Cross-Lingual Transfer for Machine Translation in Turkic Languages

AI 中文解读

核心亮点：这项研究首次系统揭示了突厥语族内部机器翻译的“血缘关系”——语言越亲近，翻译效果越好，而且翻译方向还会影响结果。通俗解读：研究人员选了五种突厥语（土耳其语、阿塞拜疆语、乌兹别克语、哈萨克语、吉尔吉斯语），它们像兄弟姐妹一样有相似之处。他们测试了用其中一种语言训练的翻译模型，能不能拿来翻译另一种语言。结果发现，关系越近的语言（比如土耳其语和突厥语族的阿塞拜疆语），互相“借用”能力越强；而哈萨克语和吉尔吉斯语这对“近亲”也配合默契。有趣的是，从A语言到B语言好使，反过来未必一样行得通。另外，如果把斯拉夫字母改成拉丁字母，某些配对的效果会提升，但也不是万能灵药。实际影响：这项研究对机器翻译的“冷门语言”很有意义。很多突厥语使用人口不多，训练数据稀缺，做高质量翻译本就困难。如今明确了哪些语言之间可以“互相帮忙”，就能用数据丰富的语言去帮助数据匮乏的语言，节省大量人工标注成本。未来你使用翻译工具时，像哈萨克语、吉尔吉斯语这类小众语言的翻译质量有望明显提升，甚至能覆盖更多方言，帮助跨境贸易、旅游交流和文化传播变得更加顺畅。

Cross-lingual transfer is central to low-resource machine translation, but its behavior within closely related language families remains insufficiently characterized. We study transfer among five Turkic languages; Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz; using pairwise transfer matrices. In this setting, each model is fine-tuned with one transfer source and evaluated on a different transfer target while the translation target remains the same. Across mT5 experiments, we find that transfer is strongest between closely related Turkic pairs, especially Turkish-Azerbaijani and Kazakh-Kyrgyz. We also show that transfer direction matters, and that the same transfer source-transfer target pair can behave differently when the translation target changes. Latinization improves BLEU and chrF in several script-mismatched settings, but its effect is not uniform across metrics. Additional analyses show that transfer sources are mostly stable across different datasets and model settings.