arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2410.15464cs.CL
收藏

一种用于解释语言模型偏见的新型可解释性指标:在东南亚多语言模型上的应用

A Novel Interpretability Metric for Explaining Bias in Language Models: Applications on Multilingual Models from Southeast Asia

School of Computer Science, University of Birmingham(英国伯明翰大学计算机科学学院);Department of Information Systems and Computer Science, Ateneo de Manila University(马尼拉大学信息系统与计算机科学系)

Lance Calvin Lim Gamboa, Mark Lee

AI总结:

本文提出基于信息论的 token 级偏见归因分数,用于解释预训练语言模型的偏见来源,并在东南亚多语言模型上验证其能揭示性别歧视、恐同偏见及相关语义诱因。

摘要:

关于预训练语言模型(PLM)偏见的研究主要集中在偏见评估和缓解上,未能解决偏见归因与可解释性问题。我们提出一种新型指标,即“偏见归因分数”(bias attribution score),该指标借鉴信息论,用于衡量 token 层面对 PLM 中偏见行为的贡献。随后,我们将该指标应用于多语言 PLM,包括偏见评估文献中尚未得到充分检验的东南亚模型,以证明其效用。结果证实,东南亚 PLM 中存在性别歧视和恐同偏见。可解释性与语义分析还表明,PLM 偏见在很大程度上由与犯罪、亲密关系、帮助以及其他话语类别相关的词语诱发,说明这些话题正是 PLM 强烈复现预训练数据中偏见的领域,在这些领域使用 PLM 应更加谨慎。

英文摘要:

Work on bias in pretrained language models (PLMs) focuses on bias evaluation and mitigation and fails to tackle the question of bias attribution and explainability. We propose a novel metric, the $\textit{bias attribution score}$, which draws from information theory to measure token-level contributions to biased behavior in PLMs. We then demonstrate the utility of this metric by applying it on multilingual PLMs, including models from Southeast Asia which have not yet been thoroughly examined in bias evaluation literature. Our results confirm the presence of sexist and homophobic bias in Southeast Asian PLMs. Interpretability and semantic analyses also reveal that PLM bias is strongly induced by words relating to crime, intimate relationships, and helping among other discursive categories, suggesting that these are topics where PLMs strongly reproduce bias from pretraining data and where PLMs should be used with more caution.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