Token-level social-bias detector built on bert-large-uncased . Given a sentence, it tags each token with one of four bias categories following a 7-label BIO scheme, highlighting which words carry bias.
GUS-Net (BERT Large) Token-level built on . Given a sentence, it tags each token with one of four bias categories following a 7-label BIO scheme, highlighting carry bias. Part of the project (a master's thesis on interpretable bias detection through transformer attention). This is the larger-capacity BERT variant of . - - multi-label token classification (per-token sigmoid, thresholded) - English - , , ## Label scheme | Index | Label | Category | |-------|-----------|-----------------------------------| | 0 | O | none | | 1 | B-STEREO | Stereotype (span start) | | 2 | I-STEREO | Stereotype (span inside) | | 3 | B-GEN | Generalisation (span start) | | 4 | I-GEN | Generalisation (span inside) | | 5 | B-UNFAIR | Unfair language (span start) | | 6 | I-UNFAIR | Unfair language (span inside) | - — a blanket generalisation about a group. - — unfair / disparaging language toward a group. - — a stereotype attributed to a group. ## Important: multi-label + per-label thresholds Outputs are (multi-label), not a softmax. F1-optimised thresholds (order ): A flat 0.5 threshold will under-detect bias — use the values above. ## Usage ## Training data Fine-tuned on the — a token-level social-bias corpus annotated for eneralisations, nfairness and tereotypes (). punctuation tokens that were annotated with bias labels in the original data were (relabelled to the non-bias class), since punctuation is not a social-bias carrier. Bias spans predicted by these models therefore exclude leading/trailing punctuation. ## Evaluation ### StereoSet (intersentence split, 2123 examples) | Metric | Score | |--------|-------| | LMS (language-modeling score, higher is better) | | | SS (stereotype score, ) | | | ICAT (bias-adjusted quality) | | Per-category SS: gender 54.96 · race 52.05 · religion 56.41 · profession 57.44. ### Token classification (GUS-Net held-out test set) Held-out 20% split (seed 42, 748 sentences) of , scored with the per-label thresholds above; each category aggregates its B-/I- labels. | Category | Precision | Recall | F1 | |----------|-----------|--------|----| | O (non-bias) | 0.839 | 0.976 | 0.902 | | GEN | 0.620 | 0.736 | 0.673 | | UNFAIR | 0.203 | 0.865 | 0.329 | | STEREO | 0.621 | 0.846 | 0.716 | | | 0.477 | 0.811 | 0.601 | ## Limitations & intended use - , not a content-moderation oracle. Predictions reflect a specific operationalisation of bias; subtle or context-dependent bias may be missed. - English only. - Labels are not error-free; treat spans as evidence to review, not ground truth. - Do use for automated decisions about individuals. ## Citation If you use these models, please cite the GUS-Net dataset and benchmark: ## License Weights released under (matching the base model). The Attention Atlas code is MIT-licensed.
Difference from the original GUS-Net dataset and models:
manually cleaned
O
69.15
50 = ideal
54.64
62.73
ethical-spectacle/gus-dataset-v1
Micro-avg
Research / auditing tool
not
bibtex @article{powers2024gusnet, title = {GUS-Net: Social Bias Classification in Text with Generalizations, Unfairness, and Stereotypes}, author = {Powers, Maximus and Raza, Shaina and Chang, Alex and Riaz, Rehana and Mavani, Umang and Jonala, Harshitha Reddy and Tiwari, Ansh and Wei, Hua}, journal = {arXiv preprint arXiv:2410.08388}, year = {2024} }
Token-level social-bias detector built on bert-large-uncased . Given a sentence, it tags each token with one of four bias categories following a 7-label BIO scheme, highlighting which words carry bias.
