# Easy Adaptation to Mitigate Gender Bias in Multilingual Text Classification

**Authors:** Xiaolei Huang  
**arXiv:** [2204.05459](https://arxiv.org/abs/2204.05459)  
**Date:** 2022-04-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2204.05459v1/

## Key points

- Study addresses demographic biases in text classification models, specifically in multilingual settings
- Existing approaches have focused on mitigating biases in monolingual data, little attention given to biases in multilingual data
- Authors treat gender as domains and propose a standard domain adaptation model to reduce gender bias and improve performance of text classifiers
- Approach evaluated on two text classification tasks: hate speech detection and rating prediction
- Results compared with three fair-aware baselines to demonstrate effectiveness of approach
- Data statistics presented in Table 1: hate speech data smaller compared to review data, both datasets have skewed label distributions, more Danish reviews due to consumer review website location
- Documents in both datasets are short, Twitter's hate speech data comparatively shorter
- Gender ratio analysis reveals relatively lower female ratio in most of the data
- Only text documents and gender information used for evaluation purposes without other user profile information or IDs
- Experimental information anonymized before training text classifiers
- Authors introduce an easy adaptation framework called "Frustratingly Easy Domain Adaptation" (FEDA)
- Applying domain adaptation techniques can help mitigate biases effectively
- Study aims to address demographic biases in multilingual text classification by treating gender as domains and using a domain adaptation model.

## Summary

This study addresses the issue of demographic biases in text classification models, specifically in multilingual settings. While existing approaches have focused on mitigating biases in monolingual data, little attention has been given to the impact of biases in multilingual data. In this work, the authors treat gender as domains (e.g., male vs. female) and propose a standard domain adaptation model to reduce gender bias and improve the performance of text classifiers. The authors evaluate their approach on two text classification tasks: hate speech detection and rating prediction. They compare their results with three fair-aware baselines to demonstrate the effectiveness of their approach. To provide context, the authors present data statistics in Table 1. The hate speech (HS) data is relatively smaller compared to the review data, and both datasets have skewed label distributions. For example, most reviews have positive labels. It is noted that the review data comes from a consumer review website in Denmark, resulting in more Danish reviews than other languages. The documents in both datasets are short, with Twitter's HS data being comparatively shorter. The gender ratio analysis reveals that most of the data has a relatively lower female ratio. In terms of ethical and privacy considerations, only text documents and gender information are used for evaluation purposes without any other user profile information such as user IDs. All experimental information has been anonymized before training text classifiers. The authors introduce an easy adaptation framework based on previous work called "Frustratingly Easy Domain Adaptation" (FEDA). They emphasize that applying domain adaptation techniques can help mitigate biases effectively. Overall, this study aims to address demographic biases in multilingual text classification by treating gender as domains and using a domain adaptation model.