A Rich-label Constituency Tree for Constituency Parsing

摘要

In this paper, we propose to improve the performance of constituency parsing by incorporate information from dependency trees. Our main contribution is a simple method to combine the information in the two type of trees, which is able to resolve discrepancies between them. This results in a new method to improve constituency parsing performance that is orthogonal to exiting methods, such as using pre-trained word vectors and using external training data. With existing dependency tree conversion tools and exiting constituency parsers, we show by experiment that our method achieves performance on top of a state-of-the-art constituency parser.