From Unaligned to Aligned: Scaling Multilingual LLMs with Multi-Way Parallel Corpora
This work addresses the challenge of improving multilingual performance in LLMs for low-resource languages, representing an incremental advancement through the use of parallel corpora.
The paper tackled the problem of scaling multilingual LLMs by addressing the limitations of unaligned data, introducing a large-scale multi-way parallel corpus called TED2025 spanning 113 languages, and found that models trained on this data consistently outperformed those on unaligned data across six benchmarks.
Continued pretraining and instruction tuning on large-scale multilingual data have proven to be effective in scaling large language models (LLMs) to low-resource languages. However, the unaligned nature of such data limits its ability to effectively capture cross-lingual semantics. In contrast, multi-way parallel data, where identical content is aligned across multiple languages, provides stronger cross-lingual consistency and offers greater potential for improving multilingual performance. In this paper, we introduce a large-scale, high-quality multi-way parallel corpus, TED2025, based on TED Talks. The corpus spans 113 languages, with up to 50 languages aligned in parallel, ensuring extensive multilingual coverage. Using this dataset, we investigate best practices for leveraging multi-way parallel data to enhance LLMs, including strategies for continued pretraining, instruction tuning, and the analysis of key influencing factors. Experiments on six multilingual benchmarks show that models trained on multiway parallel data consistently outperform those trained on unaligned multilingual data.