The fine-tuning of open-source large language models (LLMs) for machine
translation has recently received considerable attention, marking a shift
towards data-centric research from traditional neural machine translation.
However, the area of data collection for instruction fine-tuning in machine
translation remains relatively underexplored. In this paper, we present
LexMatcher, a simple yet effective method for data collection that leverages
bilingual dictionaries to generate a dataset, the design of which is driven by
the coverage of senses found in these dictionaries. The dataset comprises a
subset retrieved from an existing corpus and a smaller synthesized subset which
supplements the infrequent senses of polysemous words. Utilizing LLaMA2 as our
base model, our approach outperforms the established baselines on the WMT2022
test sets and also exhibits significant performance improvements in tasks
related to word sense disambiguation and specialized terminology translation.
These results underscore the effectiveness of LexMatcher in enhancing LLM-based
machine translation.

LexMatcher 方法使用双语词典生成数据集，利用 LLaMA2 作为基础模型并展示出在 WMT2022 测试集上优于已有基线模型的表现，以及在词义消歧和专业术语翻译等任务中显著的性能提升，突出了 LexMatcher 在增强基于 LLM 的机器翻译中的有效性。