Research On Online Adaptation Based Machine Translation Post-Edit

Batch learning was usually used for parameter optimization in traditional machine translation model. With the concept of sparse feature, batch learning could no longer meet the needs now, so online learning is gradually entering people's field of vision. Some researchers have come up with the framework of online adaptation that enables the feature weights to be adjusted and the feature function modified in real time. On the other hand, automatic post-editing of machine translation have been shown that can significantly improve the efficiency of human editors and improve the translation quality to some extent. Therefore this paper researches on online adaptation based machine translation post-edit. The main contents includes the following several aspects:(1) Research on online adaptation based machine translation post-edit model. Compare traditional batch learning with online adaptation while using the same data from the respective of the approaches. Make a comparison between automatic post-edit system and machine translation system from the angle of the systems. Explore three model parameters in online adaptation including iterations, length of the k-best list and the max step size in point of the models.(2) Research on active learning based online adaptation post-edit. This section proposed a method of active learning and applied it to the learning process of online adaptation. Firstly calculate the sentence-level BLEU score from the beginning of the learning process and at the end of the process, then annotate the label for each sentence by using the BLEU gains. Next, keep all positive sentences for active learning of the next stage. Furthermore, save the feature weights from the best test set and decode the training set in order to check the performance of the active learning method while facing to large-scale test set.(3) Research on filtering rule table based online adaptation post-edit. Come up with a method of filtering rule table in order to improve the translation quality. Respectively extract the decoding rules from the development set and test set, then explore the effects of decoding performance on each rule. Label positive instances for the rules that increases decoding gains and negative instances for the ones that decreases gains. After defining the labels, make use of TM features and features with custom to train the classification model of SVM and finally predict the table of decoding rules for the test set. In fact, this approach belongs to a way of optimizing the searching path while decoding.