Biaffine Discourse Dependency Parsing

TL;DR

This paper explores biaffine models for neural discourse dependency parsing, demonstrating improved performance and analyzing structural properties of generated trees compared to gold standards.

Contribution

It introduces a biaffine approach for discourse dependency parsing and compares algorithms, showing superior results with the Chu-Liu-Edmonds method and analyzing tree structures.

Findings

Chu-Liu-Edmonds algorithm yields deeper trees and better performance

Parser-generated trees are structurally close to gold trees

Dependency structures have limited non-projectivity features

Abstract

We provide a study of using the biaffine model for neural discourse dependency parsing and achieve significant performance improvement compared with the baseline parsers. We compare the Eisner algorithm and the Chu-Liu-Edmonds algorithm in the task and find that using the Chu-Liu-Edmonds algorithm generates deeper trees and achieves better performance. We also evaluate the structure of the output of the parser with average maximum path length and average proportion of leaf nodes and find that the dependency trees generated by the parser are close to the gold trees. As the corpus allows non-projective structures, we analyze the complexity of non-projectivity of the corpus and find that the dependency structures in this corpus have gap degree at most one and edge degree at most one.