Arc Representation for Graph-based Dependency Parsing

Anonymous

Abstract: In this paper, we address the explicit representation of arcs in syntactic dependency parsing, diverging from conventional approaches where parsing algorithms directly manipulate dependency arc scores derived from input token representations. We propose augmenting the parser with an intermediate arc representation, arguing for two main advantages. Firstly, arc vectors encapsulate richer information, enhancing the capabilities of subsequent scoring functions. Secondly, by introducing refinement layers, we enable interactions among vector representations, facilitating the consideration of global long-range dependencies. We demonstrate the efficacy of this approach through empirical evaluations on PTB and UD dependency treebanks.

Paper Type: short

Research Area: Syntax: Tagging, Chunking and Parsing / ML

Languages Studied: English, Bulgarian, Catalan, Czech, German, Spanish, French, Italian, Dutch, Norwegian, Romanian, Russian