The power of Prompts: Evaluating and Mitigating Gender Bias in MT with LLMs

Aleix Sant, Carlos Escolano, Audrey Mash, Francesca De Luca Fornaciari, Maite MeleroACL 2024

TL;DR

This paper investigates gender bias in Large Language Models (LLMs) for machine translation (MT), finding that base LLMs exhibit more bias than traditional NMT models. The authors demonstrate that carefully crafted prompts can significantly reduce this bias in instruction-tuned LLMs, bringing them closer to NMT performance levels.

What Problem Does This Paper Address?

Gender bias in machine translation is a significant issue, leading to perpetuation of stereotypes and potentially harmful consequences. While much research has focused on Neural Machine Translation (NMT) models, the rapid rise of Large Language Models (LLMs) for tasks like MT necessitates a similar investigation into their gender bias. This paper aims to bridge this gap by evaluating gender bias in LLMs and exploring methods to mitigate it, specifically through prompt engineering.

Methods

The study employs a two-pronged approach:

  1. Benchmarking: Several widely-used LLMs (Llama-2-7B, Aguila-7B, Flor-6.3B) and state-of-the-art NMT models (M2M-100-1.2B, NLLB-200-1.3B, Mt-aina-en-ca, Google Translate) are benchmarked on English-to-Catalan and English-to-Spanish translation directions. This involves evaluating both translation quality (using BLEU and COMET) and gender bias.
  2. Prompt Engineering for Mitigation: An instruction-tuned LLM (Llama-2-7B-chat) is used to explore prompt engineering techniques. This includes few-shot prompting, context-supplying, and chain-of-thought prompting, with a focus on the Gender Coreference Resolution task. The goal is to find a prompt that effectively reduces gender bias without significantly degrading translation quality.

Datasets

The following datasets were used for evaluation:

  • FLoRes-200: A multilingual general domain dataset for machine translation quality assessment.
  • WinoMT: A test set designed to evaluate gender bias in MT, specifically focusing on Gender Coreference Resolution. It includes pro- and anti-stereotypical sets.
  • Gold BUG: A large-scale corpus of real-world English sentences for gender bias evaluation, also including pro- and anti-stereotypical sets.
  • MuST-SHE: A dataset for evaluating gender bias in speech translation, adapted for MT. It includes instances with and without gender cues, testing both Gender Terms Detection and Gender Coreference Resolution.

Key Insights and Ideas

  • LLMs are more biased than NMT: Base LLMs, despite their advanced capabilities, exhibit a higher degree of gender bias compared to traditional NMT models. This is a crucial finding given the increasing adoption of LLMs.
  • Prompting is a powerful tool: The study highlights that well-designed prompts can be a highly effective strategy for mitigating gender bias in instruction-tuned LLMs. This is a novel application of prompting beyond task specification.
  • Chain-of-thought and anti-stereotypical examples are key: The most effective prompts involved a simplified chain-of-thought approach combined with 5-shot examples that were anti-stereotypical and featured increased female representation. This suggests that guiding the model through a reasoning process and providing counter-stereotypical examples is crucial for bias reduction.
  • Trade-off between bias mitigation and translation quality: While prompt engineering significantly reduces gender bias, it can lead to a slight decrease in overall translation quality. This presents an important trade-off to consider.

Key Results

  • Bias Reduction: The top-performing prompt achieved remarkable increases in WinoMT scores, reducing gender bias by 12.4% (En→Ca) and 11.7% (En→Es) compared to the baseline. This brought the LLM's performance on par with, and in some cases surpassing, NMT models in terms of gender bias manifestation.
  • Pro-stereotypical bias: All models showed a tendency to perform better on pro-stereotypical sentences, indicating a reliance on societal biases.
  • Absence of gender cues: In the absence of explicit gender cues, LLMs and NMT models predominantly predicted male terms (around 75%-94%), with female terms appearing mainly in stereotypical contexts.

Limitations and Future Work

  • Binary Gender Focus: The study exclusively focused on binary male and female genders. Additional experiments on new datasets would be needed to assess performance on non-binary scenarios.
  • Complexity of Bias: The authors acknowledge that their metrics and experiments focus on specific manifestations of gender bias, recognizing that the problem is more complex.
  • Translation Quality Trade-off: The successful bias mitigation prompt resulted in a slight loss of translation quality. Future work could explore prompts that balance bias reduction and translation quality more effectively.
  • Generalizability: While the study explored prompt generalizability across different test sets, further research could investigate this across more languages and diverse domains.

Summary updated

BibTeX
@inproceedings{sant2024power,
  author = {Aleix Sant and Carlos Escolano and Audrey Mash and Francesca De Luca Fornaciari and Maite Melero},
  title = {{The power of Prompts: Evaluating and Mitigating Gender Bias in MT with LLMs}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure