Greedy direct decoding algorithm for syntax-based tree-to-string statistical translation model

In order to effectively direct the translation process by syntax information, a greedy direct decoding algorithm is proposed for the syntax-based tree-to-string statistical translation model. The log-linear model is adopted as the framework and the feature functions are defined upon the translation model probability, the language model probability and the null translation penalty. The decoder firstly generates the initial translation gloss, and then improves the gloss by iteratively traversing the parse tree. The scoring methods for translation segments are described. The experiment was carried out on IWSLT 2004 data set. The translation results were evaluated by the BLEU metrics. Experimental results show that the greedy direct decoding algorithm gives better results than the current reverse decoding algorithm on translation quality and speed. This means that the greedy direct decoding algorithm can make more efficient use of syntactical information, thus is more suitable for the tree-to-string statistical translation model.