Forest-to-String SMT for Asian Language Translation: NAIST at WAT 2014

摘要

This paper describes the Nara Institute of Science and Technology's (NAIST) submission to the 2014 Workshop on Asian Translation's four translation tasks. All systems are based on forest-to-string (F2S) translation, in which the input sentence is first parsed using a syntactic parser, then a forest of possible syntactic analyses is translated into the target language. In addition to the baseline F2S system, we add rescoring using a recurrent neural network language model (RNNLM), which allows for more fluent output. The resulting system achieved the highest results in both automatic and manual evaluation for all four of the language pairs targeted by the workshop.