses in these tasks.\n", "date": "2023-11-12", "categories": []}
{"id": "2311.04789", "title": "Determination of toxic comments and unintended model bias minimization\n  using Deep learning approach", "abstract": "  Online conversations can be toxic and subjected to threats, abuse, or\nharassment. To identify toxic text comments, several deep learning and machine\nlearning models have been proposed throughout the years. However, recent\nstudies demonstrate that because of the imbalances in the training data, some\nmodels are more likely to show unintended biases including gender bias and\nidentity bias. In this research, our aim is to detect toxic comment and reduce\nthe unintended bias concerning identity features such as race, gender, sex,\nreligion by fine-tuning an attention based model called BERT(Bidirectional\nEncoder Representation from Transformers). We apply weighted loss to address\nthe issue of unbalanced data and compare the performance of a fine-tuned BERT\nmodel with a traditional Logistic Regression model in terms of classification\nand bias minimization. The Logistic Regression model with the TFIDF vectorizer\nachieve 57.1% accuracy, and fine-tuned BERT model's accuracy is 89%. Code is\navailable at\nhttps://github.com/zim10/Determine_Toxic_comment_and_identity_bias.git\n", "date": "2023-11-08", "categories": []}
{"id": "2311.03969", "title": "Factoring Hate Speech: A New Annotation Framework to Study Hate Speech\n  in Social Media", "abstract": "  In this work we propose a novel annotation scheme which factors hate speech\ninto five separate discursive categories. To evaluate our scheme, we construct\na corpus of over 2.9M Twitter posts containing hateful expressions directed at\nJews, and annotate a sample dataset of 1,050 tweets. We present a statistical\nanalysis of the annotated dataset as well as discuss annotation examples, and\nconclude by discussing promising directions for future work.\n", "date": "2023-11-07", "categories": []}
{"id": "2307.06775", "title": "A Novel Site-Agnostic Multimodal Deep Learning Model to Identify\n  Pro-Eating Disorder Content on Social Media", "abstract": "  Over the last decade, there has been a vast increase in eating disorder\ndiagnoses and eating disorder-attributed deaths, reaching their zenith during\nthe Covid-19 pandemic. This immense growth derived in part from the stressors\nof the pandemic but also from increased exposure to social media, which is rife\nwith content that promotes eating disorders. This study aimed to create a\nmultimodal deep learning model that can determine if a given social media post\npromotes eating disorders based on a combination of visual and textual data. A\nlabeled dataset of Tweets was collected from Twitter, recently rebranded as X,\nupon which twelve deep learning models were trained and evaluated. Based on\nmodel performance, the most effective deep learning model was the multimodal\nfusion of the RoBERTa natural language processing model and the MaxViT image\nclassification model, attaining accuracy and F1 scores of 95.9% and 0.959,\nrespectively. The RoBERTa and MaxViT fusion model, deployed to classify an\nunlabeled dataset of posts from the social media sites Tumblr and Reddit,\ngenerated results akin to those of previous research studies that did not\nemploy artificial intelligence-based techniques, indicating that deep learning\nmodels can develop insights congruent to those of researchers. Additionally,\nthe model was used to conduct a time-series analysis of yet unseen Tweets from\neight Twitter hashtags, uncovering that, since 2014, the relative abundance of\ncontent that promotes eating disorders has decreased drastically within those\ncommunities. Despite this reduction, by 2018, content that promotes eating\ndisorders had either stopped declining or increased in ampleness anew on those\nhashtags.\n", "date": "2023-07-06", "categories": []}
{"id": "2305.13677", "title": "Towards Legally Enforceable Hate Speech Detection for Public Forums", "abstract": "  Hate speech causes widespread and deep-seated societal issues. Proper\nenforcement of hate speech laws is key for protecting groups of people against\nharmful and discriminatory language. However, determining what constitutes hate\nspeech is a complex task that is highly open to subjective interpretations.\nExisting works do not align their systems with enforceable definitions of hate\nspeech, which can make their outputs inconsistent with the goals of regulators.\nThis research introduces a new perspective and task for enforceable hate speech\ndetection centred around legal definitions, and a dataset annotated on\nviolations of eleven possible definitions by legal experts. Given the challenge\nof identifying clear, legally enforceable instances of hate speech, we augment\nthe dataset with expert-generated samples and an automatically mined challenge\nset. We experiment with grounding the model decision in these definitions using\nzero-shot and few-shot prompting. We then report results on several large\nlanguage models (LLMs). With this task definition, automatic hate speech\ndetection can be more closely aligned to enforceable laws, and hence assist in\nmore rigorous enforcement of legal protections against harmful speech in public\nforums.\n", "date": "2023-05-23", "categories": []}
{"id": "2311.00161", "title": "Beyond Denouncing Hate: Strategies for Countering Implied Biases and\n  Stereotypes in Language", "abstract": "  Counterspeech, i.e., responses to counteract potential harms of hateful\nspeech, has become an increasingly popular solution to address online hate\nspeech without censorship. However, properly countering hateful language\nrequires countering and dispelling the underlying inaccurate stereotypes\nimplied by such language. In this work, we draw from psychology and philosophy\nliterature to craft six psychologically inspired strategies to challenge the\nunderlying stereotypical implications of hateful language. We first examine the\nconvincingness of each of these strategies through a user study, and then\ncompare their usages in both human- and machine-generated counterspeech\ndatasets. Our results show that human-written counterspeech uses countering\nstrategies that are more specific to the implied stereotype (e.g., counter\nexamples to the stereotype, external factors about the stereotype's origins),\nwhereas machine-generated counterspeech uses less specific strategies (e.g.,\ngenerally denouncing the hatefulness of speech). Furthermore, machine-generated\ncounterspeech often employs strategies that humans deem less convincing\ncompared to human-produced counterspeech. Our findings point to the importance\nof accounting for the underlying stereotypical implications of speech when\ngenerating counterspeech and for better machine reasoning about\nanti-stereotypical examples.\n", "date": "2023-10-31", "categories": []}
{"id": "2310.12860", "title": "Probing LLMs for hate speech detection: strengths and vulnerabilities", "abstract": "  Recently efforts have been made by social media platforms as well as\nresearchers to detect hateful or toxic language using large language models.\nHowever, none of these works aim to use explanation, additional context and\nvictim community information in the detection process. We utilise different\nprompt variation, input information and evaluate large language models in zero\nshot setting (without adding any in-context examples). We select three large\nlanguage models (GPT-3.5, text-davinci and Flan-T5) and three datasets -\nHateXplain, implicit hate and ToxicSpans. We find that on average including the\ntarget information in the pipeline improves the model performance substantially\n(~20-30%) over the baseline across the datasets. There is also a considerable\neffect of adding the rationales/explanations into the pipeline (~10-20%) over\nthe baseline across the datasets. In addition, we further provide a typology of\nthe error cases where these large language models fail to (i) classify and (ii)\nexplain the reason for the decisions they take. Such vulnerable points\nautomatically constitute 'jailbreak' prompts for these models and industry\nscale safeguard techniques need to be developed to make the models robust\nagainst such prompts.\n", "date": "2023-10-19", "categories": []}
{"id": "2303.03387", "title": "CoSyn: Detecting Implicit Hate Speech in Online Conversations Using a\n  Context Synergized Hyperbolic Network", "abstract": "  The tremendous growth of social media users interacting in online\nconversations has led to significant growth in hate speech, affecting people\nfrom various demographics. Most of the prior works focus on detecting explicit\nhate speech, which is overt and leverages hateful phrases, with very little\nwork focusing on detecting hate speech that is implicit or denotes hatred\nthrough indirect or coded language. In this paper, we present CoSyn, a\ncontext-synergized neural network that explicitly incorporates user- and\nconversational context for detecting implicit hate speech in online\nconversations. CoSyn introduces novel ways to encode these external contexts\nand employs a novel context interaction mechanism that clearly captures the\ninterplay between them, making independent assessments of the amounts of\ninformation to be retrieved from these noisy contexts. Additionally, it carries\nout all these operations in the hyperbolic space to account for the scale-free\ndynamics of social media. We demonstrate the effectiveness of CoSyn on 6 hate\nspeech datasets and show that CoSyn outperforms all our baselines in detecting\nimplicit hate speech with absolute improvements in the range of 1.24% - 57.8%.\n", "date": "2023-03-02", "categories": []}
{"id": "2310.15439", "title": "K-HATERS: A Hate Speech Detection Corpus in Korean with Target-Specific\n  Ratings", "abstract": "  Numerous datasets have been proposed to combat the spread of online hate.\nDespite these efforts, a majority of these resources are English-centric,\nprimarily focusing on overt forms of hate. This research gap calls for\ndeveloping high-quality corpora in diverse languages that also encapsulate more\nsubtle hate expressions. This study introduces K-HATERS, a new corpus for hate\nspeech detection in Korean, comprising approximately 192K news comments with\ntarget-specific offensiveness ratings. This resource is the largest offensive\nlanguage corpus in Korean and is the first to offer target-specific ratings on\na three-point Likert scale, enabling the detection of hate expressions in\nKorean across varying degrees of offensiveness. We conduct experiments showing\nthe effectiveness of the proposed corpus, including a comparison with existing\ndatasets. Additionally, to address potential noise and bias in human\nannotations, we explore a novel idea of adopting the Cognitive Reflection Test,\nwhich is widely used in social science for assessing an individual's cognitive\nability, as a proxy of labeling quality. Findings indicate that annotations\nfrom individuals with the lowest test scores tend to yield detection models\nthat make biased predictions toward specific target groups and are less\naccurate. This study contributes to the NLP research on hate speech detection\nand resource construction. The code and dataset can be accessed at\nhttps://github.com/ssu-humane/K-HATERS.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.13985", "title": "HateRephrase: Zero- and Few-Shot Reduction of Hate Intensity in Online\n  Posts using Large Language Models", "abstract": "  Hate speech has become pervasive in today's digital age. Although there has\nbeen considerable research to detect hate speech or generate counter speech to\ncombat hateful views, these approaches still cannot completely eliminate the\npotential harmful societal consequences of hate speech -- hate speech, even\nwhen detected, can often not be taken down or is often not taken down enough;\nand hate speech unfortunately spreads quickly, often much faster than any\ngenerated counter speech.\n  This paper investigates a relatively new yet simple and effective approach of\nsuggesting a rephrasing of potential hate speech content even before the post\nis made. We show that Large Language Models (LLMs) perform well on this task,\noutperforming state-of-the-art baselines such as BART-Detox. We develop 4\ndifferent prompts based on task description, hate definition, few-shot\ndemonstrations and chain-of-thoughts for comprehensive experiments and conduct\nexperiments on open-source LLMs such as LLaMA-1, LLaMA-2 chat, Vicuna as well\nas OpenAI's GPT-3.5. We propose various evaluation metrics to measure the\nefficacy of the generated text and ensure the generated text has reduced hate\nintensity without drastically changing the semantic meaning of the original\ntext.\n  We find that LLMs with a few-shot demonstrations prompt work the best in\ngenerating acceptable hate-rephrased text with semantic meaning similar to the\noriginal text. Overall, we find that GPT-3.5 outperforms the baseline and\nopen-source models for all the different kinds of prompts. We also perform\nhuman evaluations and interestingly, find that the rephrasings generated by\nGPT-3.5 outperform even the human-generated ground-truth rephrasings in the\ndataset. We also conduct detailed ablation studies to investigate why LLMs work\nsatisfactorily on this task and conduct a failure analysis to understand the\ngaps.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14325", "title": "Towards Harmful Erotic Content Detection through Coreference-Driven\n  Contextual Analysis", "abstract": "  Adult content detection still poses a great challenge for automation.\nExisting classifiers primarily focus on distinguishing between erotic and\nnon-erotic texts. However, they often need more nuance in assessing the\npotential harm. Unfortunately, the content of this nature falls beyond the\nreach of generative models due to its potentially harmful nature. Ethical\nrestrictions prohibit large language models (LLMs) from analyzing and\nclassifying harmful erotics, let alone generating them to create synthetic\ndatasets for other neural models. In such instances where data is scarce and\nchallenging, a thorough analysis of the structure of such texts rather than a\nlarge model may offer a viable solution. Especially given that harmful erotic\nnarratives, despite appearing similar to harmless ones, usually reveal their\nharmful nature first through contextual information hidden in the non-sexual\nparts of the narrative.\n  This paper introduces a hybrid neural and rule-based context-aware system\nthat leverages coreference resolution to identify harmful contextual cues in\nerotic content. Collaborating with professional moderators, we compiled a\ndataset and developed a classifier capable of distinguishing harmful from\nnon-harmful erotic content. Our hybrid model, tested on Polish text,\ndemonstrates a promising accuracy of 84% and a recall of 80%. Models based on\nRoBERTa and Longformer without explicit usage of coreference chains achieved\nsignificantly weaker results, underscoring the importance of coreference\nresolution in detecting such nuanced content as harmful erotics. This approach\nalso offers the potential for enhanced visual explainability, supporting\nmoderators in evaluating predictions and taking necessary actions to address\nharmful content.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.10781", "title": "BanglaNLP at BLP-2023 Task 1: Benchmarking different Transformer Models\n  for Violence Inciting Text Detection in Bengali", "abstract": "  This paper presents the system that we have developed while solving this\nshared task on violence inciting text detection in Bangla. We explain both the\ntraditional and the recent approaches that we have used to make our models\nlearn. Our proposed system helps to classify if the given text contains any\nthreat. We studied the impact of data augmentation when there is a limited\ndataset available. Our quantitative results show that finetuning a\nmultilingual-e5-base model performed the best in our task compared to other\ntransformer-based architectures. We obtained a macro F1 of 68.11\\% in the test\nset and our performance in this shared task is ranked at 23 in the leaderboard.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.08368", "title": "Mapping Memes to Words for Multimodal Hateful Meme Classification", "abstract": "  Multimodal image-text memes are prevalent on the internet, serving as a\nunique form of communication that combines visual and textual elements to\nconvey humor, ideas, or emotions. However, some memes take a malicious turn,\npromoting hateful content and perpetuating discrimination. Detecting hateful\nmemes within this multimodal context is a challenging task that requires\nunderstanding the intertwined meaning of text and images. In this work, we\naddress this issue by proposing a novel approach named ISSUES for multimodal\nhateful meme classification. ISSUES leverages a pre-trained CLIP\nvision-language model and the textual inversion technique to effectively\ncapture the multimodal semantic content of the memes. The experiments show that\nour method achieves state-of-the-art results on the Hateful Memes Challenge and\nHarMeme datasets. The code and the pre-trained models are publicly available at\nhttps://github.com/miccunifi/ISSUES.\n", "date": "2023-10-12", "categories": []}
{"id": "2306.08804", "title": "PEACE: Cross-Platform Hate Speech Detection- A Causality-guided\n  Framework", "abstract": "  Hate speech detection refers to the task of detecting hateful content that\naims at denigrating an individual or a group based on their religion, gender,\nsexual orientation, or other characteristics. Due to the different policies of\nthe platforms, different groups of people express hate in different ways.\nFurthermore, due to the lack of labeled data in some platforms it becomes\nchallenging to build hate speech detection models. To this end, we revisit if\nwe can learn a generalizable hate speech detection model for the cross platform\nsetting, where we train the model on the data from one (source) platform and\ngeneralize the model across multiple (target) platforms. Existing\ngeneralization models rely on linguistic cues or auxiliary information, making\nthem biased towards certain tags or certain kinds of words (e.g., abusive\nwords) on the source platform and thus not applicable to the target platforms.\nInspired by social and psychological theories, we endeavor to explore if there\nexist inherent causal cues that can be leveraged to learn generalizable\nrepresentations for detecting hate speech across these distribution shifts. To\nthis end, we propose a causality-guided framework, PEACE, that identifies and\nleverages two intrinsic causal cues omnipresent in hateful content: the overall\nsentiment and the aggression in the text. We conduct extensive experiments\nacross multiple platforms (representing the distribution shift) showing if\ncausal cues can help cross-platform generalization.\n", "date": "2023-06-15", "categories": []}
{"id": "2306.14939", "title": "The Art of Embedding Fusion: Optimizing Hate Speech Detection", "abstract": "  Hate speech detection is a challenging natural language processing task that\nrequires capturing linguistic and contextual nuances. Pre-trained language\nmodels (PLMs) offer rich semantic representations of text that can improve this\ntask. However there is still limited knowledge about ways to effectively\ncombine representations across PLMs and leverage their complementary strengths.\nIn this work, we shed light on various combination techniques for several PLMs\nand comprehensively analyze their effectiveness. Our findings show that\ncombining embeddings leads to slight improvements but at a high computational\ncost and the choice of combination has marginal effect on the final outcome. We\nalso make our codebase public at\nhttps://github.com/aflah02/The-Art-of-Embedding-Fusion-Optimizing-Hate-Speech-Detection .\n", "date": "2023-06-26", "categories": []}
{"id": "2310.04465", "title": "How toxic is antisemitism? Potentials and limitations of automated\n  toxicity scoring for antisemitic online content", "abstract": "  The Perspective API, a popular text toxicity assessment service by Google and\nJigsaw, has found wide adoption in several application areas, notably content\nmoderation, monitoring, and social media research. We examine its potentials\nand limitations for the detection of antisemitic online content that, by\ndefinition, falls under the toxicity umbrella term. Using a manually annotated\nGerman-language dataset comprising around 3,600 posts from Telegram and\nTwitter, we explore as how toxic antisemitic texts are rated and how the\ntoxicity scores differ regarding different subforms of antisemitism and the\nstance expressed in the texts. We show that, on a basic level, Perspective API\nrecognizes antisemitic content as toxic, but shows critical weaknesses with\nrespect to non-explicit forms of antisemitism and texts taking a critical\nstance towards it. Furthermore, using simple text manipulations, we demonstrate\nthat the use of widespread antisemitic codes can substantially reduce API\nscores, making it rather easy to bypass content moderation based on the\nservice's results.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.05192", "title": "HOD: A Benchmark Dataset for Harmful Object Detection", "abstract": "  Recent multi-media data such as images and videos have been rapidly spread\nout on various online services such as social network services (SNS). With the\nexplosive growth of online media services, the number of image content that may\nharm users is also growing exponentially. Thus, most recent online platforms\nsuch as Facebook and Instagram have adopted content filtering systems to\nprevent the prevalence of harmful content and reduce the possible risk of\nadverse effects on users. Unfortunately, computer vision research on detecting\nharmful content has not yet attracted attention enough. Users of each platform\nstill manually click the report button to recognize patterns of harmful content\nthey dislike when exposed to harmful content. However, the problem with manual\nreporting is that users are already exposed to harmful content. To address\nthese issues, our research goal in this work is to develop automatic harmful\nobject detection systems for online services. We present a new benchmark\ndataset for harmful object detection. Unlike most related studies focusing on a\nsmall subset of object categories, our dataset addresses various categories.\nSpecifically, our proposed dataset contains more than 10,000 images across 6\ncategories that might be harmful, consisting of not only normal cases but also\nhard cases that are difficult to detect. Moreover, we have conducted extensive\nexperiments to evaluate the effectiveness of our proposed dataset. We have\nutilized the recently proposed state-of-the-art (SOTA) object detection\narchitectures and demonstrated our proposed dataset can be greatly useful for\nthe real-time harmful object detection task. The whole source codes and\ndatasets are publicly accessible at\nhttps://github.com/poori-nuna/HOD-Benchmark-Dataset.\n", "date": "2023-10-08", "categories": []}
{"id": "2309.16905", "title": "Towards a Unified Framework for Adaptable Problematic Content Detection\n  via Continual Learning", "abstract": "  Detecting problematic content, such as hate speech, is a multifaceted and\never-changing task, influenced by social dynamics, user populations, diversity\nof sources, and evolving language. There has been significant efforts, both in\nacademia and in industry, to develop annotated resources that capture various\naspects of problematic content. Due to researchers' diverse objectives, the\nannotations are inconsistent and hence, reports of progress on detection of\nproblematic content are fragmented. This pattern is expected to persist unless\nwe consolidate resources considering the dynamic nature of the problem. We\npropose integrating the available resources, and leveraging their dynamic\nnature to break this pattern. In this paper, we introduce a continual learning\nbenchmark and framework for problematic content detection comprising over 84\nrelated tasks encompassing 15 annotation schemas from 8 sources. Our benchmark\ncreates a novel measure of progress: prioritizing the adaptability of\nclassifiers to evolving tasks over excelling in specific tasks. To ensure the\ncontinuous relevance of our framework, we designed it so that new tasks can\neasily be integrated into the benchmark. Our baseline results demonstrate the\npotential of continual learning in capturing the evolving content and adapting\nto novel manifestations of problematic content.\n", "date": "2023-09-29", "categories": []}
{"id": "2310.02876", "title": "Hate Speech Detection in Limited Data Contexts using Synthetic Data\n  Generation", "abstract": "  A growing body of work has focused on text classification methods for\ndetecting the increasing amount of hate speech posted online. This progress has\nbeen limited to only a select number of highly-resourced languages causing\ndetection systems to either under-perform or not exist in limited data\ncontexts. This is majorly caused by a lack of training data which is expensive\nto collect and curate in these settings. In this work, we propose a data\naugmentation approach that addresses the problem of lack of data for online\nhate speech detection in limited data contexts using synthetic data generation\ntechniques. Given a handful of hate speech examples in a high-resource language\nsuch as English, we present three methods to synthesize new examples of hate\nspeech data in a target language that retains the hate sentiment in the\noriginal examples but transfers the hate targets. We apply our approach to\ngenerate training data for hate speech classification tasks in Hindi and\nVietnamese. Our findings show that a model trained on synthetic data performs\ncomparably to, and in some cases outperforms, a model trained only on the\nsamples available in the target domain. This method can be adopted to bootstrap\nhate speech detection models from scratch in limited data contexts. As the\ngrowth of social media within these contexts continues to outstrip response\nefforts, this work furthers our capacities for detection, understanding, and\nresponse to hate speech.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.02249", "title": "Harnessing Pre-Trained Sentence Transformers for Offensive Language\n  Detection in Indian Languages", "abstract": "  In our increasingly interconnected digital world, social media platforms have\nemerged as powerful channels for the dissemination of hate speech and offensive\ncontent. This work delves into the domain of hate speech detection, placing\nspecific emphasis on three low-resource Indian languages: Bengali, Assamese,\nand Gujarati. The challenge is framed as a text classification task, aimed at\ndiscerning whether a tweet contains offensive or non-offensive content.\nLeveraging the HASOC 2023 datasets, we fine-tuned pre-trained BERT and SBERT\nmodels to evaluate their effectiveness in identifying hate speech. Our findings\nunderscore the superiority of monolingual sentence-BERT models, particularly in\nthe Bengali language, where we achieved the highest ranking. However, the\nperformance in Assamese and Gujarati languages signifies ongoing opportunities\nfor enhancement. Our goal is to foster inclusive online spaces by countering\nhate speech proliferation.\n", "date": "2023-10-03", "categories": []}
{"id": "2309.15827", "title": "How We Define Harm Impacts Data Annotations: Explaining How Annotators\n  Distinguish Hateful, Offensive, and Toxic Comments", "abstract": "  Computational social science research has made advances in machine learning\nand natural language processing that support content moderators in detecting\nharmful content. These advances often rely on training datasets annotated by\ncrowdworkers for harmful content. In designing instructions for annotation\ntasks to generate training data for these algorithms, researchers often treat\nthe harm concepts that we train algorithms to detect - 'hateful', 'offensive',\n'toxic', 'racist', 'sexist', etc. - as interchangeable. In this work, we\nstudied whether the way that researchers define 'harm' affects annotation\noutcomes. Using Venn diagrams, information gain comparisons, and content\nanalyses, we reveal that annotators do not use the concepts 'hateful',\n'offensive', and 'toxic' interchangeably. We identify that features of harm\ndefinitions and annotators' individual characteristics explain much of how\nannotators use these terms differently. Our results offer empirical evidence\ndiscouraging the common practice of using harm concepts interchangeably in\ncontent moderation research. Instead, researchers should make specific choices\nabout which harm concepts to analyze based on their research goals. Recognizing\nthat researchers are often resource constrained, we also encourage researchers\nto provide information to bound their findings when their concepts of interest\ndiffer from concepts that off-the-shelf harmful content detection algorithms\nidentify. Finally, we encourage algorithm providers to ensure their instruments\ncan adapt to contextually-specific content detection goals (e.g., soliciting\ninstrument users' feedback).\n", "date": "2023-09-12", "categories": []}
{"id": "2206.12983", "title": "Explainable and High-Performance Hate and Offensive Speech Detection", "abstract": "  The spread of information through social media platforms can create\nenvironments possibly hostile to vulnerable communities and silence certain\ngroups in society. To mitigate such instances, several models have been\ndeveloped to detect hate and offensive speech. Since detecting hate and\noffensive speech in social media platforms could incorrectly exclude\nindividuals from social media platforms, which can reduce trust, there is a\nneed to create explainable and interpretable models. Thus, we build an\nexplainable and interpretable high performance model based on the XGBoost\nalgorithm, trained on Twitter data. For unbalanced Twitter data, XGboost\noutperformed the LSTM, AutoGluon, and ULMFiT models on hate speech detection\nwith an F1 score of 0.75 compared to 0.38 and 0.37, and 0.38 respectively. When\nwe down-sampled the data to three separate classes of approximately 5000\ntweets, XGBoost performed better than LSTM, AutoGluon, and ULMFiT; with F1\nscores for hate speech detection of 0.79 vs 0.69, 0.77, and 0.66 respectively.\nXGBoost also performed better than LSTM, AutoGluon, and ULMFiT in the\ndown-sampled version for offensive speech detection with F1 score of 0.83 vs\n0.88, 0.82, and 0.79 respectively. We use Shapley Additive Explanations (SHAP)\non our XGBoost models' outputs to makes it explainable and interpretable\ncompared to LSTM, AutoGluon and ULMFiT that are black-box models.\n", "date": "2022-06-26", "categories": []}
{"id": "2309.13354", "title": "Lexical Squad@Multimodal Hate Speech Event Detection 2023: Multimodal\n  Hate Speech Detection using Fused Ensemble Approach", "abstract": "  With a surge in the usage of social media postings to express opinions,\nemotions, and ideologies, there has been a significant shift towards the\ncalibration of social media as a rapid medium of conveying viewpoints and\noutlooks over the globe. Concurrently, the emergence of a multitude of\nconflicts between two entities has given rise to a stream of social media\ncontent containing propaganda, hate speech, and inconsiderate views. Thus, the\nissue of monitoring social media postings is rising swiftly, attracting major\nattention from those willing to solve such problems. One such problem is Hate\nSpeech detection. To mitigate this problem, we present our novel ensemble\nlearning approach for detecting hate speech, by classifying text-embedded\nimages into two labels, namely \"Hate Speech\" and \"No Hate Speech\". We have\nincorporated state-of-art models including InceptionV3, BERT, and XLNet. Our\nproposed ensemble model yielded promising results with 75.21 and 74.96 as\naccuracy and F-1 score (respectively). We also present an empirical evaluation\nof the text-embedded images to elaborate on how well the model was able to\npredict and classify. We release our codebase here\n(https://github.com/M0hammad-Kashif/MultiModalHateSpeech).\n", "date": "2023-09-23", "categories": []}
