1CS 294-5: StatisticalNatural Language ProcessingSemantics IDan KleinFeedback Your comments: Like lectures, prefer to have slides Assignments educational, but too much Java hacking Sections useful, want more of them Readings not so useful? My comments: I’m really impressed with the quality of the work! I’ve enjoyed this class immenselySome Honors (HW1) Speech Recognition: HUB WER < 6.7x Arlo Faria Dave Latham Preslav Nakov Generative PNP > 84% Dave Latham Preslav NakovCourse Updates One more missed class: Nov 1 (sorry!) In exchange, a bunch of sections: Oct 27: Agenda-based parsing Nov 10: The EM algorithm Nov 17: Machine translation TBD: Java tricks? Too late? TBD: CRFs and M3Ns Fernando Pereira visit and talk on Oct 27 (next Wednesday!)Semantics Once we’ve got a syntactic parse, then what?I’d like to buy a flight from Chicago to Denver for under $200Information Extraction Information extraction is basically role-filling The slots are particular to the application Air reservation: departure_city, arrival_city, departure_time Financial: acquired_company, hired_employee Classic information extraction systems (e.g. MUC entries), maximally distilled: Use verbs to identify which frame is present Fill the slots using syntactic and semantic cues Frames can extend across sentences (integration)I’d like to buy a flight from Chicago to Denver for under $200PURCHASE_REQUESTDEPT_CITY: ChicagoARRV_CITY: DenverDEPT_TIME: ???DATE: ???PRICE_LIMIT: $2002Semantic Roles Semantic roles: Verbs (and some nouns) express events Arguments fill roles in those events Semantic role theory models how roles pattern, how they relate to the syntax Granularity of roles Proto-agent, proto-patient (think subject and object) Fillmore’s case theory had 9 (agent, patient, location, experiencer, etc) Can subdivide them forever! Extreme view: each verb has its own set of roles buyer, bought_thing, seller, sold_thing PropBank works like this Middle view: roles are particular to a “semantic frame” like transaction Frames can be evoked by various verbs, but not too many FrameNet (here at Berkeley!) works like thisSo where’s the model? Not much work on frame filling … aside from years of IE systems, of course First broad coverage PropBank / FrameNet system was Gildeaand Jurafsky 02 How does it work? Go node by node, predicting the roles P(role|verb) is the baseline How to do better? (You tell me!)Is this Semantics? It’s certainly a step closer! You could imagine extending such a model to make inferences between sentences Can extract relational data You can do IE with such a system (sort of) It’s part of lexical semantics What’s missing? Quantifiers, negation, coordination, reference ambiguity, modality, tense and aspect… … most of what you learn about in an intro semantics course!Modeling Compositional Semantics We have no statistical model of compositional semantics In applications which extract structured data, the last step is always rule-driven For the rest of today and next class, we’re going to sketch a logical approach to compositional semantics … at least you’ll know what we’re trying to replace … this is an extension of the lambda-translation approach from the second class (except this time deeper and more interactive)Phenomena to Model Proper names Simple verbs Quantifiers Subject quantifiers Object quantifiers Reverse scope Generalized quantifiers Adjectives and adverbs wh-movement (easy and hard!) Conjunction and plurals Tenses Propositional
View Full Document