Reasoning about Entailment with Neural Attention

Reasoning about Entailment with Neural Attention

ID:41220310

大小:1.22 MB

页数:9页

时间:2019-08-19

Reasoning about Entailment with Neural Attention_第1页
Reasoning about Entailment with Neural Attention_第2页
Reasoning about Entailment with Neural Attention_第3页
Reasoning about Entailment with Neural Attention_第4页
Reasoning about Entailment with Neural Attention_第5页
资源描述:

《Reasoning about Entailment with Neural Attention》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、ReasoningaboutEntailmentwithNeuralAttentionTimRocktaschel¨EdwardGrefenstette&KarlMoritzHermannUniversityCollegeLondonGoogleDeepMindt.rocktaschel@cs.ucl.ac.ukfetg,kmhg@google.comToma´sKoˇciskˇy´PhilBlunsomUniversityofOxfordGoogleDeepMind&UniversityofOxfordtomas@kocisky.e

2、upblunsom@google.comAbstractAutomaticallyrecognizingentailmentrelationsbetweenpairsofnaturallanguagesentenceshassofarbeenthedominionofclassifiersemployinghandengineeredfeaturesderivedfromnaturallanguageprocessingpipelines.End-to-enddiffer-entiableneuralarchitectureshavef

3、ailedtoapproachstate-of-the-artperformanceuntilveryrecently.Inthispaper,weproposeaneuralmodelthatreadstwosen-tencestodetermineentailmentusinglongshort-termmemoryunits.Weextendthismodelwithaword-by-wordneuralattentionmechanismthatencouragesrea-soningoverentailmentsofpair

4、sofwordsandphrases.Furthermore,wepresentaqualitativeanalysisofattentionweightsproducedbythismodel,demonstratingsuchreasoningcapabilities.Onalargeentailmentdatasetthismodeloutperformsthepreviousbestneuralmodelandaclassifierwithengineeredfeaturesbyasub-stantialmargin.Itist

5、hefirstgenericend-to-enddifferentiablesystemthatachievesstate-of-the-artaccuracyonatextualentailmentdataset.1IntroductionTheabilitytodeterminethesemanticrelationshipbetweentwosentencesisanintegralpartofma-chinesthatunderstandandreasonwithnaturallanguage.Recognizingtextua

6、lentailment(RTE)isthetaskofdeterminingwhethertwonaturallanguagesentencesare(i)contradictingeachother,(ii)notrelated,orwhether(iii)thefirstsentence(calledpremise)entailsthesecondsentence(calledarXiv:1509.06664v1[cs.CL]22Sep2015hypothesis).Thistaskisimportantsincemanynatur

7、allanguageprocessing(NLP)problems,suchasinformationextraction,relationextraction,textsummarizationormachinetranslation,relyonitexplicitlyorimplicitlyandcouldbenefitfrommoreaccurateRTEsystems[Daganetal.,2006].State-of-the-artsystemsforRTEsofarreliedheavilyonengineeredNLPp

8、ipelines,extensiveman-ualcreationoffeatures,aswellasvariousexternalresourcesandspecializedsubcomponentssuchasn

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。