An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Language Model Inference

TL;DR

An empirical study of five CVA methods on four generative LLMs across four typologically-diverse languages and four natural language understanding tasks finds that CVA substantially contributes to LLM inference speedups of up to 271.5\%.

Abstract

The development of state-of-the-art generative large language models (LLMs) disproportionately relies on English-centric tokenizers, vocabulary and pre-training data. Despite the fact that some LLMs have multilingual capabilities, recent studies have shown that their inference efficiency deteriorates when generating text in languages other than English. This results in increased inference time and costs. Cross-lingual vocabulary adaptation (CVA) methods have been proposed for adapting models to a target language aiming to improve downstream performance. However, the effectiveness of these methods on increasing inference efficiency of generative LLMs has yet to be explored. In this paper, we perform an empirical study of five CVA methods on four generative LLMs (including monolingual and multilingual models) across four typologically-diverse languages and four natural language understanding tasks. We find that CVA substantially contributes to LLM inference speedups of up to 271.5\%. We also show that adapting LLMs that have been pre-trained on more balanced multilingual data results in downstream performance comparable to the original models.

Figures (10)

• Figure 1: Example of overfragmentation when applying the Mistral-7B tokenizer to non-English text.
• Figure 2: Relative speedup ratios to each base model/tokenizer when prompted in English and a target language. Dotted lines denote the average speedup ratio across tasks in each setting.
• Figure 3: Performance difference between English and in-language prompts. Positive and negative values indicate better performance using in-language or English prompts respectively.
• Figure 4: Kendall's $\tau$ correlation between the number of LAPT steps and performance (in-language prompting).
• Figure 5: Performance changes in span with respect to LoRA rank $r$.