Adversarial Learning for Bias Mitigation in Machine Translation

Abstract

Natural language processing (NLP) systems often contain significant biases regarding sensitive attributes, such as gender, that worsen system performance and perpetuate harmful stereotypes. Recent research has found that adversarial neural networks can help to mitigate bias in word embeddings on the task of completing analogies, without impairing model performance. This model-agnostic method, which requires no cumbersome data modifications, aims to mitigate both the biases present in datasets and those amplified during training. However, this strategy still needs further development for use in downstream applications, particularly for use with large language models and language tasks in which gender or another protected variable like gender must be deduced from the data itself. To that end, this work proposes two methods, one based on the structure of sentence encodings and one based on the use of gendered pronouns, to measure gender representation in machine translation. It then presents an adversarial learning framework that uses these measures to mitigate gender bias in English-French translation on the language model T5. I found that this method successfully mitigated gender bias in the model’s translated output with minimal effect on translation quality. The results suggest that adversarial learning is a promising technique for use with large language models and complex, realistic applications.