Paper: Rich Prior Knowledge in Learning for Natural Language Processing

ACL ID P11-5005
Title Rich Prior Knowledge in Learning for Natural Language Processing
Venue Annual Meeting of the Association of Computational Linguistics
Session Tutorial Abstracts
Year 2011
Authors

We presentan approachto grammarinduction that utilizessyntacticuniversalsto improve dependency parsingacrossa rangeof languages. Our methoduses a single set of manually-specifiedlanguage-independent rulesthatidentifysyntacticdependenciesbetweenpairsof syntacticcategoriesthatcommonlyoccuracrosslanguages.Duringinfer- monlyoccuracrosslanguages.Duringinferenceof theprobabilisticmodel,weuseposteriorexpectation constraintstorequirethata minimumproportionof thedependencieswe inferbeinstancesoftheserules.Wealsoautomaticallyrefinethesyntacticcategoriesgiven maticallyrefinethesyntacticcategoriesgiven in ourcoarselytaggedinput. Acrosssixlanguagesourapproachoutperformsstate-of-the- guagesourapproachoutperformsstate-of-theartunsupervisedmethodsbyasignificantmar- artunsupervisedmethodsbyasignificantmarg...