Sluice resolution without hand-crafted features over brittle syntax trees

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

Sluice resolution in English is the problemof finding antecedents of wh-fronted ellipses.Previous work has relied on handcraftedfeatures over syntax trees that scalepoorly to other languages and domains;in particular, to dialogue, which is one ofthe most interesting applications of sluiceresolution. Syntactic information is arguablyimportant for sluice resolution, butwe show that multi-task learning with partialparsing as auxiliary tasks effectivelycloses the gap and buys us an additional9% error reduction over previous work.Since we are not directly relying on featuresfrom partial parsers, our system ismore robust to domain shifts, giving a26% error reduction on embedded sluicesin dialogue.
Original languageEnglish
Title of host publicationProceedings, 16th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies : (Long Papers)
Volume1
PublisherAssociation for Computational Linguistics
Publication date2018
Pages236–241
Publication statusPublished - 2018
Event16th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies - New Orleans, United States
Duration: 1 Jun 20186 Jun 2018

Conference

Conference16th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies
LandUnited States
ByNew Orleans
Periode01/06/201806/06/2018

ID: 214755903