Driving Semantic Parsing from the World’s Response James Clarke , Dan Goldwasser, Ming-Wei Chang, Dan Roth Cognitive Computation Group University of Illinois at Urbana-Champaign CoNLL 2010 Clarke, Goldwasser, Chang, Roth 1
What is Semantic Parsing? Meaning Representation make(coffee, sugar=0, milk=0.3) I’d like a coffee with no sugar and just a little milk Clarke, Goldwasser, Chang, Roth 2
What is Semantic Parsing? Meaning Representation make(coffee, sugar=0, milk=0.3) I’d like a coffee with no sugar and just a little milk Clarke, Goldwasser, Chang, Roth 2
Supervised Learning Problem meaning text Training Model algorithm Challenges: Structured Prediction problem Model part of the structure as hidden? Clarke, Goldwasser, Chang, Roth 3
Lots of previous work Multiple approaches to the problem: K RISP (Kate & Mooney 2006) SVM-based parser using string kernels. Zettlemoyer & Collins 2005; Zettlemoyer & Collins 2007 Probabilistic parser based on relaxed CCG grammars. W ASP (Wong & Mooney 2006; Wong & Mooney 2007) Based on Synchronous CFG. Ge & Mooney 2009 Integrated syntactic and semantic parser. Clarke, Goldwasser, Chang, Roth 4
Lots of previous work Multiple approaches to the problem: K RISP (Kate & Mooney 2006) SVM-based parser using string kernels. Zettlemoyer & Collins 2005; Zettlemoyer & Collins 2007 Probabilistic parser based on relaxed CCG grammars. W ASP (Wong & Mooney 2006; Wong & Mooney 2007) Based on Synchronous CFG. Ge & Mooney 2009 Integrated syntactic and semantic parser. Assumption : A training set consisting of natural language and meaning representation pairs. Clarke, Goldwasser, Chang, Roth 4
Using the World’s response Meaning Representation make(coffee, sugar=0, milk=0.3) I’d like a coffee with no sugar and just a little milk Clarke, Goldwasser, Chang, Roth 5
Using the World’s response Meaning Representation make(coffee, sugar=0, milk=0.3) I’d like a coffee with no sugar and just a little milk Good! Bad! Clarke, Goldwasser, Chang, Roth 5
Using the World’s response Meaning Representation make(coffee, sugar=0, milk=0.3) I’d like a coffee with no sugar and just a little milk Good! Bad! Question: Can we use feedback based on the response to provide supervision? Clarke, Goldwasser, Chang, Roth 5
This work We aim to : Reduce the burden of annotation for semantic parsing. We focus on : Using the World’s response to learn a semantic parser. Developing new training algorithms to support this learning paradigm. A lightweight semantic parsing model that doesn’t require annotated data. This results in : Learning a semantic parser using zero annotated meaning representations. Clarke, Goldwasser, Chang, Roth 6
Outline Semantic Parsing 1 Learning 2 D IRECT Approach A GGRESSIVE Approach Semantic Parsing Model 3 Experiments 4 Clarke, Goldwasser, Chang, Roth 7
Outline Semantic Parsing 1 Learning 2 D IRECT Approach A GGRESSIVE Approach Semantic Parsing Model 3 Experiments 4 Clarke, Goldwasser, Chang, Roth 8
Semantic Parsing What is the largest state that borders Texas? I NPUT x H IDDEN y O UTPUT z largest(state(next_to(texas))) Clarke, Goldwasser, Chang, Roth 9
Semantic Parsing What is the largest state that borders Texas? I NPUT x H IDDEN y O UTPUT z largest(state(next_to(texas))) F : X → Z w T Φ( x , y , z ) ˆ z = F w ( x ) = arg max y ∈Y , z ∈Z Clarke, Goldwasser, Chang, Roth 9
Semantic Parsing What is the largest state that borders Texas? I NPUT x H IDDEN y O UTPUT z largest(state(next_to(texas))) F : X → Z w T Φ( x , y , z ) ˆ z = F w ( x ) = arg max y ∈Y , z ∈Z Model The nature of inference and feature functions. Learning Strategy How we obtain the weights. Clarke, Goldwasser, Chang, Roth 9
Semantic Parsing What is the largest state that borders Texas? I NPUT x H IDDEN y O UTPUT z largest(state(next_to(texas))) Response r New Mexico F : X → Z w T Φ( x , y , z ) ˆ z = F w ( x ) = arg max y ∈Y , z ∈Z Model The nature of inference and feature functions. Learning Strategy How we obtain the weights. Clarke, Goldwasser, Chang, Roth 9
Outline Semantic Parsing 1 Learning 2 D IRECT Approach A GGRESSIVE Approach Semantic Parsing Model 3 Experiments 4 Clarke, Goldwasser, Chang, Roth 10
Learning Inputs : Natural language sentences. Feedback : X × Z → { + 1 , − 1 } . Zero meaning representations. Clarke, Goldwasser, Chang, Roth 11
