Evaluation of feature combination for effective structural disambiguation

In this paper, we present the useful features of a syntactic constituent for a probabilistic parsing model and analyze the combination of the features in order to disambiguate parse trees effectively. Unlike most of previous works focusing on the features of a single head, the features of a functional head, the features of a content head, and the features of size are utilized in this study. Experimental results show that the combination of different features such as the functional head feature and the size feature is prefered to the combination of similar features such as the functional head feature and the content head feature. Besides, it is remarkable that the function feature is more useful than the combination of the content feature and the size feature.