Efficient accurate syntactic direct translation models: one tree at a time

A challenging aspect of Statistical Machine Translation from Arabic to English lies in bringing the Arabic source morpho-syntax to bear on the lexical as well as word-order choices of the English target string. In this article, we extend the feature-rich discriminative Direct Translation Model 2 (DTM2) with a novel linear-time parsing algorithm based on an eager, incremental interpretation of Combinatory Categorial Grammar. This way we can reap the benefits of a target syntactic enhancement that leads to more grammatical output while also enabling dynamic decoding without the risk of blowing up decoding space and time requirements. Our model defines a mix of model parameters, some of which involve DTM2 source morpho-syntactic features, and others are novel target side syntactic features. Alongside translation features extracted from the derived parse tree, we explore syntactic features extracted from the incremental derivation process. Our empirical experiments show that our model significantly outperforms the state-of-the-art DTM2 system.