Debiasing Multilingual Word Embeddings: A Case Study of Three Indian Languages

  • 2021-07-21 16:12:51
  • Srijan Bansal, Vishal Garimella, Ayush Suhane, Animesh Mukherjee
  • 5

Abstract

In this paper, we advance the current state-of-the-art method for debiasingmonolingual word embeddings so as to generalize well in a multilingual setting.We consider different methods to quantify bias and different debiasingapproaches for monolingual as well as multilingual settings. We demonstrate thesignificance of our bias-mitigation approach on downstream NLP applications.Our proposed methods establish the state-of-the-art performance for debiasingmultilingual embeddings for three Indian languages - Hindi, Bengali, and Teluguin addition to English. We believe that our work will open up new opportunitiesin building unbiased downstream NLP applications that are inherently dependenton the quality of the word embeddings used.

 

Quick Read (beta)

loading the full paper ...