An Empirical Study of Many-Shot In-Context Learning for Machine Translation of Low-Resource Languages

TL;DR

This study empirically evaluates many-shot in-context learning for English-to-low-resource language translation, highlighting how retrieval-based example selection enhances data efficiency and model performance.

Contribution

It demonstrates that BM25-based retrieval significantly improves data efficiency in many-shot ICL for low-resource machine translation tasks.

Findings

Many-shot ICL effectiveness increases with more examples.

BM25 retrieval matches large example sets with fewer retrieved examples.

Retrieving 50 examples can be as effective as using 250 many-shot examples.

Abstract

In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks from a few examples, making it promising for languages underrepresented in pre-training. Recent work on many-shot ICL suggests that modern LLMs can further benefit from larger ICL examples enabled by their long context windows. However, such gains depend on careful example selection, and the inference cost can be prohibitive for low-resource language communities. In this paper, we present an empirical study of many-shot ICL for machine translation from English into ten truly low-resource languages recently added to FLORES+. We analyze the effects of retrieving more informative examples, using out-of-domain data, and ordering examples by length. Our findings show that many-shot ICL becomes more effective as the number of examples increases. More importantly, we show that BM25-based retrieval substantially improves data efficiency: 50 retrieved examples roughly match 250 many-shot examples, while 250 retrieved examples perform similarly to 1,000 many-shot examples.