Learning Inputs : Natural language sentences. Feedback : X × Z → { + 1 , − 1 } . Zero meaning representations. � + 1 if execute ( z ) = r Feedback ( x , z ) = − 1 otherwise Clarke, Goldwasser, Chang, Roth 11
Learning Inputs : Natural language sentences. Feedback : X × Z → { + 1 , − 1 } . Zero meaning representations. Goal : A weight vector that scores the correct meaning representation higher than all other meaning representations. Response Driven Learning : Feedback predict apply to Meaning Input text World Representation Clarke, Goldwasser, Chang, Roth 11
Learning Strategies x 1 repeat for all input sentences do x 2 Solve the inference problem Query Feedback function x 3 end for Learn a new w using feedback until Convergence . . . x n Clarke, Goldwasser, Chang, Roth 12
Learning Strategies y 1 x 1 z 1 repeat for all input sentences do x 2 y 2 z 2 Solve the inference problem Query Feedback function x 3 y 3 z 3 end for Learn a new w using feedback until Convergence . . . . . . . . . y , z = arg max w T Φ( x , y , z ) x n y n z n Clarke, Goldwasser, Chang, Roth 12
Learning Strategies x 1 y 1 z 1 + 1 repeat x 2 y 2 z 2 for all input sentences do − 1 Solve the inference problem Query Feedback function end for x 3 y 3 z 3 − 1 Learn a new w using feedback until Convergence . . . . . . . . . . . . y n x n z n − 1 Clarke, Goldwasser, Chang, Roth 12
Learning Strategies x 1 y 1 z 1 + 1 repeat x 2 y 2 z 2 for all input sentences do − 1 Solve the inference problem Query Feedback function end for x 3 y 3 z 3 − 1 Learn a new w using feedback until Convergence . . . . . . . . . . . . y n x n z n − 1 Clarke, Goldwasser, Chang, Roth 12
Outline Semantic Parsing 1 Learning 2 D IRECT Approach A GGRESSIVE Approach Semantic Parsing Model 3 Experiments 4 Clarke, Goldwasser, Chang, Roth 13
D IRECT Approach Binary Learning Feedback predict apply to Meaning Input text World Representation D IRECT Learn a binary classifier to discriminate between good and bad meaning representations. Clarke, Goldwasser, Chang, Roth 14
D IRECT Approach x 1 y 1 z 1 + 1 x 2 y 2 z 2 − 1 Use ( x , y , z ) as a training example with label from x 3 y 3 z 3 − 1 feedback. . . . . . . . . . . . . y n x n z n − 1 Clarke, Goldwasser, Chang, Roth 15
D IRECT Approach x 1 , y 1 , z 1 + 1 x 2 , y 2 , z 2 − 1 Use ( x , y , z ) as a training example with label from x 3 , y 3 , z 3 − 1 feedback. Find w such that f · w T Φ( x , y , z ) > 0 . . . . . . x n , y n , z n − 1 Clarke, Goldwasser, Chang, Roth 15
D IRECT Approach Each point represented by Φ( x , y , x ) normalized by | x | Clarke, Goldwasser, Chang, Roth 16
D IRECT Approach w Learn a binary classifier to discriminate between good and bad meaning representations. Clarke, Goldwasser, Chang, Roth 16
D IRECT Approach repeat for all input sentences do Solve the inference problem Query Feedback function end for Learn a new w using feedback until Convergence Clarke, Goldwasser, Chang, Roth 17
D IRECT Approach x 1 repeat for all input sentences do x 2 Solve the inference problem Query Feedback function x 3 end for Learn a new w using feedback until Convergence . . . x n Clarke, Goldwasser, Chang, Roth 17
D IRECT Approach y ′ z ′ x 1 1 1 repeat x 2 y ′ z ′ 2 for all input sentences do 2 Solve the inference problem Query Feedback function end for x 3 y ′ z ′ 3 3 Learn a new w using feedback until Convergence . . . . . . . . . y , z = arg max w T Φ( x , y , z ) x n y ′ z ′ n n Clarke, Goldwasser, Chang, Roth 17
D IRECT Approach y ′ z ′ x 1 + 1 1 1 repeat x 2 y ′ z ′ 2 for all input sentences do + 1 2 Solve the inference problem Query Feedback function end for x 3 y ′ z ′ + 1 3 3 Learn a new w using feedback until Convergence . . . . . . . . . . . . y ′ z ′ x n − 1 n n Clarke, Goldwasser, Chang, Roth 17
D IRECT Approach x 1 , y ′ 1 , z ′ + 1 1 repeat x 2 , y ′ 2 , z ′ for all input sentences do + 1 2 Solve the inference problem Query Feedback function x 3 , y ′ 3 , z ′ end for + 1 3 Learn a new w using feedback until Convergence . . . . . . x n , y ′ n , z ′ − 1 n Clarke, Goldwasser, Chang, Roth 17
D IRECT Approach w Clarke, Goldwasser, Chang, Roth 18
D IRECT Approach w Clarke, Goldwasser, Chang, Roth 18
D IRECT Approach w Clarke, Goldwasser, Chang, Roth 18
D IRECT Approach w Repeat until convergence! Clarke, Goldwasser, Chang, Roth 18
Outline Semantic Parsing 1 Learning 2 D IRECT Approach A GGRESSIVE Approach Semantic Parsing Model 3 Experiments 4 Clarke, Goldwasser, Chang, Roth 19
Recommend
More recommend