{"id": "2309.13561", "title": "Cordyceps@LT-EDI: Patching Language-Specific Homophobia/Transphobia\n  Classifiers with a Multilingual Understanding", "abstract": "  Detecting transphobia, homophobia, and various other forms of hate speech is\ndifficult. Signals can vary depending on factors such as language, culture,\ngeographical region, and the particular online platform. Here, we present a\njoint multilingual (M-L) and language-specific (L-S) approach to homophobia and\ntransphobic hate speech detection (HSD). M-L models are needed to catch words,\nphrases, and concepts that are less common or missing in a particular language\nand subsequently overlooked by L-S models. Nonetheless, L-S models are better\nsituated to understand the cultural and linguistic context of the users who\ntypically write in a particular language. Here we construct a simple and\nsuccessful way to merge the M-L and L-S approaches through simple weight\ninterpolation in such a way that is interpretable and data-driven. We\ndemonstrate our system on task A of the 'Shared Task on Homophobia/Transphobia\nDetection in social media comments' dataset for homophobia and transphobic HSD.\nOur system achieves the best results in three of five languages and achieves a\n0.997 macro average F1-score on Malayalam texts.\n", "date": "2023-09-24", "categories": []}
{"id": "2309.14146", "title": "Examining Temporal Bias in Abusive Language Detection", "abstract": "  The use of abusive language online has become an increasingly pervasive\nproblem that damages both individuals and society, with effects ranging from\npsychological harm right through to escalation to real-life violence and even\ndeath. Machine learning models have been developed to automatically detect\nabusive language, but these models can suffer from temporal bias, the\nphenomenon in which topics, language use or social norms change over time. This\nstudy aims to investigate the nature and impact of temporal bias in abusive\nlanguage detection across various languages and explore mitigation methods. We\nevaluate the performance of models on abusive data sets from different time\nperiods. Our results demonstrate that temporal bias is a significant challenge\nfor abusive language detection, with models trained on historical data showing\na significant drop in performance over time. We also present an extensive\nlinguistic analysis of these abusive data sets from a diachronic perspective,\naiming to explore the reasons for language evolution and performance decline.\nThis study sheds light on the pervasive issue of temporal bias in abusive\nlanguage detection across languages, offering crucial insights into language\nevolution and temporal bias mitigation.\n", "date": "2023-09-25", "categories": []}
{"id": "2309.05150", "title": "Faster, Lighter, More Accurate: A Deep Learning Ensemble for Content\n  Moderation", "abstract": "  To address the increasing need for efficient and accurate content moderation,\nwe propose an efficient and lightweight deep classification ensemble structure.\nOur approach is based on a combination of simple visual features, designed for\nhigh-accuracy classification of violent content with low false positives. Our\nensemble architecture utilizes a set of lightweight models with narrowed-down\ncolor features, and we apply it to both images and videos.\n  We evaluated our approach using a large dataset of explosion and blast\ncontents and compared its performance to popular deep learning models such as\nResNet-50. Our evaluation results demonstrate significant improvements in\nprediction accuracy, while benefiting from 7.64x faster inference and lower\ncomputation cost.\n  While our approach is tailored to explosion detection, it can be applied to\nother similar content moderation and violence detection use cases as well.\nBased on our experiments, we propose a \"think small, think many\" philosophy in\nclassification scenarios. We argue that transforming a single, large,\nmonolithic deep model into a verification-based step model ensemble of multiple\nsmall, simple, and lightweight models with narrowed-down visual features can\npossibly lead to predictions with higher accuracy.\n", "date": "2023-09-10", "categories": []}
{"id": "2309.02724", "title": "Offensive Hebrew Corpus and Detection using BERT", "abstract": "  Offensive language detection has been well studied in many languages, but it\nis lagging behind in low-resource languages, such as Hebrew. In this paper, we\npresent a new offensive language corpus in Hebrew. A total of 15,881 tweets\nwere retrieved from Twitter. Each was labeled with one or more of five classes\n(abusive, hate, violence, pornographic, or none offensive) by Arabic-Hebrew\nbilingual speakers. The annotation process was challenging as each annotator is\nexpected to be familiar with the Israeli culture, politics, and practices to\nunderstand the context of each tweet. We fine-tuned two Hebrew BERT models,\nHeBERT and AlephBERT, using our proposed dataset and another published dataset.\nWe observed that our data boosts HeBERT performance by 2% when combined with\nD_OLaH. Fine-tuning AlephBERT on our data and testing on D_OLaH yields 69%\naccuracy, while fine-tuning on D_OLaH and testing on our data yields 57%\naccuracy, which may be an indication to the generalizability our data offers.\nOur dataset and fine-tuned models are available on GitHub and Huggingface.\n", "date": "2023-09-06", "categories": []}
{"id": "2309.02912", "title": "On the Challenges of Building Datasets for Hate Speech Detection", "abstract": "  Detection of hate speech has been formulated as a standalone application of\nNLP and different approaches have been adopted for identifying the target\ngroups, obtaining raw data, defining the labeling process, choosing the\ndetection algorithm, and evaluating the performance in the desired setting.\nHowever, unlike other downstream tasks, hate speech suffers from the lack of\nlarge-sized, carefully curated, generalizable datasets owing to the highly\nsubjective nature of the task. In this paper, we first analyze the issues\nsurrounding hate speech detection through a data-centric lens. We then outline\na holistic framework to encapsulate the data creation pipeline across seven\nbroad dimensions by taking the specific example of hate speech towards sexual\nminorities. We posit that practitioners would benefit from following this\nframework as a form of best practice when creating hate speech datasets in the\nfuture.\n", "date": "2023-09-06", "categories": []}
{"id": "2309.01413", "title": "Hateful Messages: A Conversational Data Set of Hate Speech produced by\n  Adolescents on Discord", "abstract": "  With the rise of social media, a rise of hateful content can be observed.\nEven though the understanding and definitions of hate speech varies, platforms,\ncommunities, and legislature all acknowledge the problem. Therefore,\nadolescents are a new and active group of social media users. The majority of\nadolescents experience or witness online hate speech. Research in the field of\nautomated hate speech classification has been on the rise and focuses on\naspects such as bias, generalizability, and performance. To increase\ngeneralizability and performance, it is important to understand biases within\nthe data. This research addresses the bias of youth language within hate speech\nclassification and contributes by providing a modern and anonymized hate speech\nyouth language data set consisting of 88.395 annotated chat messages. The data\nset consists of publicly available online messages from the chat platform\nDiscord. ~6,42% of the messages were classified by a self-developed annotation\nschema as hate speech. For 35.553 messages, the user profiles provided age\nannotations setting the average author age to under 20 years old.\n", "date": "2023-09-04", "categories": []}
{"id": "2309.02311", "title": "Weigh Your Own Words: Improving Hate Speech Counter Narrative Generation\n  via Attention Regularization", "abstract": "  Recent computational approaches for combating online hate speech involve the\nautomatic generation of counter narratives by adapting Pretrained\nTransformer-based Language Models (PLMs) with human-curated data. This process,\nhowever, can produce in-domain overfitting, resulting in models generating\nacceptable narratives only for hatred similar to training data, with little\nportability to other targets or to real-world toxic language. This paper\nintroduces novel attention regularization methodologies to improve the\ngeneralization capabilities of PLMs for counter narratives generation.\nOverfitting to training-specific terms is then discouraged, resulting in more\ndiverse and richer narratives. We experiment with two attention-based\nregularization techniques on a benchmark English dataset. Regularized models\nproduce better counter narratives than state-of-the-art approaches in most\ncases, both in terms of automatic metrics and human evaluation, especially when\nhateful targets are not present in the training data. This work paves the way\nfor better and more flexible counter-speech generation models, a task for which\ndatasets are highly challenging to produce.\n", "date": "2023-09-05", "categories": []}
{"id": "2308.15448", "title": "Vulgar Remarks Detection in Chittagonian Dialect of Bangla", "abstract": "  The negative effects of online bullying and harassment are increasing with\nInternet popularity, especially in social media. One solution is using natural\nlanguage processing (NLP) and machine learning (ML) methods for the automatic\ndetection of harmful remarks, but these methods are limited in low-resource\nlanguages like the Chittagonian dialect of Bangla.This study focuses on\ndetecting vulgar remarks in social media using supervised ML and deep learning\nalgorithms.Logistic Regression achieved promising accuracy (0.91) while simple\nRNN with Word2vec and fastTex had lower accuracy (0.84-0.90), highlighting the\nissue that NN algorithms require more data.\n", "date": "2023-08-29", "categories": []}
{"id": "2308.14683", "title": "Fine-Tuning Llama 2 Large Language Models for Detecting Online Sexual\n  Predatory Chats and Abusive Texts", "abstract": "  Detecting online sexual predatory behaviours and abusive language on social\nmedia platforms has become a critical area of research due to the growing\nconcerns about online safety, especially for vulnerable populations such as\nchildren and adolescents. Researchers have been exploring various techniques\nand approaches to develop effective detection systems that can identify and\nmitigate these risks. Recent development of large language models (LLMs) has\nopened a new opportunity to address this problem more effectively. This paper\nproposes an approach to detection of online sexual predatory chats and abusive\nlanguage using the open-source pretrained Llama 2 7B-parameter model, recently\nreleased by Meta GenAI. We fine-tune the LLM using datasets with different\nsizes, imbalance degrees, and languages (i.e., English, Roman Urdu and Urdu).\nBased on the power of LLMs, our approach is generic and automated without a\nmanual search for a synergy between feature extraction and classifier design\nsteps like conventional methods in this domain. Experimental results show a\nstrong performance of the proposed approach, which performs proficiently and\nconsistently across three distinct datasets with five sets of experiments. This\nstudy's outcomes indicate that the proposed method can be implemented in\nreal-world applications (even with non-English languages) for flagging sexual\npredators, offensive or toxic content, hate speech, and discriminatory language\nin online discussions and comments to maintain respectful internet or digital\ncommunities. Furthermore, it can be employed for solving text classification\nproblems with other potential applications such as sentiment analysis, spam and\nphishing detection, sorting legal documents, fake news detection, language\nidentification, user intent recognition, text-based product categorization,\nmedical record analysis, and resume screening.\n", "date": "2023-08-28", "categories": []}
{"id": "2308.10370", "title": "cantnlp@LT-EDI-2023: Homophobia/Transphobia Detection in Social Media\n  Comments using Spatio-Temporally Retrained Language Models", "abstract": "  This paper describes our multiclass classification system developed as part\nof the LTEDI@RANLP-2023 shared task. We used a BERT-based language model to\ndetect homophobic and transphobic content in social media comments across five\nlanguage conditions: English, Spanish, Hindi, Malayalam, and Tamil. We\nretrained a transformer-based crosslanguage pretrained language model,\nXLMRoBERTa, with spatially and temporally relevant social media language data.\nWe also retrained a subset of models with simulated script-mixed social media\nlanguage data with varied performance. We developed the best performing\nseven-label classification system for Malayalam based on weighted macro\naveraged F1 score (ranked first out of six) with variable performance for other\nlanguage and class-label conditions. We found the inclusion of this\nspatio-temporal data improved the classification performance for all language\nand task conditions when compared with the baseline. The results suggests that\ntransformer-based language classification systems are sensitive to\nregister-specific and language-specific retraining.\n", "date": "2023-08-20", "categories": []}
{"id": "2308.09722", "title": "A Trustable LSTM-Autoencoder Network for Cyberbullying Detection on\n  Social Media Using Synthetic Data", "abstract": "  Social media cyberbullying has a detrimental effect on human life. As online\nsocial networking grows daily, the amount of hate speech also increases. Such\nterrible content can cause depression and actions related to suicide. This\npaper proposes a trustable LSTM-Autoencoder Network for cyberbullying detection\non social media using synthetic data. We have demonstrated a cutting-edge\nmethod to address data availability difficulties by producing\nmachine-translated data. However, several languages such as Hindi and Bangla\nstill lack adequate investigations due to a lack of datasets. We carried out\nexperimental identification of aggressive comments on Hindi, Bangla, and\nEnglish datasets using the proposed model and traditional models, including\nLong Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM),\nLSTM-Autoencoder, Word2vec, Bidirectional Encoder Representations from\nTransformers (BERT), and Generative Pre-trained Transformer 2 (GPT-2) models.\nWe employed evaluation metrics such as f1-score, accuracy, precision, and\nrecall to assess the models performance. Our proposed model outperformed all\nthe models on all datasets, achieving the highest accuracy of 95%. Our model\nachieves state-of-the-art results among all the previous works on the dataset\nwe used in this paper.\n", "date": "2023-08-15", "categories": []}
{"id": "2308.09810", "title": "An Image is Worth a Thousand Toxic Words: A Metamorphic Testing\n  Framework for Content Moderation Software", "abstract": "  The exponential growth of social media platforms has brought about a\nrevolution in communication and content dissemination in human society.\nNevertheless, these platforms are being increasingly misused to spread toxic\ncontent, including hate speech, malicious advertising, and pornography, leading\nto severe negative consequences such as harm to teenagers' mental health.\nDespite tremendous efforts in developing and deploying textual and image\ncontent moderation methods, malicious users can evade moderation by embedding\ntexts into images, such as screenshots of the text, usually with some\ninterference. We find that modern content moderation software's performance\nagainst such malicious inputs remains underexplored. In this work, we propose\nOASIS, a metamorphic testing framework for content moderation software. OASIS\nemploys 21 transform rules summarized from our pilot study on 5,000 real-world\ntoxic contents collected from 4 popular social media applications, including\nTwitter, Instagram, Sina Weibo, and Baidu Tieba. Given toxic textual contents,\nOASIS can generate image test cases, which preserve the toxicity yet are likely\nto bypass moderation. In the evaluation, we employ OASIS to test five\ncommercial textual content moderation software from famous companies (i.e.,\nGoogle Cloud, Microsoft Azure, Baidu Cloud, Alibaba Cloud and Tencent Cloud),\nas well as a state-of-the-art moderation research model. The results show that\nOASIS achieves up to 100% error finding rates. Moreover, through retraining the\nmodels with the test cases generated by OASIS, the robustness of the moderation\nmodel can be improved without performance degradation.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.08088", "title": "Pro-Cap: Leveraging a Frozen Vision-Language Model for Hateful Meme\n  Detection", "abstract": "  Hateful meme detection is a challenging multimodal task that requires\ncomprehension of both vision and language, as well as cross-modal interactions.\nRecent studies have tried to fine-tune pre-trained vision-language models\n(PVLMs) for this task. However, with increasing model sizes, it becomes\nimportant to leverage powerful PVLMs more efficiently, rather than simply\nfine-tuning them. Recently, researchers have attempted to convert meme images\ninto textual captions and prompt language models for predictions. This approach\nhas shown good performance but suffers from non-informative image captions.\nConsidering the two factors mentioned above, we propose a probing-based\ncaptioning approach to leverage PVLMs in a zero-shot visual question answering\n(VQA) manner. Specifically, we prompt a frozen PVLM by asking hateful\ncontent-related questions and use the answers as image captions (which we call\nPro-Cap), so that the captions contain information critical for hateful content\ndetection. The good performance of models with Pro-Cap on three benchmarks\nvalidates the effectiveness and generalization of the proposed method.\n", "date": "2023-08-16", "categories": []}
{"id": "2308.05596", "title": "You Only Prompt Once: On the Capabilities of Prompt Learning on Large\n  Language Models to Tackle Toxic Content", "abstract": "  The spread of toxic content online is an important problem that has adverse\neffects on user experience online and in our society at large. Motivated by the\nimportance and impact of the problem, research focuses on developing solutions\nto detect toxic content, usually leveraging machine learning (ML) models\ntrained on human-annotated datasets. While these efforts are important, these\nmodels usually do not generalize well and they can not cope with new trends\n(e.g., the emergence of new toxic terms). Currently, we are witnessing a shift\nin the approach to tackling societal issues online, particularly leveraging\nlarge language models (LLMs) like GPT-3 or T5 that are trained on vast corpora\nand have strong generalizability. In this work, we investigate how we can use\nLLMs and prompt learning to tackle the problem of toxic content, particularly\nfocusing on three tasks; 1) Toxicity Classification, 2) Toxic Span Detection,\nand 3) Detoxification. We perform an extensive evaluation over five model\narchitectures and eight datasets demonstrating that LLMs with prompt learning\ncan achieve similar or even better performance compared to models trained on\nthese specific tasks. We find that prompt learning achieves around 10\\%\nimprovement in the toxicity classification task compared to the baselines,\nwhile for the toxic span detection task we find better performance to the best\nbaseline (0.643 vs. 0.640 in terms of $F_1$-score). Finally, for the\ndetoxification task, we find that prompt learning can successfully reduce the\naverage toxicity score (from 0.775 to 0.213) while preserving semantic meaning.\n", "date": "2023-08-10", "categories": []}
{"id": "2308.04180", "title": "Studying Socially Unacceptable Discourse Classification (SUD) through\n  different eyes: \"Are we on the same page ?\"", "abstract": "  We study Socially Unacceptable Discourse (SUD) characterization and detection\nin online text. We first build and present a novel corpus that contains a large\nvariety of manually annotated texts from different online sources used so far\nin state-of-the-art Machine learning (ML) SUD detection solutions. This global\ncontext allows us to test the generalization ability of SUD classifiers that\nacquire knowledge around the same SUD categories, but from different contexts.\nFrom this perspective, we can analyze how (possibly) different annotation\nmodalities influence SUD learning by discussing open challenges and open\nresearch directions. We also provide several data insights which can support\ndomain experts in the annotation task.\n", "date": "2023-08-08", "categories": []}
{"id": "2307.16609", "title": "Noisy Self-Training with Data Augmentations for Offensive and Hate\n  Speech Detection Tasks", "abstract": "  Online social media is rife with offensive and hateful comments, prompting\nthe need for their automatic detection given the sheer amount of posts created\nevery second. Creating high-quality human-labelled datasets for this task is\ndifficult and costly, especially because non-offensive posts are significantly\nmore frequent than offensive ones. However, unlabelled data is abundant,\neasier, and cheaper to obtain. In this scenario, self-training methods, using\nweakly-labelled examples to increase the amount of training data, can be\nemployed. Recent \"noisy\" self-training approaches incorporate data augmentation\ntechniques to ensure prediction consistency and increase robustness against\nnoisy data and adversarial attacks. In this paper, we experiment with default\nand noisy self-training using three different textual data augmentation\ntechniques across five different pre-trained BERT architectures varying in\nsize. We evaluate our experiments on two offensive/hate-speech datasets and\ndemonstrate that (i) self-training consistently improves performance regardless\nof model size, resulting in up to +1.5% F1-macro on both datasets, and (ii)\nnoisy self-training with textual data augmentations, despite being successfully\napplied in similar settings, decreases performance on offensive and hate-speech\ndomains when compared to the default method, even with state-of-the-art\naugmentations such as backtranslation.\n", "date": "2023-07-31", "categories": []}
{"id": "2307.13829", "title": "ARC-NLP at Multimodal Hate Speech Event Detection 2023: Multimodal\n  Methods Boosted by Ensemble Learning, Syntactical and Entity Features", "abstract": "  Text-embedded images can serve as a means of spreading hate speech,\npropaganda, and extremist beliefs. Throughout the Russia-Ukraine war, both\nopposing factions heavily relied on text-embedded images as a vehicle for\nspreading propaganda and hate speech. Ensuring the effective detection of hate\nspeech and propaganda is of utmost importance to mitigate the negative effect\nof hate speech dissemination. In this paper, we outline our methodologies for\ntwo subtasks of Multimodal Hate Speech Event Detection 2023. For the first\nsubtask, hate speech detection, we utilize multimodal deep learning models\nboosted by ensemble learning and syntactical text attributes. For the second\nsubtask, target detection, we employ multimodal deep learning models boosted by\nnamed entity features. Through experimentation, we demonstrate the superior\nperformance of our models compared to all textual, visual, and text-visual\nbaselines employed in multimodal hate speech detection. Furthermore, our models\nachieve the first place in both subtasks on the final leaderboard of the shared\ntask.\n", "date": "2023-07-25", "categories": []}
{"id": "2307.11519", "title": "Multi-modal Hate Speech Detection using Machine Learning", "abstract": "  With the continuous growth of internet users and media content, it is very\nhard to track down hateful speech in audio and video. Converting video or audio\ninto text does not detect hate speech accurately as human sometimes uses\nhateful words as humorous or pleasant in sense and also uses different voice\ntones or show different action in the video. The state-ofthe-art hate speech\ndetection models were mostly developed on a single modality. In this research,\na combined approach of multimodal system has been proposed to detect hate\nspeech from video contents by extracting feature images, feature values\nextracted from the audio, text and used machine learning and Natural language\nprocessing.\n", "date": "2023-06-15", "categories": []}
{"id": "2307.10213", "title": "Mitigating Bias in Conversations: A Hate Speech Classifier and Debiaser\n  with Prompts", "abstract": "  Discriminatory language and biases are often present in hate speech during\nconversations, which usually lead to negative impacts on targeted groups such\nas those based on race, gender, and religion. To tackle this issue, we propose\nan approach that involves a two-step process: first, detecting hate speech\nusing a classifier, and then utilizing a debiasing component that generates\nless biased or unbiased alternatives through prompts. We evaluated our approach\non a benchmark dataset and observed reduction in negativity due to hate speech\ncomments. The proposed method contributes to the ongoing efforts to reduce\nbiases in online discourse and promote a more inclusive and fair environment\nfor communication.\n", "date": "2023-07-14", "categories": []}
{"id": "2306.15745", "title": "Identity Construction in a Misogynist Incels Forum", "abstract": "  Online communities of involuntary celibates (incels) are a prominent source\nof misogynist hate speech. In this paper, we use quantitative text and network\nanalysis approaches to examine how identity groups are discussed on\nincels-dot-is, the largest black-pilled incels forum. We find that this\ncommunity produces a wide range of novel identity terms and, while terms for\nwomen are most common, mentions of other minoritized identities are increasing.\nAn analysis of the associations made with identity groups suggests an\nessentialist ideology where physical appearance, as well as gender and racial\nhierarchies, determine human value. We discuss implications for research into\nautomated misogynist hate speech detection.\n", "date": "2023-06-27", "categories": []}
{"id": "2307.05578", "title": "Hate Speech Detection via Dual Contrastive Learning", "abstract": "  The fast spread of hate speech on social media impacts the Internet\nenvironment and our society by increasing prejudice and hurting people.\nDetecting hate speech has aroused broad attention in the field of natural\nlanguage processing. Although hate speech detection has been addressed in\nrecent work, this task still faces two inherent unsolved challenges. The first\nchallenge lies in the complex semantic information conveyed in hate speech,\nparticularly the interference of insulting words in hate speech detection. The\nsecond challenge is the imbalanced distribution of hate speech and non-hate\nspeech, which may significantly deteriorate the performance of models. To\ntackle these challenges, we propose a novel dual contrastive learning (DCL)\nframework for hate speech detection. Our framework jointly optimizes the\nself-supervised and the supervised contrastive learning loss for capturing\nspan-level information beyond the token-level emotional semantics used in\nexisting models, particularly detecting speech containing abusive and insulting\nwords. Moreover, we integrate the focal loss into the dual contrastive learning\nframework to alleviate the problem of data imbalance. We conduct experiments on\ntwo publicly available English datasets, and experimental results show that the\nproposed model outperforms the state-of-the-art models and precisely detects\nhate speeches.\n", "date": "2023-07-10", "categories": []}
{"id": "2210.09014", "title": "Addressing contingency in algorithmic (mis)information classification:\n  Toward a responsible machine learning agenda", "abstract": "  Machine learning (ML) enabled classification models are becoming increasingly\npopular for tackling the sheer volume and speed of online misinformation and\nother content that could be identified as harmful. In building these models,\ndata scientists need to take a stance on the legitimacy, authoritativeness and\nobjectivity of the sources of ``truth\" used for model training and testing.\nThis has political, ethical and epistemic implications which are rarely\naddressed in technical papers. Despite (and due to) their reported high\naccuracy and performance, ML-driven moderation systems have the potential to\nshape online public debate and create downstream negative impacts such as undue\ncensorship and the reinforcing of false beliefs. Using collaborative\nethnography and theoretical insights from social studies of science and\nexpertise, we offer a critical analysis of the process of building ML models\nfor (mis)information classification: we identify a series of algorithmic\ncontingencies--key moments during model development that could lead to\ndifferent future outcomes, uncertainty and harmful effects as these tools are\ndeployed by social media platforms. We conclude by offering a tentative path\ntoward reflexive and responsible development of ML tools for moderating\nmisinformation and other harmful content online.\n", "date": "2022-10-05", "categories": []}
{"id": "2212.06573", "title": "On the Evolution of (Hateful) Memes by Means of Multimodal Contrastive\n  Learning", "abstract": "  The dissemination of hateful memes online has adverse effects on social media\nplatforms and the real world. Detecting hateful memes is challenging, one of\nthe reasons being the evolutionary nature of memes; new hateful memes can\nemerge by fusing hateful connotations with other cultural ideas or symbols. In\nthis paper, we propose a framework that leverages multimodal contrastive\nlearning models, in particular OpenAI's CLIP, to identify targets of hateful\ncontent and systematically investigate the evolution of hateful memes. We find\nthat semantic regularities exist in CLIP-generated embeddings that describe\nsemantic relationships within the same modality (images) or across modalities\n(images and text). Leveraging this property, we study how hateful memes are\ncreated by combining visual elements from multiple images or fusing textual\ninformation with a hateful image. We demonstrate the capabilities of our\nframework for analyzing the evolution of hateful memes by focusing on\nantisemitic memes, particularly the Happy Merchant meme. Using our framework on\na dataset extracted from 4chan, we find 3.3K variants of the Happy Merchant\nmeme, with some linked to specific countries, persons, or organizations. We\nenvision that our framework can be used to aid human moderators by flagging new\nvariants of hateful memes so that moderators can manually verify them and\nmitigate the problem of hateful content online.\n", "date": "2022-12-13", "categories": []}
