Learning Compositional Semantics for Introduction Open Domain - PowerPoint PPT Presentation

Learning Compositional Semantics Phong Le, Willem Zuidema Learning Compositional Semantics for Introduction Open Domain Semantic Parsing Meaning Representation Semantic Composition Experimental results Groningen Meaning Phong Le, Willem Zuidema Bank Geoquery Conclusion Institute for Logic, Language and Computation University of Amsterdam October 31, 2012

Learning Outline Compositional Semantics Phong Le, Willem Zuidema Introduction Introduction Meaning Representation Semantic Composition Experimental results Meaning Representation Groningen Meaning Bank Geoquery Conclusion Semantic Composition Experimental results Groningen Meaning Bank Geoquery Conclusion

Learning Does Google understand what I mean? Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition Experimental results Groningen Meaning Bank Geoquery Conclusion

Learning Even people misunderstand... Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition Experimental results Groningen Meaning Bank Geoquery Conclusion

Learning What should we do? Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition Experimental results Semantic Parsing (or Semantic Analysis) Groningen Meaning Bank Translate natural language sentences into their computer Geoquery Conclusion executable meaning representations . Example Which states border Arizona ? answer(A,(state(A),const(B,stateid(arizona)),next to(A,B)))

Learning Common Strategy Compositional Semantics Phong Le, Willem Zuidema Introduction Principle of Compositionality Meaning Representation “The meaning of a whole is a function of the meanings Semantic Composition of the parts and of the way they are syntactically Experimental results Groningen Meaning combined.” Bank Geoquery Conclusion

Learning Traditional approach: with lambda calculus Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Lambda calculus Representation is an elegant tool for semantic composition in a bottom up manner Semantic Composition Experimental results Groningen Meaning Bank Geoquery John :- λ x . john ( x ) Conclusion walks :- λ P .λ y . walks ( y ) ∧ P y John walks :- ( λ P .λ y . walks ( y ) ∧ P y ) ( λ x . john ( x )) :- λ y . walks ( y ) ∧ ( λ x . john ( x )) y :- λ y . walks ( y ) ∧ john ( y )

Learning Why learning semantic parsing? Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition Speech recognition and syntactic analysis have had significant Experimental results development under the umbrella of machine learning, thanks to Groningen Meaning Bank ◮ the power of machine learning tools (e.g. Hidden Markov Geoquery Conclusion Model, Expectation Maximization) ◮ large corpora (e.g. WSJ) How about semantic parsing? a complicated story...

Learning Domain-dependent semantic parsing Compositional Semantics Phong Le, Willem Geoquery Zuidema Introduction Meaning Representation Semantic Composition Experimental results Groningen Meaning Bank Geoquery Conclusion Features closed world, simple present tense, wh-question No need to handle anaphora, possibility/necessity, tense, event,...

Learning Learning approaches Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning ◮ Supervised Representation ◮ fully supervised (MRs are available) Semantic Composition ◮ Structured learning with CCG Experimental results Groningen Meaning ◮ Syntax-based Machine translation Bank ◮ Kernel-based approach Geoquery ◮ Integrating syntax and semantics Conclusion ◮ weakly supervised (response-driven) ◮ Clarke et al. (2010) ◮ Liang et al. (2011) ◮ Semi-supervised ◮ Kernel-based approach ◮ Unsupervised ◮ Confidence driven semantic parsing

Learning Open-domain semantic parsing Compositional Semantics Phong Le, Willem Zuidema Introduction Learning open-domain semantic parsing Meaning Representation is still largely unexplored, because of many difficulties Semantic Composition ◮ need to handle various linguistics phenomena and syntactic Experimental results Groningen Meaning structures Bank Geoquery Conclusion In addition: presupposition, anaphora, etc. ◮ lack large standard corpora

Learning In this paper Compositional Semantics Phong Le, Willem Zuidema We want to bridge this gap! Introduction Meaning by introducing a new learning open-domain semantic parsing Representation approach: Dependency-based Semantic Composition using Graphs Semantic Composition (DeSCoG) Experimental results Groningen Meaning Bank Outline Geoquery Conclusion ◮ Meaning representation with graph-based variant of Discourse Representation Structures ◮ remove the need of the lambda calculus ◮ Semantic composition ◮ use existing state-of-the-art syntactic dependency parsers ◮ with a probability model ◮ Experimental results on ◮ Groningen Meaning Bank ◮ Geoquery

Learning Why abandon the lambda calculus? Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition How to learn lexicon? Experimental results Given Groningen Meaning Bank John walks :- λ y . walks ( y ) ∧ john ( y ) Geoquery Conclusion how to find lambda forms for John and walks ? Notorious problem!!! ⇒ Easy for composition, but difficult for learning lexicon! Our idea Not so difficult for composition, but easy for learning lexicon!

