Investigating Cross-Linguistic Gender Bias in Hindi-English Across Domains

TL;DR

This paper investigates how gender bias varies across different domains in Hindi-English language models, focusing on measuring and understanding bias in Indic languages using domain-specific embeddings.

Contribution

It introduces a method to quantify cross-domain gender bias in Hindi-English embeddings, addressing a gap in bias research for Indic languages.

Findings

Bias varies significantly across domains.

Domain-specific embeddings reveal insights into gender bias.

The approach outperforms existing models in measuring bias.

Abstract

Measuring, evaluating and reducing Gender Bias has come to the forefront with newer and improved language embeddings being released every few months. But could this bias vary from domain to domain? We see a lot of work to study these biases in various embedding models but limited work has been done to debias Indic languages. We aim to measure and study this bias in Hindi language, which is a higher-order language (gendered) with reference to English, a lower-order language. To achieve this, we study the variations across domains to quantify if domain embeddings allow us some insight into Gender bias for this pair of Hindi-English model. We will generate embeddings in four different corpora and compare results by implementing different metrics like with pre-trained State of the Art Indic-English translation model, which has performed better at many NLP tasks than existing models.