ko-social-bias-detector

Korean Social-Bias Detector v1.2

A Korean one-sentence social-bias detector using a fine-tuned KLUE-RoBERTa multitask classifier plus deterministic target, context, and minimal-edit logic.

Public API

from pipeline import detect_bias

result = detect_bias("여성은 원래 지도자 역할에 적합하지 않다.")

Important score semantics

base_social_score_1_100 is the trained model's raw_bias_probability × social_target_probability.

It is not calibrated confidence for the final label. The final decision may also be changed by:

• a guarded high-precision rule rescue;
• a deterministic context guard.

Inspect decision_basis to see how the final label was reached.

Output spans

• target.surface_text: target with Korean particle.
• target.core_text: target without the particle.
• evidence_atoms: minimal bias-triggering expressions.
• rewrite_spans: character-offset spans permitted for local revision.

Apply multiple edits from right to left and recheck rewritten text with the detector before accepting it.

Limitations

The Korean span layer is rule-assisted. A future release should replace it with a manually labeled token-role model and broader independent holdout evaluation.