EXPLOITING RICH FEATURE REPRESENTATION FOR SMT -BEST RERANKING by Yu Tong Thesis Supervisor: Derek F. Wong and Lidia S. Chao Master of Science in Computer Science In the field of Statistical Machine Translation, there is one issue that the poor quality of translation result which generated by machine translation system. One effective way to improve the translation quality is reranking the translation candidates which are from -best list. The sentences in -best lists represent the translation results of one source sentence. The translation result which is considered as the best during the process of decoder may be not optimal. The traditional method is to rerank -best list through training a reranking model by using the standard model features generated from translation system. For a rerankig model, standard model features can not well represent and distinguish the similar translation candidates from -best list. In this thesis, to solve this problem, instead of using standard model features, we focus on extracting semantic and syntactic features to represent the similar translation hypotheses for training a better reranking model. Our work focuses on investigating the performance on different kinds of features working on reranking tasks. We describe our work and evaluate it on different language pairs with different reranking algorithms. Applying our proposed method to do the experiment, the improvements up to three percent in the Bi-Lingual Evaluation Understudy (BLEU) score are obtained.