Hate Speech Detection (DistilBERT, fine-tuned + counterfactually augmented)
DistilBERT fine-tuned for 3-class hate speech / offensive language detection, trained on the Davidson et al. (2017) hate speech and offensive language dataset (24,783 tweets) plus a targeted Counterfactual Data Augmentation (CDA) set addressing two specific failure patterns found during live testing.
Labels
- 0: Hate Speech
- 1: Offensive Language
- 2: Neither
Performance (held-out test set, n=3718)
- Macro-F1: 0.7504
- Accuracy: 0.8784
This is ~1.6 macro-F1 points below the pre-augmentation version (0.7663) -- a deliberate, evidence-based trade-off, not a regression we're unaware of. See "Counterfactual Data Augmentation" below for why.
Counterfactual Data Augmentation
Live API testing surfaced two real problems not caught by earlier offline evaluation:
- Insult-severity non-monotonicity: harsher personal insults sometimes scored LOWER on Hate Speech than milder ones (the pre-augmentation model got this right on only 1/6 held-out severity examples).
- Identity-term confidence instability: neutral sentences mentioning "queer" swung Neither-confidence by up to 60 percentage points versus a no-identity-term baseline -- a textbook case of the bias CDA (Dixon et al., 2018) was designed to fix.
After adding 114 targeted, hand-labeled examples (severity-graded personal insults + genuine group-targeted hate + benign identity-term sentences, severity examples oversampled 3x) and retraining:
- Severity: 1/6 -> 5/6 correct on held-out (unseen) test phrasings
- "queer" gap: -60.2pp -> -6.3pp
- Identity-term spread (best-to-worst term): 73.4pp -> 18.6pp
- AAVE dialect bias (see below): confirmed no regression, still 0% gap
A 2x-oversampling variant was also tested and rejected despite a HIGHER macro-F1 (0.7769): it only partially fixed identity calibration ("queer" still at -45.4pp) and reintroduced a real AAVE false positive ("Ain't nobody got time for that" -> Offensive Language). Chasing the higher benchmark number there would have meant shipping a worse model on the property that actually mattered -- documented explicitly so this trade-off is visible, not buried.
Bias audit (AAVE dialect)
Tested against a documented failure mode in this dataset family (Sap et al., 2019): over-flagging African-American Vernacular English (AAVE) as offensive. On 20 hand-constructed AAVE/Standard-English sentence pairs with equivalent (benign) meaning, this model shows a 0% false-positive-rate gap between dialects, versus a 20-point gap in a TF-IDF+LogReg baseline tested the same way (see repo for full methodology, including statistical tests).
Known limitations
Qualitative error analysis (see repo, backed by LIME explanations) found the model's decisions are driven heavily by the presence of specific slur tokens as near-decisive lexical triggers, rather than deeper contextual understanding of intent or target -- the CDA above measurably improved this for the specific patterns tested, but should not be read as a complete fix for all instances of this behavior.
Intended use
Content-moderation assistance / research and portfolio demonstration. NOT intended as a sole or automatic basis for real moderation decisions against real people. Predictions are probabilistic and can be wrong -- see the full disclaimer and bias-audit results in the source repository before any other use.
- Downloads last month
- 66