Bias Attribution in Filipino Language Models: Extending a Bias Interpretability Metric for Application on Agglutinative Languages

2025-06-08

AI Summary

This study addresses the lack of interpretability in bias attribution for agglutinative languages—exemplified by Filipino—within language models. We adapt information-theoretic bias attribution scores to accommodate agglutinative morphology, integrating Filipino tokenization and morphological analysis, inter-layer gradient computation, and cross-lingual model comparisons. We systematically evaluate bias sources across monolingual Filipino and three multilingual models. Results reveal that bias in Filipino models stems predominantly from entity-type tokens (persons, objects, relations), contrasting sharply with English models where action-oriented themes dominate—highlighting fundamental cross-linguistic differences in bias mechanisms. This work extends the linguistic applicability of interpretable bias attribution methods and introduces an entity-driven paradigm for non-English bias analysis. It further provides a reusable methodological framework for fairness research in low-resource languages.

Technology Category

• Natural Language Processing: Lexical Semantics and Morphology
• Philosophy and Ethics of AI: Bias, Fairness & Equity
• Machine Learning: Ethics, Bias, and Fairness

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Emerging research on bias attribution and interpretability have revealed how tokens contribute to biased behavior in language models processing English texts. We build on this line of inquiry by adapting the information-theoretic bias attribution score metric for implementation on models handling agglutinative languages, particularly Filipino. We then demonstrate the effectiveness of our adapted method by using it on a purely Filipino model and on three multilingual models: one trained on languages worldwide and two on Southeast Asian data. Our results show that Filipino models are driven towards bias by words pertaining to people, objects, and relationships, entity-based themes that stand in contrast to the action-heavy nature of bias-contributing themes in English (i.e., criminal, sexual, and prosocial behaviors). These findings point to differences in how English and non-English models process inputs linked to sociodemographic groups and bias.

Problem

• Extend bias attribution metric for agglutinative languages like Filipino
• Analyze bias sources in Filipino vs. English language models
• Compare bias themes in monolingual and multilingual Filipino models

Innovation

• Adapting bias attribution score for agglutinative languages
• Testing method on Filipino and multilingual models
• Identifying entity-based bias themes in Filipino models