Summary of An Empirical Study on Cross-lingual Vocabulary Adaptation For Efficient Language Model Inference, by Atsuki Yamaguchi et al.
An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Language Model Inference
by Atsuki Yamaguchi, Aline Villavicencio, Nikolaos Aletras
First submitted to arxiv on: 16 Feb 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The development of state-of-the-art generative large language models (LLMs) is heavily reliant on English-centric tokenizers, vocabulary, and pre-training data. While some LLMs possess multilingual capabilities, recent studies have demonstrated that their inference efficiency declines significantly when generating text in languages other than English. This leads to increased inference time and costs. Cross-lingual vocabulary adaptation (CVA) methods have been proposed for adapting models to a target language with the goal of improving downstream performance. However, the effectiveness of these methods on increasing inference efficiency of generative LLMs remains unexplored. 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. Our findings indicate that CVA substantially contributes to LLM inference speedups of up to 271.5%. Additionally, we show that adapting LLMs pre-trained on more balanced multilingual data results in downstream performance comparable to the original models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Generative large language models are very smart computers that can create new text. But they often have trouble understanding languages other than English. This makes it hard and slow for them to generate text in those languages. To fix this, researchers have developed special methods called cross-lingual vocabulary adaptation (CVA). These methods help adapt the model to a new language. The goal is to make the model work better and faster in that language. In this study, we tested five different CVA methods on four types of generative models across four languages. We found that these methods can greatly improve how fast the models generate text. We also learned that if the models are trained on more balanced data from multiple languages, they will perform just as well in their original language. |
Keywords
» Artificial intelligence » Inference » Language understanding