{"id": "2307.03274", "title": "It is not Sexually Suggestive, It is Educative. Separating Sex Education\n  from Suggestive Content on TikTok Videos", "abstract": "  We introduce SexTok, a multi-modal dataset composed of TikTok videos labeled\nas sexually suggestive (from the annotator's point of view), sex-educational\ncontent, or neither. Such a dataset is necessary to address the challenge of\ndistinguishing between sexually suggestive content and virtual sex education\nvideos on TikTok. Children's exposure to sexually suggestive videos has been\nshown to have adversarial effects on their development. Meanwhile, virtual sex\neducation, especially on subjects that are more relevant to the LGBTQIA+\ncommunity, is very valuable. The platform's current system removes or penalizes\nsome of both types of videos, even though they serve different purposes. Our\ndataset contains video URLs, and it is also audio transcribed. To validate its\nimportance, we explore two transformer-based models for classifying the videos.\nOur preliminary results suggest that the task of distinguishing between these\ntypes of videos is learnable but challenging. These experiments suggest that\nthis dataset is meaningful and invites further study on the subject.\n", "date": "2023-07-06", "categories": []}
{"id": "2307.03385", "title": "AI-UPV at EXIST 2023 -- Sexism Characterization Using Large Language\n  Models Under The Learning with Disagreements Regime", "abstract": "  With the increasing influence of social media platforms, it has become\ncrucial to develop automated systems capable of detecting instances of sexism\nand other disrespectful and hateful behaviors to promote a more inclusive and\nrespectful online environment. Nevertheless, these tasks are considerably\nchallenging considering different hate categories and the author's intentions,\nespecially under the learning with disagreements regime. This paper describes\nAI-UPV team's participation in the EXIST (sEXism Identification in Social\nneTworks) Lab at CLEF 2023. The proposed approach aims at addressing the task\nof sexism identification and characterization under the learning with\ndisagreements paradigm by training directly from the data with disagreements,\nwithout using any aggregated label. Yet, performances considering both soft and\nhard evaluations are reported. The proposed system uses large language models\n(i.e., mBERT and XLM-RoBERTa) and ensemble strategies for sexism identification\nand classification in English and Spanish. In particular, our system is\narticulated in three different pipelines. The ensemble approach outperformed\nthe individual large language models obtaining the best performances both\nadopting a soft and a hard label evaluation. This work describes the\nparticipation in all the three EXIST tasks, considering a soft evaluation, it\nobtained fourth place in Task 2 at EXIST and first place in Task 3, with the\nhighest ICM-Soft of -2.32 and a normalized ICM-Soft of 0.79. The source code of\nour approaches is publicly available at\nhttps://github.com/AngelFelipeMP/Sexism-LLM-Learning-With-Disagreement.\n", "date": "2023-07-07", "categories": []}
{"id": "2307.01680", "title": "Robust Hate Speech Detection in Social Media: A Cross-Dataset Empirical\n  Evaluation", "abstract": "  The automatic detection of hate speech online is an active research area in\nNLP. Most of the studies to date are based on social media datasets that\ncontribute to the creation of hate speech detection models trained on them.\nHowever, data creation processes contain their own biases, and models\ninherently learn from these dataset-specific biases. In this paper, we perform\na large-scale cross-dataset comparison where we fine-tune language models on\ndifferent hate speech detection datasets. This analysis shows how some datasets\nare more generalisable than others when used as training data. Crucially, our\nexperiments show how combining hate speech detection datasets can contribute to\nthe development of robust hate speech detection models. This robustness holds\neven when controlling by data size and compared with the best individual\ndatasets.\n", "date": "2023-07-04", "categories": []}
{"id": "2306.14764", "title": "Uncovering Political Hate Speech During Indian Election Campaign: A New\n  Low-Resource Dataset and Baselines", "abstract": "  The detection of hate speech in political discourse is a critical issue, and\nthis becomes even more challenging in low-resource languages. To address this\nissue, we introduce a new dataset named IEHate, which contains 11,457 manually\nannotated Hindi tweets related to the Indian Assembly Election Campaign from\nNovember 1, 2021, to March 9, 2022. We performed a detailed analysis of the\ndataset, focusing on the prevalence of hate speech in political communication\nand the different forms of hateful language used. Additionally, we benchmark\nthe dataset using a range of machine learning, deep learning, and\ntransformer-based algorithms. Our experiments reveal that the performance of\nthese models can be further improved, highlighting the need for more advanced\ntechniques for hate speech detection in low-resource languages. In particular,\nthe relatively higher score of human evaluation over algorithms emphasizes the\nimportance of utilizing both human and automated approaches for effective hate\nspeech moderation. Our IEHate dataset can serve as a valuable resource for\nresearchers and practitioners working on developing and evaluating hate speech\ndetection techniques in low-resource languages. Overall, our work underscores\nthe importance of addressing the challenges of identifying and mitigating hate\nspeech in political discourse, particularly in the context of low-resource\nlanguages. The dataset and resources for this work are made available at\nhttps://github.com/Farhan-jafri/Indian-Election.\n", "date": "2023-06-26", "categories": []}
{"id": "2306.15732", "title": "A Weakly Supervised Classifier and Dataset of White Supremacist Language", "abstract": "  We present a dataset and classifier for detecting the language of white\nsupremacist extremism, a growing issue in online hate speech. Our weakly\nsupervised classifier is trained on large datasets of text from explicitly\nwhite supremacist domains paired with neutral and anti-racist data from similar\ndomains. We demonstrate that this approach improves generalization performance\nto new domains. Incorporating anti-racist texts as counterexamples to white\nsupremacist language mitigates bias.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.01105", "title": "Revisiting Hate Speech Benchmarks: From Data Curation to System\n  Deployment", "abstract": "  Social media is awash with hateful content, much of which is often veiled\nwith linguistic and topical diversity. The benchmark datasets used for hate\nspeech detection do not account for such divagation as they are predominantly\ncompiled using hate lexicons. However, capturing hate signals becomes\nchallenging in neutrally-seeded malicious content. Thus, designing models and\ndatasets that mimic the real-world variability of hate warrants further\ninvestigation.\n  To this end, we present GOTHate, a large-scale code-mixed crowdsourced\ndataset of around 51k posts for hate speech detection from Twitter. GOTHate is\nneutrally seeded, encompassing different languages and topics. We conduct\ndetailed comparisons of GOTHate with the existing hate speech datasets,\nhighlighting its novelty. We benchmark it with 10 recent baselines. Our\nextensive empirical and benchmarking experiments suggest that GOTHate is hard\nto classify in a text-only setup. Thus, we investigate how adding endogenous\nsignals enhances the hate speech detection task. We augment GOTHate with the\nuser's timeline information and ego network, bringing the overall data source\ncloser to the real-world setup for understanding hateful content. Our proposed\nsolution HEN-mBERT is a modular, multilingual, mixture-of-experts model that\nenriches the linguistic subspace with latent endogenous signals from history,\ntopology, and exemplars. HEN-mBERT transcends the best baseline by 2.5% and 5%\nin overall macro-F1 and hate class F1, respectively. Inspired by our\nexperiments, in partnership with Wipro AI, we are developing a semi-automated\npipeline to detect hateful content as a part of their mission to tackle online\nharm.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.03722", "title": "Evaluating the Effectiveness of Natural Language Inference for Hate\n  Speech Detection in Languages with Limited Labeled Data", "abstract": "  Most research on hate speech detection has focused on English where a\nsizeable amount of labeled training data is available. However, to expand hate\nspeech detection into more languages, approaches that require minimal training\ndata are needed. In this paper, we test whether natural language inference\n(NLI) models which perform well in zero- and few-shot settings can benefit hate\nspeech detection performance in scenarios where only a limited amount of\nlabeled data is available in the target language. Our evaluation on five\nlanguages demonstrates large performance improvements of NLI fine-tuning over\ndirect fine-tuning in the target language. However, the effectiveness of\nprevious work that proposed intermediate fine-tuning on English data is hard to\nmatch. Only in settings where the English training data does not match the test\ndomain, can our customised NLI-formulation outperform intermediate fine-tuning\non English. Based on our extensive experiments, we propose a set of\nrecommendations for hate speech detection in languages where minimal labeled\ntraining data is available.\n", "date": "2023-06-06", "categories": []}
{"id": "2306.01985", "title": "COBRA Frames: Contextual Reasoning about Effects and Harms of Offensive\n  Statements", "abstract": "  Warning: This paper contains content that may be offensive or upsetting.\nUnderstanding the harms and offensiveness of statements requires reasoning\nabout the social and situational context in which statements are made. For\nexample, the utterance \"your English is very good\" may implicitly signal an\ninsult when uttered by a white man to a non-white colleague, but uttered by an\nESL teacher to their student would be interpreted as a genuine compliment. Such\ncontextual factors have been largely ignored by previous approaches to toxic\nlanguage detection. We introduce COBRA frames, the first context-aware\nformalism for explaining the intents, reactions, and harms of offensive or\nbiased statements grounded in their social and situational context. We create\nCOBRACORPUS, a dataset of 33k potentially offensive statements paired with\nmachine-generated contexts and free-text explanations of offensiveness, implied\nbiases, speaker intents, and listener reactions. To study the contextual\ndynamics of offensiveness, we train models to generate COBRA explanations, with\nand without access to the context. We find that explanations by\ncontext-agnostic models are significantly worse than by context-aware ones,\nespecially in situations where the context inverts the statement's\noffensiveness (29% accuracy drop). Our work highlights the importance and\nfeasibility of contextualized NLP by modeling social factors.\n", "date": "2023-06-03", "categories": []}
{"id": "2207.08230", "title": "A Context-Sensitive Word Embedding Approach for The Detection of Troll\n  Tweets", "abstract": "  In this study, we aimed to address the growing concern of trolling behavior\non social media by developing and evaluating a set of model architectures for\nthe automatic detection of troll tweets. Utilizing deep learning techniques and\npre-trained word embedding methods such as BERT, ELMo, and GloVe, we evaluated\nthe performance of each architecture using metrics such as classification\naccuracy, F1 score, AUC, and precision. Our results indicate that BERT and ELMo\nembedding methods performed better than the GloVe method, likely due to their\nability to provide contextualized word embeddings that better capture the\nnuances and subtleties of language use in online social media. Additionally, we\nfound that CNN and GRU encoders performed similarly in terms of F1 score and\nAUC, suggesting their effectiveness in extracting relevant information from\ninput text. The best-performing method was found to be an ELMo-based\narchitecture that employed a GRU classifier, with an AUC score of 0.929. This\nresearch highlights the importance of utilizing contextualized word embeddings\nand appropriate encoder methods in the task of troll tweet detection, which can\nassist social-based systems in improving their performance in identifying and\naddressing trolling behavior on their platforms.\n", "date": "2022-07-17", "categories": []}
{"id": "2306.03907", "title": "CL-UZH at SemEval-2023 Task 10: Sexism Detection through Incremental\n  Fine-Tuning and Multi-Task Learning with Label Descriptions", "abstract": "  The widespread popularity of social media has led to an increase in hateful,\nabusive, and sexist language, motivating methods for the automatic detection of\nsuch phenomena. The goal of the SemEval shared task \\textit{Towards Explainable\nDetection of Online Sexism} (EDOS 2023) is to detect sexism in English social\nmedia posts (subtask A), and to categorize such posts into four coarse-grained\nsexism categories (subtask B), and eleven fine-grained subcategories (subtask\nC). In this paper, we present our submitted systems for all three subtasks,\nbased on a multi-task model that has been fine-tuned on a range of related\ntasks and datasets before being fine-tuned on the specific EDOS subtasks. We\nimplement multi-task learning by formulating each task as binary pairwise text\nclassification, where the dataset and label descriptions are given along with\nthe input text. The results show clear improvements over a fine-tuned\nDeBERTa-V3 serving as a baseline leading to $F_1$-scores of 85.9\\% in subtask A\n(rank 13/84), 64.8\\% in subtask B (rank 19/69), and 44.9\\% in subtask C\n(26/63).\n", "date": "2023-06-06", "categories": []}
{"id": "2306.00021", "title": "Explaining Hate Speech Classification with Model Agnostic Methods", "abstract": "  There have been remarkable breakthroughs in Machine Learning and Artificial\nIntelligence, notably in the areas of Natural Language Processing and Deep\nLearning. Additionally, hate speech detection in dialogues has been gaining\npopularity among Natural Language Processing researchers with the increased use\nof social media. However, as evidenced by the recent trends, the need for the\ndimensions of explainability and interpretability in AI models has been deeply\nrealised. Taking note of the factors above, the research goal of this paper is\nto bridge the gap between hate speech prediction and the explanations generated\nby the system to support its decision. This has been achieved by first\npredicting the classification of a text and then providing a posthoc, model\nagnostic and surrogate interpretability approach for explainability and to\nprevent model bias. The bidirectional transformer model BERT has been used for\nprediction because of its state of the art efficiency over other Machine\nLearning models. The model agnostic algorithm LIME generates explanations for\nthe output of a trained classifier and predicts the features that influence the\nmodel decision. The predictions generated from the model were evaluated\nmanually, and after thorough evaluation, we observed that the model performs\nefficiently in predicting and explaining its prediction. Lastly, we suggest\nfurther directions for the expansion of the provided research work.\n", "date": "2023-05-30", "categories": []}
{"id": "2306.00410", "title": "Towards hate speech detection in low-resource languages: Comparing ASR\n  to acoustic word embeddings on Wolof and Swahili", "abstract": "  We consider hate speech detection through keyword spotting on radio\nbroadcasts. One approach is to build an automatic speech recognition (ASR)\nsystem for the target low-resource language. We compare this to using acoustic\nword embedding (AWE) models that map speech segments to a space where matching\nwords have similar vectors. We specifically use a multilingual AWE model\ntrained on labelled data from well-resourced languages to spot keywords in data\nin the unseen target language. In contrast to ASR, the AWE approach only\nrequires a few keyword exemplars. In controlled experiments on Wolof and\nSwahili where training and test data are from the same domain, an ASR model\ntrained on just five minutes of data outperforms the AWE approach. But in an\nin-the-wild test on Swahili radio broadcasts with actual hate speech keywords,\nthe AWE model (using one minute of template data) is more robust, giving\nsimilar performance to an ASR system trained on 30 hours of labelled data.\n", "date": "2023-06-01", "categories": []}
{"id": "2305.18585", "title": "Exploiting Explainability to Design Adversarial Attacks and Evaluate\n  Attack Resilience in Hate-Speech Detection Models", "abstract": "  The advent of social media has given rise to numerous ethical challenges,\nwith hate speech among the most significant concerns. Researchers are\nattempting to tackle this problem by leveraging hate-speech detection and\nemploying language models to automatically moderate content and promote civil\ndiscourse. Unfortunately, recent studies have revealed that hate-speech\ndetection systems can be misled by adversarial attacks, raising concerns about\ntheir resilience. While previous research has separately addressed the\nrobustness of these models under adversarial attacks and their\ninterpretability, there has been no comprehensive study exploring their\nintersection. The novelty of our work lies in combining these two critical\naspects, leveraging interpretability to identify potential vulnerabilities and\nenabling the design of targeted adversarial attacks. We present a comprehensive\nand comparative analysis of adversarial robustness exhibited by various\nhate-speech detection models. Our study evaluates the resilience of these\nmodels against adversarial attacks using explainability techniques. To gain\ninsights into the models' decision-making processes, we employ the Local\nInterpretable Model-agnostic Explanations (LIME) framework. Based on the\nexplainability results obtained by LIME, we devise and execute targeted attacks\non the text by leveraging the TextAttack tool. Our findings enhance the\nunderstanding of the vulnerabilities and strengths exhibited by\nstate-of-the-art hate-speech detection models. This work underscores the\nimportance of incorporating explainability in the development and evaluation of\nsuch models to enhance their resilience against adversarial attacks.\nUltimately, this work paves the way for creating more robust and reliable\nhate-speech detection systems, fostering safer online environments and\npromoting ethical discourse on social media platforms.\n", "date": "2023-05-29", "categories": []}
{"id": "2305.17984", "title": "minOffense: Inter-Agreement Hate Terms for Stable Rules, Concepts,\n  Transitivities, and Lattices", "abstract": "  Hate speech classification has become an important problem due to the spread\nof hate speech on social media platforms. For a given set of Hate Terms lists\n(HTs-lists) and Hate Speech data (HS-data), it is challenging to understand\nwhich hate term contributes the most for hate speech classification. This paper\ncontributes two approaches to quantitatively measure and qualitatively\nvisualise the relationship between co-occurring Hate Terms (HTs). Firstly, we\npropose an approach for the classification of hate-speech by producing a Severe\nHate Terms list (Severe HTs-list) from existing HTs-lists. To achieve our goal,\nwe proposed three metrics (Hatefulness, Relativeness, and Offensiveness) to\nmeasure the severity of HTs. These metrics assist to create an Inter-agreement\nHTs-list, which explains the contribution of an individual hate term toward\nhate speech classification. Then, we used the Offensiveness metric values of\nHTs above a proposed threshold minimum Offense (minOffense) to generate a new\nSevere HTs-list. To evaluate our approach, we used three hate speech datasets\nand six hate terms lists. Our approach shown an improvement from 0.845 to 0.923\n(best) as compared to the baseline. Secondly, we also proposed Stable Hate Rule\n(SHR) mining to provide ordered co-occurrence of various HTs with minimum\nStability (minStab). The SHR mining detects frequently co-occurring HTs to form\nStable Hate Rules and Concepts. These rules and concepts are used to visualise\nthe graphs of Transitivities and Lattices formed by HTs.\n", "date": "2023-05-29", "categories": []}
{"id": "2305.15551", "title": "Malicious or Benign? Towards Effective Content Moderation for Children's\n  Videos", "abstract": "  Online video platforms receive hundreds of hours of uploads every minute,\nmaking manual content moderation impossible. Unfortunately, the most vulnerable\nconsumers of malicious video content are children from ages 1-5 whose attention\nis easily captured by bursts of color and sound. Scammers attempting to\nmonetize their content may craft malicious children's videos that are\nsuperficially similar to educational videos, but include scary and disgusting\ncharacters, violent motions, loud music, and disturbing noises. Prominent video\nhosting platforms like YouTube have taken measures to mitigate malicious\ncontent on their platform, but these videos often go undetected by current\ncontent moderation tools that are focused on removing pornographic or\ncopyrighted content. This paper introduces our toolkit Malicious or Benign for\npromoting research on automated content moderation of children's videos. We\npresent 1) a customizable annotation tool for videos, 2) a new dataset with\ndifficult to detect test cases of malicious content and 3) a benchmark suite of\nstate-of-the-art video classification models.\n", "date": "2023-05-24", "categories": []}
{"id": "2305.12542", "title": "ToxBuster: In-game Chat Toxicity Buster with BERT", "abstract": "  Detecting toxicity in online spaces is challenging and an ever more pressing\nproblem given the increase in social media and gaming consumption. We introduce\nToxBuster, a simple and scalable model trained on a relatively large dataset of\n194k lines of game chat from Rainbow Six Siege and For Honor, carefully\nannotated for different kinds of toxicity. Compared to the existing\nstate-of-the-art, ToxBuster achieves 82.95% (+7) in precision and 83.56% (+57)\nin recall. This improvement is obtained by leveraging past chat history and\nmetadata. We also study the implication towards real-time and post-game\nmoderation as well as the model transferability from one game to another.\n", "date": "2023-05-21", "categories": []}
{"id": "2305.13276", "title": "Evaluating ChatGPT's Performance for Multilingual and Emoji-based Hate\n  Speech Detection", "abstract": "  Hate speech is a severe issue that affects many online platforms. So far,\nseveral studies have been performed to develop robust hate speech detection\nsystems. Large language models like ChatGPT have recently shown a great promise\nin performing several tasks, including hate speech detection. However, it is\ncrucial to comprehend the limitations of these models to build robust hate\nspeech detection systems. To bridge this gap, our study aims to evaluate the\nstrengths and weaknesses of the ChatGPT model in detecting hate speech at a\ngranular level across 11 languages. Our evaluation employs a series of\nfunctionality tests that reveals various intricate failures of the model which\nthe aggregate metrics like macro F1 or accuracy are not able to unfold. In\naddition, we investigate the influence of complex emotions, such as the use of\nemojis in hate speech, on the performance of the ChatGPT model. Our analysis\nhighlights the shortcomings of the generative models in detecting certain types\nof hate speech and highlighting the need for further research and improvements\nin the workings of these models.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.13589", "title": "BiasX: \"Thinking Slow\" in Toxic Content Moderation with Explanations of\n  Implied Social Biases", "abstract": "  Toxicity annotators and content moderators often default to mental shortcuts\nwhen making decisions. This can lead to subtle toxicity being missed, and\nseemingly toxic but harmless content being over-detected. We introduce BiasX, a\nframework that enhances content moderation setups with free-text explanations\nof statements' implied social biases, and explore its effectiveness through a\nlarge-scale crowdsourced user study. We show that indeed, participants\nsubstantially benefit from explanations for correctly identifying subtly\n(non-)toxic content. The quality of explanations is critical: imperfect\nmachine-generated explanations (+2.4% on hard toxic examples) help less\ncompared to expert-written human explanations (+7.2%). Our results showcase the\npromise of using free-text explanations to encourage more thoughtful toxicity\nmoderation.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.13623", "title": "Validating Multimedia Content Moderation Software via Semantic Fusion", "abstract": "  The exponential growth of social media platforms, such as Facebook and\nTikTok, has revolutionized communication and content publication in human\nsociety. Users on these platforms can publish multimedia content that delivers\ninformation via the combination of text, audio, images, and video. Meanwhile,\nthe multimedia content release facility has been increasingly exploited to\npropagate toxic content, such as hate speech, malicious advertisements, and\npornography. To this end, content moderation software has been widely deployed\non these platforms to detect and blocks toxic content. However, due to the\ncomplexity of content moderation models and the difficulty of understanding\ninformation across multiple modalities, existing content moderation software\ncan fail to detect toxic content, which often leads to extremely negative\nimpacts.\n  We introduce Semantic Fusion, a general, effective methodology for validating\nmultimedia content moderation software. Our key idea is to fuse two or more\nexisting single-modal inputs (e.g., a textual sentence and an image) into a new\ninput that combines the semantics of its ancestors in a novel manner and has\ntoxic nature by construction. This fused input is then used for validating\nmultimedia content moderation software. We realized Semantic Fusion as DUO, a\npractical content moderation software testing tool. In our evaluation, we\nemploy DUO to test five commercial content moderation software and two\nstate-of-the-art models against three kinds of toxic content. The results show\nthat DUO achieves up to 100% error finding rate (EFR) when testing moderation\nsoftware. In addition, we leverage the test cases generated by DUO to retrain\nthe two models we explored, which largely improves model robustness while\nmaintaining the accuracy on the original test set.\n", "date": "2023-05-23", "categories": []}
{"id": "2304.01890", "title": "Sociocultural knowledge is needed for selection of shots in hate speech\n  detection tasks", "abstract": "  We introduce HATELEXICON, a lexicon of slurs and targets of hate speech for\nthe countries of Brazil, Germany, India and Kenya, to aid training and\ninterpretability of models. We demonstrate how our lexicon can be used to\ninterpret model predictions, showing that models developed to classify extreme\nspeech rely heavily on target words when making predictions. Further, we\npropose a method to aid shot selection for training in low-resource settings\nvia HATELEXICON. In few-shot learning, the selection of shots is of paramount\nimportance to model performance. In our work, we simulate a few-shot setting\nfor German and Hindi, using HASOC data for training and the Multilingual\nHateCheck (MHC) as a benchmark. We show that selecting shots based on our\nlexicon leads to models performing better on MHC than models trained on shots\nsampled randomly. Thus, when given only a few training examples, using our\nlexicon to select shots containing more sociocultural information leads to\nbetter few-shot performance.\n", "date": "2023-04-04", "categories": []}
{"id": "2304.00020", "title": "SemiMemes: A Semi-supervised Learning Approach for Multimodal Memes\n  Analysis", "abstract": "  The prevalence of memes on social media has created the need to sentiment\nanalyze their underlying meanings for censoring harmful content. Meme censoring\nsystems by machine learning raise the need for a semi-supervised learning\nsolution to take advantage of the large number of unlabeled memes available on\nthe internet and make the annotation process less challenging. Moreover, the\napproach needs to utilize multimodal data as memes' meanings usually come from\nboth images and texts. This research proposes a multimodal semi-supervised\nlearning approach that outperforms other multimodal semi-supervised learning\nand supervised learning state-of-the-art models on two datasets, the Multimedia\nAutomatic Misogyny Identification and Hateful Memes dataset. Building on the\ninsights gained from Contrastive Language-Image Pre-training, which is an\neffective multimodal learning technique, this research introduces SemiMemes, a\nnovel training method that combines auto-encoder and classification task to\nmake use of the resourceful unlabeled data.\n", "date": "2023-03-31", "categories": []}
{"id": "2206.08406", "title": "Predicting Hate Intensity of Twitter Conversation Threads", "abstract": "  Tweets are the most concise form of communication in online social media,\nwherein a single tweet has the potential to make or break the discourse of the\nconversation. Online hate speech is more accessible than ever, and stifling its\npropagation is of utmost importance for social media companies and users for\ncongenial communication. Most of the research barring a recent few has focused\non classifying an individual tweet regardless of the tweet thread/context\nleading up to that point. One of the classical approaches to curb hate speech\nis to adopt a reactive strategy after the hate speech postage. The ex-post\nfacto strategy results in neglecting subtle posts that do not show the\npotential to instigate hate speech on their own but may portend in the\nsubsequent discussion ensuing in the post's replies. In this paper, we propose\nDRAGNET++, which aims to predict the intensity of hatred that a tweet can bring\nin through its reply chain in the future. It uses the semantic and propagating\nstructure of the tweet threads to maximize the contextual information leading\nup to and the fall of hate intensity at each subsequent tweet. We explore three\npublicly available Twitter datasets -- Anti-Racism contains the reply tweets of\na collection of social media discourse on racist remarks during US political\nand Covid-19 background; Anti-Social presents a dataset of 40 million tweets\namidst the COVID-19 pandemic on anti-social behaviours; and Anti-Asian presents\nTwitter datasets collated based on anti-Asian behaviours during COVID-19\npandemic. All the curated datasets consist of structural graph information of\nthe Tweet threads. We show that DRAGNET++ outperforms all the state-of-the-art\nbaselines significantly. It beats the best baseline by an 11% margin on the\nPerson correlation coefficient and a decrease of 25% on RMSE for the\nAnti-Racism dataset with a similar performance on the other two datasets.\n", "date": "2022-06-16", "categories": []}
