Adaptive k-Nearest-Neighbor Machine Translation with Domain Relevance for Chinese-Mongolian Husbandry Domain

Currently, there is a significant lack of Mongolian translation resources for technical exchanges in the husbandry industry. Manual translation not only suffers from low efficiency but also struggles to ensure the accuracy of specialized terminology. Therefore, this paper proposed an adaptive <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{k}$</tex> -Nearest-Neighbor machine translation with domain relevance (Adaptive kNN-MT-DR) method, aiming to improve the adaptability of Chinese-Mongolian neural machine translation (NMT) in the low-resource husbandry domain. Firstly, a vocabulary with varying degrees of domain relevance was constructed to adjust the adaptive kNN probability distribution during the translation inference phase, improving the accuracy of terminology translation in the husbandry field. Then, a sample repository was built by randomly selecting a portion of commondomain corpora and mixing them with husbandry corpora. To validate our proposed method, we collected Chinese corpora related to the husbandry industry and constructed parallel Chinese-Mongolian husbandry corpora through manual translation. Experimental results demonstrate that our method achieved improvements of 4.67 BLEU points and 3.52 BLEU points compared to the kNN-NMT and Adaptive kNN-MT methods, respectively.