Learning interpretable word embeddings via bidirectional alignment of dimensions with semantic concepts

作者：

Highlights：

• The method improves interpretability of word embeddings while retaining task performance.

• The method utilizes both directions of the embedding dimensions.

• The method can gather gender information in a single dimension which helps gender debiasing.

摘要

•The method improves interpretability of word embeddings while retaining task performance.•The method utilizes both directions of the embedding dimensions.•The method can gather gender information in a single dimension which helps gender debiasing.

论文关键词：Word embeddings,Interpretability,Word semantics

论文评审过程：Received 26 October 2021, Revised 6 February 2022, Accepted 1 March 2022, Available online 22 March 2022, Version of Record 22 March 2022.

论文官网地址：https://doi.org/10.1016/j.ipm.2022.102925