Improving Dependency Parsing Using Sentence Clause Charts

Speaker:
Vincent Kríž
Abstract:
We propose a method for improving the dependency parsing of complex sentences. This method assumes segmentation of input sentences into clauses and does not require to re-train a parser of one's choice. We represent a sentence clause structure using clause charts that provide a layer of embedding for each clause in the sentence. Then we formulate a parsing strategy as a two-stage process where (i) coordinaed and subordinated clauses of the sentence are parsed separately with respect to the sentence clause chart and (ii) their dependency trees become subtrees of the final tree of the sentence. The object language is Czech and the parser used is a maximum spanning tree parser trained on the Prague Dependency Treebank. We have achieved an average 0.97% improvement in the unlabeled attachment score.
Length:
00:53:09
Date:
10/10/2016
views:

Images:
Attachments: (video, slides, etc.)
73 MB
34 downloads
546 MB
29 downloads
82 MB
34 downloads
114 MB
33 downloads
233 MB
31 downloads