{"id": "2305.08636", "title": "AdamR at SemEval-2023 Task 10: Solving the Class Imbalance Problem in\n  Sexism Detection with Ensemble Learning", "abstract": "  The Explainable Detection of Online Sexism task presents the problem of\nexplainable sexism detection through fine-grained categorisation of sexist\ncases with three subtasks. Our team experimented with different ways to combat\nclass imbalance throughout the tasks using data augmentation and loss\nalteration techniques. We tackled the challenge by utilising ensembles of\nTransformer models trained on different datasets, which are tested to find the\nbalance between performance and interpretability. This solution ranked us in\nthe top 40\\% of teams for each of the tracks.\n", "date": "2023-05-15", "categories": []}
{"id": "2305.07639", "title": "Efficient Neural Network based Classification and Outlier Detection for\n  Image Moderation using Compressed Sensing and Group Testing", "abstract": "  Popular social media platforms employ neural network based image moderation\nengines to classify images uploaded on them as having potentially objectionable\ncontent. Such moderation engines must answer a large number of queries with\nheavy computational cost, even though the actual number of images with\nobjectionable content is usually a tiny fraction. Inspired by recent work on\nNeural Group Testing, we propose an approach which exploits this fact to reduce\nthe overall computational cost of such engines using the technique of\nCompressed Sensing (CS). We present the quantitative matrix-pooled neural\nnetwork (QMPNN), which takes as input $n$ images, and a $m \\times n$ binary\npooling matrix with $m < n$, whose rows indicate $m$ pools of images i.e.\nselections of $r$ images out of $n$. The QMPNN efficiently outputs the product\nof this matrix with the unknown sparse binary vector indicating whether each\nimage is objectionable or not, i.e. it outputs the number of objectionable\nimages in each pool. For suitable matrices, this is decoded using CS decoding\nalgorithms to predict which images were objectionable. The computational cost\nof running the QMPNN and the CS algorithms is significantly lower than the cost\nof using a neural network with the same number of parameters separately on each\nimage to classify the images, which we demonstrate via extensive experiments.\nOur technique is inherently resilient to moderate levels of errors in the\nprediction from the QMPNN. Furthermore, we present pooled deep outlier\ndetection, which brings CS and group testing techniques to deep outlier\ndetection, to provide for the case when the objectionable images do not belong\nto a set of pre-defined classes. This technique enables efficient automated\nmoderation of off-topic images shared on topical forums dedicated to sharing\nimages of a certain single class, many of which are currently human-moderated.\n", "date": "2023-05-12", "categories": []}
{"id": "2305.06892", "title": "IUST_NLP at SemEval-2023 Task 10: Explainable Detecting Sexism with\n  Transformers and Task-adaptive Pretraining", "abstract": "  This paper describes our system on SemEval-2023 Task 10: Explainable\nDetection of Online Sexism (EDOS). This work aims to design an automatic system\nfor detecting and classifying sexist content in online spaces. We propose a set\nof transformer-based pre-trained models with task-adaptive pretraining and\nensemble learning. The main contributions of our system include analyzing the\nperformance of different transformer-based pre-trained models and combining\nthese models, as well as providing an efficient method using large amounts of\nunlabeled data for model adaptive pretraining. We have also explored several\nother strategies. On the test dataset, our system achieves F1-scores of 83%,\n64%, and 47% on subtasks A, B, and C, respectively.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.06159", "title": "A Review of Vision-Language Models and their Performance on the Hateful\n  Memes Challenge", "abstract": "  Moderation of social media content is currently a highly manual task, yet\nthere is too much content posted daily to do so effectively. With the advent of\na number of multimodal models, there is the potential to reduce the amount of\nmanual labor for this task. In this work, we aim to explore different models\nand determine what is most effective for the Hateful Memes Challenge, a\nchallenge by Meta designed to further machine learning research in content\nmoderation. Specifically, we explore the differences between early fusion and\nlate fusion models in classifying multimodal memes containing text and images.\nWe first implement a baseline using unimodal models for text and images\nseparately using BERT and ResNet-152, respectively. The outputs from these\nunimodal models were then concatenated together to create a late fusion model.\nIn terms of early fusion models, we implement ConcatBERT, VisualBERT, ViLT,\nCLIP, and BridgeTower. It was found that late fusion performed significantly\nworse than early fusion models, with the best performing model being CLIP which\nachieved an AUROC of 70.06. The code for this work is available at\nhttps://github.com/bzhao18/CS-7643-Project.\n", "date": "2023-05-09", "categories": []}
{"id": "2305.04927", "title": "Detecting and Reasoning of Deleted Tweets before they are Posted", "abstract": "  Social media platforms empower us in several ways, from information\ndissemination to consumption. While these platforms are useful in promoting\ncitizen journalism, public awareness etc., they have misuse potentials.\nMalicious users use them to disseminate hate-speech, offensive content, rumor\netc. to gain social and political agendas or to harm individuals, entities and\norganizations. Often times, general users unconsciously share information\nwithout verifying it, or unintentionally post harmful messages. Some of such\ncontent often get deleted either by the platform due to the violation of terms\nand policies, or users themselves for different reasons, e.g., regrets. There\nis a wide range of studies in characterizing, understanding and predicting\ndeleted content. However, studies which aims to identify the fine-grained\nreasons (e.g., posts are offensive, hate speech or no identifiable reason)\nbehind deleted content, are limited. In this study we address this gap, by\nidentifying deleted tweets, particularly within the Arabic context, and\nlabeling them with a corresponding fine-grained disinformation category. We\nthen develop models that can predict the potentiality of tweets getting\ndeleted, as well as the potential reasons behind deletion. Such models can help\nin moderating social media posts before even posting.\n", "date": "2023-05-05", "categories": []}
{"id": "2303.04222", "title": "SemEval-2023 Task 10: Explainable Detection of Online Sexism", "abstract": "  Online sexism is a widespread and harmful phenomenon. Automated tools can\nassist the detection of sexism at scale. Binary detection, however, disregards\nthe diversity of sexist content, and fails to provide clear explanations for\nwhy something is sexist. To address this issue, we introduce SemEval Task 10 on\nthe Explainable Detection of Online Sexism (EDOS). We make three main\ncontributions: i) a novel hierarchical taxonomy of sexist content, which\nincludes granular vectors of sexism to aid explainability; ii) a new dataset of\n20,000 social media comments with fine-grained labels, along with larger\nunlabelled datasets for model adaptation; and iii) baseline models as well as\nan analysis of the methods, results and errors for participant submissions to\nour task.\n", "date": "2023-03-07", "categories": []}
{"id": "2305.03915", "title": "HateMM: A Multi-Modal Dataset for Hate Video Classification", "abstract": "  Hate speech has become one of the most significant issues in modern society,\nhaving implications in both the online and the offline world. Due to this, hate\nspeech research has recently gained a lot of traction. However, most of the\nwork has primarily focused on text media with relatively little work on images\nand even lesser on videos. Thus, early stage automated video moderation\ntechniques are needed to handle the videos that are being uploaded to keep the\nplatform safe and healthy. With a view to detect and remove hateful content\nfrom the video sharing platforms, our work focuses on hate video detection\nusing multi-modalities. To this end, we curate ~43 hours of videos from\nBitChute and manually annotate them as hate or non-hate, along with the frame\nspans which could explain the labelling decision. To collect the relevant\nvideos we harnessed search keywords from hate lexicons. We observe various cues\nin images and audio of hateful videos. Further, we build deep learning\nmulti-modal models to classify the hate videos and observe that using all the\nmodalities of the videos improves the overall hate speech detection performance\n(accuracy=0.798, macro F1-score=0.790) by ~5.7% compared to the best uni-modal\nmodel in terms of macro F1 score. In summary, our work takes the first step\ntoward understanding and modeling hateful videos on video hosting platforms\nsuch as BitChute.\n", "date": "2023-05-06", "categories": []}
{"id": "2305.04356", "title": "Stanford MLab at SemEval-2023 Task 10: Exploring GloVe- and\n  Transformer-Based Methods for the Explainable Detection of Online Sexism", "abstract": "  In this paper, we discuss the methods we applied at SemEval-2023 Task 10:\nTowards the Explainable Detection of Online Sexism. Given an input text, we\nperform three classification tasks to predict whether the text is sexist and\nclassify the sexist text into subcategories in order to provide an additional\nexplanation as to why the text is sexist. We explored many different types of\nmodels, including GloVe embeddings as the baseline approach, transformer-based\ndeep learning models like BERT, RoBERTa, and DeBERTa, ensemble models, and\nmodel blending. We explored various data cleaning and augmentation methods to\nimprove model performance. Pre-training transformer models yielded significant\nimprovements in performance, and ensembles and blending slightly improved\nrobustness in the F1 score.\n", "date": "2023-05-07", "categories": []}
{"id": "2305.04446", "title": "Facilitating Fine-grained Detection of Chinese Toxic Language:\n  Hierarchical Taxonomy, Resources, and Benchmarks", "abstract": "  The widespread dissemination of toxic online posts is increasingly damaging\nto society. However, research on detecting toxic language in Chinese has lagged\nsignificantly. Existing datasets lack fine-grained annotation of toxic types\nand expressions, and ignore the samples with indirect toxicity. In addition, it\nis crucial to introduce lexical knowledge to detect the toxicity of posts,\nwhich has been a challenge for researchers. In this paper, we facilitate the\nfine-grained detection of Chinese toxic language. First, we built Monitor Toxic\nFrame, a hierarchical taxonomy to analyze toxic types and expressions. Then, a\nfine-grained dataset ToxiCN is presented, including both direct and indirect\ntoxic samples. We also build an insult lexicon containing implicit profanity\nand propose Toxic Knowledge Enhancement (TKE) as a benchmark, incorporating the\nlexical feature to detect toxic language. In the experimental stage, we\ndemonstrate the effectiveness of TKE. After that, a systematic quantitative and\nqualitative analysis of the findings is given.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.01050", "title": "SafeWebUH at SemEval-2023 Task 11: Learning Annotator Disagreement in\n  Derogatory Text: Comparison of Direct Training vs Aggregation", "abstract": "  Subjectivity and difference of opinion are key social phenomena, and it is\ncrucial to take these into account in the annotation and detection process of\nderogatory textual content. In this paper, we use four datasets provided by\nSemEval-2023 Task 11 and fine-tune a BERT model to capture the disagreement in\nthe annotation. We find individual annotator modeling and aggregation lowers\nthe Cross-Entropy score by an average of 0.21, compared to the direct training\non the soft labels. Our findings further demonstrate that annotator metadata\ncontributes to the average 0.029 reduction in the Cross-Entropy score.\n", "date": "2023-05-01", "categories": []}
{"id": "2301.10871", "title": "Qualitative Analysis of a Graph Transformer Approach to Addressing Hate\n  Speech: Adapting to Dynamically Changing Content", "abstract": "  Our work advances an approach for predicting hate speech in social media,\ndrawing out the critical need to consider the discussions that follow a post to\nsuccessfully detect when hateful discourse may arise. Using graph transformer\nnetworks, coupled with modelling attention and BERT-level natural language\nprocessing, our approach can capture context and anticipate upcoming\nanti-social behaviour. In this paper, we offer a detailed qualitative analysis\nof this solution for hate speech detection in social networks, leading to\ninsights into where the method has the most impressive outcomes in comparison\nwith competitors and identifying scenarios where there are challenges to\nachieving ideal performance. Included is an exploration of the kinds of posts\nthat permeate social media today, including the use of hateful images. This\nsuggests avenues for extending our model to be more comprehensive. A key\ninsight is that the focus on reasoning about the concept of context positions\nus well to be able to support multi-modal analysis of online posts. We conclude\nwith a reflection on how the problem we are addressing relates especially well\nto the theme of dynamic change, a critical concern for all AI solutions for\nsocial impact. We also comment briefly on how mental health well-being can be\nadvanced with our work, through curated content attuned to the extent of hate\nin posts.\n", "date": "2023-01-25", "categories": []}
{"id": "2305.00076", "title": "HausaNLP at SemEval-2023 Task 10: Transfer Learning, Synthetic Data and\n  Side-Information for Multi-Level Sexism Classification", "abstract": "  We present the findings of our participation in the SemEval-2023 Task 10:\nExplainable Detection of Online Sexism (EDOS) task, a shared task on offensive\nlanguage (sexism) detection on English Gab and Reddit dataset. We investigated\nthe effects of transferring two language models: XLM-T (sentiment\nclassification) and HateBERT (same domain -- Reddit) for multi-level\nclassification into Sexist or not Sexist, and other subsequent\nsub-classifications of the sexist data. We also use synthetic classification of\nunlabelled dataset and intermediary class information to maximize the\nperformance of our models. We submitted a system in Task A, and it ranked 49th\nwith F1-score of 0.82. This result showed to be competitive as it only\nunder-performed the best system by 0.052% F1-score.\n", "date": "2023-04-28", "categories": []}
{"id": "2304.14599", "title": "Antisemitic Messages? A Guide to High-Quality Annotation and a Labeled\n  Dataset of Tweets", "abstract": "  One of the major challenges in automatic hate speech detection is the lack of\ndatasets that cover a wide range of biased and unbiased messages and that are\nconsistently labeled. We propose a labeling procedure that addresses some of\nthe common weaknesses of labeled datasets. We focus on antisemitic speech on\nTwitter and create a labeled dataset of 6,941 tweets that cover a wide range of\ntopics common in conversations about Jews, Israel, and antisemitism between\nJanuary 2019 and December 2021 by drawing from representative samples with\nrelevant keywords. Our annotation process aims to strictly apply a commonly\nused definition of antisemitism by forcing annotators to specify which part of\nthe definition applies, and by giving them the option to personally disagree\nwith the definition on a case-by-case basis. Labeling tweets that call out\nantisemitism, report antisemitism, or are otherwise related to antisemitism\n(such as the Holocaust) but are not actually antisemitic can help reduce false\npositives in automated detection. The dataset includes 1,250 tweets (18%) that\nare antisemitic according to the International Holocaust Remembrance Alliance\n(IHRA) definition of antisemitism. It is important to note, however, that the\ndataset is not comprehensive. Many topics are still not covered, and it only\nincludes tweets collected from Twitter between January 2019 and December 2021.\nAdditionally, the dataset only includes tweets that were written in English.\nDespite these limitations, we hope that this is a meaningful contribution to\nimproving the automated detection of antisemitic speech.\n", "date": "2023-04-28", "categories": []}
{"id": "2304.12847", "title": "NLP-LTU at SemEval-2023 Task 10: The Impact of Data Augmentation and\n  Semi-Supervised Learning Techniques on Text Classification Performance on an\n  Imbalanced Dataset", "abstract": "  In this paper, we propose a methodology for task 10 of SemEval23, focusing on\ndetecting and classifying online sexism in social media posts. The task is\ntackling a serious issue, as detecting harmful content on social media\nplatforms is crucial for mitigating the harm of these posts on users. Our\nsolution for this task is based on an ensemble of fine-tuned transformer-based\nmodels (BERTweet, RoBERTa, and DeBERTa). To alleviate problems related to class\nimbalance, and to improve the generalization capability of our model, we also\nexperiment with data augmentation and semi-supervised learning. In particular,\nfor data augmentation, we use back-translation, either on all classes, or on\nthe underrepresented classes only. We analyze the impact of these strategies on\nthe overall performance of the pipeline through extensive experiments. while\nfor semi-supervised learning, we found that with a substantial amount of\nunlabelled, in-domain data available, semi-supervised learning can enhance the\nperformance of certain models. Our proposed method (for which the source code\nis available on Github attains an F1-score of 0.8613 for sub-taskA, which\nranked us 10th in the competition\n", "date": "2023-04-25", "categories": []}
{"id": "2303.09314", "title": "TOT: Topology-Aware Optimal Transport For Multimodal Hate Detection", "abstract": "  Multimodal hate detection, which aims to identify harmful content online such\nas memes, is crucial for building a wholesome internet environment. Previous\nwork has made enlightening exploration in detecting explicit hate remarks.\nHowever, most of their approaches neglect the analysis of implicit harm, which\nis particularly challenging as explicit text markers and demographic visual\ncues are often twisted or missing. The leveraged cross-modal attention\nmechanisms also suffer from the distributional modality gap and lack logical\ninterpretability. To address these semantic gaps issues, we propose TOT: a\ntopology-aware optimal transport framework to decipher the implicit harm in\nmemes scenario, which formulates the cross-modal aligning problem as solutions\nfor optimal transportation plans. Specifically, we leverage an optimal\ntransport kernel method to capture complementary information from multiple\nmodalities. The kernel embedding provides a non-linear transformation ability\nto reproduce a kernel Hilbert space (RKHS), which reflects significance for\neliminating the distributional modality gap. Moreover, we perceive the topology\ninformation based on aligned representations to conduct bipartite graph path\nreasoning. The newly achieved state-of-the-art performance on two publicly\navailable benchmark datasets, together with further visual analysis,\ndemonstrate the superiority of TOT in capturing implicit cross-modal alignment.\n", "date": "2023-02-27", "categories": []}
{"id": "2304.03518", "title": "SSS at SemEval-2023 Task 10: Explainable Detection of Online Sexism\n  using Majority Voted Fine-Tuned Transformers", "abstract": "  This paper describes our submission to Task 10 at SemEval 2023-Explainable\nDetection of Online Sexism (EDOS), divided into three subtasks. The recent rise\nin social media platforms has seen an increase in disproportionate levels of\nsexism experienced by women on social media platforms. This has made detecting\nand explaining online sexist content more important than ever to make social\nmedia safer and more accessible for women. Our approach consists of\nexperimenting and finetuning BERT-based models and using a Majority Voting\nensemble model that outperforms individual baseline model scores. Our system\nachieves a macro F1 score of 0.8392 for Task A, 0.6092 for Task B, and 0.4319\nfor Task C.\n", "date": "2023-04-07", "categories": []}
{"id": "2304.11223", "title": "A Group-Specific Approach to NLP for Hate Speech Detection", "abstract": "  Automatic hate speech detection is an important yet complex task, requiring\nknowledge of common sense, stereotypes of protected groups, and histories of\ndiscrimination, each of which may constantly evolve. In this paper, we propose\na group-specific approach to NLP for online hate speech detection. The approach\nconsists of creating and infusing historical and linguistic knowledge about a\nparticular protected group into hate speech detection models, analyzing\nhistorical data about discrimination against a protected group to better\npredict spikes in hate speech against that group, and critically evaluating\nhate speech detection models through lenses of intersectionality and ethics. We\ndemonstrate this approach through a case study on NLP for detection of\nantisemitic hate speech. The case study synthesizes the current\nEnglish-language literature on NLP for antisemitism detection, introduces a\nnovel knowledge graph of antisemitic history and language from the 20th century\nto the present, infuses information from the knowledge graph into a set of\ntweets over Logistic Regression and uncased DistilBERT baselines, and suggests\nthat incorporating context from the knowledge graph can help models pick up\nsubtle stereotypes.\n", "date": "2023-04-21", "categories": []}
{"id": "2209.08697", "title": "Quantifying How Hateful Communities Radicalize Online Users", "abstract": "  While online social media offers a way for ignored or stifled voices to be\nheard, it also allows users a platform to spread hateful speech. Such speech\nusually originates in fringe communities, yet it can spill over into mainstream\nchannels. In this paper, we measure the impact of joining fringe hateful\ncommunities in terms of hate speech propagated to the rest of the social\nnetwork. We leverage data from Reddit to assess the effect of joining one type\nof echo chamber: a digital community of like-minded users exhibiting hateful\nbehavior. We measure members' usage of hate speech outside the studied\ncommunity before and after they become active participants. Using Interrupted\nTime Series (ITS) analysis as a causal inference method, we gauge the spillover\neffect, in which hateful language from within a certain community can spread\noutside that community by using the level of out-of-community hate word usage\nas a proxy for learned hate. We investigate four different Reddit\nsub-communities (subreddits) covering three areas of hate speech: racism,\nmisogyny and fat-shaming. In all three cases we find an increase in hate speech\noutside the originating community, implying that joining such community leads\nto a spread of hate speech throughout the platform. Moreover, users are found\nto pick up this new hateful speech for months after initially joining the\ncommunity. We show that the harmful speech does not remain contained within the\ncommunity. Our results provide new evidence of the harmful effects of echo\nchambers and the potential benefit of moderating them to reduce adoption of\nhateful speech.\n", "date": "2022-09-19", "categories": []}
{"id": "2304.04610", "title": "Attention at SemEval-2023 Task 10: Explainable Detection of Online\n  Sexism (EDOS)", "abstract": "  In this paper, we have worked on interpretability, trust, and understanding\nof the decisions made by models in the form of classification tasks. The task\nis divided into 3 subtasks. The first task consists of determining Binary\nSexism Detection. The second task describes the Category of Sexism. The third\ntask describes a more Fine-grained Category of Sexism. Our work explores\nsolving these tasks as a classification problem by fine-tuning\ntransformer-based architecture. We have performed several experiments with our\narchitecture, including combining multiple transformers, using domain adaptive\npretraining on the unlabelled dataset provided by Reddit and Gab, Joint\nlearning, and taking different layers of transformers as input to a\nclassification head. Our system (with team name Attention) was able to achieve\na macro F1 score of 0.839 for task A, 0.5835 macro F1 score for task B and\n0.3356 macro F1 score for task C at the Codalab SemEval Competition. Later we\nimproved the accuracy of Task B to 0.6228 and Task C to 0.3693 in the test set.\n", "date": "2023-04-10", "categories": []}
{"id": "2212.05612", "title": "Multimodal and Explainable Internet Meme Classification", "abstract": "  In the current context where online platforms have been effectively\nweaponized in a variety of geo-political events and social issues, Internet\nmemes make fair content moderation at scale even more difficult. Existing work\non meme classification and tracking has focused on black-box methods that do\nnot explicitly consider the semantics of the memes or the context of their\ncreation. In this paper, we pursue a modular and explainable architecture for\nInternet meme understanding. We design and implement multimodal classification\nmethods that perform example- and prototype-based reasoning over training\ncases, while leveraging both textual and visual SOTA models to represent the\nindividual cases. We study the relevance of our modular and explainable models\nin detecting harmful memes on two existing tasks: Hate Speech Detection and\nMisogyny Classification. We compare the performance between example- and\nprototype-based methods, and between text, vision, and multimodal models,\nacross different categories of harmfulness (e.g., stereotype and\nobjectification). We devise a user-friendly interface that facilitates the\ncomparative analysis of examples retrieved by all of our models for any given\nmeme, informing the community about the strengths and limitations of these\nexplainable methods.\n", "date": "2022-12-11", "categories": []}
{"id": "2304.02906", "title": "MemeFier: Dual-stage Modality Fusion for Image Meme Classification", "abstract": "  Hate speech is a societal problem that has significantly grown through the\nInternet. New forms of digital content such as image memes have given rise to\nspread of hate using multimodal means, being far more difficult to analyse and\ndetect compared to the unimodal case. Accurate automatic processing, analysis\nand understanding of this kind of content will facilitate the endeavor of\nhindering hate speech proliferation through the digital world. To this end, we\npropose MemeFier, a deep learning-based architecture for fine-grained\nclassification of Internet image memes, utilizing a dual-stage modality fusion\nmodule. The first fusion stage produces feature vectors containing modality\nalignment information that captures non-trivial connections between the text\nand image of a meme. The second fusion stage leverages the power of a\nTransformer encoder to learn inter-modality correlations at the token level and\nyield an informative representation. Additionally, we consider external\nknowledge as an additional input, and background image caption supervision as a\nregularizing component. Extensive experiments on three widely adopted\nbenchmarks, i.e., Facebook Hateful Memes, Memotion7k and MultiOFF, indicate\nthat our approach competes and in some cases surpasses state-of-the-art. Our\ncode is available on https://github.com/ckoutlis/memefier.\n", "date": "2023-04-06", "categories": []}
{"id": "2304.01241", "title": "Detection of Homophobia & Transphobia in Dravidian Languages: Exploring\n  Deep Learning Methods", "abstract": "  The increase in abusive content on online social media platforms is impacting\nthe social life of online users. Use of offensive and hate speech has been\nmaking so-cial media toxic. Homophobia and transphobia constitute offensive\ncomments against LGBT+ community. It becomes imperative to detect and handle\nthese comments, to timely flag or issue a warning to users indulging in such\nbehaviour. However, automated detection of such content is a challenging task,\nmore so in Dravidian languages which are identified as low resource languages.\nMotivated by this, the paper attempts to explore applicability of different\ndeep learning mod-els for classification of the social media comments in\nMalayalam and Tamil lan-guages as homophobic, transphobic and\nnon-anti-LGBT+content. The popularly used deep learning models- Convolutional\nNeural Network (CNN), Long Short Term Memory (LSTM) using GloVe embedding and\ntransformer-based learning models (Multilingual BERT and IndicBERT) are applied\nto the classification problem. Results obtained show that IndicBERT outperforms\nthe other imple-mented models, with obtained weighted average F1-score of 0.86\nand 0.77 for Malayalam and Tamil, respectively. Therefore, the present work\nconfirms higher performance of IndicBERT on the given task in selected\nDravidian languages.\n", "date": "2023-04-03", "categories": []}
{"id": "2304.00913", "title": "LAHM : Large Annotated Dataset for Multi-Domain and Multilingual Hate\n  Speech Identification", "abstract": "  Current research on hate speech analysis is typically oriented towards\nmonolingual and single classification tasks. In this paper, we present a new\nmultilingual hate speech analysis dataset for English, Hindi, Arabic, French,\nGerman and Spanish languages for multiple domains across hate speech - Abuse,\nRacism, Sexism, Religious Hate and Extremism. To the best of our knowledge,\nthis paper is the first to address the problem of identifying various types of\nhate speech in these five wide domains in these six languages. In this work, we\ndescribe how we created the dataset, created annotations at high level and low\nlevel for different domains and how we use it to test the current\nstate-of-the-art multilingual and multitask learning approaches. We evaluate\nour dataset in various monolingual, cross-lingual and machine translation\nclassification settings and compare it against open source English datasets\nthat we aggregated and merged for this task. Then we discuss how this approach\ncan be used to create large scale hate-speech datasets and how to leverage our\nannotations in order to improve hate speech detection and classification in\ngeneral.\n", "date": "2023-04-03", "categories": []}
{"id": "2304.01179", "title": "Hate Speech Targets Detection in Parler using BERT", "abstract": "  Online social networks have become a fundamental component of our everyday\nlife. Unfortunately, these platforms are also a stage for hate speech. Popular\nsocial networks have regularized rules against hate speech. Consequently,\nsocial networks like Parler and Gab advocating and claiming to be free speech\nplatforms have evolved. These platforms have become a district for hate speech\nagainst diverse targets. We present in our paper a pipeline for detecting hate\nspeech and its targets and use it for creating Parler hate targets'\ndistribution. The pipeline consists of two models; one for hate speech\ndetection and the second for target classification, both based on BERT with\nBack-Translation and data pre-processing for improved results. The source code\nused in this work, as well as other relevant sources, are available at:\nhttps://github.com/NadavSc/HateRecognition.git\n", "date": "2023-04-03", "categories": []}
