Headed-Span-Based Projective Dependency Parsing

Anonymous

Abstract: We propose a new paradigm for projective dependency parsing based on headed spans. In a projective dependency tree, the subtree rooted at each word covers a contiguous sequence (i.e., a span) in the surface order. We call a span marked with a root word headed span. A projective dependency tree can be represented as a collection of headed spans. We decompose the score of a dependency tree into the scores of the headed spans and design a novel $O(n^3)$ dynamic programming algorithm to enable global training and exact inference. The advantages of our headed-span-based dependency parsing include that it captures subtree information more adequately than first-order graph-based methods and performs global optimization in decoding (in contrast to transition-based methods). We evaluate our method on PTB, CTB, and UD and it achieves competitive results in comparison with previous methods.