Coherence in Text and DialogueWhy is Coherence a Computational Issue?Question AnsweringText ExtractionText SummarizationDiscourseTwo Kinds of KnowledgeTask/Domain Knowledge - ScriptsTask/Domain Knowledge – Pre and Post ConditionsPre and Post Conditions May Involve BeliefsLinguistic CuesThe Structure of Text – RelationsThe Structure of Text – Hanging TogetherThe Structure of Dialogue – Typed vs. SpokenThe Structure of Dialogue – Speech ActsSpeech Acts – Three LevelsSpeech Acts – An ExampleSpeech ActsIndirect Speech ActsPlanning Speech ActsConversational ImplicatureGrice’s Maxims at WorkSlide 23ExamplesInterpretation as ReasoningScalar ImplicaturePresuppositionsCoherence in Text and DialogueRead J & M Chapters 18.2, 18.3, 19Why is Coherence a Computational Issue?•Anaphora resolution•Question answering•Text extraction•Text summarization•Discourse•GenerationQuestion AnsweringJohn arrived at the station later than he’d hoped. He rushed to the ticket booth, then ran to the platform. He got there just before the conductor closed the doors.Did John get on the train?Did John have enough money for the ticket?Or a news wire example: http://www.csi.uottawa.ca/tanka/QA/sample.htmlText ExtractionThe alphas have a long-standing hatred of the betas. Their leaders have decided that the time has come to launch an attack. They are considering the Thanksgiving weekend, but that’s long ways away. They know they’ll have to count on the fact that they are much better shots than the betas are.attack (status <planned, imminent, active, completed>)(attacker )(attackee )(location )(mode-of-attack )(time-of-attack )(credit-claimer )Text SummarizationGenerating long summaries:http://www1.cs.columbia.edu/nlp/newsblaster/ Suppose Google (http://channels.netscape.com/ns/search/default.jsp) returned a summary of a page, not just the first line:domestic robotsDiscourseWhat we don’t want:http://www.dcs.shef.ac.uk/~walker/Seinf.mpg What we do want:I want to see The Matrix tonight.It isn’t playing anywhere in town.Two Kinds of Knowledge•Task/domain knowledge:John offered to go stop by Bill’s house to pick up the report. He didn’t have a key but the door was unlocked.•Linguistic/social conventions that govern the way we use language:•Local linguistic clues: John went to the mall but he didn’t buy anything. * John went to the mall so he didn’t buy anything.•Conventions: Do you know what time it is?Can we come up with a single mechanism for representing and reasoning with all of these things?Task/Domain Knowledge - ScriptsScripts or frames represent facts about the world as we know it:Example: a train-taking scriptScene 1 (arrival): S arrives at stationScene 2: (ticket purchase): S goes to ticket window S tells C destination and class C tells S the price S hands over money C hands over ticketScene 3: (boarding): S goes to platform S gets on train P closes doors train pulls away Scene 4: travellingetc.Task/Domain Knowledge – Pre and Post ConditionsActions are especially important, particularly in interactive dialogue systems. Actions can be represented as:•Preconditions•Postconditions•Body (sequence of subactions)The first two may be able to be represented in FOPC. Sometimes the latter is too, or in a separate action language. But we may need a higher order system.Pre and Post Conditions May Involve BeliefsDo you know what time it is?(Action (Tell, P, R, A) **P tells R A(precondition: Know(P, A) Contact(P, R))(postcondition: Know(R, A) or maybeKnow (R, Believe(P, A)))Know(Tim, time(now, 2:00)) or Know(Tim Believe (I, time(now, 2:00)))Linguistic CuesJohn went to the mall but he didn’t buy anything.John went to the mall so he didn’t go to the movie.John went to the mall and he went to the movie.John went to the mall after he went to the movie.John went to the mall so that he could go to the movie.John went to the mall even though he didn’t have any money.The Structure of Text – RelationsThe pieces have to hang together, typically in some tree like structure that represents relations among the pieces. Some relations:Result: John forgot to set his alarm. So he missed the exam.Explanation: John missed the exam. He’d forgotten to set his alarm.Parallel/contrast: John always brings ham and cheese. Mary always brings peanut butter.Elaboration: John went to the Friday night dance. He brought Mary and arrived in a white limousine.Sequence: The police arrested a tall man for the burglary. They brought him to the station and read him his rights. Then they locked him up until morning.The Structure of Text – Hanging TogetherThe police are searching for a tall white man. They canvassed the neighborhood and interviewed everyone they could find. Then they examined several security cameras in the area. They think the robbery was caught on at least one of them. They are also requesting cell phone records at the time of the robbery. They are reasonably certain that they will be able to find their man.The Structure of Dialogue – Typed vs. SpokenIn addition to all of the issues we must confront with typed dialogues, spoken dialogue management must also consider:•Conventions for interruptions and turn taking.•The need for grounding to confirm success over a noisy channel•False starts•UngrammaticalityWhen’s the next um or any cheap flights to Boston?Let’s see. There’s one tomorrow at 8:00.The Structure of Dialogue – Speech ActsTalking (or writing) is just one kind of action that we can take when we want to achieve a goal. So we should analyze each utterance of a dialogue just as we would analyze any other action. Some utterances have an effect just by virtue of being uttered:I name this ship the Titanic.I second that motion.I bet you five dollars that it will snow.I hereby announce my candidacy for governor.Speech Acts – Three LevelsBut most utterances are actions because of the effect they have on their listeners. We can analyze utterances on three levels:Locutionary act: the utterance of some sentence with some particular linguistic meaning.Illocutionary act: the act of asking or answering or telling or ordering promising or whatever.Perlocutionary act: the production of some specific effect on the addressee as a result of the act.Speech Acts – An ExampleI’m freezing.Locutionary force: DCL(e time(e, now) cold(e) AE(e, speaker))Illocutionary force:
View Full Document
Unlocking...