Why Corpus-Based Statistics-Oriented Machine Translation

Rule-based approaches have been the dominant paradigm in developing MT systems. Such approaches, however, suffer from difficulties in knowledge acquisition to meet the wide variety and time-changing characteristics of the real text. To attack this problem, some statistical translation models and supporting tools had been developed in the last few years. However, a simple statistical model often results in a large parameter space and thus requires a large training corpus. Therefore, it is required to introduce language models that take advantages of well-justified linguistic knowledge to make stochastic MT systems practical. A stochastic model that emphasizes the adoption of well-justified linguistic knowledge in developing the model is called a corpus-based statistics-oriented approach. In this paper, corpus-based statistics-oriented paradigm is proposed, its characteristics is compared with other methodologies. The recent progress in some corpus-based statistics-oriented models for MT are also reviewed.