Feature Exploration for Cross-Lingual Pronoun Prediction

Abstract

We explore a large number of features for cross-lingual pronoun prediction for translation between English and German/ French. We find that features related to German/French are more informative than features related to English, regardless of the translation direction. Our most useful features are local context, dependency head features, and source pronouns. We also find that it is sometimes more successful to employ a 2-step procedure that first makes a binary choice between pronouns and other, then classifies pronouns. For the pronoun/other distinction POS ngrams were very useful.