Abstract
We present a method for learning to perform syntactical reordering in machine translation. In our approach, source sentences are parsed into parse trees aimed at reordering source parse trees into reordered parse trees closer to target language structure. The method involves aligning words, parsing source sentences into parse trees, determining tree nodes reordering operation, and training a probability model using tree node features via machine learning model. At run-time, we parse the test sentence to obtain the parse trees, estimating reordering operation for each tree node using the trained model, and returning the sequence of words in reordered source parse tree to obtain reordered source sentence. We submit reordered source sentence to a state-of-the-art machine translation system for evaluation. We describe the implementation of the method using parallel Hong Kong corpus. The experiment results show that phrase-based machine translation model with our reordering model outperforms machine translation model without our reordering model in terms of BLEU score. Our methodology is clearly a step forward for producing more fluent and grammatical translation.