Tree Topological Features for Unlexicalized Parsing

As unlexicalized parsing lacks word token information, it is important to investigate novel parsing features to improve the accuracy. This paper studies a set of tree topological (TT) features. They quantitatively describe the tree shape dominated by each non-terminal node. The features are useful in capturing linguistic notions such as grammatical weight and syntactic branching, which are factors important to syntactic processing but overlooked in the parsing literature. By using an ensemble classifierbased model, TT features can significantly improve the parsing accuracy of our unlexicalized parser. Further, the ease of estimating TT feature values makes them easy to be incorporated into virtually any mainstream parsers.