Research On Auto-Construction Of EBMT Translation Model

The Example-Based Machine Translation (EBMT) system can be developed in a short period with relatively better translation result for a given domain. In translation searching process, many EBMT systems can only rely on a heuristic guidance which is given by human. Therefore that process will not be objective and always lean much on the intuition of system developer and may be overfitting to a special domain. It is hard to maintain the performance when the system is transplanted to other domains. Moreover, the general EBMT system can not model the rich features in translation process.This thesis tries to construct the translation model of EBMT automatically by using machine learning algorithm (here, Maximum Entropy). In order to adequately incorporate different kinds of information which can be explored from examples, this thesis introduces a log-linear translation model into EBMT. In addition, a high dimensional feature space is formally constructed to include general features of different aspects.In order to get the translations space as full as possible, a controllable search algorithm has been proposed in this thesis. The translation auto-evaluation metric BLEU is suitable for whole document translation evaluation. But many problems will arise when the evaluation objective is single sentence. To meet the need of single sentence evaluation in the research, this thesis presents an amendment to BLEU metric, which is also useful for related research topics.The preliminary experimental results indicate that the maximum entropy framework based EBMT translation model proposed in this thesis outperforms the base system and the rich features can supply much useful information for translation process. In contribution analysis, the result shows that word-level features are the most useful information.