Semi-automatic Filtering of Translation Errors in Triangle Corpus

We are developing a multilingual machine translation system to provide foreign tourists with a multilingual speech translation service in the Winter Olympic Games that will be held in Korea in 2018. For a knowledge learning to make the multilingual expansibility possible, we needed large bilingual corpus. In Korea there were a lot of Korean-English bilingual corpus, but Korean-French bilingual corpus and Korean-Spanish bilingual corpus lacked absolutely. Korean-English-French and KoreanEnglish-Spanish triangle corpus were constructed by crowdsourcing translation using the existing large Korean-English corpus. But we found a lot of translation errors from the triangle corpora. This paper aims at filtering of translation errors in large triangle corpus constructed by crowdsourcing translation to reduce the translation loss of triangle corpus with English as a pivot language. Experiment shows that our method improves +0.34 BLEU points over the baseline system.