Automatic Construction of Translation Knowledge for Corpus-based Machine Translation

Many machine translation (MT) systems that utilize the knowledge automatically acquired from bilingual corpora have been proposed in conjunction with efforts to accumulate corpora. We call this approach corpus-based machine translation in this thesis. This thesis focuses on automatic construction of the translation knowledge needed for corpus-based MT and discusses the following three tasks. 1. Proposing a knowledge acquisition method from bilingual corpora. 2. Applying the acquired knowledge to an actual MT engine and measuring the MT quality. 3. Identifying the inherent problems of the corpus-based MT that decrease MT quality and proposing solutions. A feature of this thesis is not only investigating the first task but also investigating the second and third tasks. In order to clarify features of corpus-based MT, this thesis identifies inherent problems by translating sentences using acquired knowledge and proposes solutions. For the first task, this thesis proposes a hierarchical phrase alignment (HPA) method. This method automatically extracts equivalent phrases, which are corresponding expressions between bilingual sentences. HPA employs parsers. Previous methods extract correspondences after determining the parsing trees of the bilingual sentence, while HPA simultaneously extracts the best parsing trees and corresponding phrases by utilizing the structural similarity measure called the phrase correspondence score. HPA has two features. One is the ability to resolve ∗Doctor’s Thesis, Department of Information Processing, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DT0261006, May 10, 2004.