First the worst: Finding better gender translations during beam search

TL;DR

This paper proposes inference-time techniques such as reranking and gender constraints during beam search to reduce gender bias in neural machine translation, achieving significant accuracy improvements without extra training data.

Contribution

It introduces inference-based methods to mitigate gender bias in NMT, avoiding the need for retraining or additional data, and demonstrates their effectiveness.

Findings

Significant gains in WinoMT accuracy with proposed methods

Gender diversity in nbest lists improved during decoding

No additional bilingual data required for bias reduction

Abstract

Neural machine translation inference procedures like beam search generate the most likely output under the model. This can exacerbate any demographic biases exhibited by the model. We focus on gender bias resulting from systematic errors in grammatical gender translation, which can lead to human referents being misrepresented or misgendered. Most approaches to this problem adjust the training data or the model. By contrast, we experiment with simply adjusting the inference procedure. We experiment with reranking nbest lists using gender features obtained automatically from the source sentence, and applying gender constraints while decoding to improve nbest list gender diversity. We find that a combination of these techniques allows large gains in WinoMT accuracy without requiring additional bilingual data or an additional NMT model.