{"id": "2211.08447", "title": "SexWEs: Domain-Aware Word Embeddings via Cross-lingual Semantic\n  Specialisation for Chinese Sexism Detection in Social Media", "abstract": "  The goal of sexism detection is to mitigate negative online content targeting\ncertain gender groups of people. However, the limited availability of labeled\nsexism-related datasets makes it problematic to identify online sexism for\nlow-resource languages. In this paper, we address the task of automatic sexism\ndetection in social media for one low-resource language -- Chinese. Rather than\ncollecting new sexism data or building cross-lingual transfer learning models,\nwe develop a cross-lingual domain-aware semantic specialisation system in order\nto make the most of existing data. Semantic specialisation is a technique for\nretrofitting pre-trained distributional word vectors by integrating external\nlinguistic knowledge (such as lexico-semantic relations) into the specialised\nfeature space. To do this, we leverage semantic resources for sexism from a\nhigh-resource language (English) to specialise pre-trained word vectors in the\ntarget language (Chinese) to inject domain knowledge. We demonstrate the\nbenefit of our sexist word embeddings (SexWEs) specialised by our framework via\nintrinsic evaluation of word similarity and extrinsic evaluation of sexism\ndetection. Compared with other specialisation approaches and Chinese baseline\nword vectors, our SexWEs shows an average score improvement of 0.033 and 0.064\nin both intrinsic and extrinsic evaluations, respectively. The ablative results\nand visualisation of SexWEs also prove the effectiveness of our framework on\nretrofitting word vectors in low-resource languages.\n", "date": "2022-11-15", "categories": []}
{"id": "2303.17927", "title": "Cross-Cultural Transfer Learning for Chinese Offensive Language\n  Detection", "abstract": "  Detecting offensive language is a challenging task. Generalizing across\ndifferent cultures and languages becomes even more challenging: besides\nlexical, syntactic and semantic differences, pragmatic aspects such as cultural\nnorms and sensitivities, which are particularly relevant in this context, vary\ngreatly. In this paper, we target Chinese offensive language detection and aim\nto investigate the impact of transfer learning using offensive language\ndetection data from different cultural backgrounds, specifically Korean and\nEnglish. We find that culture-specific biases in what is considered offensive\nnegatively impact the transferability of language models (LMs) and that LMs\ntrained on diverse cultural data are sensitive to different features in Chinese\noffensive language detection. In a few-shot learning scenario, however, our\nstudy shows promising prospects for non-English offensive language detection\nwith limited resources. Our findings highlight the importance of cross-cultural\ntransfer learning in improving offensive language detection and promoting\ninclusive digital spaces.\n", "date": "2023-03-31", "categories": []}
{"id": "2303.12808", "title": "PACO: Provocation Involving Action, Culture, and Oppression", "abstract": "  In India, people identify with a particular group based on certain attributes\nsuch as religion. The same religious groups are often provoked against each\nother. Previous studies show the role of provocation in increasing tensions\nbetween India's two prominent religious groups: Hindus and Muslims. With the\nadvent of the Internet, such provocation also surfaced on social media\nplatforms such as WhatsApp.\n  By leveraging an existing dataset of Indian WhatsApp posts, we identified\nthree categories of provoking sentences against Indian Muslims. Further, we\nlabeled 7,000 sentences for three provocation categories and called this\ndataset PACO. We leveraged PACO to train a model that can identify provoking\nsentences from a WhatsApp post. Our best model is fine-tuned RoBERTa and\nachieved a 0.851 average AUC score over five-fold cross-validation.\nAutomatically identifying provoking sentences could stop provoking text from\nreaching out to the masses, and can prevent possible discrimination or violence\nagainst the target religious group.\n  Further, we studied the provocative speech through a pragmatic lens, by\nidentifying the dialog acts and impoliteness super-strategies used against the\nreligious group.\n", "date": "2023-03-19", "categories": []}
{"id": "2302.07736", "title": "Is ChatGPT better than Human Annotators? Potential and Limitations of\n  ChatGPT in Explaining Implicit Hate Speech", "abstract": "  Recent studies have alarmed that many online hate speeches are implicit. With\nits subtle nature, the explainability of the detection of such hateful speech\nhas been a challenging problem. In this work, we examine whether ChatGPT can be\nused for providing natural language explanations (NLEs) for implicit hateful\nspeech detection. We design our prompt to elicit concise ChatGPT-generated NLEs\nand conduct user studies to evaluate their qualities by comparison with\nhuman-written NLEs. We discuss the potential and limitations of ChatGPT in the\ncontext of implicit hateful speech research.\n", "date": "2023-02-11", "categories": []}
{"id": "2303.09402", "title": "ToxVis: Enabling Interpretability of Implicit vs. Explicit Toxicity\n  Detection Models with Interactive Visualization", "abstract": "  The rise of hate speech on online platforms has led to an urgent need for\neffective content moderation. However, the subjective and multi-faceted nature\nof hateful online content, including implicit hate speech, poses significant\nchallenges to human moderators and content moderation systems. To address this\nissue, we developed ToxVis, a visually interactive and explainable tool for\nclassifying hate speech into three categories: implicit, explicit, and\nnon-hateful. We fine-tuned two transformer-based models using RoBERTa, XLNET,\nand GPT-3 and used deep learning interpretation techniques to provide\nexplanations for the classification results. ToxVis enables users to input\npotentially hateful text and receive a classification result along with a\nvisual explanation of which words contributed most to the decision. By making\nthe classification process explainable, ToxVis provides a valuable tool for\nunderstanding the nuances of hateful content and supporting more effective\ncontent moderation. Our research contributes to the growing body of work aimed\nat mitigating the harms caused by online hate speech and demonstrates the\npotential for combining state-of-the-art natural language processing models\nwith interpretable deep learning techniques to address this critical issue.\nFinally, ToxVis can serve as a resource for content moderators, social media\nplatforms, and researchers working to combat the spread of hate speech online.\n", "date": "2023-03-01", "categories": []}
{"id": "2303.07484", "title": "Deep Learning Approach for Classifying the Aggressive Comments on Social\n  Media: Machine Translated Data Vs Real Life Data", "abstract": "  Aggressive comments on social media negatively impact human life. Such\noffensive contents are responsible for depression and suicidal-related\nactivities. Since online social networking is increasing day by day, the hate\ncontent is also increasing. Several investigations have been done on the domain\nof cyberbullying, cyberaggression, hate speech, etc. The majority of the\ninquiry has been done in the English language. Some languages (Hindi and\nBangla) still lack proper investigations due to the lack of a dataset. This\npaper particularly worked on the Hindi, Bangla, and English datasets to detect\naggressive comments and have shown a novel way of generating machine-translated\ndata to resolve data unavailability issues. A fully machine-translated English\ndataset has been analyzed with the models such as the Long Short term memory\nmodel (LSTM), Bidirectional Long-short term memory model (BiLSTM),\nLSTM-Autoencoder, word2vec, Bidirectional Encoder Representations from\nTransformers (BERT), and generative pre-trained transformer (GPT-2) to make an\nobservation on how the models perform on a machine-translated noisy dataset. We\nhave compared the performance of using the noisy data with two more datasets\nsuch as raw data, which does not contain any noises, and semi-noisy data, which\ncontains a certain amount of noisy data. We have classified both the raw and\nsemi-noisy data using the aforementioned models. To evaluate the performance of\nthe models, we have used evaluation metrics such as F1-score,accuracy,\nprecision, and recall. We have achieved the highest accuracy on raw data using\nthe gpt2 model, semi-noisy data using the BERT model, and fully\nmachine-translated data using the BERT model. Since many languages do not have\nproper data availability, our approach will help researchers create\nmachine-translated datasets for several analysis purposes.\n", "date": "2023-03-13", "categories": []}
{"id": "2210.00465", "title": "Assessing the impact of contextual information in hate speech detection", "abstract": "  In recent years, hate speech has gained great relevance in social networks\nand other virtual media because of its intensity and its relationship with\nviolent acts against members of protected groups. Due to the great amount of\ncontent generated by users, great effort has been made in the research and\ndevelopment of automatic tools to aid the analysis and moderation of this\nspeech, at least in its most threatening forms. One of the limitations of\ncurrent approaches to automatic hate speech detection is the lack of context.\nMost studies and resources are performed on data without context; that is,\nisolated messages without any type of conversational context or the topic being\ndiscussed. This restricts the available information to define if a post on a\nsocial network is hateful or not. In this work, we provide a novel corpus for\ncontextualized hate speech detection based on user responses to news posts from\nmedia outlets on Twitter. This corpus was collected in the Rioplatense\ndialectal variety of Spanish and focuses on hate speech associated with the\nCOVID-19 pandemic. Classification experiments using state-of-the-art techniques\nshow evidence that adding contextual information improves hate speech detection\nperformance for two proposed tasks (binary and multi-label prediction). We make\nour code, models, and corpus available for further research.\n", "date": "2022-10-02", "categories": []}
{"id": "2303.02513", "title": "Model-Agnostic Meta-Learning for Multilingual Hate Speech Detection", "abstract": "  Hate speech in social media is a growing phenomenon, and detecting such toxic\ncontent has recently gained significant traction in the research community.\nExisting studies have explored fine-tuning language models (LMs) to perform\nhate speech detection, and these solutions have yielded significant\nperformance. However, most of these studies are limited to detecting hate\nspeech only in English, neglecting the bulk of hateful content that is\ngenerated in other languages, particularly in low-resource languages.\nDeveloping a classifier that captures hate speech and nuances in a low-resource\nlanguage with limited data is extremely challenging. To fill the research gap,\nwe propose HateMAML, a model-agnostic meta-learning-based framework that\neffectively performs hate speech detection in low-resource languages. HateMAML\nutilizes a self-supervision strategy to overcome the limitation of data\nscarcity and produces better LM initialization for fast adaptation to an unseen\ntarget language (i.e., cross-lingual transfer) or other hate speech datasets\n(i.e., domain generalization). Extensive experiments are conducted on five\ndatasets across eight different low-resource languages. The results show that\nHateMAML outperforms the state-of-the-art baselines by more than 3% in the\ncross-domain multilingual transfer setting. We also conduct ablation studies to\nanalyze the characteristics of HateMAML.\n", "date": "2023-03-04", "categories": []}
{"id": "2302.12840", "title": "HULAT at SemEval-2023 Task 10: Data augmentation for pre-trained\n  transformers applied to the detection of sexism in social media", "abstract": "  This paper describes our participation in SemEval-2023 Task 10, whose goal is\nthe detection of sexism in social media. We explore some of the most popular\ntransformer models such as BERT, DistilBERT, RoBERTa, and XLNet. We also study\ndifferent data augmentation techniques to increase the training dataset. During\nthe development phase, our best results were obtained by using RoBERTa and data\naugmentation for tasks B and C. However, the use of synthetic data does not\nimprove the results for task C. We participated in the three subtasks. Our\napproach still has much room for improvement, especially in the two\nfine-grained classifications. All our code is available in the repository\nhttps://github.com/isegura/hulat_edos.\n", "date": "2023-02-24", "categories": []}
{"id": "2302.09243", "title": "A Federated Approach for Hate Speech Detection", "abstract": "  Hate speech detection has been the subject of high research attention, due to\nthe scale of content created on social media. In spite of the attention and the\nsensitive nature of the task, privacy preservation in hate speech detection has\nremained under-studied. The majority of research has focused on centralised\nmachine learning infrastructures which risk leaking data. In this paper, we\nshow that using federated machine learning can help address privacy the\nconcerns that are inherent to hate speech detection while obtaining up to 6.81%\nimprovement in terms of F1-score.\n", "date": "2023-02-18", "categories": []}
{"id": "2302.08777", "title": "Hate Speech and Offensive Language Detection using an Emotion-aware\n  Shared Encoder", "abstract": "  The rise of emergence of social media platforms has fundamentally altered how\npeople communicate, and among the results of these developments is an increase\nin online use of abusive content. Therefore, automatically detecting this\ncontent is essential for banning inappropriate information, and reducing\ntoxicity and violence on social media platforms. The existing works on hate\nspeech and offensive language detection produce promising results based on\npre-trained transformer models, however, they considered only the analysis of\nabusive content features generated through annotated datasets. This paper\naddresses a multi-task joint learning approach which combines external\nemotional features extracted from another corpora in dealing with the\nimbalanced and scarcity of labeled datasets. Our analysis are using two\nwell-known Transformer-based models, BERT and mBERT, where the later is used to\naddress abusive content detection in multi-lingual scenarios. Our model jointly\nlearns abusive content detection with emotional features by sharing\nrepresentations through transformers' shared encoder. This approach increases\ndata efficiency, reduce overfitting via shared representations, and ensure fast\nlearning by leveraging auxiliary information. Our findings demonstrate that\nemotional knowledge helps to more reliably identify hate speech and offensive\nlanguage across datasets. Our hate speech detection Multi-task model exhibited\n3% performance improvement over baseline models, but the performance of\nmulti-task models were not significant for offensive language detection task.\nMore interestingly, in both tasks, multi-task models exhibits less false\npositive errors compared to single task scenario.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.05703", "title": "HateProof: Are Hateful Meme Detection Systems really Robust?", "abstract": "  Exploiting social media to spread hate has tremendously increased over the\nyears. Lately, multi-modal hateful content such as memes has drawn relatively\nmore traction than uni-modal content. Moreover, the availability of implicit\ncontent payloads makes them fairly challenging to be detected by existing\nhateful meme detection systems. In this paper, we present a use case study to\nanalyze such systems' vulnerabilities against external adversarial attacks. We\nfind that even very simple perturbations in uni-modal and multi-modal settings\nperformed by humans with little knowledge about the model can make the existing\ndetection models highly vulnerable. Empirically, we find a noticeable\nperformance drop of as high as 10% in the macro-F1 score for certain attacks.\nAs a remedy, we attempt to boost the model's robustness using contrastive\nlearning as well as an adversarial training-based method - VILLA. Using an\nensemble of the above two approaches, in two of our high resolution datasets,\nwe are able to (re)gain back the performance to a large extent for certain\nattacks. We believe that ours is a first step toward addressing this crucial\nproblem in an adversarial setting and would inspire more such investigations in\nthe future.\n", "date": "2023-02-11", "categories": []}
{"id": "2302.04156", "title": "Prompting for Multimodal Hateful Meme Classification", "abstract": "  Hateful meme classification is a challenging multimodal task that requires\ncomplex reasoning and contextual background knowledge. Ideally, we could\nleverage an explicit external knowledge base to supplement contextual and\ncultural information in hateful memes. However, there is no known explicit\nexternal knowledge base that could provide such hate speech contextual\ninformation. To address this gap, we propose PromptHate, a simple yet effective\nprompt-based model that prompts pre-trained language models (PLMs) for hateful\nmeme classification. Specifically, we construct simple prompts and provide a\nfew in-context examples to exploit the implicit knowledge in the pre-trained\nRoBERTa language model for hateful meme classification. We conduct extensive\nexperiments on two publicly available hateful and offensive meme datasets. Our\nexperimental results show that PromptHate is able to achieve a high AUC of\n90.96, outperforming state-of-the-art baselines on the hateful meme\nclassification task. We also perform fine-grained analyses and case studies on\nvarious prompt settings and demonstrate the effectiveness of the prompts on\nhateful meme classification.\n", "date": "2023-02-08", "categories": []}
{"id": "2301.13668", "title": "Automated Sentiment and Hate Speech Analysis of Facebook Data by\n  Employing Multilingual Transformer Models", "abstract": "  In recent years, there has been a heightened consensus within academia and in\nthe public discourse that Social Media Platforms (SMPs), amplify the spread of\nhateful and negative sentiment content. Researchers have identified how hateful\ncontent, political propaganda, and targeted messaging contributed to real-world\nharms including insurrections against democratically elected governments,\ngenocide, and breakdown of social cohesion due to heightened negative discourse\ntowards certain communities in parts of the world. To counter these issues,\nSMPs have created semi-automated systems that can help identify toxic speech.\nIn this paper we analyse the statistical distribution of hateful and negative\nsentiment contents within a representative Facebook dataset (n= 604,703)\nscrapped through 648 public Facebook pages which identify themselves as\nproponents (and followers) of far-right Hindutva actors. These pages were\nidentified manually using keyword searches on Facebook and on CrowdTangleand\nclassified as far-right Hindutva pages based on page names, page descriptions,\nand discourses shared on these pages. We employ state-of-the-art, open-source\nXLM-T multilingual transformer-based language models to perform sentiment and\nhate speech analysis of the textual contents shared on these pages over a\nperiod of 5.5 years. The result shows the statistical distributions of the\npredicted sentiment and the hate speech labels; top actors, and top page\ncategories. We further discuss the benchmark performances and limitations of\nthese pre-trained language models.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.11579", "title": "Down the Rabbit Hole: Detecting Online Extremism, Radicalisation, and\n  Politicised Hate Speech", "abstract": "  Social media is a modern person's digital voice to project and engage with\nnew ideas and mobilise communities $\\unicode{x2013}$ a power shared with\nextremists. Given the societal risks of unvetted content-moderating algorithms\nfor Extremism, Radicalisation, and Hate speech (ERH) detection, responsible\nsoftware engineering must understand the who, what, when, where, and why such\nmodels are necessary to protect user safety and free expression. Hence, we\npropose and examine the unique research field of ERH context mining to unify\ndisjoint studies. Specifically, we evaluate the start-to-finish design process\nfrom socio-technical definition-building and dataset collection strategies to\ntechnical algorithm design and performance. Our 2015-2021 51-study Systematic\nLiterature Review (SLR) provides the first cross-examination of textual,\nnetwork, and visual approaches to detecting extremist affiliation, hateful\ncontent, and radicalisation towards groups and movements. We identify\nconsensus-driven ERH definitions and propose solutions to existing ideological\nand geographic biases, particularly due to the lack of research in\nOceania/Australasia. Our hybridised investigation on Natural Language\nProcessing, Community Detection, and visual-text models demonstrates the\ndominating performance of textual transformer-based algorithms. We conclude\nwith vital recommendations for ERH context mining researchers and propose an\nuptake roadmap with guidelines for researchers, industries, and governments to\nenable a safer cyberspace.\n", "date": "2023-01-27", "categories": []}
{"id": "2208.07522", "title": "Reliable Decision from Multiple Subtasks through Threshold Optimization:\n  Content Moderation in the Wild", "abstract": "  Social media platforms struggle to protect users from harmful content through\ncontent moderation. These platforms have recently leveraged machine learning\nmodels to cope with the vast amount of user-generated content daily. Since\nmoderation policies vary depending on countries and types of products, it is\ncommon to train and deploy the models per policy. However, this approach is\nhighly inefficient, especially when the policies change, requiring dataset\nre-labeling and model re-training on the shifted data distribution. To\nalleviate this cost inefficiency, social media platforms often employ\nthird-party content moderation services that provide prediction scores of\nmultiple subtasks, such as predicting the existence of underage personnel, rude\ngestures, or weapons, instead of directly providing final moderation decisions.\nHowever, making a reliable automated moderation decision from the prediction\nscores of the multiple subtasks for a specific target policy has not been\nwidely explored yet. In this study, we formulate real-world scenarios of\ncontent moderation and introduce a simple yet effective threshold optimization\nmethod that searches the optimal thresholds of the multiple subtasks to make a\nreliable moderation decision in a cost-effective way. Extensive experiments\ndemonstrate that our approach shows better performance in content moderation\ncompared to existing threshold optimization methods and heuristics.\n", "date": "2022-08-16", "categories": []}
{"id": "2301.10186", "title": "ViHOS: Hate Speech Spans Detection for Vietnamese", "abstract": "  The rise in hateful and offensive language directed at other users is one of\nthe adverse side effects of the increased use of social networking platforms.\nThis could make it difficult for human moderators to review tagged comments\nfiltered by classification systems. To help address this issue, we present the\nViHOS (Vietnamese Hate and Offensive Spans) dataset, the first human-annotated\ncorpus containing 26k spans on 11k comments. We also provide definitions of\nhateful and offensive spans in Vietnamese comments as well as detailed\nannotation guidelines. Besides, we conduct experiments with various\nstate-of-the-art models. Specifically, XLM-R$_{Large}$ achieved the best\nF1-scores in Single span detection and All spans detection, while\nPhoBERT$_{Large}$ obtained the highest in Multiple spans detection. Finally,\nour error analysis demonstrates the difficulties in detecting specific types of\nspans in our data for future research.\n  Disclaimer: This paper contains real comments that could be considered\nprofane, offensive, or abusive.\n", "date": "2023-01-24", "categories": []}
{"id": "2301.04248", "title": "Predicting Hateful Discussions on Reddit using Graph Transformer\n  Networks and Communal Context", "abstract": "  We propose a system to predict harmful discussions on social media platforms.\nOur solution uses contextual deep language models and proposes the novel idea\nof integrating state-of-the-art Graph Transformer Networks to analyze all\nconversations that follow an initial post. This framework also supports\nadapting to future comments as the conversation unfolds. In addition, we study\nwhether a community-specific analysis of hate speech leads to more effective\ndetection of hateful discussions. We evaluate our approach on 333,487 Reddit\ndiscussions from various communities. We find that community-specific modeling\nimproves performance two-fold and that models which capture wider-discussion\ncontext improve accuracy by 28\\% (35\\% for the most hateful content) compared\nto limited context models.\n", "date": "2023-01-10", "categories": []}
{"id": "2212.10405", "title": "AnnoBERT: Effectively Representing Multiple Annotators' Label Choices to\n  Improve Hate Speech Detection", "abstract": "  Supervised approaches generally rely on majority-based labels. However, it is\nhard to achieve high agreement among annotators in subjective tasks such as\nhate speech detection. Existing neural network models principally regard labels\nas categorical variables, while ignoring the semantic information in diverse\nlabel texts. In this paper, we propose AnnoBERT, a first-of-its-kind\narchitecture integrating annotator characteristics and label text with a\ntransformer-based model to detect hate speech, with unique representations\nbased on each annotator's characteristics via Collaborative Topic Regression\n(CTR) and integrate label text to enrich textual representations. During\ntraining, the model associates annotators with their label choices given a\npiece of text; during evaluation, when label information is not available, the\nmodel predicts the aggregated label given by the participating annotators by\nutilising the learnt association. The proposed approach displayed an advantage\nin detecting hate speech, especially in the minority class and edge cases with\nannotator disagreement. Improvement in the overall performance is the largest\nwhen the dataset is more label-imbalanced, suggesting its practical value in\nidentifying real-world hate speech, as the volume of hate speech in-the-wild is\nextremely small on social media, when compared with normal (non-hate) speech.\nThrough ablation studies, we show the relative contributions of annotator\nembeddings and label text to the model performance, and tested a range of\nalternative annotator embeddings and label text combinations.\n", "date": "2022-12-20", "categories": []}
{"id": "2301.01874", "title": "Critical Perspectives: A Benchmark Revealing Pitfalls in PerspectiveAPI", "abstract": "  Detecting \"toxic\" language in internet content is a pressing social and\ntechnical challenge. In this work, we focus on PERSPECTIVE from Jigsaw, a\nstate-of-the-art tool that promises to score the \"toxicity\" of text, with a\nrecent model update that claims impressive results (Lees et al., 2022). We seek\nto challenge certain normative claims about toxic language by proposing a new\nbenchmark, Selected Adversarial SemanticS, or SASS. We evaluate PERSPECTIVE on\nSASS, and compare to low-effort alternatives, like zero-shot and few-shot GPT-3\nprompt models, in binary classification settings. We find that PERSPECTIVE\nexhibits troubling shortcomings across a number of our toxicity categories.\nSASS provides a new tool for evaluating performance on previously undetected\ntoxic language that avoids common normative pitfalls. Our work leads us to\nemphasize the importance of questioning assumptions made by tools already in\ndeployment for toxicity detection in order to anticipate and prevent disparate\nharms.\n", "date": "2023-01-05", "categories": []}
{"id": "2212.02108", "title": "Human-in-the-Loop Hate Speech Classification in a Multilingual Context", "abstract": "  The shift of public debate to the digital sphere has been accompanied by a\nrise in online hate speech. While many promising approaches for hate speech\nclassification have been proposed, studies often focus only on a single\nlanguage, usually English, and do not address three key concerns:\npost-deployment performance, classifier maintenance and infrastructural\nlimitations. In this paper, we introduce a new human-in-the-loop BERT-based\nhate speech classification pipeline and trace its development from initial data\ncollection and annotation all the way to post-deployment. Our classifier,\ntrained using data from our original corpus of over 422k examples, is\nspecifically developed for the inherently multilingual setting of Switzerland\nand outperforms with its F1 score of 80.5 the currently best-performing\nBERT-based multilingual classifier by 5.8 F1 points in German and 3.6 F1 points\nin French. Our systematic evaluations over a 12-month period further highlight\nthe vital importance of continuous, human-in-the-loop classifier maintenance to\nensure robust hate speech classification post-deployment.\n", "date": "2022-12-05", "categories": []}
{"id": "2212.13949", "title": "RevealED: Uncovering Pro-Eating Disorder Content on Twitter Using Deep\n  Learning", "abstract": "  The Covid-19 pandemic induced a vast increase in adolescents diagnosed with\neating disorders and hospitalized due to eating disorders. This immense growth\nstemmed partially from the stress of the pandemic but also from increased\nexposure to content that promotes eating disorders via social media, which,\nwithin the last decade, has become plagued by pro-eating disorder content. This\nstudy aimed to create a deep learning model capable of determining whether a\ngiven social media post promotes eating disorders based solely on image data.\nTweets from hashtags that have been documented to promote eating disorders\nalong with Tweets from unrelated hashtags were collected. After prepossessing,\nthese images were labeled as either pro-eating disorder or not based on which\nTwitter hashtag they were scraped from. Several deep-learning models were\ntrained on the scraped dataset and were evaluated based on their accuracy, F1\nscore, precision, and recall. Ultimately, the Vision Transformer model was\ndetermined to be the most accurate, attaining an F1 score of 0.877 and an\naccuracy of 86.7% on the test set. The model, which was applied to unlabeled\nTwitter image data scraped from \"#selfie\", uncovered seasonal fluctuations in\nthe relative abundance of pro-eating disorder content, which reached its peak\nin the summertime. These fluctuations correspond not only to the seasons, but\nalso to stressors, such as the Covid-19 pandemic. Moreover, the Twitter image\ndata indicated that the relative amount of pro-eating disorder content has been\nsteadily rising over the last five years and is likely to continue increasing\nin the future.\n", "date": "2022-12-28", "categories": []}
