A Causal View of Entity Bias in (Large) Language Models
Fei Wang, Wenjie Mo, Yiwei Wang, Wenxuan Zhou, Muhao Chen · 2023
Entity bias widely affects pretrained (large) language models, causing them to rely on (biased) parametric knowledge to make unfaithful predictions.Although causality-inspired methods 1 Our code is available at https://github.com/ luka-group/Causal-View-of-Entity-Bias Assume subject_entity can be any of Bill Gates, Jeff Bezos, and Steve Jobs, while object_entity can be any of Google, Microsoft, and Meta.Context: subject_entity went to object_entity Building 99.Question: What's the relation between subject_entity and object_entity in the given context?Option: founder, visitor.Answer with one word: visitor (GPT-3.5)✔ Context: Bill Gates went to Microsoft Building 99.