Multistack Decoding in Statistical Machine Translation

Abstract

In a machine translation system, decoding is the process of finding the most likely translation according to previously learned parameters. The success of any such system is highly dependent on the quality of its decoder. Stack decoders were the first decoders developed for statistical machine translation, and they represent a good compromise between search thoroughness and efficiency. In this thesis I review the models used in IBM's Candide system, and then describe a multistack decoder developed for use with the models. I also compare the performance of the multistack decoder to that of two other decoding algorithms.