{"id": "2212.14100", "title": "Leveraging World Knowledge in Implicit Hate Speech Detection", "abstract": "  While much attention has been paid to identifying explicit hate speech,\nimplicit hateful expressions that are disguised in coded or indirect language\nare pervasive and remain a major challenge for existing hate speech detection\nsystems. This paper presents the first attempt to apply Entity Linking (EL)\ntechniques to both explicit and implicit hate speech detection, where we show\nthat such real world knowledge about entity mentions in a text does help models\nbetter detect hate speech, and the benefit of adding it into the model is more\npronounced when explicit entity triggers (e.g., rally, KKK) are present. We\nalso discuss cases where real world knowledge does not add value to hate speech\ndetection, which provides more insights into understanding and modeling the\nsubtleties of hate speech.\n", "date": "2022-12-28", "categories": []}
{"id": "2212.14727", "title": "Countering Malicious Content Moderation Evasion in Online Social\n  Networks: Simulation and Detection of Word Camouflage", "abstract": "  Content moderation is the process of screening and monitoring user-generated\ncontent online. It plays a crucial role in stopping content resulting from\nunacceptable behaviors such as hate speech, harassment, violence against\nspecific groups, terrorism, racism, xenophobia, homophobia, or misogyny, to\nmention some few, in Online Social Platforms. These platforms make use of a\nplethora of tools to detect and manage malicious information; however,\nmalicious actors also improve their skills, developing strategies to surpass\nthese barriers and continuing to spread misleading information. Twisting and\ncamouflaging keywords are among the most used techniques to evade platform\ncontent moderation systems. In response to this recent ongoing issue, this\npaper presents an innovative approach to address this linguistic trend in\nsocial networks through the simulation of different content evasion techniques\nand a multilingual Transformer model for content evasion detection. In this\nway, we share with the rest of the scientific community a multilingual public\ntool, named \"pyleetspeak\" to generate/simulate in a customizable way the\nphenomenon of content evasion through automatic word camouflage and a\nmultilingual Named-Entity Recognition (NER) Transformer-based model tuned for\nits recognition and detection. The multilingual NER model is evaluated in\ndifferent textual scenarios, detecting different types and mixtures of\ncamouflage techniques, achieving an overall weighted F1 score of 0.8795. This\narticle contributes significantly to countering malicious information by\ndeveloping multilingual tools to simulate and detect new methods of evasion of\ncontent on social networks, making the fight against information disorders more\neffective.\n", "date": "2022-12-27", "categories": []}
{"id": "2212.13205", "title": "Personalized Prediction of Offensive News Comments by Considering the\n  Characteristics of Commenters", "abstract": "  When reading news articles on social networking services and news sites,\nreaders can view comments marked by other people on these articles. By reading\nthese comments, a reader can understand the public opinion about the news, and\nit is often helpful to grasp the overall picture of the news. However, these\ncomments often contain offensive language that readers do not prefer to read.\nThis study aims to predict such offensive comments to improve the quality of\nthe experience of the reader while reading comments. By considering the\ndiversity of the readers' values, the proposed method predicts offensive news\ncomments for each reader based on the feedback from a small number of news\ncomments that the reader rated as \"offensive\" in the past. In addition, we used\na machine learning model that considers the characteristics of the commenters\nto make predictions, independent of the words and topics in news comments. The\nexperimental results of the proposed method show that prediction can be\npersonalized even when the amount of readers' feedback data used in the\nprediction is limited. In particular, the proposed method, which considers the\ncommenters' characteristics, has a low probability of false detection of\noffensive comments.\n", "date": "2022-12-26", "categories": []}
{"id": "2212.10039", "title": "A Twitter BERT Approach for Offensive Language Detection in Marathi", "abstract": "  Automated offensive language detection is essential in combating the spread\nof hate speech, particularly in social media. This paper describes our work on\nOffensive Language Identification in low resource Indic language Marathi. The\nproblem is formulated as a text classification task to identify a tweet as\noffensive or non-offensive. We evaluate different mono-lingual and\nmulti-lingual BERT models on this classification task, focusing on BERT models\npre-trained with social media datasets. We compare the performance of MuRIL,\nMahaTweetBERT, MahaTweetBERT-Hateful, and MahaBERT on the HASOC 2022 test set.\nWe also explore external data augmentation from other existing Marathi hate\nspeech corpus HASOC 2021 and L3Cube-MahaHate. The MahaTweetBERT, a BERT model,\npre-trained on Marathi tweets when fine-tuned on the combined dataset (HASOC\n2021 + HASOC 2022 + MahaHate), outperforms all models with an F1 score of 98.43\non the HASOC 2022 test set. With this, we also provide a new state-of-the-art\nresult on HASOC 2022 / MOLD v2 test set.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.06655", "title": "The Hateful Memes Challenge Next Move", "abstract": "  State-of-the-art image and text classification models, such as Convolutional\nNeural Networks and Transformers, have long been able to classify their\nrespective unimodal reasoning satisfactorily with accuracy close to or\nexceeding human accuracy. However, images embedded with text, such as hateful\nmemes, are hard to classify using unimodal reasoning when difficult examples,\nsuch as benign confounders, are incorporated into the data set. We attempt to\ngenerate more labeled memes in addition to the Hateful Memes data set from\nFacebook AI, based on the framework of a winning team from the Hateful Meme\nChallenge. To increase the number of labeled memes, we explore semi-supervised\nlearning using pseudo-labels for newly introduced, unlabeled memes gathered\nfrom the Memotion Dataset 7K. We find that the semi-supervised learning task on\nunlabeled data required human intervention and filtering and that adding a\nlimited amount of new data yields no extra classification performance.\n", "date": "2022-12-13", "categories": []}
{"id": "2212.07424", "title": "Hope Speech Detection on Social Media Platforms", "abstract": "  Since personal computers became widely available in the consumer market, the\namount of harmful content on the internet has significantly expanded. In simple\nterms, harmful content is anything online which causes a person distress or\nharm. It may include hate speech, violent content, threats, non-hope speech,\netc. The online content must be positive, uplifting and supportive. Over the\npast few years, many studies have focused on solving this problem through hate\nspeech detection, but very few focused on identifying hope speech. This paper\ndiscusses various machine learning approaches to identify a sentence as Hope\nSpeech, Non-Hope Speech, or a Neutral sentence. The dataset used in the study\ncontains English YouTube comments and is released as a part of the shared task\n\"EACL-2021: Hope Speech Detection for Equality, Diversity, and Inclusion\".\nInitially, the dataset obtained from the shared task had three classes: Hope\nSpeech, non-Hope speech, and not in English; however, upon deeper inspection,\nwe discovered that dataset relabeling is required. A group of undergraduates\nwas hired to help perform the entire dataset's relabeling task. We experimented\nwith conventional machine learning models (such as Na\\\"ive Bayes, logistic\nregression and support vector machine) and pre-trained models (such as BERT) on\nrelabeled data. According to the experimental results, the relabeled data has\nachieved a better accuracy for Hope speech identification than the original\ndata set.\n", "date": "2022-11-14", "categories": []}
{"id": "2210.04267", "title": "Spread Love Not Hate: Undermining the Importance of Hateful Pre-training\n  for Hate Speech Detection", "abstract": "  Pre-training large neural language models, such as BERT, has led to\nimpressive gains on many natural language processing (NLP) tasks. Although this\nmethod has proven to be effective for many domains, it might not always provide\ndesirable benefits. In this paper, we study the effects of hateful pre-training\non low-resource hate speech classification tasks. While previous studies on the\nEnglish language have emphasized its importance, we aim to augment their\nobservations with some non-obvious insights. We evaluate different variations\nof tweet-based BERT models pre-trained on hateful, non-hateful, and mixed\nsubsets of a 40M tweet dataset. This evaluation is carried out for the Indian\nlanguages Hindi and Marathi. This paper is empirical evidence that hateful\npre-training is not the best pre-training option for hate speech detection. We\nshow that pre-training on non-hateful text from the target domain provides\nsimilar or better results. Further, we introduce HindTweetBERT and\nMahaTweetBERT, the first publicly available BERT models pre-trained on Hindi\nand Marathi tweets, respectively. We show that they provide state-of-the-art\nperformance on hate speech classification tasks. We also release hateful BERT\nfor the two languages and a gold hate speech evaluation benchmark HateEval-Hi\nand HateEval-Mr consisting of manually labeled 2000 tweets each. The models and\ndata are available at https://github.com/l3cube-pune/MarathiNLP .\n", "date": "2022-10-09", "categories": []}
{"id": "2212.04533", "title": "Deep Architectures for Content Moderation and Movie Content Rating", "abstract": "  Rating a video based on its content is an important step for classifying\nvideo age categories. Movie content rating and TV show rating are the two most\ncommon rating systems established by professional committees. However, manually\nreviewing and evaluating scene/film content by a committee is a tedious work\nand it becomes increasingly difficult with the ever-growing amount of online\nvideo content. As such, a desirable solution is to use computer vision based\nvideo content analysis techniques to automate the evaluation process. In this\npaper, related works are summarized for action recognition, multi-modal\nlearning, movie genre classification, and sensitive content detection in the\ncontext of content moderation and movie content rating. The project page is\navailable at https://github.com/fcakyon/content-moderation-deep-learning.\n", "date": "2022-12-08", "categories": []}
{"id": "2210.10839", "title": "How Hate Speech Varies by Target Identity: A Computational Analysis", "abstract": "  This paper investigates how hate speech varies in systematic ways according\nto the identities it targets. Across multiple hate speech datasets annotated\nfor targeted identities, we find that classifiers trained on hate speech\ntargeting specific identity groups struggle to generalize to other targeted\nidentities. This provides empirical evidence for differences in hate speech by\ntarget identity; we then investigate which patterns structure this variation.\nWe find that the targeted demographic category (e.g. gender/sexuality or\nrace/ethnicity) appears to have a greater effect on the language of hate speech\nthan does the relative social power of the targeted identity group. We also\nfind that words associated with hate speech targeting specific identities often\nrelate to stereotypes, histories of oppression, current social movements, and\nother social contexts specific to identities. These experiments suggest the\nimportance of considering targeted identity, as well as the social contexts\nassociated with these identities, in automated hate speech classification.\n", "date": "2022-10-19", "categories": []}
{"id": "2212.00668", "title": "Leveraging Large-scale Multimedia Datasets to Refine Content Moderation\n  Models", "abstract": "  The sheer volume of online user-generated content has rendered content\nmoderation technologies essential in order to protect digital platform\naudiences from content that may cause anxiety, worry, or concern. Despite the\nefforts towards developing automated solutions to tackle this problem, creating\naccurate models remains challenging due to the lack of adequate task-specific\ntraining data. The fact that manually annotating such data is a highly\ndemanding procedure that could severely affect the annotators' emotional\nwell-being is directly related to the latter limitation. In this paper, we\npropose the CM-Refinery framework that leverages large-scale multimedia\ndatasets to automatically extend initial training datasets with hard examples\nthat can refine content moderation models, while significantly reducing the\ninvolvement of human annotators. We apply our method on two model adaptation\nstrategies designed with respect to the different challenges observed while\ncollecting data, i.e. lack of (i) task-specific negative data or (ii) both\npositive and negative data. Additionally, we introduce a diversity criterion\napplied to the data collection process that further enhances the generalization\nperformance of the refined models. The proposed method is evaluated on the Not\nSafe for Work (NSFW) and disturbing content detection tasks on benchmark\ndatasets achieving 1.32% and 1.94% accuracy improvements compared to the state\nof the art, respectively. Finally, it significantly reduces human involvement,\nas 92.54% of data are automatically annotated in case of disturbing content\nwhile no human intervention is required for the NSFW task.\n", "date": "2022-12-01", "categories": []}
{"id": "2211.16525", "title": "Proactive Moderation of Online Discussions: Existing Practices and the\n  Potential for Algorithmic Support", "abstract": "  To address the widespread problem of uncivil behavior, many online discussion\nplatforms employ human moderators to take action against objectionable content,\nsuch as removing it or placing sanctions on its authors. This reactive paradigm\nof taking action against already-posted antisocial content is currently the\nmost common form of moderation, and has accordingly underpinned many recent\nefforts at introducing automation into the moderation process. Comparatively\nless work has been done to understand other moderation paradigms -- such as\nproactively discouraging the emergence of antisocial behavior rather than\nreacting to it -- and the role algorithmic support can play in these paradigms.\nIn this work, we investigate such a proactive framework for moderation in a\ncase study of a collaborative setting: Wikipedia Talk Pages. We employ a mixed\nmethods approach, combining qualitative and design components for a holistic\nanalysis. Through interviews with moderators, we find that despite a lack of\ntechnical and social support, moderators already engage in a number of\nproactive moderation behaviors, such as preemptively intervening in\nconversations to keep them on track. Further, we explore how automation could\nassist with this existing proactive moderation workflow by building a prototype\ntool, presenting it to moderators, and examining how the assistance it provides\nmight fit into their workflow. The resulting feedback uncovers both strengths\nand drawbacks of the prototype tool and suggests concrete steps towards further\ndeveloping such assisting technology so it can most effectively support\nmoderators in their existing proactive moderation workflow.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.17163", "title": "Misogyny classification of German newspaper forum comments", "abstract": "  This paper presents work on detecting misogyny in the comments of a large\nAustrian German language newspaper forum. We describe the creation of a corpus\nof 6600 comments which were annotated with 5 levels of misogyny. The forum\nmoderators were involved as experts in the creation of the annotation\nguidelines and the annotation of the comments. We also describe the results of\ntraining transformer-based classification models for both binarized and\noriginal label classification of that corpus.\n", "date": "2022-11-30", "categories": []}
{"id": "2211.14369", "title": "The Naughtyformer: A Transformer Understands Offensive Humor", "abstract": "  Jokes are intentionally written to be funny, but not all jokes are created\nthe same. Some jokes may be fit for a classroom of kindergarteners, but others\nare best reserved for a more mature audience. While recent work has shown\nimpressive results on humor detection in text, here we instead investigate the\nmore nuanced task of detecting humor subtypes, especially of the less innocent\nvariety. To that end, we introduce a novel jokes dataset filtered from Reddit\nand solve the subtype classification task using a finetuned Transformer dubbed\nthe Naughtyformer. Moreover, we show that our model is significantly better at\ndetecting offensiveness in jokes compared to state-of-the-art methods.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.15262", "title": "HERDPhobia: A Dataset for Hate Speech against Fulani in Nigeria", "abstract": "  Social media platforms allow users to freely share their opinions about\nissues or anything they feel like. However, they also make it easier to spread\nhate and abusive content. The Fulani ethnic group has been the victim of this\nunfortunate phenomenon. This paper introduces the HERDPhobia - the first\nannotated hate speech dataset on Fulani herders in Nigeria - in three\nlanguages: English, Nigerian-Pidgin, and Hausa. We present a benchmark\nexperiment using pre-trained languages models to classify the tweets as either\nhateful or non-hateful. Our experiment shows that the XML-T model provides\nbetter performance with 99.83% weighted F1. We released the dataset at\nhttps://github.com/hausanlp/HERDPhobia for further research.\n", "date": "2022-11-28", "categories": []}
{"id": "2211.15532", "title": "YZR-net : Self-supervised Hidden representations Invariant to\n  Transformations for profanity detection", "abstract": "  On current {\\it e-}learning platforms, live classes are an important tool\nthat provides students with an opportunity to get more involved while learning\nnew concepts. In such classes, the element of interaction with teachers and\nfellow peers helps in removing learning silos and gives each student a chance\nto experience some aspects relevant to offline learning in this era of virtual\nclasses. One common way of interaction in a class is through the chats /\nmessaging framework, where the teacher can broadcast messages as well as get\ninstant feedback from the students in the live class. This freedom of\ninteraction is a crucial aspect for any student's learning growth but misuse of\nit can have serious repercussions. Some miscreants use this framework to send\nprofane messages which can have a negative impact on other students as well as\nthe teacher of the class. These rare but high impact situations obviate the\nneed for automatic detection mechanisms that prevent the posting of such chats\non any platform. In this work we develop YZR-Net which is a self-supervised\nframework that is able to robustly detect profane words used in a chat even if\nthe student tries to add clever modifications to fool the system. The matching\nmechanism on token / word level allows us to maintain a compact as well as\ndynamic profane vocabulary which can be updated without retraining the\nunderlying model. Our profanity detection framework is language independent and\ncan handle abuses in both English as well as its transliterated counterpart\nHinglish (Hindi language words written in English).\n", "date": "2022-11-22", "categories": []}
{"id": "2211.12570", "title": "Predicting the Type and Target of Offensive Social Media Posts in\n  Marathi", "abstract": "  The presence of offensive language on social media is very common motivating\nplatforms to invest in strategies to make communities safer. This includes\ndeveloping robust machine learning systems capable of recognizing offensive\ncontent online. Apart from a few notable exceptions, most research on automatic\noffensive language identification has dealt with English and a few other high\nresource languages such as French, German, and Spanish. In this paper we\naddress this gap by tackling offensive language identification in Marathi, a\nlow-resource Indo-Aryan language spoken in India. We introduce the Marathi\nOffensive Language Dataset v.2.0 or MOLD 2.0 and present multiple experiments\non this dataset. MOLD 2.0 is a much larger version of MOLD with expanded\nannotation to the levels B (type) and C (target) of the popular OLID taxonomy.\nMOLD 2.0 is the first hierarchical offensive language dataset compiled for\nMarathi, thus opening new avenues for research in low-resource Indo-Aryan\nlanguages. Finally, we also introduce SeMOLD, a larger dataset annotated\nfollowing the semi-supervised methods presented in SOLID.\n", "date": "2022-11-22", "categories": []}
{"id": "2211.06116", "title": "How Much Hate with #china? A Preliminary Analysis on China-related\n  Hateful Tweets Two Years After the Covid Pandemic Began", "abstract": "  Following the outbreak of a global pandemic, online content is filled with\nhate speech. Donald Trump's ''Chinese Virus'' tweet shifted the blame for the\nspread of the Covid-19 virus to China and the Chinese people, which triggered a\nnew round of anti-China hate both online and offline. This research intends to\nexamine China-related hate speech on Twitter during the two years following the\nburst of the pandemic (2020 and 2021). Through Twitter's API, in total\n2,172,333 tweets hashtagged #china posted during the time were collected. By\nemploying multiple state-of-the-art pretrained language models for hate speech\ndetection, we identify a wide range of hate of various types, resulting in an\nautomatically labeled anti-China hate speech dataset. We identify a hateful\nrate in #china tweets of 2.5% in 2020 and 1.9% in 2021. This is well above the\naverage rate of online hate speech on Twitter at 0.6% identified in Gao et al.,\n2017. We further analyzed the longitudinal development of #china tweets and\nthose identified as hateful in 2020 and 2021 through visualizing the daily\nnumber and hate rate over the two years. Our keyword analysis of hate speech in\n#china tweets reveals the most frequently mentioned terms in the hateful #china\ntweets, which can be used for further social science studies.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.10163", "title": "Overview of the HASOC Subtrack at FIRE 2022: Offensive Language\n  Identification in Marathi", "abstract": "  The widespread of offensive content online has become a reason for great\nconcern in recent years, motivating researchers to develop robust systems\ncapable of identifying such content automatically. With the goal of carrying\nout a fair evaluation of these systems, several international competitions have\nbeen organized, providing the community with important benchmark data and\nevaluation methods for various languages. Organized since 2019, the HASOC (Hate\nSpeech and Offensive Content Identification) shared task is one of these\ninitiatives. In its fourth iteration, HASOC 2022 included three subtracks for\nEnglish, Hindi, and Marathi. In this paper, we report the results of the HASOC\n2022 Marathi subtrack which provided participants with a dataset containing\ndata from Twitter manually annotated using the popular OLID taxonomy. The\nMarathi track featured three additional subtracks, each corresponding to one\nlevel of the taxonomy: Task A - offensive content identification (offensive vs.\nnon-offensive); Task B - categorization of offensive types (targeted vs.\nuntargeted), and Task C - offensive target identification (individual vs. group\nvs. others). Overall, 59 runs were submitted by 10 teams. The best systems\nobtained an F1 of 0.9745 for Subtrack 3A, an F1 of 0.9207 for Subtrack 3B, and\nF1 of 0.9607 for Subtrack 3C. The best performing algorithms were a mixture of\ntraditional and deep learning approaches.\n", "date": "2022-11-18", "categories": []}
{"id": "2211.06452", "title": "Cross-Platform and Cross-Domain Abusive Language Detection with\n  Supervised Contrastive Learning", "abstract": "  The prevalence of abusive language on different online platforms has been a\nmajor concern that raises the need for automated cross-platform abusive\nlanguage detection. However, prior works focus on concatenating data from\nmultiple platforms, inherently adopting Empirical Risk Minimization (ERM)\nmethod. In this work, we address this challenge from the perspective of domain\ngeneralization objective. We design SCL-Fish, a supervised contrastive learning\nintegrated meta-learning algorithm to detect abusive language on unseen\nplatforms. Our experimental analysis shows that SCL-Fish achieves better\nperformance over ERM and the existing state-of-the-art models. We also show\nthat SCL-Fish is data-efficient and achieves comparable performance with the\nlarge-scale pre-trained models upon finetuning for the abusive language\ndetection task.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.06516", "title": "Bandits for Online Calibration: An Application to Content Moderation on\n  Social Media Platforms", "abstract": "  We describe the current content moderation strategy employed by Meta to\nremove policy-violating content from its platforms. Meta relies on both\nhandcrafted and learned risk models to flag potentially violating content for\nhuman review. Our approach aggregates these risk models into a single ranking\nscore, calibrating them to prioritize more reliable risk models. A key\nchallenge is that violation trends change over time, affecting which risk\nmodels are most reliable. Our system additionally handles production challenges\nsuch as changing risk models and novel risk models. We use a contextual bandit\nto update the calibration in response to such trends. Our approach increases\nMeta's top-line metric for measuring the effectiveness of its content\nmoderation strategy by 13%.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.06053", "title": "CoRAL: a Context-aware Croatian Abusive Language Dataset", "abstract": "  In light of unprecedented increases in the popularity of the internet and\nsocial media, comment moderation has never been a more relevant task.\nSemi-automated comment moderation systems greatly aid human moderators by\neither automatically classifying the examples or allowing the moderators to\nprioritize which comments to consider first. However, the concept of\ninappropriate content is often subjective, and such content can be conveyed in\nmany subtle and indirect ways. In this work, we propose CoRAL -- a language and\nculturally aware Croatian Abusive dataset covering phenomena of implicitness\nand reliance on local and global context. We show experimentally that current\nmodels degrade when comments are not explicit and further degrade when language\nskill and context knowledge are required to interpret the comment.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.00243", "title": "Why Is It Hate Speech? Masked Rationale Prediction for Explainable Hate\n  Speech Detection", "abstract": "  In a hate speech detection model, we should consider two critical aspects in\naddition to detection performance-bias and explainability. Hate speech cannot\nbe identified based solely on the presence of specific words: the model should\nbe able to reason like humans and be explainable. To improve the performance\nconcerning the two aspects, we propose Masked Rationale Prediction (MRP) as an\nintermediate task. MRP is a task to predict the masked human\nrationales-snippets of a sentence that are grounds for human judgment-by\nreferring to surrounding tokens combined with their unmasked rationales. As the\nmodel learns its reasoning ability based on rationales by MRP, it performs hate\nspeech detection robustly in terms of bias and explainability. The proposed\nmethod generally achieves state-of-the-art performance in various metrics,\ndemonstrating its effectiveness for hate speech detection.\n", "date": "2022-11-01", "categories": []}
{"id": "2210.16398", "title": "System Demo: Tool and Infrastructure for Offensive Language Error\n  Analysis (OLEA) in English", "abstract": "  The automatic detection of offensive language is a pressing societal need.\nMany systems perform well on explicit offensive language but struggle to detect\nmore complex, nuanced, or implicit cases of offensive and hateful language.\nOLEA is an open-source Python library that provides easy-to-use tools for error\nanalysis in the context of detecting offensive language in English. OLEA also\nprovides an infrastructure for re-distribution of new datasets and analysis\nmethods requiring very little coding.\n", "date": "2022-10-28", "categories": []}
{"id": "2210.15870", "title": "\"It's Not Just Hate'': A Multi-Dimensional Perspective on Detecting\n  Harmful Speech Online", "abstract": "  Well-annotated data is a prerequisite for good Natural Language Processing\nmodels. Too often, though, annotation decisions are governed by optimizing time\nor annotator agreement. We make a case for nuanced efforts in an\ninterdisciplinary setting for annotating offensive online speech. Detecting\noffensive content is rapidly becoming one of the most important real-world NLP\ntasks. However, most datasets use a single binary label, e.g., for hate or\nincivility, even though each concept is multi-faceted. This modeling choice\nseverely limits nuanced insights, but also performance. We show that a more\nfine-grained multi-label approach to predicting incivility and hateful or\nintolerant content addresses both conceptual and performance issues. We release\na novel dataset of over 40,000 tweets about immigration from the US and UK,\nannotated with six labels for different aspects of incivility and intolerance.\nOur dataset not only allows for a more nuanced understanding of harmful speech\nonline, models trained on it also outperform or match performance on benchmark\ndatasets.\n", "date": "2022-10-28", "categories": []}
{"id": "2210.11762", "title": "Detecting Unintended Social Bias in Toxic Language Datasets", "abstract": "  With the rise of online hate speech, automatic detection of Hate Speech,\nOffensive texts as a natural language processing task is getting popular.\nHowever, very little research has been done to detect unintended social bias\nfrom these toxic language datasets. This paper introduces a new dataset\nToxicBias curated from the existing dataset of Kaggle competition named \"Jigsaw\nUnintended Bias in Toxicity Classification\". We aim to detect social biases,\ntheir categories, and targeted groups. The dataset contains instances annotated\nfor five different bias categories, viz., gender, race/ethnicity, religion,\npolitical, and LGBTQ. We train transformer-based models using our curated\ndatasets and report baseline performance for bias identification, target\ngeneration, and bias implications. Model biases and their mitigation are also\ndiscussed in detail. Our study motivates a systematic extraction of social bias\ndata from toxic language datasets. All the codes and dataset used for\nexperiments in this work are publicly available\n", "date": "2022-10-21", "categories": []}
