Machine Translation Utilizing Similar Translation Retrieval

Takuya Tamura, Yizhen Wei, Takehito Utsuro, Masaaki Nagata

Abstract: Neural Fuzzy Repair (NFR) system enables an NMT system to improve translation accuracy using similar translations searched from Translation Memory. This paper compared edit distance and sentence-BERT (SBERT) as the similarity measures used in the search for similar translations, and showed that SBERT outperformed edit distance in the case of small corpus sizes. This paper also studied a method to automatically select the most appropriate translation from more than one candidates. Compared to the naive method based on the number of tokens, the method based on the inner product of SBERT’s sentence embedding achieved significant improvements. These results prove the effectiveness of the SBERT-based approach in the NFR system.