Empowering Cross-lingual Abilities of Instruction-tuned Large Language Models by Translation-following demonstrations

TL;DR

This paper introduces CrossAlpaca, a method to enhance the cross-lingual abilities of instruction-tuned large language models by using translation-following demonstrations to improve semantic alignment across languages.

Contribution

It proposes a novel approach combining instruction tuning with translation-following demonstrations to improve multilingual capabilities of LLMs.

Findings

CrossAlpaca outperforms monolingually tuned instruction models on multilingual benchmarks.

Translation-following demonstrations significantly enhance semantic alignment across languages.

The approach improves performance in six different languages on various question-answering and reasoning tasks.

Abstract

The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instructiontuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations.