A Corpus-based Two-Way Design for Parameterized MT Systems: Rationale, Architecture and Training Issues

In many conventional MT systems, the translation output of a machine translation system is strongly affected by the sentence patterns of the source language due to the one-way processing steps from analysis to transfer and then to generation, which tends to produce literal translation that is not natural to the native speakers. The literal translation, however, is usually not suitable for direct publication to the public unless a great deal of post-editing efforts is made. In this paper, we will propose a training paradigm for acquiring the transfer and translation knowledge in a corpus-based parameterized MT system from a bilingual corpus with a two-way training method. In such a training paradigm, the knowledge is acquired from both the source sentences and the target sentences. It is thus possible to avoid the translated output from being affected by the source sentence patterns. Training methods for adapting the parameter set to the various specific user styles are also suggested for the particular needs in restricted domains. Because it provides a flexible way to adapt the system to the various domains (or sublanguages), it is expected to be a promising paradigm for producing high quality translation according to user preferred styles.