AndreCVieira/persuasion-lens-pt

PersuasionLens-PT: Cross-Lingual Detection of Persuasion Techniques in Portuguese

Model Description

PersuasionLens-PT is a span-level multi-label NER model for detecting and classifying 23 persuasion techniques in European Portuguese news articles. It was developed as part of a Master's thesis at FCUP / INESC TEC, following the CLEF-2024 CheckThat! Lab Task 3 framework.

The model tackles a zero-resource cross-lingual scenario: no native Portuguese training data exists, so it learns entirely from six other languages (English, Italian, French, German, Polish, Russian) and machine-translated data, then transfers to Portuguese at inference time.

Note: PersuasionLens corresponds to Model 2 (final system) in the associated Master's thesis, with the IT→PT-only Stage 2 fine-tuning configuration.

Key Features

• Detects 23 persuasion techniques at the text-span level (not document or sentence level)
• Cross-lingual transfer from 6 source languages to European Portuguese
• Custom multi-label sigmoid architecture (not standard BIO tagging)
• Surpasses the CLEF-2024 competition winner by +25% relative F1

Model Details

• Architecture: XLMRMultiLabelNER (custom classification head on XLM-RoBERTa)
• Base Model: xlm-roberta-base (Conneau et al., 2020)
• Model Type: Multi-label Token Classification
• Parameters: ~278M (base) + custom head
• Output: 46 binary outputs per token (B + I for each of 23 techniques)
• Loss Function: BCEWithLogitsLoss with per-output pos_weight
• Max Sequence Length: 512 tokens (sliding window with 50% overlap at inference)
• Global Classification Threshold: 0.70
• Framework: PyTorch

Important: This model uses a custom architecture and is not compatible with AutoModelForTokenClassification. It must be loaded using the custom XLMRMultiLabelNER class provided in the GitHub repository.

Intended Uses

Primary Use Cases

• Detecting persuasion and manipulation techniques in Portuguese news articles
• Supporting media literacy and fact-checking workflows
• Research on cross-lingual transfer learning for low-resource NLP tasks
• Studying disinformation patterns in Portuguese-language media

Out-of-Scope Uses

• General-purpose NER for Portuguese (use domain-specific models instead)
• Real-time production classification without adaptation
• Legal or policy decision-making without human review
• Languages other than Portuguese (though the cross-lingual architecture could be adapted)

Persuasion Techniques

The model detects 23 techniques from the SemEval/CLEF taxonomy (Piskorski et al., 2023), grouped into four macro-categories:

Category	Techniques
Justification	Appeal to Authority, Appeal to Fear/Prejudice, Appeal to Hypocrisy, Appeal to Popularity, Appeal to Time, Appeal to Values, Flag Waving
Simplification	Causal Oversimplification, Consequential Oversimplification, False Dilemma, Straw Man
Call	Appeal to (Strong) Emotions, Conversation Killer, Slogans
Distraction	Red Herring, Whataboutism
Reputation	Doubt, Guilt by Association, Loaded Language, Name Calling/Labelling, Obfuscation/Vagueness/Confusion, Questioning the Reputation, Repetition

Performance

Evaluated on the CLEF-2024 Portuguese test set using the official character-level evaluation script with partial span matching:

Overall Results

System	F1-micro	F1-macro
PersuasionLens-PT (this model)	0.134	0.103
PersuasionMultiSpan (organisers, post-competition)	0.132	0.120
UniBO (CLEF-2024 1st place)	0.107	0.073
Baseline (zero-shot)	0.002	—

Training Strategy Comparison

Configuration	F1-micro	F1-macro
Stage 1 (all 6 languages) + Stage 2 (IT→PT only)	0.134	0.103
Stage 1 (all 6 languages) + Stage 2 (IT→PT + FR→PT)	0.128	0.099
Stage 1 (all 6 languages) + Stage 2 (IT→PT + FR→PT + RU→PT)	0.117	0.095

Usage

Requirements

pip install torch transformers

Loading the Model

This model requires the custom XLMRMultiLabelNER class. Clone the repository first:

git clone https://github.com/andrevieira1203/MastersThesis.git

import torch
from transformers import AutoTokenizer

# Load tokenizer from HuggingFace
tokenizer = AutoTokenizer.from_pretrained("AndreCVieira/persuasion-lens-pt")

# Load the custom model class (from the GitHub repository)
from model import XLMRMultiLabelNER  # adjust import path as needed

model = XLMRMultiLabelNER.from_custom_checkpoint("path/to/model_directory")
model.eval()

Web Demo

A Streamlit demo application is available in the GitHub repository:

streamlit run app_streamlit.py 

Training Procedure

Two-Stage Training

1. Stage 1 (Multilingual Pre-training): Fine-tuning on all 6 source languages (EN, IT, FR, DE, PL, RU) with original + machine-translated data
2. Stage 2 (Cross-lingual Fine-tuning): Further fine-tuning on Italian→Portuguese translated data only, guided by embedding similarity analysis

Architecture Design

The key architectural innovation is replacing the standard BIO+softmax formulation with a multi-label sigmoid head:

• Each token produces 46 independent binary predictions (B and I for each of 23 techniques)
• This allows a single token to belong to multiple overlapping persuasion technique spans
• Per-output pos_weight in BCEWithLogitsLoss addresses the severe class imbalance (most tokens are non-persuasive)

Language Selection

An embedding analysis using paraphrase-multilingual-MiniLM-L12-v2 revealed Italian as the closest language to Portuguese (39.5% nearest neighbours), informing the Stage 2 data selection.

Translation

• Italian, French: Google Translate with character-level offset alignment
• Russian: DeepL with character-level offset alignment

File Structure

persuasion-lens-pt/
├── model.pt                    # Model weights (PyTorch state dict)
├── multilabel_config.json      # Multi-label head configuration (23 techniques, thresholds)
├── config.json                 # XLM-RoBERTa base config
├── tokenizer.json              # Tokenizer
├── tokenizer_config.json       # Tokenizer config
└── special_tokens_map.json     # Special tokens

Limitations and Biases

Limitations

1. No native Portuguese training data: The model learns Portuguese entirely through cross-lingual transfer; native annotated data would likely improve performance
2. Custom architecture: Not compatible with standard HuggingFace pipelines (AutoModelForTokenClassification); requires the custom class
3. Context length: Limited to 512 tokens per window (mitigated by sliding window with overlap)
4. Pickle format: The model.pt file uses PyTorch's pickle-based serialisation

Potential Biases

• Translation artifacts may introduce systematic errors not present in native Portuguese text
• The CLEF training data reflects specific European media landscapes and may not generalise to all Portuguese-language contexts
• Performance varies across the 23 techniques; rare techniques are harder to detect

Citation

If you use this model, please cite:

@mastersthesis{vieira2026persuasionlens,
  title={Language Models for the Detection of Manipulative Discourse and Disinformation in Text},
  author={Vieira, André},
  year={2026},
  school={Faculdade de Ciências da Universidade do Porto (FCUP)},
  type={Master's Thesis}
}

Acknowledgments

• Supervisors: Nuno Guimarães and Alípio Jorge (FCUP / INESC TEC)
• Built on XLM-RoBERTa by Conneau et al. (2020)
• Evaluation framework from CLEF-2024 CheckThat! Lab Task 3 (Piskorski et al., 2024)
• GitHub repository: andrevieira1203/MastersThesis