{"id": "2210.11359", "title": "Data-Efficient Strategies for Expanding Hate Speech Detection into\n  Under-Resourced Languages", "abstract": "  Hate speech is a global phenomenon, but most hate speech datasets so far\nfocus on English-language content. This hinders the development of more\neffective hate speech detection models in hundreds of languages spoken by\nbillions across the world. More data is needed, but annotating hateful content\nis expensive, time-consuming and potentially harmful to annotators. To mitigate\nthese issues, we explore data-efficient strategies for expanding hate speech\ndetection into under-resourced languages. In a series of experiments with mono-\nand multilingual models across five non-English languages, we find that 1) a\nsmall amount of target-language fine-tuning data is needed to achieve strong\nperformance, 2) the benefits of using more such data decrease exponentially,\nand 3) initial fine-tuning on readily-available English data can partially\nsubstitute target-language data and improve model generalisability. Based on\nthese findings, we formulate actionable recommendations for hate speech\ndetection in low-resource language settings.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.09340", "title": "Transferring Knowledge via Neighborhood-Aware Optimal Transport for\n  Low-Resource Hate Speech Detection", "abstract": "  The concerning rise of hateful content on online platforms has increased the\nattention towards automatic hate speech detection, commonly formulated as a\nsupervised classification task. State-of-the-art deep learning-based approaches\nusually require a substantial amount of labeled resources for training.\nHowever, annotating hate speech resources is expensive, time-consuming, and\noften harmful to the annotators. This creates a pressing need to transfer\nknowledge from the existing labeled resources to low-resource hate speech\ncorpora with the goal of improving system performance. For this,\nneighborhood-based frameworks have been shown to be effective. However, they\nhave limited flexibility. In our paper, we propose a novel training strategy\nthat allows flexible modeling of the relative proximity of neighbors retrieved\nfrom a resource-rich corpus to learn the amount of transfer. In particular, we\nincorporate neighborhood information with Optimal Transport, which permits\nexploiting the geometry of the data embedding space. By aligning the joint\nembedding and label distributions of neighbors, we demonstrate substantial\nimprovements over strong baselines, in low-resource scenarios, on different\npublicly available hate speech corpora.\n", "date": "2022-10-17", "categories": []}
{"id": "2210.05916", "title": "Hate-CLIPper: Multimodal Hateful Meme Classification based on\n  Cross-modal Interaction of CLIP Features", "abstract": "  Hateful memes are a growing menace on social media. While the image and its\ncorresponding text in a meme are related, they do not necessarily convey the\nsame meaning when viewed individually. Hence, detecting hateful memes requires\ncareful consideration of both visual and textual information. Multimodal\npre-training can be beneficial for this task because it effectively captures\nthe relationship between the image and the text by representing them in a\nsimilar feature space. Furthermore, it is essential to model the interactions\nbetween the image and text features through intermediate fusion. Most existing\nmethods either employ multimodal pre-training or intermediate fusion, but not\nboth. In this work, we propose the Hate-CLIPper architecture, which explicitly\nmodels the cross-modal interactions between the image and text representations\nobtained using Contrastive Language-Image Pre-training (CLIP) encoders via a\nfeature interaction matrix (FIM). A simple classifier based on the FIM\nrepresentation is able to achieve state-of-the-art performance on the Hateful\nMemes Challenge (HMC) dataset with an AUROC of 85.8, which even surpasses the\nhuman performance of 82.65. Experiments on other meme datasets such as\nPropaganda Memes and TamilMemes also demonstrate the generalizability of the\nproposed approach. Finally, we analyze the interpretability of the FIM\nrepresentation and show that cross-modal interactions can indeed facilitate the\nlearning of meaningful concepts. The code for this work is available at\nhttps://github.com/gokulkarthik/hateclipper.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.07934", "title": "Codes, Patterns and Shapes of Contemporary Online Antisemitism and\n  Conspiracy Narratives -- an Annotation Guide and Labeled German-Language\n  Dataset in the Context of COVID-19", "abstract": "  Over the course of the COVID-19 pandemic, existing conspiracy theories were\nrefreshed and new ones were created, often interwoven with antisemitic\nnarratives, stereotypes and codes. The sheer volume of antisemitic and\nconspiracy theory content on the Internet makes data-driven algorithmic\napproaches essential for anti-discrimination organizations and researchers\nalike. However, the manifestation and dissemination of these two interrelated\nphenomena is still quite under-researched in scholarly empirical research of\nlarge text corpora. Algorithmic approaches for the detection and classification\nof specific contents usually require labeled datasets, annotated based on\nconceptually sound guidelines. While there is a growing number of datasets for\nthe more general phenomenon of hate speech, the development of corpora and\nannotation guidelines for antisemitic and conspiracy content is still in its\ninfancy, especially for languages other than English. We contribute to closing\nthis gap by developing an annotation guide for antisemitic and conspiracy\ntheory online content in the context of the COVID-19 pandemic. We provide\nworking definitions, including specific forms of antisemitism such as encoded\nand post-Holocaust antisemitism. We use these to annotate a German-language\ndataset consisting of ~3,700 Telegram messages sent between 03/2020 and\n12/2021.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.06351", "title": "A Keyword Based Approach to Understanding the Overpenalization of\n  Marginalized Groups by English Marginal Abuse Models on Twitter", "abstract": "  Harmful content detection models tend to have higher false positive rates for\ncontent from marginalized groups. In the context of marginal abuse modeling on\nTwitter, such disproportionate penalization poses the risk of reduced\nvisibility, where marginalized communities lose the opportunity to voice their\nopinion on the platform. Current approaches to algorithmic harm mitigation, and\nbias detection for NLP models are often very ad hoc and subject to human bias.\nWe make two main contributions in this paper. First, we design a novel\nmethodology, which provides a principled approach to detecting and measuring\nthe severity of potential harms associated with a text-based model. Second, we\napply our methodology to audit Twitter's English marginal abuse model, which is\nused for removing amplification eligibility of marginally abusive content.\nWithout utilizing demographic labels or dialect classifiers, we are still able\nto detect and measure the severity of issues related to the over-penalization\nof the speech of marginalized communities, such as the use of reclaimed speech,\ncounterspeech, and identity related terms. In order to mitigate the associated\nharms, we experiment with adding additional true negative examples and find\nthat doing so provides improvements to our fairness metrics without large\ndegradations in model performance.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.05480", "title": "T5 for Hate Speech, Augmented Data and Ensemble", "abstract": "  We conduct relatively extensive investigations of automatic hate speech (HS)\ndetection using different state-of-the-art (SoTA) baselines over 11 subtasks of\n6 different datasets. Our motivation is to determine which of the recent SoTA\nmodels is best for automatic hate speech detection and what advantage methods\nlike data augmentation and ensemble may have on the best model, if any. We\ncarry out 6 cross-task investigations. We achieve new SoTA on two subtasks -\nmacro F1 scores of 91.73% and 53.21% for subtasks A and B of the HASOC 2020\ndataset, where previous SoTA are 51.52% and 26.52%, respectively. We achieve\nnear-SoTA on two others - macro F1 scores of 81.66% for subtask A of the OLID\n2019 dataset and 82.54% for subtask A of the HASOC 2021 dataset, where SoTA are\n82.9% and 83.05%, respectively. We perform error analysis and use two\nexplainable artificial intelligence (XAI) algorithms (IG and SHAP) to reveal\nhow two of the models (Bi-LSTM and T5) make the predictions they do by using\nexamples. Other contributions of this work are 1) the introduction of a simple,\nnovel mechanism for correcting out-of-class (OOC) predictions in T5, 2) a\ndetailed description of the data augmentation methods, 3) the revelation of the\npoor data annotations in the HASOC 2021 dataset by using several examples and\nXAI (buttressing the need for better quality control), and 4) the public\nrelease of our model checkpoints and codes to foster transparency.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.03479", "title": "Hate Speech and Offensive Language Detection in Bengali", "abstract": "  Social media often serves as a breeding ground for various hateful and\noffensive content. Identifying such content on social media is crucial due to\nits impact on the race, gender, or religion in an unprejudiced society.\nHowever, while there is extensive research in hate speech detection in English,\nthere is a gap in hateful content detection in low-resource languages like\nBengali. Besides, a current trend on social media is the use of Romanized\nBengali for regular interactions. To overcome the existing research's\nlimitations, in this study, we develop an annotated dataset of 10K Bengali\nposts consisting of 5K actual and 5K Romanized Bengali tweets. We implement\nseveral baseline models for the classification of such hateful posts. We\nfurther explore the interlingual transfer mechanism to boost classification\nperformance. Finally, we perform an in-depth error analysis by looking into the\nmisclassified posts by the models. While training actual and Romanized datasets\nseparately, we observe that XLM-Roberta performs the best. Further, we witness\nthat on joint training and few-shot training, MuRIL outperforms other models by\ninterpreting the semantic expressions better. We make our code and dataset\npublic for others.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.00910", "title": "Hypothesis Engineering for Zero-Shot Hate Speech Detection", "abstract": "  Standard approaches to hate speech detection rely on sufficient available\nhate speech annotations. Extending previous work that repurposes natural\nlanguage inference (NLI) models for zero-shot text classification, we propose a\nsimple approach that combines multiple hypotheses to improve English NLI-based\nzero-shot hate speech detection. We first conduct an error analysis for vanilla\nNLI-based zero-shot hate speech detection and then develop four strategies\nbased on this analysis. The strategies use multiple hypotheses to predict\nvarious aspects of an input text and combine these predictions into a final\nverdict. We find that the zero-shot baseline used for the initial error\nanalysis already outperforms commercial systems and fine-tuned BERT-based hate\nspeech detection models on HateCheck. The combination of the proposed\nstrategies further increases the zero-shot accuracy of 79.4% on HateCheck by\n7.9 percentage points (pp), and the accuracy of 69.6% on ETHOS by 10.0pp.\n", "date": "2022-10-03", "categories": []}
{"id": "2208.10684", "title": "K-MHaS: A Multi-label Hate Speech Detection Dataset in Korean Online\n  News Comment", "abstract": "  Online hate speech detection has become an important issue due to the growth\nof online content, but resources in languages other than English are extremely\nlimited. We introduce K-MHaS, a new multi-label dataset for hate speech\ndetection that effectively handles Korean language patterns. The dataset\nconsists of 109k utterances from news comments and provides a multi-label\nclassification using 1 to 4 labels, and handles subjectivity and\nintersectionality. We evaluate strong baseline experiments on K-MHaS using\nKorean-BERT-based language models with six different metrics. KR-BERT with a\nsub-character tokenizer outperforms others, recognizing decomposed characters\nin each hate speech class.\n", "date": "2022-08-23", "categories": []}
{"id": "2209.14091", "title": "Offensive Language Detection on Twitter", "abstract": "  Detection of offensive language in social media is one of the key challenges\nfor social media. Researchers have proposed many advanced methods to accomplish\nthis task. In this report, we try to use the learnings from their approach and\nincorporate our ideas to improve upon them. We have successfully achieved an\naccuracy of 74% in classifying offensive tweets. We also list upcoming\nchallenges in the abusive content detection in the social media world.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.12320", "title": "Can We Automate the Analysis of Online Child Sexual Exploitation\n  Discourse?", "abstract": "  Social media's growing popularity raises concerns around children's online\nsafety. Interactions between minors and adults with predatory intentions is a\nparticularly grave concern. Research into online sexual grooming has often\nrelied on domain experts to manually annotate conversations, limiting both\nscale and scope. In this work, we test how well-automated methods can detect\nconversational behaviors and replace an expert human annotator. Informed by\npsychological theories of online grooming, we label $6772$ chat messages sent\nby child-sex offenders with one of eleven predatory behaviors. We train\nbag-of-words and natural language inference models to classify each behavior,\nand show that the best performing models classify behaviors in a manner that is\nconsistent, but not on-par, with human annotation.\n", "date": "2022-09-25", "categories": []}
{"id": "2209.09975", "title": "Power of Explanations: Towards automatic debiasing in hate speech\n  detection", "abstract": "  Hate speech detection is a common downstream application of natural language\nprocessing (NLP) in the real world. In spite of the increasing accuracy,\ncurrent data-driven approaches could easily learn biases from the imbalanced\ndata distributions originating from humans. The deployment of biased models\ncould further enhance the existing social biases. But unlike handling tabular\ndata, defining and mitigating biases in text classifiers, which deal with\nunstructured data, are more challenging. A popular solution for improving\nmachine learning fairness in NLP is to conduct the debiasing process with a\nlist of potentially discriminated words given by human annotators. In addition\nto suffering from the risks of overlooking the biased terms, exhaustively\nidentifying bias with human annotators are unsustainable since discrimination\nis variable among different datasets and may evolve over time. To this end, we\npropose an automatic misuse detector (MiD) relying on an explanation method for\ndetecting potential bias. And built upon that, an end-to-end debiasing\nframework with the proposed staged correction is designed for text classifiers\nwithout any external resources required.\n", "date": "2022-09-07", "categories": []}
{"id": "2209.10193", "title": "Is More Data Better? Re-thinking the Importance of Efficiency in Abusive\n  Language Detection with Transformers-Based Active Learning", "abstract": "  Annotating abusive language is expensive, logistically complex and creates a\nrisk of psychological harm. However, most machine learning research has\nprioritized maximizing effectiveness (i.e., F1 or accuracy score) rather than\ndata efficiency (i.e., minimizing the amount of data that is annotated). In\nthis paper, we use simulated experiments over two datasets at varying\npercentages of abuse to demonstrate that transformers-based active learning is\na promising approach to substantially raise efficiency whilst still maintaining\nhigh effectiveness, especially when abusive content is a smaller percentage of\nthe dataset. This approach requires a fraction of labeled data to reach\nperformance equivalent to training over the full dataset.\n", "date": "2022-09-21", "categories": []}
{"id": "2209.08681", "title": "Domain Classification-based Source-specific Term Penalization for Domain\n  Adaptation in Hate-speech Detection", "abstract": "  State-of-the-art approaches for hate-speech detection usually exhibit poor\nperformance in out-of-domain settings. This occurs, typically, due to\nclassifiers overemphasizing source-specific information that negatively impacts\nits domain invariance. Prior work has attempted to penalize terms related to\nhate-speech from manually curated lists using feature attribution methods,\nwhich quantify the importance assigned to input terms by the classifier when\nmaking a prediction. We, instead, propose a domain adaptation approach that\nautomatically extracts and penalizes source-specific terms using a domain\nclassifier, which learns to differentiate between domains, and\nfeature-attribution scores for hate-speech classes, yielding consistent\nimprovements in cross-domain evaluation.\n", "date": "2022-09-18", "categories": []}
{"id": "2209.06505", "title": "BERT-based Ensemble Approaches for Hate Speech Detection", "abstract": "  With the freedom of communication provided in online social media, hate\nspeech has increasingly generated. This leads to cyber conflicts affecting\nsocial life at the individual and national levels. As a result, hateful content\nclassification is becoming increasingly demanded for filtering hate content\nbefore being sent to the social networks. This paper focuses on classifying\nhate speech in social media using multiple deep models that are implemented by\nintegrating recent transformer-based language models such as BERT, and neural\nnetworks. To improve the classification performances, we evaluated with several\nensemble techniques, including soft voting, maximum value, hard voting and\nstacking. We used three publicly available Twitter datasets (Davidson,\nHatEval2019, OLID) that are generated to identify offensive languages. We fused\nall these datasets to generate a single dataset (DHO dataset), which is more\nbalanced across different labels, to perform multi-label classification. Our\nexperiments have been held on Davidson dataset and the DHO corpora. The later\ngave the best overall results, especially F1 macro score, even it required more\nresources (time execution and memory). The experiments have shown good results\nespecially the ensemble models, where stacking gave F1 score of 97% on Davidson\ndataset and aggregating ensembles 77% on the DHO dataset.\n", "date": "2022-09-14", "categories": []}
{"id": "2209.07353", "title": "Measuring Geographic Performance Disparities of Offensive Language\n  Classifiers", "abstract": "  Text classifiers are applied at scale in the form of one-size-fits-all\nsolutions. Nevertheless, many studies show that classifiers are biased\nregarding different languages and dialects. When measuring and discovering\nthese biases, some gaps present themselves and should be addressed. First,\n``Does language, dialect, and topical content vary across geographical\nregions?'' and secondly ``If there are differences across the regions, do they\nimpact model performance?''. We introduce a novel dataset called GeoOLID with\nmore than 14 thousand examples across 15 geographically and demographically\ndiverse cities to address these questions. We perform a comprehensive analysis\nof geographical-related content and their impact on performance disparities of\noffensive language detection models. Overall, we find that current models do\nnot generalize across locations. Likewise, we show that while offensive\nlanguage models produce false positives on African American English, model\nperformance is not correlated with each city's minority population proportions.\nWarning: This paper contains offensive language.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.05294", "title": "A Review of Challenges in Machine Learning based Automated Hate Speech\n  Detection", "abstract": "  The spread of hate speech on social media space is currently a serious issue.\nThe undemanding access to the enormous amount of information being generated on\nthese platforms has led people to post and react with toxic content that\noriginates violence. Though efforts have been made toward detecting and\nrestraining such content online, it is still challenging to identify it\naccurately. Deep learning based solutions have been at the forefront of\nidentifying hateful content. However, the factors such as the context-dependent\nnature of hate speech, the intention of the user, undesired biases, etc. make\nthis process overcritical. In this work, we deeply explore a wide range of\nchallenges in automatic hate speech detection by presenting a hierarchical\norganization of these problems. We focus on challenges faced by machine\nlearning or deep learning based solutions to hate speech identification. At the\ntop level, we distinguish between data level, model level, and human level\nchallenges. We further provide an exhaustive analysis of each level of the\nhierarchy with examples. This survey will help researchers to design their\nsolutions more efficiently in the domain of hate speech detection.\n", "date": "2022-09-12", "categories": []}
{"id": "2209.03711", "title": "What Did I Just Hear? Detecting Pornographic Sounds in Adult Videos\n  Using Neural Networks", "abstract": "  Audio-based pornographic detection enables efficient adult content filtering\nwithout sacrificing performance by exploiting distinct spectral\ncharacteristics. To improve it, we explore pornographic sound modeling based on\ndifferent neural architectures and acoustic features. We find that CNN trained\non log mel spectrogram achieves the best performance on Pornography-800\ndataset. Our experiment results also show that log mel spectrogram allows\nbetter representations for the models to recognize pornographic sounds.\nFinally, to classify whole audio waveforms rather than segments, we employ\nvoting segment-to-audio technique that yields the best audio-level detection\nresults.\n", "date": "2022-09-08", "categories": []}
{"id": "2208.13595", "title": "Combating high variance in Data-Scarce Implicit Hate Speech\n  Classification", "abstract": "  Hate speech classification has been a long-standing problem in natural\nlanguage processing. However, even though there are numerous hate speech\ndetection methods, they usually overlook a lot of hateful statements due to\nthem being implicit in nature. Developing datasets to aid in the task of\nimplicit hate speech classification comes with its own challenges; difficulties\nare nuances in language, varying definitions of what constitutes hate speech,\nand the labor-intensive process of annotating such data. This had led to a\nscarcity of data available to train and test such systems, which gives rise to\nhigh variance problems when parameter-heavy transformer-based models are used\nto address the problem. In this paper, we explore various optimization and\nregularization techniques and develop a novel RoBERTa-based model that achieves\nstate-of-the-art performance.\n", "date": "2022-08-29", "categories": []}
{"id": "2206.02114", "title": "Speech Detection Task Against Asian Hate: BERT the Central, While\n  Data-Centric Studies the Crucial", "abstract": "  With the COVID-19 pandemic continuing, hatred against Asians is intensifying\nin countries outside Asia, especially among the Chinese. There is an urgent\nneed to detect and prevent hate speech towards Asians effectively. In this\nwork, we first create COVID-HATE-2022, an annotated dataset including 2,025\nannotated tweets fetched in early February 2022, which are labeled based on\nspecific criteria, and we present the comprehensive collection of scenarios of\nhate and non-hate tweets in the dataset. Second, we fine-tune the BERT model\nbased on the relevant datasets and demonstrate several strategies related to\nthe \"cleaning\" of the tweets. Third, we investigate the performance of advanced\nfine-tuning strategies with various model-centric and data-centric approaches,\nand we show that both strategies generally improve the performance, while\ndata-centric ones outperform the others, and it demonstrates the feasibility\nand effectiveness of the data-centric approaches in the associated tasks.\n", "date": "2022-06-05", "categories": []}
{"id": "2208.04489", "title": "Exploring Hate Speech Detection with HateXplain and BERT", "abstract": "  Hate Speech takes many forms to target communities with derogatory comments,\nand takes humanity a step back in societal progress. HateXplain is a recently\npublished and first dataset to use annotated spans in the form of rationales,\nalong with speech classification categories and targeted communities to make\nthe classification more humanlike, explainable, accurate and less biased. We\ntune BERT to perform this task in the form of rationales and class prediction,\nand compare our performance on different metrics spanning across accuracy,\nexplainability and bias. Our novelty is threefold. Firstly, we experiment with\nthe amalgamated rationale class loss with different importance values.\nSecondly, we experiment extensively with the ground truth attention values for\nthe rationales. With the introduction of conservative and lenient attentions,\nwe compare performance of the model on HateXplain and test our hypothesis.\nThirdly, in order to improve the unintended bias in our models, we use masking\nof the target community words and note the improvement in bias and\nexplainability metrics. Overall, we are successful in achieving model\nexplanability, bias removal and several incremental improvements on the\noriginal BERT implementation.\n", "date": "2022-08-09", "categories": []}
{"id": "2208.01099", "title": "Parsimonious Argument Annotations for Hate Speech Counter-narratives", "abstract": "  We present an enrichment of the Hateval corpus of hate speech tweets (Basile\net. al 2019) aimed to facilitate automated counter-narrative generation.\nComparably to previous work (Chung et. al. 2019), manually written\ncounter-narratives are associated to tweets. However, this information alone\nseems insufficient to obtain satisfactory language models for counter-narrative\ngeneration. That is why we have also annotated tweets with argumentative\ninformation based on Wagemanns (2016), that we believe can help in building\nconvincing and effective counter-narratives for hate speech against particular\ngroups.\n  We discuss adequacies and difficulties of this annotation process and present\nseveral baselines for automatic detection of the annotated elements.\nPreliminary results show that automatic annotators perform close to human\nannotators to detect some aspects of argumentation, while others only reach low\nor moderate level of inter-annotator agreement.\n", "date": "2022-08-01", "categories": []}
{"id": "2208.01312", "title": "BEIKE NLP at SemEval-2022 Task 4: Prompt-Based Paragraph Classification\n  for Patronizing and Condescending Language Detection", "abstract": "  PCL detection task is aimed at identifying and categorizing language that is\npatronizing or condescending towards vulnerable communities in the general\nmedia.Compared to other NLP tasks of paragraph classification, the negative\nlanguage presented in the PCL detection task is usually more implicit and\nsubtle to be recognized, making the performance of common text-classification\napproaches disappointed. Targeting the PCL detection problem in SemEval-2022\nTask 4, in this paper, we give an introduction to our team's solution, which\nexploits the power of prompt-based learning on paragraph classification. We\nreformulate the task as an appropriate cloze prompt and use pre-trained Masked\nLanguage Models to fill the cloze slot. For the two subtasks, binary\nclassification and multi-label classification, DeBERTa model is adopted and\nfine-tuned to predict masked label words of task-specific prompts. On the\nevaluation dataset, for binary classification, our approach achieves an\nF1-score of 0.6406; for multi-label classification, our approach achieves an\nmacro-F1-score of 0.4689 and ranks first in the leaderboard.\n", "date": "2022-08-02", "categories": []}
{"id": "2207.11897", "title": "AI Powered Anti-Cyber Bullying System using Machine Learning Algorithm\n  of Multinomial Naive Bayes and Optimized Linear Support Vector Machine", "abstract": "  \"Unless and until our society recognizes cyber bullying for what it is, the\nsuffering of thousands of silent victims will continue.\" ~ Anna Maria Chavez.\nThere had been series of research on cyber bullying which are unable to provide\nreliable solution to cyber bullying. In this research work, we were able to\nprovide a permanent solution to this by developing a model capable of detecting\nand intercepting bullying incoming and outgoing messages with 92% accuracy. We\nalso developed a chatbot automation messaging system to test our model leading\nto the development of Artificial Intelligence powered anti-cyber bullying\nsystem using machine learning algorithm of Multinomial Naive Bayes (MNB) and\noptimized linear Support Vector Machine (SVM). Our model is able to detect and\nintercept bullying outgoing and incoming bullying messages and take immediate\naction.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.10639", "title": "Session-based Cyberbullying Detection in Social Media: A Survey", "abstract": "  Cyberbullying is a pervasive problem in online social media, where a bully\nabuses a victim through a social media session. By investigating cyberbullying\nperpetrated through social media sessions, recent research has looked into\nmining patterns and features for modeling and understanding the two defining\ncharacteristics of cyberbullying: repetitive behavior and power imbalance. In\nthis survey paper, we define the Session-based Cyberbullying Detection\nframework that encapsulates the different steps and challenges of the problem.\nBased on this framework, we provide a comprehensive overview of session-based\ncyberbullying detection in social media, delving into existing efforts from a\ndata and methodological perspective. Our review leads us to propose\nevidence-based criteria for a set of best practices to create session-based\ncyberbullying datasets. In addition, we perform benchmark experiments comparing\nthe performance of state-of-the-art session-based cyberbullying detection\nmodels as well as large pre-trained language models across two different\ndatasets. Through our review, we also put forth a set of open challenges as\nfuture research directions.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.10032", "title": "Detecting Harmful Online Conversational Content towards LGBTQIA+\n  Individuals", "abstract": "  Online discussions, panels, talk page edits, etc., often contain harmful\nconversational content i.e., hate speech, death threats and offensive language,\nespecially towards certain demographic groups. For example, individuals who\nidentify as members of the LGBTQIA+ community and/or BIPOC (Black, Indigenous,\nPeople of Color) are at higher risk for abuse and harassment online. In this\nwork, we first introduce a real-world dataset that will enable us to study and\nunderstand harmful online conversational content. Then, we conduct several\nexploratory data analysis experiments to gain deeper insights from the dataset.\nWe later describe our approach for detecting harmful online Anti-LGBTQIA+\nconversational content, and finally, we implement two baseline machine learning\nmodels (i.e., Support Vector Machine and Logistic Regression), and fine-tune 3\npre-trained large language models (BERT, RoBERTa, and HateBERT). Our findings\nverify that large language models can achieve very promising performance on\ndetecting online Anti-LGBTQIA+ conversational content detection tasks.\n", "date": "2022-06-15", "categories": []}
