文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Hierarchical Language Models for One-Stage Speech Interpretation
侵权投诉

Hierarchical Language Models for One-Stage Speech Interpretation

This paper presents a robust semantic model for one-stage interpretation of natural speech. Our semantic analysis uses no explicit syntactic and morphologic knowledge, which seems sufficient for narrow application domains. In contrast to previous approache

HierarchicalLanguageModelsforOne-StageSpeechInterpretation

MatthiasThomae,TiborFabian,RobertLieb,GüntherRuske

InstituteforHuman-MachineCommunicationTechnischeUniversitätMünchen,Germany

eurospeech05@thomae-privat.de,{fab,lie,rus}@mmk.ei.tum.de

Abstract

Thispaperpresentsarobustsemanticmodelforone-stageinterpretationofnaturalspeech.Oursemanticanalysisusesnoexplicitsyntacticandmorphologicknowledge,whichseemssuf cientfornarrowapplicationdomains.Incontrasttopre-viousapproaches,oursemanticmodelisembeddedinauni-form,hierarchical,stochasticmodelingframeworktogetherwithacoustic-phoneticandlexicalknowledge,andsemanticrepresentationsarecomputeddirectlyfromacousticobserva-tionsthroughaone-stagedecodingprocess.Thedecoderpro-ducesahierarchical(tree-)structureofwordsandsemanticcat-egorysymbolsbyuseoftheso-calledhierarchicallanguagemodel(HLM).WediscussgenerationofHLMbymixingrule-basedanddata-drivenlanguagemodel(LM)generationtech-niques,namelyweightedregularexpressions,n-gramsandex-actLM.DifferentHLMcon gurationswithvaryingdiscount-ingtechniques,n-gramordersandscalingfactorsareexamined.Experimentswereconductedwithanairportinformationdia-logueapplication.TheevaluationresultsarebasedonHLMperplexityandourpreviouslypublishedsemantictreeaccuracy.

1.Introduction

In[1]weintroducedaOne-stageDecoderforInterpretationofNaturalSpeech(ODINS),whichtightlyintegratesautomaticspeechrecognition(ASR)andnaturallanguageunderstand-ing(NLU)techniquesinaone-stagedecodingprocess.Ex-perimentscon rmedthattheone-stageapproachcanbefavor-ableovertypicallyusedmulti-stagemethodsbyavoidingerrorscausedbydecisionsinearlyprocessingstages.Incontrasttotypicalhybridspeechunderstandingapproaches,ODINSoper-atesonasingle,uniformmodelforacoustic-phonetic,lexicalandsyntactic-semanticknowledge.Uniformityisappealingbe-causeitenablestheintegrationofdifferenttypesofknowledgemodelswithouthavingtomodifythedecoder,aslongastheex-pressivepoweroftheunderlyingframeworkissuf cient.Fur-thermore,decodercomplexityispotentiallylower,asonlyonealgorithmicframeworkneedstobeconsidered,insteadofmany.Onthedownside,ahybridmodelingapproachmaylenditselftomoreef cientdecoding.

Ouruniformmodelingapproachisrealizedasastochas-ticcontext-freegrammarrepresentationthroughweighedtransi-tionnetworkhierarchies(WTNH).WTNHconsistoftransitionnetworkswhosenodeseitherrepresentterminalsymbolsorre-fertoothertransitionnetworks.WTNHaresimilartostochas-ticrecursivetransitionnetworks,buttheyemploytheMooremachinerepresentationinsteadoftheMealyone.Asanotherfundamentaldifference,WTNHarelogicallydividedintohier-

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email