Domain adaptation of a speech translation system for lectures by utilizing frequently appearing parallel phrases in-domain

This paper describes our scheme to translate spoken English lectures into Japanese consisting of an English automatic speech recognition system (ASR) that utilizes a deep neural network (DNN) and an English to Japanese phrase-based statistical machine translation system (SMT). We focused on domain adaptation of the acoustic and translation models. For domain adaptation of the translation model, frequently appearing English-phrases consisting of multiple words are extracted from transcripts of in-domain lectures based on n-gram words or a part of syntax tree. Then we translated the English phrases into Japanese-phrases by hand semi-automatically. These phrase pairs of source and target language are used to learn an SMT model for domain adaptation. An adaptation method directly inserts these phrase pairs into a phrase table or adds them to a parallel corpus. In the experiments, n-gram and syntax tree based methods are compared whilst extracting frequent English-phrases. Furthermore, the adapted phrase table and the parallel corpus are compared. When the frequent English and Japanese phrase pairs based on syntax tree were added to the phrase table, the baseline model was improved.