Improving Dependency Parsing with Interlinear Glossed Text and Syntactic Projection

Producing annotated corpora for resource-poor languages can be prohibitively expensive, while obtaining parallel, unannotated corpora may be more easily achieved. We propose a method of augmenting a discriminative dependency parser using syntactic projection information. This modification will allow the parser to take advantage of unannotated parallel corpora where high-quality automatic annotation tools exist for one of the languages. We use corpora of interlinear glossed text—short bitexts commonly found in linguistic papers on resource-poor languages with an additional gloss line that supports word alignment—and demonstrate this technique on eight different languages, including resource-poor languages such as Welsh, Yaqui, and Hausa. We find that incorporating syntactic projection information in a discriminative parser generally outperforms deterministic syntactic projection. While this paper uses small IGT corpora for word alignment, our method can be adapted to larger parallel corpora by using statistical word alignment instead.