Pruning Multilingual Large Language Models for Multilingual Inference

TL;DR

This paper proposes a pruning method that leverages large magnitude features in multilingual language models to improve their zero-shot performance in non-English languages, addressing performance disparities.

Contribution

It introduces a novel pruning approach based on large magnitude features to enhance multilingual models' non-English inference capabilities.

Findings

Pruning based on large magnitude features improves non-English zero-shot performance.

The method retains translation-relevant features, boosting multilingual inference.

Empirical results show significant performance gains in non-English tasks.

Abstract

Multilingual large language models (MLLMs), trained on multilingual balanced data, demonstrate better zero-shot learning performance in non-English languages compared to large language models trained on English-dominant data. However, the disparity in performance between English and non-English languages remains a challenge yet to be fully addressed. A distinctive characteristic of MLLMs is their high-quality translation capabilities, indicating an acquired proficiency in aligning between languages. This study explores how to enhance the zero-shot performance of MLLMs in non-English languages by leveraging their alignment capability between English and non-English languages. To achieve this, we first analyze the behavior of MLLMs when performing translation and reveal that there are large magnitude features that play a critical role in the translation process. Inspired by these findings, we retain the weights associated with operations involving the large magnitude features and prune other weights to force MLLMs to rely on these features for tasks beyond translation. We empirically demonstrate that this pruning strategy can enhance the MLLMs' performance in non-English language.