Learning Why use existing syntactic dependency parsers? Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition Experimental results ◮ dependency structures encode predicate-argument relations Groningen Meaning Bank which are strongly related to semantics Geoquery Conclusion ◮ the total complexity is reduced significantly compared with parsing syntax and semantics simultaneously ◮ prior knowledge of syntax is particularly helpful when sentences are long and complex

Learning Outline Compositional Semantics Phong Le, Willem Zuidema Introduction Introduction Meaning Representation Semantic Composition Experimental results Meaning Representation Groningen Meaning Bank Geoquery Conclusion Semantic Composition Experimental results Groningen Meaning Bank Geoquery Conclusion

Learning Compositional Semantics Phong Le, Willem Zuidema Discourse Representation Structure (DRS) Introduction is used to represent a mental representation of the hearer as the Meaning Representation discourse unfolds. Semantic Composition Example Experimental results Groningen Meaning Mary loves a man. Bank Geoquery Conclusion x, y mary (x) man (y) love (x,y) Our goal is to assign as-good-as-possible DRS to unseen sentences.

Learning How to evaluate success? Compositional Semantics Phong Le, Willem Zuidema x Introduction jones (x) Meaning 1. If Jones sees a ball, he will kick it. Representation y ⇒ Semantic Composition kick (x,y) ball (y) Experimental results see (x,y) Groningen Meaning Bank Geoquery u Conclusion jones (u) 2. Jones will see a ball or a cake. v t ∨ ball (v) cake (t) see 1 (u,v) see 2 (u,t) The best alignment A is A ( x ) = u , A ( y ) = v , A ( jones ) = jones , A ( ball ) = ball , A ( see ) = see 2 A ( outerbox ) = outerbox , A ( leftbox ⇒ ) = leftbox ∨ A ( rightbox ⇒ ) = rightbox ∨

Learning Compositional Semantics Phong Le, Willem Zuidema x Introduction jones (x) Meaning Representation 1. If Jones sees a ball, he will kick it. y Semantic Composition ⇒ kick (x,y) ball (y) Experimental results Groningen Meaning see (x,y) Bank Geoquery Conclusion u jones ∗∗ ( u ) 2. Jones will see a ball or a cake. v t ∨ ball ∗ ( v ) cake (t) see ∗ 1 (u,v) see 2 (u,t)

Learning Compositional Semantics Phong Le, Willem Zuidema x Introduction jones (x) Meaning Representation 1. If Jones sees a ball, he will kick it. y Semantic Composition ⇒ kick (x,y) ball (y) Experimental results Groningen Meaning see (x,y) Bank Geoquery Conclusion u jones ∗∗ ( u ) 2. Jones will see a ball or a cake. v t ∨ ball ∗ ( v ) cake (t) see ∗ 1 (u,v) see 2 (u,t) Ω( DRS 1 , DRS 2) = 4 recall = Ω( DRS 1 , DRS 2) 10 , prec = Ω( DRS 1 , DRS 2) 4 4 Ω( DRS 1 , DRS 1) = Ω( DRS 2 , DRS 2) = 12 , fscore = 0 . 36

Learning Does it fit our intuition? Compositional Semantics Phong Le, Willem x Zuidema jones (x) Introduction 1. If Jones sees a ball, he will kick it. Meaning y Representation ⇒ kick (x,y) ball (y) Semantic Composition see (x,y) Experimental results Groningen Meaning Bank Geoquery u Conclusion jones ( u ) 2. Jones will see a ball or a cake. v t ∨ ball ( v ) cake (t) see 1 (u,v) see 2 (u,t) which one is more similar to l jones (l) 3 If Jones sees a ball, he will see a cake. h k ⇒ ball (h) cake (k) see 1 (l,h) see 2 (l,k)

Learning Does it fit our intuition? Compositional Semantics Phong Le, Willem Zuidema Introduction Meaning Representation Semantic Composition Experimental results Human intuition Groningen Meaning Bank DRS1 is more similar to DRS3 than DRS2 to DRS3 Geoquery Conclusion The measure f-score( DRS 1 , DRS 3 ) = 16 22 = 0 . 73 and f-score( DRS 2 , DRS 3 ) = 12 24 = 0 . 5; hence f-score( DRS 1 , DRS 3 ) > f-score( DRS 2 , DRS 3 )