AAAI2022 | Word Embeddings via Causal Inference: Gender Bias Reducing and Semantic Information Preserving

Address of thesis ：https://arxiv.org/pdf/2112.05194.pdf

Word vectors are widely used in various naturallanguageprocessing tasks , However, many studies have shown that there are various biases in word vectors （ Such as gender or race ）. This paper presents a method based on causal reasoning , It can effectively eliminate gender bias in word vectors and retain other semantics . Specifically , This paper proposes two causal reasoning frameworks , Respectively used to eliminate Potential proxy bias and Unresolved bias. Experiments show that , The word vector obtained by this method can not only better eliminate the bias in the word vector , It can also improve the performance of semantic related tasks .