6.863J Natural Language ProcessingLecture 15: The meaning of it allInstructor: Robert C. [email protected]•863J/9•611J SP04 Lecture 15The Menu Bar• Administrivia:• Lab 3b out; due April 12• Lab 4a on lexical semantics, out April 12Agenda:What does this all mean?Frege’s principle of compositionalityRepresentation and lambda calculus6•863J/9•611J SP04 Lecture 15Cognition as computation• Computation manipulates formal symbols• The symbols are represented• The symbol manipulation is purely syntactic• The symbol manipulation is semantically invariant6•863J/9•611J SP04 Lecture 15Our general view• Syntactic representations to…• Semantic representations to…• Conceptual representations…6•863J/9•611J SP04 Lecture 15We know…• What syntactic representations are• We know much less about semantic or conceptual representations, but…• Assume: they are the representations and vehicle for reasoning…• So…must preserve what?• Should be built up compositionally• Why?6•863J/9•611J SP04 Lecture 15Compositionality, Turing, and all that• Brown cow Æ• Meaning(Brown) & Meaning(cow) & some mode of composition• Why?• Cf: Purple cow6•863J/9•611J SP04 Lecture 15Easy case• Bob sleeps• Bob likes ice-cream• Event: likes(Bob, ice-cream)6•863J/9•611J SP04 Lecture 15Hard case(But the Accord was redesigned for the 2003 model year.)The roomier, faster, and sleeker sedan’s sales stabilized last year,falling by just 1,230 units -- a strong showing in a market that saw combined total passenger car sales fall by 471,000 units.6•863J/9•611J SP04 Lecture 15The envelope please…the(x1,e1&e3&e5&e7) & more’(e1,x1,y1,e2) & roomy’(e2,x1) & more’(e3,x1,y1,e4) & fast’(e4,x1) & more’(e5,x1,y1,e6) & sleek’(e6,x1) & sedan’(e7,x1) & poss(x1,z1) & sale(z1,x2) & Plur(z1,s1) & stabilize’(e8,s1) & Past(e8) & at-time(e8,y2) & last(y2,u1) & year(y2) & fall’(e9,s1) & by(e9,s2) & just(e6) & card’(e6,s2,1230) & unit(u2) & Plur(u2,s2) & Appos(e8,e11) & a(e11,e10&e11) & strong’(e10,e11) & show’(e11,x3,x4) & in(e10,m) & a(m,e12&e13) & market’(e12,m) & see’(e13,m,e14) & Past(e13) & combine(x5,s3) & total(s3) & passenger(p) & nn(p,c) & car(c) & nn(c,z2) & sale(z2,x6) & Plur(z2,s3) & fall’(e14,s3) & by(e14,s4) & card(s4,471000) & unit(u3) & Plur(u3,s4)6•863J/9•611J SP04 Lecture 15Why: recover meaning from structure –syntax-directed translationSNP VPVNPBoblikesice-cream=λy.likes(y, ice-cream)VP(NP)=likes (Bob , , ice-cream)ice-creamBobλyλx.likes(x, y)6•863J/9•611J SP04 Lecture 15How: function applicationSNP VPVNPBoblikesice-cream=λy.likes(y, ice-cream)VP(NP)=likes (Bob , , ice-cream)ice-creamBobλyλx.likes(x, y)6•863J/9•611J SP04 Lecture 15What’s meaning? What’s semantics –2 ends of the spectrum• Answer 1: whatever it is, it’s mapping (translation) between representations And it depends on all of the text• Answer 2: whatever it is, our answer depends on a much more focused task-specific question, viz., information extraction from texts• Perhaps call this ‘natural language engineering’• These two ends of the spectrum have different characteristics, and difft uses• Deep vs. Shallow?6•863J/9•611J SP04 Lecture 15What Counts as Understanding?some notions• We understand statement if we know how to determine its truth• What are exact conditions under which it would be true?• necessary + sufficient• Equivalently, derive all its consequences • what else must be true if we accept the statement?• Philosophers tend to use this definition• We understand statement if we can use it to answer questions [very similar to above – requires reasoning]• Easy: John ate pizza. What was eaten by John?• Hard: White’s first move is P-Q4. Can Black checkmate?• Constructing a procedure to get the answer is enough6•863J/9•611J SP04 Lecture 15What Counts as Understanding?• Be able to translate• Depends on target language• English to English? bah humbug!• English to French? reasonable• English to Chinese? requires deeper understanding• English to logic? deepest all humans are mortal = ∀x [human(x) ⇒mortal(x)]• Assume we have logic-manipulating rules to tell us how to act, draw conclusions, answer questions …6•863J/9•611J SP04 Lecture 15Answer 1: translation – from ‘syntactic’ rep to ‘semantic’ rep, aka “Deep”• Mirrors the progamming language approach• When is it used?• DB Q&A (but answer 2 can be used here…when and how?)• Text understanding: when all the text is relevant - voice, inference, paraphrase, important• Intentions, beliefs, desires (non-extensional= not just sets of items)6•863J/9•611J SP04 Lecture 15What requirements must meaning representations fulfill?• Verifiability: The system should allow us to compare representations to facts in a Knowledge Base (KB)• Cat(Huey)• Ambiguity: The system should allow us to represent meanings unambiguously• ‘German teachers’ has 2 representations• Vagueness: The system should allow us to represent vagueness• He lives somewhere in the south of France.6•863J/9•611J SP04 Lecture 15Requirements: Canonical Form• Inputs that mean the same thing have the same representation.• Huey eats kibble.• Kibble, Huey will eat.• What Huey eats is kibble.• It’s kibble that Huey eats.• Alternatives• Four different semantic representations• Store all possible meaning representations in Knowledge Base6•863J/9•611J SP04 Lecture 15Requirements: Semantic Ambiguity• Parallel to syntactic ambiguity•Happy [cats and dogs] live on the farm•[Happy cats] and dogs live on the farm• Independent of syntactic structure•Every boy loves a dog• “all boys love a single dog”• “foreach boy, there is a dog he loves”6•863J/9•611J SP04 Lecture 15Requirements: Inference• Draw valid conclusions based on the meaning representation of inputs and its store of background knowledge.Does Huey eat kibble?thing(kibble)Eat(Huey,x) ^ thing(x)6•863J/9•611J SP04 Lecture 15Word Senses & Ambiguity• Q: Can the basic unit of meaning rep be a word?• A: No, words have different senses• Example: gohas many senses (to move, depart, pass, vanish, reach, extend, …)• Senses are organized into an ontology6•863J/9•611J SP04 Lecture 15Requirements: Word Senses• Ontology• Example: Aristotle’s classes• substance (physical objects)•
View Full Document
Unlocking...