RomanSetu: Efficiently unlocking multilingual capabilities of Large Language Models via Romanization

TL;DR

This paper introduces a method to extend Large Language Models to non-English languages using romanized text, improving cross-lingual performance and reducing token complexity, thus enabling better multilingual NLP capabilities.

Contribution

The study demonstrates that training LLMs on romanized text enhances multilingual understanding and outperforms native script models, offering a new approach for low-resource language support.

Findings

Romanized text reduces token fertility by 2x-4x.

Romanized embeddings align more closely with English translations.

Outperforms native script models on various NLP tasks.

Abstract

This study addresses the challenge of extending Large Language Models (LLMs) to non-English languages that use non-Roman scripts. We propose an approach that utilizes the romanized form of text as an interface for LLMs, hypothesizing that its frequent informal use and shared tokens with English enhance cross-lingual alignment. Our approach involves the continual pretraining of an English LLM like Llama 2 on romanized text of non-English, non-Roman script languages, followed by instruction tuning on romanized data. The results indicate that romanized text not only reduces token fertility by 2x-4x but also matches or outperforms native script representation across various NLU, NLG, and MT tasks. Moreover, the embeddings computed on romanized text exhibit closer alignment with their English translations than those from the native script. Our approach presents a promising direction for leveraging the power of English LLMs in languages traditionally underrepresented in NLP. Our code is available on https://github.com/AI4Bharat/romansetu.