Statistical machine translation (SMT) is one of the most important research fields in natural language processing. In recent years, SMT has shown considerable success, and phrase-based translation models have been suggested to be the state of art by recent empirical evaluations. Now most of SMT systems are based on maximum entropy (ME) model. This thesis is about the design and implementation of an SMT decoder and the building of an SMT experiment platform. The main work is summarized as follows: (1) Minimum Error Rate Training in Statistical Machine Translation Minimum Error Rate (MER) Training improves the performance of the SMT system by directly using the evaluation criteria as the training criteria. The implementation of MER provides a tool for the experiment platform. (2) The Design and Implementation of A Statistical Machine Translation Decoder The efficiency and expansion are considered in the design and implementation of SMT decoder. The decoder is the basic of the experiment platform for SMT. (3) The Experiment Platform for Statistical Machine Translation An SMT experiment platform is built based on the above work and previous technologies. The platform provides plenty of functions and affords a good environment for the researchers of SMT.
修改评论