Finite Automata with Predicates for NLP

《Finite Automata with Predicates for NLP》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、FiniteAutomatawithredi atesforGertjanvanoordandDaleGerdemannune21,2000Abstra tFiniteautomataarewidelyusedinnaturallanguagepro essing.Wepresentanextensiontoniteautomata,inwhi hatomi symbolsarerepla edbyarbitrarypredi atesoversymbols.Aftermotivatingtheex-tension,weshowhowoperati

2、onssu hasinterse tion, omposition,deter-minizationandminimizationareimplemented.1
ntrodu tionFiniteautomataarewidelyusedinnaturallanguagepro essing.Wepresentanextensiontoniteautomata,inwhi hatomi symbolsarerepla edbyarbi-trarypredi atesoversymbols.Below,weshowhowvariousoperationsons

3、u hextendedautomataareimplemented.Theextensionismotivatedasfollows.
tisoftenmorenaturaltothinkofsymbolsintermsofpredi atesor lasses.Forinstan e,in omputationalphonologyitisnaturaltotalkaboutvowelsand onsonantsratherthanusetheunionoperatortoenumerateea hofthephonemesinthese lasses.And

4、innite-statesyntax,itisnaturaltousethesetofnounsorverbsoreveninnitesetssu hasthesetofunknownwordsorthesetofunknownwordsendingin-ion,et .Aprimitiveformofpredi atessu hashara ter lassesandrangeshavebeenusedinthepast.Anothermotivationforpredi atesonsymbolsis ompa tness.Withpred-i ates

5、,potentiallylargesetsoftransitionsarerepla edbyasingletransition.Forexample,ifanautomatonhastransitionsfromstateptostateqoverallASC

symbolsex eptforasymbol,forwhi hthereisatransitionfromptor,thenthereare128transitionsleavingp.Usingpredi ates,thereareonlytwotransitionsleavingp(onela

6、beledbyapredi atewhi hisonlytrueof,andonelabeledby:).Considerationsofsizebe omemoreimportantforappli ationswithlargealphabets,su hastheUni odealphabet.Evenlargeralphabetsmaysurfa einappli ationsinwhi hthesymbolsarewords.Typi alele troni di tio-narieshaveatleast200wordsandevent

7、hislargesizealphabetisnotenoughtohandleunrestri tedtexts.Realisti ally,robustsynta ti parsingrequiresaninnitealphabet.Theuseofpredi atesoversymbolssubsumesthetreatmentofthequestionmarkoperator(akatheanyorotheroperator)whi hisalreadypresentinthe al ulus(aplananday,1994;arttunen,19

8、91;vanoorda