Gender Bias Detection and Evaluation in Low-Resource Languages: A Comprehensive Analysis of Gujarati NLP Datasets

Abstract

It's a big concern when NLP algorithms are biased against women, especially for languages like Gujarati where there isn't a lot of data and much of it comes via translation. In this paper well-known fairness tools to look at gender bias in two Gujarati datasets BUG and CORGI. In this article 500 phrases in total and determined that the BUG dataset has a lot of bias whereas the CORGI sample has some bias. Neither dataset meets fairness norms. This research focus on that Gujarati grammar exacerbates bias by necessitating the inclusion of gender in several sentence components thereby amplifying even minor biases present in the original text. Another crucial thing to remember is that translation mistakes don't produce prejudice. The translations are correct although they carry the same bias from the original data. So for the low resource language to build a fairer NLP system needs better data, proper fairness checks and also a proper handling of features related to language specific.