Arc Representation for Graph-based Dependency Parsing
Abstract: In this paper, we address the explicit representation of arcs in graph-based syntactic dependency parsing, departing from conventional approaches where parsing algorithms compute dependency arc scores directly from input token representations.
We propose to augment the parser with an intermediate arc representation, arguing for two main advantages.
Firstly, arc vectors encapsulate richer information, improving the capabilities of scoring functions.
Secondly, by introducing refinement layers, we allow interactions between arc representations, facilitating interactions between arcs.
We demonstrate the efficacy of this approach through evaluations on PTB and UD treebanks.
Our approach
achieves an LAS error rate reduction of 1.0\% on the PTB test set, and 1.7\% on UD, over the best SOTA model.
Paper Type: Short
Research Area: Machine Learning for NLP
Research Area Keywords: Syntax: Tagging, Chunking and Parsing / ML
Languages Studied: English, Bulgarian, Catalan, Czech, German, Spanish, French, Italian, Dutch, Norwegian, Romanian, Russian
Submission Number: 5931
Loading