XBMU-MC: A Multilingual Parallel Corpus

<p indent="0mm">Multilingual parallel corpus is a basic resource for natural language processing to break through language barriers and improve the efficiency and quality of cross-language tasks. To address the relative scarcity of parallel corpora for low-resource languages (Chinese-Tibetan, Chinese-Uyghur, Chinese-Mongolian), this study constructs a parallel corpus—XBMU-MC (Northwest Minzu University-Multilingual Corpus) for multilingual machine translation and cross-language information retrieval tasks. The original corpus consists of manually constructed specific bilingual texts and web-crawled publicly available multilingual data. The manually constructed corpus consists of Chinese texts written by researchers in the fields of culture, science and technology, and society, and the target language translations are obtained through machine translation and manual proofreading; while the web-crawled corpus captures the original texts from mainstream Tibetan, Mongolian, and Uyghur news websites to expand the scale of the corpus and the scope of the domains covered. After that, data enhancement techniques are used to expand the existing corpus, and the data are strictly screened to ensure the alignment consistency and translation accuracy between the source and target languages for each parallel corpus pair. Finally, 21,579 high-quality samples were selected and stored in JSON format, including three attributes: instruction, input and output. Finally, the multilingual translation comparison experiments between the XBMU-MC dataset and the open-source MMDS dataset based on the GLM4-9B, Qwen3-8B, and Deepseek-R1-8B models show an average increase in BLEU scores of 21.16, 20.51, and 15.37 percentage points. This dataset can help mitigate the problem of insufficient parallel corpus for low-resource languages to a certain extent, effectively support model training and multilingual translation tasks, and can also be used as a corpus basis for cross-lingual large language models in cross-language alignment, instruction fine-tuning and intelligent information processing applications in ethnic languages, helping to improve the performance of related tasks.