Predicting Pronoun Translation Using Syntactic, Morphological and Contextual Features from Parallel Data

We describe the systems submitted to the shared task on pronoun prediction organized within the Second DiscoMT Workshop. The systems are trained on linguistically motivated features extracted from both sides of an English-French parallel corpus and their parses. We have used a parser that integrates morphological disambiguation and which handles the REPLACE_XX placeholders explicitly. In particular, we compare the relevance of three groups of features: a) syntactic (from the English parse), b) morphological (from the French morphological analysis) and c) contextual (from the French sentence) for French pronoun prediction. A discussion on the role of these sets of features for each pronoun class is included.