{"id": "2207.08557", "title": "AlexU-AIC at Arabic Hate Speech 2022: Contrast to Classify", "abstract": "  Online presence on social media platforms such as Facebook and Twitter has\nbecome a daily habit for internet users. Despite the vast amount of services\nthe platforms offer for their users, users suffer from cyber-bullying, which\nfurther leads to mental abuse and may escalate to cause physical harm to\nindividuals or targeted groups. In this paper, we present our submission to the\nArabic Hate Speech 2022 Shared Task Workshop (OSACT5 2022) using the associated\nArabic Twitter dataset. The shared task consists of 3 sub-tasks, sub-task A\nfocuses on detecting whether the tweet is offensive or not. Then, For offensive\nTweets, sub-task B focuses on detecting whether the tweet is hate speech or\nnot. Finally, For hate speech Tweets, sub-task C focuses on detecting the\nfine-grained type of hate speech among six different classes. Transformer\nmodels proved their efficiency in classification tasks, but with the problem of\nover-fitting when fine-tuned on a small or an imbalanced dataset. We overcome\nthis limitation by investigating multiple training paradigms such as\nContrastive learning and Multi-task learning along with Classification\nfine-tuning and an ensemble of our top 5 performers. Our proposed solution\nachieved 0.841, 0.817, and 0.476 macro F1-average in sub-tasks A, B, and C\nrespectively.\n", "date": "2022-07-18", "categories": []}
{"id": "2207.06710", "title": "Overview of Abusive and Threatening Language Detection in Urdu at FIRE\n  2021", "abstract": "  With the growth of social media platform influence, the effect of their\nmisuse becomes more and more impactful. The importance of automatic detection\nof threatening and abusive language can not be overestimated. However, most of\nthe existing studies and state-of-the-art methods focus on English as the\ntarget language, with limited work on low- and medium-resource languages. In\nthis paper, we present two shared tasks of abusive and threatening language\ndetection for the Urdu language which has more than 170 million speakers\nworldwide. Both are posed as binary classification tasks where participating\nsystems are required to classify tweets in Urdu into two classes, namely: (i)\nAbusive and Non-Abusive for the first task, and (ii) Threatening and\nNon-Threatening for the second. We present two manually annotated datasets\ncontaining tweets labelled as (i) Abusive and Non-Abusive, and (ii) Threatening\nand Non-Threatening. The abusive dataset contains 2400 annotated tweets in the\ntrain part and 1100 annotated tweets in the test part. The threatening dataset\ncontains 6000 annotated tweets in the train part and 3950 annotated tweets in\nthe test part. We also provide logistic regression and BERT-based baseline\nclassifiers for both tasks. In this shared task, 21 teams from six countries\nregistered for participation (India, Pakistan, China, Malaysia, United Arab\nEmirates, and Taiwan), 10 teams submitted their runs for Subtask A, which is\nAbusive Language Detection and 9 teams submitted their runs for Subtask B,\nwhich is Threatening Language detection, and seven teams submitted their\ntechnical reports. The best performing system achieved an F1-score value of\n0.880 for Subtask A and 0.545 for Subtask B. For both subtasks, m-Bert based\ntransformer model showed the best performance.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.04003", "title": "No Time Like the Present: Effects of Language Change on Automated\n  Comment Moderation", "abstract": "  The spread of online hate has become a significant problem for newspapers\nthat host comment sections. As a result, there is growing interest in using\nmachine learning and natural language processing for (semi-) automated abusive\nlanguage detection to avoid manual comment moderation costs or having to shut\ndown comment sections altogether. However, much of the past work on abusive\nlanguage detection assumes that classifiers operate in a static language\nenvironment, despite language and news being in a state of constant flux. In\nthis paper, we show using a new German newspaper comments dataset that the\nclassifiers trained with naive ML techniques like a random-test train split\nwill underperform on future data, and that a time stratified evaluation split\nis more appropriate. We also show that classifier performance rapidly degrades\nwhen evaluated on data from a different period than the training data. Our\nfindings suggest that it is necessary to consider the temporal dynamics of\nlanguage when developing an abusive language detection system or risk deploying\na model that will quickly become defunct.\n", "date": "2022-07-08", "categories": []}
{"id": "2206.15455", "title": "Hate Speech Criteria: A Modular Approach to Task-Specific Hate Speech\n  Definitions", "abstract": "  \\textbf{Offensive Content Warning}: This paper contains offensive language\nonly for providing examples that clarify this research and do not reflect the\nauthors' opinions. Please be aware that these examples are offensive and may\ncause you distress.\n  The subjectivity of recognizing \\textit{hate speech} makes it a complex task.\nThis is also reflected by different and incomplete definitions in NLP. We\npresent \\textit{hate speech} criteria, developed with perspectives from law and\nsocial science, with the aim of helping researchers create more precise\ndefinitions and annotation guidelines on five aspects: (1) target groups, (2)\ndominance, (3) perpetrator characteristics, (4) type of negative group\nreference, and the (5) type of potential consequences/effects. Definitions can\nbe structured so that they cover a more broad or more narrow phenomenon. As\nsuch, conscious choices can be made on specifying criteria or leaving them\nopen. We argue that the goal and exact task developers have in mind should\ndetermine how the scope of \\textit{hate speech} is defined. We provide an\noverview of the properties of English datasets from \\url{hatespeechdata.com}\nthat may help select the most suitable dataset for a specific scenario.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.09917", "title": "Multilingual HateCheck: Functional Tests for Multilingual Hate Speech\n  Detection Models", "abstract": "  Hate speech detection models are typically evaluated on held-out test sets.\nHowever, this risks painting an incomplete and potentially misleading picture\nof model performance because of increasingly well-documented systematic gaps\nand biases in hate speech datasets. To enable more targeted diagnostic\ninsights, recent research has thus introduced functional tests for hate speech\ndetection models. However, these tests currently only exist for\nEnglish-language content, which means that they cannot support the development\nof more effective models in other languages spoken by billions across the\nworld. To help address this issue, we introduce Multilingual HateCheck (MHC), a\nsuite of functional tests for multilingual hate speech detection models. MHC\ncovers 34 functionalities across ten languages, which is more languages than\nany other hate speech dataset. To illustrate MHC's utility, we train and test a\nhigh-performing multilingual hate speech detection model, and reveal critical\nmodel weaknesses for monolingual and cross-lingual applications.\n", "date": "2022-06-20", "categories": []}
{"id": "2206.10280", "title": "muBoost: An Effective Method for Solving Indic Multilingual Text\n  Classification Problem", "abstract": "  Text Classification is an integral part of many Natural Language Processing\ntasks such as sarcasm detection, sentiment analysis and many more such\napplications. Many e-commerce websites, social-media/entertainment platforms\nuse such models to enhance user-experience to generate traffic and thus,\nrevenue on their platforms. In this paper, we are presenting our solution to\nMultilingual Abusive Comment Identification Problem on Moj, an Indian\nvideo-sharing social networking service, powered by ShareChat. The problem\ndealt with detecting abusive comments, in 13 regional Indic languages such as\nHindi, Telugu, Kannada etc., on the videos on Moj platform. Our solution\nutilizes the novel muBoost, an ensemble of CatBoost classifier models and\nMultilingual Representations for Indian Languages (MURIL) model, to produce\nSOTA performance on Indic text classification tasks. We were able to achieve a\nmean F1-score of 89.286 on the test data, an improvement over baseline MURIL\nmodel with a F1-score of 87.48.\n", "date": "2022-06-21", "categories": []}
{"id": "2206.08407", "title": "Deep Multi-Task Models for Misogyny Identification and Categorization on\n  Arabic Social Media", "abstract": "  The prevalence of toxic content on social media platforms, such as hate\nspeech, offensive language, and misogyny, presents serious challenges to our\ninterconnected society. These challenging issues have attracted widespread\nattention in Natural Language Processing (NLP) community. In this paper, we\npresent the submitted systems to the first Arabic Misogyny Identification\nshared task. We investigate three multi-task learning models as well as their\nsingle-task counterparts. In order to encode the input text, our models rely on\nthe pre-trained MARBERT language model. The overall obtained results show that\nall our submitted models have achieved the best performances (top three ranked\nsubmissions) in both misogyny identification and categorization tasks.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.08445", "title": "Enriching Abusive Language Detection with Community Context", "abstract": "  Uses of pejorative expressions can be benign or actively empowering. When\nmodels for abuse detection misclassify these expressions as derogatory, they\ninadvertently censor productive conversations held by marginalized groups. One\nway to engage with non-dominant perspectives is to add context around\nconversations. Previous research has leveraged user- and thread-level features,\nbut it often neglects the spaces within which productive conversations take\nplace. Our paper highlights how community context can improve classification\noutcomes in abusive language detection. We make two main contributions to this\nend. First, we demonstrate that online communities cluster by the nature of\ntheir support towards victims of abuse. Second, we establish how community\ncontext improves accuracy and reduces the false positive rates of\nstate-of-the-art abusive language classifiers. These findings suggest a\npromising direction for context-aware models in abusive language research.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.06423", "title": "Hate Speech and Counter Speech Detection: Conversational Context Does\n  Matter", "abstract": "  Hate speech is plaguing the cyberspace along with user-generated content.\nThis paper investigates the role of conversational context in the annotation\nand detection of online hate and counter speech, where context is defined as\nthe preceding comment in a conversation thread. We created a context-aware\ndataset for a 3-way classification task on Reddit comments: hate speech,\ncounter speech, or neutral. Our analyses indicate that context is critical to\nidentify hate and counter speech: human judgments change for most comments\ndepending on whether we show annotators the context. A linguistic analysis\ndraws insights into the language people use to express hate and counter speech.\nExperimental results show that neural networks obtain significantly better\nresults if context is taken into account. We also present qualitative error\nanalyses shedding light into (a) when and why context is beneficial and (b) the\nremaining errors made by our best model when context is taken into account.\n", "date": "2022-06-13", "categories": []}
{"id": "2206.04162", "title": "Improved two-stage hate speech classification for twitter based on Deep\n  Neural Networks", "abstract": "  Hate speech is a form of online harassment that involves the use of abusive\nlanguage, and it is commonly seen in social media posts. This sort of\nharassment mainly focuses on specific group characteristics such as religion,\ngender, ethnicity, etc and it has both societal and economic consequences\nnowadays. The automatic detection of abusive language in text postings has\nalways been a difficult task, but it is lately receiving much interest from the\nscientific community. This paper addresses the important problem of discerning\nhateful content in social media. The model we propose in this work is an\nextension of an existing approach based on LSTM neural network architectures,\nwhich we appropriately enhanced and fine-tuned to detect certain forms of\nhatred language, such as racism or sexism, in a short text. The most\nsignificant enhancement is the conversion to a two-stage scheme consisting of\nRecurrent Neural Network (RNN) classifiers. The output of all One-vs-Rest (OvR)\nclassifiers from the first stage are combined and used to train the second\nstage classifier, which finally determines the type of harassment. Our study\nincludes a performance comparison of several proposed alternative methods for\nthe second stage evaluated on a public corpus of 16k tweets, followed by a\ngeneralization study on another dataset. The reported results show the superior\nclassification quality of the proposed scheme in the task of hate speech\ndetection as compared to the current state-of-the-art.\n", "date": "2022-06-08", "categories": []}
{"id": "2206.04007", "title": "Proactively Reducing the Hate Intensity of Online Posts via Hate Speech\n  Normalization", "abstract": "  Curbing online hate speech has become the need of the hour; however, a\nblanket ban on such activities is infeasible for several geopolitical and\ncultural reasons. To reduce the severity of the problem, in this paper, we\nintroduce a novel task, hate speech normalization, that aims to weaken the\nintensity of hatred exhibited by an online post. The intention of hate speech\nnormalization is not to support hate but instead to provide the users with a\nstepping stone towards non-hate while giving online platforms more time to\nmonitor any improvement in the user's behavior.\n  To this end, we manually curated a parallel corpus - hate texts and their\nnormalized counterparts (a normalized text is less hateful and more benign). We\nintroduce NACL, a simple yet efficient hate speech normalization model that\noperates in three stages - first, it measures the hate intensity of the\noriginal sample; second, it identifies the hate span(s) within it; and finally,\nit reduces hate intensity by paraphrasing the hate spans. We perform extensive\nexperiments to measure the efficacy of NACL via three-way evaluation\n(intrinsic, extrinsic, and human-study). We observe that NACL outperforms six\nbaselines - NACL yields a score of 0.1365 RMSE for the intensity prediction,\n0.622 F1-score in the span identification, and 82.27 BLEU and 80.05 perplexity\nfor the normalized text generation. We further show the generalizability of\nNACL across other platforms (Reddit, Facebook, Gab). An interactive prototype\nof NACL was put together for the user study. Further, the tool is being\ndeployed in a real-world setting at Wipro AI as a part of its mission to tackle\nharmful content on online platforms.\n", "date": "2022-06-08", "categories": []}
{"id": "2206.01889", "title": "Initial Study into Application of Feature Density and\n  Linguistically-backed Embedding to Improve Machine Learning-based\n  Cyberbullying Detection", "abstract": "  In this research, we study the change in the performance of machine learning\n(ML) classifiers when various linguistic preprocessing methods of a dataset\nwere used, with the specific focus on linguistically-backed embeddings in\nConvolutional Neural Networks (CNN). Moreover, we study the concept of Feature\nDensity and confirm its potential to comparatively predict the performance of\nML classifiers, including CNN. The research was conducted on a Formspring\ndataset provided in a Kaggle competition on automatic cyberbullying detection.\nThe dataset was re-annotated by objective experts (psychologists), as the\nimportance of professional annotation in cyberbullying research has been\nindicated multiple times. The study confirmed the effectiveness of Neural\nNetworks in cyberbullying detection and the correlation between classifier\nperformance and Feature Density while also proposing a new approach of training\nvarious linguistically-backed embeddings for Convolutional Neural Networks.\n", "date": "2022-06-04", "categories": []}
{"id": "2206.02206", "title": "Performance Comparison of Simple Transformer and Res-CNN-BiLSTM for\n  Cyberbullying Classification", "abstract": "  The task of text classification using Bidirectional based LSTM architectures\nis computationally expensive and time consuming to train. For this,\ntransformers were discovered which effectively give good performance as\ncompared to the traditional deep learning architectures. In this paper we\npresent a performance based comparison between simple transformer based network\nand Res-CNN-BiLSTM based network for cyberbullying text classification problem.\nThe results obtained show that transformer we trained with 0.65 million\nparameters has significantly being able to beat the performance of\nRes-CNN-BiLSTM with 48.82 million parameters for faster training speeds and\nmore generalized metrics. The paper also compares the 1-dimensional character\nlevel embedding network and 100-dimensional glove embedding network with\ntransformer.\n", "date": "2022-06-05", "categories": []}
{"id": "2206.00372", "title": "BD-SHS: A Benchmark Dataset for Learning to Detect Online Bangla Hate\n  Speech in Different Social Contexts", "abstract": "  Social media platforms and online streaming services have spawned a new breed\nof Hate Speech (HS). Due to the massive amount of user-generated content on\nthese sites, modern machine learning techniques are found to be feasible and\ncost-effective to tackle this problem. However, linguistically diverse datasets\ncovering different social contexts in which offensive language is typically\nused are required to train generalizable models. In this paper, we identify the\nshortcomings of existing Bangla HS datasets and introduce a large manually\nlabeled dataset BD-SHS that includes HS in different social contexts. The\nlabeling criteria were prepared following a hierarchical annotation process,\nwhich is the first of its kind in Bangla HS to the best of our knowledge. The\ndataset includes more than 50,200 offensive comments crawled from online social\nnetworking sites and is at least 60% larger than any existing Bangla HS\ndatasets. We present the benchmark result of our dataset by training different\nNLP models resulting in the best one achieving an F1-score of 91.0%. In our\nexperiments, we found that a word embedding trained exclusively using 1.47\nmillion comments from social media and streaming sites consistently resulted in\nbetter modeling of HS detection in comparison to other pre-trained embeddings.\nOur dataset and all accompanying codes is publicly available at\ngithub.com/naurosromim/hate-speech-dataset-for-Bengali-social-media\n", "date": "2022-06-01", "categories": []}
{"id": "2206.00524", "title": "Vietnamese Hate and Offensive Detection using PhoBERT-CNN and Social\n  Media Streaming Data", "abstract": "  Society needs to develop a system to detect hate and offense to build a\nhealthy and safe environment. However, current research in this field still\nfaces four major shortcomings, including deficient pre-processing techniques,\nindifference to data imbalance issues, modest performance models, and lacking\npractical applications. This paper focused on developing an intelligent system\ncapable of addressing these shortcomings. Firstly, we proposed an efficient\npre-processing technique to clean comments collected from Vietnamese social\nmedia. Secondly, a novel hate speech detection (HSD) model, which is the\ncombination of a pre-trained PhoBERT model and a Text-CNN model, was proposed\nfor solving tasks in Vietnamese. Thirdly, EDA techniques are applied to deal\nwith imbalanced data to improve the performance of classification models.\nBesides, various experiments were conducted as baselines to compare and\ninvestigate the proposed model's performance against state-of-the-art methods.\nThe experiment results show that the proposed PhoBERT-CNN model outperforms\nSOTA methods and achieves an F1-score of 67,46% and 98,45% on two benchmark\ndatasets, ViHSD and HSD-VLSP, respectively. Finally, we also built a streaming\nHSD application to demonstrate the practicality of our proposed system.\n", "date": "2022-06-01", "categories": []}
{"id": "2606.27314", "title": "Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection", "abstract": "To avoid moderation and surveillance on social media, some users routinely invent indirect linguistic expressions (ILE) that camouflage sensitive meanings. Such expressions surface as algospeak, euphemisms, and adversarial obfuscation, depending on intent and context, and they involve recurring encoding mechanisms. We propose a comprehensive, mechanism-oriented taxonomy of ILE that abstracts away from communicative goals and instead categorizes the underlying operations through which meaning is encoded and recovered. We evaluate the taxonomy by incorporating it into LLM prompts and comparing it with four existing taxonomies and a no-taxonomy baseline, using 2,000 manually annotated TikTok and Bluesky posts. The proposed taxonomy attains the strongest document- and span-level performance across the three LLMs, achieving an improvement of 4.7% in accuracy and 5.4% in F1 over the best-performing benchmark. The empirical results reveal the importance of a comprehensive, mechanism-oriented taxonomy as a stable scaffold for detecting emerging coded language and a useful input to content moderation. Disclaimer: This paper contains content that may be profane, vulgar, or offensive.", "date": "2026-06-25", "categories": []}
{"id": "2606.27629", "title": "Cross-Platform Chinese Offensive Comment Detection via Dual-Threshold Hard Example Mining", "abstract": "Cross-platform deployment of offensive comment detection for Chinese social media suffers performance degradation. The paper proposes a dual-threshold hard mining method to address this. First, the clean-Chinese-base RoBERTa is finetuned on COLD to establish a binary baseline for fair comparison. Second, a three-class fine-labeled test set covering Weibo, Xiaohongshu, Tieba, and Zhihu is constructed, domain distances from the source are quantified using Jaccard and Proxy-A Distance, as well as the degradation bottleneck of the baseline under domain shift is systematically revealed. Herein, a dual threshold hard example mining strategy is proposed. High- and low-confidence error-prone samples are filtered from unlabeled corpora by prediction confidence. The model is secondarily finetuned under implicit contexts with merely a small set of manually labeled hard examples, realizing low-cost cross-platform domain adaptation. Experiments reveal significant performance gains of the optimized model across four platforms.", "date": "2026-06-26", "categories": []}
{"id": "2606.27981", "title": "ToxiREX: A Dataset on Toxic REasoning in ConteXt", "abstract": "We introduce a new, contextual, multilingual dataset called ToxiREX: Toxic REasoning in ConteXt. The dataset consists of threads of Reddit comments and structured characterizations of what the comments imply, following a systematic toxic reasoning schema developed in a previous paper. Using the schema allows us to capture and explain implicit and context-dependent toxicity, while supporting mappings to existing toxicity taxonomies. The dataset includes comments in six languages (English, Arabic, Turkish, Spanish, German, and Dutch), collected from posts connected to specific major events (e.g. the 2023 Turkey earthquakes; the Russian invasion of Ukraine). We describe the context-preserving preprocessing of the threads. We create a training set of 125 thousand comments which is annotated by a commercially available LLM, and a test set of just under three thousand comments that is annotated by native speakers. We show that apparent disagreements in the test set annotations often reflect defensible alternative interpretations rather than noise. Finally, we provide baseline results by prompting and fine-tuning language models. To produce these results, we develop evaluation strategies for our hierarchical, schema-based predictions. While models perform better than random, there remains a lot of room for improvement, showing the task to be challenging. ToxiREX is the first dataset to simultaneously incorporate multiple languages, conversational context, and implicit toxicity, while using the toxic reasoning schema for rich, structured annotations. Dataset available at: https://github.com/cltl/toxirex", "date": "2026-06-26", "categories": []}
{"id": "2606.28725", "title": "DriftGuard: Safety-Aware Multi-Monitor Detection and Selective Adaptation for Evolving Toxicity Moderation", "abstract": "Automated toxicity moderation systems operate in dynamic online environments where harmful behavior evolves through coded language, shifting targets, and strategic adaptation to enforcement. Existing drift detection methods often focus on global distributional change, but such signals may miss safety-relevant shifts that emerge in localized harm subspaces or high-risk model-error regions. This paper introduces DriftGuard, a safety-aware adaptive moderation framework that combines multi-monitor drift detection with selective model updating. The framework tracks global text drift, identity-harm drift, model uncertainty, toxic-risk drift, and false-negative-risk drift. When safety-relevant change is detected, the model is updated using a hard-mix adaptation set that prioritizes likely false negatives, identity-related high-risk examples, false-positive-risk examples, and uncertain boundary cases. Experiments on Civil Comments temporal shift and Jigsaw-to-DynaHate cross-dataset shift show that safety-aware monitors detect risks missed by global drift alone. Hard-mix adaptation improves toxic recall and accuracy over no-update and random-balanced baselines, raising toxic recall to 0.8777 on Civil Comments and from 0.7107 to 0.8523 on DynaHate. Bootstrap analysis further shows stable DynaHate safety gains, with toxic recall increasing by 0.1418 and false-negative prevalence decreasing by 0.0781. Overall, DriftGuard links safety-aware drift detection to targeted, lightweight model updating for more robust adaptive toxicity moderation.", "date": "2026-06-27", "categories": []}
{"id": "2606.28772", "title": "Majority Vote Silences Minority Values: Annotator Disagreement at the Hate/Offensive Boundary in HateXplain", "abstract": "Hate speech annotation pipelines routinely collapse annotator disagreement into majority vote labels before training. We show that this aggregation is not neutral: 42.6% of all annotator disagreement in HateXplain concentrates specifically at the hate/offensive boundary, a pattern consistent with annotators applying different thresholds for where hate begins (chi-squared = 135.199, df = 2, p < 0.0001). Both a hard-label BERT model (Model A) and a soft-label model (Model B) drop 22 percentage points in accuracy from agreed posts (~80%) to disagreement posts (~58%), confirmed at p < 0.0001. A per-annotator multi-head model (Model C) widens this gap further to 28 points while collapsing offensive disagreement accuracy to 0.245. Critically, Model A expresses significantly higher confidence on boundary case errors than Model C (0.710 vs. 0.495, p < 0.0001), meaning standard evaluation metrics will not detect the failure. Three downstream interventions of increasing sophistication all fail to recover boundary accuracy. We argue the problem is structural. Majority vote presents a contested judgment as ground truth, and models inherit that false certainty. The intervention must be upstream in annotation design.", "date": "2026-06-27", "categories": []}
{"id": "2606.30905", "title": "How Human Feedback Shapes AI-generated Community Notes", "abstract": "Community Notes, a bridging-based crowd-sourced fact-checking system, has emerged as a new mechanism for moderating misleading information on social media and has been adopted by major platforms including X, Facebook, Instagram, Threads, and TikTok. Since its introduction, there has been an open question about what role AI could play in scaling and optimizing the system. Recently, X extended its Community Notes system by introducing Collaborative Notes: notes initially drafted by an LLM and iteratively refined based on feedback from human contributors. In this work, we systematically analyze the complete corpus of 19,146 collaborative notes and 211,850 instances of human feedback. First, we develop a taxonomy of human suggestions for improving AI-generated note drafts and find that suggestions involving factual corrections and additional context are most likely to be incorporated, while subjective policy judgments rarely are. Second, we examine changes in helpfulness across versions of collaborative notes and find that human feedback leads to more helpful notes, with the greatest impact coming from suggestions that challenge the main claim in the previous draft, particularly when submitted by more active contributors. Finally, we find that although collaborative notes improve through human feedback, they reach helpful status and are shown on the platform at lower rates than human-only or AI-only notes, with limited human participation emerging as a key bottleneck. Nevertheless, rather than serving as a weaker substitute, collaborative notes tend to play a complementary role, predominantly targeting posts that do not attract human-only or AI-only notes. Our analysis provides an initial description of efforts to use AI to improve crowdsourced content moderation in a real-world moderation system and outlines pathways for future improvements to such features.", "date": "2026-06-29", "categories": []}
{"id": "2607.00143", "title": "Hate Speech Detection in Turkish and Arabic Languages: A Comprehensive Study", "abstract": "Online hate speech has been linked to a global rise in violence against minorities, including incidents such as mass shootings, lynchings, and ethnic cleansing. Societies grappling with this issue, particularly when hate speech targets specific groups based on religion, race, ethnicity, culture, nationality, or migration status, face the challenge of balancing freedom of expression with the need for effective content moderation on widely used online platforms. In response to this challenge, we introduce a comprehensive hate speech dataset covering five distinct topics in Turkish: refugees, the Israel-Palestine conflict, anti-Greek sentiment in Turkey, ethnic or religious communities (Alevis, Armenians, Arabs, Jews, and Kurds), and LGBTI+, alongside one topic in Arabic (refugees). In addition, we develop state-of-the-art BERT-based models to address multiple dimensions of hate speech analysis, including hate category classification, hate intensity prediction, target identification, and hate speech span detection, enabling a comprehensive understanding of hateful content in online discourse.", "date": "2026-06-30", "categories": []}