Improving the parsing of French coordination through annotation standards and targeted features
Assaf Urieli · 2014
In the present study we explore various methods for improving the transition-based parsing of coordinated structures in French. Features targeting syntactic parallelism in coordinated struc-tures are used as additional features when training the statistical model, but also as an efficient means to find and correct annotation errors in training corpora. In terms of annotation, we com-pare four different annotations for coordinated structures, demonstrate the importance of globally unambiguous annotation for punctuation, and discuss the decision process of a transition-based parser for coordination, explaining why certain annotations consistently out-perform others. We compare the gains provided by different annotation standards, by targeted features, and by using a wider beam. Our best configuration gives a 37.28 % reduction in the coordination error rate, when compared to the baseline SPMRL test corpus for French after manual corrections. 1