Exploiting Multiple Resources for Japanese to English Patent Translation

This paper describes the development of a Japanese to English translation system using multiple resources and NTCIR-10 Patent translation collection. The MT system is based on different training data, the Wiktionary as a bilingual dictionary and Moses decoder. Due to the lack of parallel data on the patent domain, additional training data of the general domain was extracted from Wikipedia. Experiments using NTCIR-10 Patent translation data collection showed an improvement of the BLEU score when using a 5-grams language model and when adding the data extracted from Wikipedia but no improvement when adding the Wiktionary.