Most statistical machine translation systems employ a word based alignment model. In this paper we demonstrate that wordbased alignment is a major cause of translation errors. We propose a new alignment model based on shallow phrase structures and the structures can be automatically acquired from parallel corpus. This new model achieved over 10% error reduction for our spoken language translation task.

PDF file

In: COLING/ACL98

Publisher: Association for Computational Linguistics
All copyrights reserved by ACL 2007

Details