Scholarlink

This paper introduces a three-stage framework to enhance Tibetan-Chinese machine translation in low-resource, multi-domain settings. First, we expand the vocabulary with carefully selected Tibetan tokens to reduce token fragmentation and improve semantic representation. Second, we propose a cross-lingual embedding initialization method that leverages token-level alignment to transfer semantic information from Chinese to Tibetan, enabling better representation in data-scarce conditions. Third, we construct a domain-specific fine-tuning pipeline by integrating curated bilingual lexicons and parallel corpora across various fields, including Buddhism, law, and Tibetan medicine. Training samples incorporate both translation pair annotations and contextual sentence examples to improve the model’s handling of specialized terminology. The framework is evaluated across five domains: Colloquial, News, Law, Buddhism, and Tibetan Medicine. Experiments demonstrate significant improvements over strong baselines and commercial systems. For instance, in Tibetan \(\rightarrow \) Chinese translation, BLEU scores in the Buddhism and Tibetan Medicine domains increased from 7.18 and 3.82 to 16.40 and 8.54, respectively. Similar gains are observed in the reverse direction. These results highlight the effectiveness of the proposed alignment and supervision strategies. Overall, this work provides a practical solution for low-resource translation and contributes to the preservation and accessibility of underrepresented languages and cultural knowledge.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-Enhanced Translation for Low-Resource Languages: Cross-Lingual Alignment and Multi-domain Adaptation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qifeng Su,
  • Zhicong Wu,
  • Xiaodong Shi

摘要

This paper introduces a three-stage framework to enhance Tibetan-Chinese machine translation in low-resource, multi-domain settings. First, we expand the vocabulary with carefully selected Tibetan tokens to reduce token fragmentation and improve semantic representation. Second, we propose a cross-lingual embedding initialization method that leverages token-level alignment to transfer semantic information from Chinese to Tibetan, enabling better representation in data-scarce conditions. Third, we construct a domain-specific fine-tuning pipeline by integrating curated bilingual lexicons and parallel corpora across various fields, including Buddhism, law, and Tibetan medicine. Training samples incorporate both translation pair annotations and contextual sentence examples to improve the model’s handling of specialized terminology. The framework is evaluated across five domains: Colloquial, News, Law, Buddhism, and Tibetan Medicine. Experiments demonstrate significant improvements over strong baselines and commercial systems. For instance, in Tibetan \(\rightarrow \) Chinese translation, BLEU scores in the Buddhism and Tibetan Medicine domains increased from 7.18 and 3.82 to 16.40 and 8.54, respectively. Similar gains are observed in the reverse direction. These results highlight the effectiveness of the proposed alignment and supervision strategies. Overall, this work provides a practical solution for low-resource translation and contributes to the preservation and accessibility of underrepresented languages and cultural knowledge.