Learning Joint Semantic Parsers from Disjoint Data Hao Peng 1 , Sam - PowerPoint PPT Presentation

Learning Joint Semantic Parsers from Disjoint Data Hao Peng 1 , Sam Thomson 2 , Swabha Swayamdipta 2 � Noah A. Smith 1 1 University of Washington 2 Carnegie Mellon University @NAACL June 4, 2018

Motivations almost ❖ Larger data Better performance ❖ Overlaps among di ff erent theories

Overview Learning Joint Semantic Parsers from Disjoint Data FrameNet vs. semantic dependencies Di ff erent structures; no parallel annotations

Overview Joint decoding Latent variables Learning Joint Semantic Parsers from Disjoint Data FrameNet vs. semantic dependencies Di ff erent structures; no parallel annotations

Outline ❖ Parsing semantic spans and dependencies ❖ Joint parsing ❖ Learning with latent variables ❖ Empirical results

Parsing FrameNet Structures Input: Target: token span A few books fell in the room . fall.v Lexical unit: lemma.pos Baker et al., (1998)

Parsing FrameNet Structures Input: Target: token span A few books fell in the room . fall.v Lexical unit: lemma.pos Output: Frame who what A few books fell in the room . when fall.v where Motion Theme Place … Directional Arguments: span + semantic roles Baker et al., (1998)

Parsing FrameNet Structures Input: A few books fell in the room . fall.v Score: � � A few books fell in the room . F fall.v Motion Theme Place Directional = � � � � � � f frame + f arg + f arg Motion Directional Theme Place

Parsing FrameNet Structures Input: A few books fell in the room . fall.v Score: � � A few books fell in the room . F fall.v Motion Theme Place Directional = � � � � � � f frame + f arg + f arg Motion Directional Theme Place BiLSTM+MLPs

Parsing FrameNet Structures Decoding: Dynamic program Kong et al., (2016); Swayamdipta et al., (2017) � � A few books fell in the room . F max fall.v arg1? arg2? arg3? frame? frame, args • non-overlapping s.t. • consistency • …

Parsing Semantic Dependencies Input: A few books fell in the room . Output: MRS-derived dependencies (DM) top arg2 mwe arg1 arg1 arg1 BV A few books fell in the room . who what role label when where head modifier … Oepen et al., (2015)

Parsing Semantic Dependencies Input: A few books fell in the room . Score: top � � arg2 G mwe arg1 arg1 arg1 BV A few books fell in the room . = X role � � BiLSTM+MLPs g head mod labeled arcs

Parsing Semantic Dependencies Decoding: Linear program AD 3 ; Martins et al., (2011) � compound � arg2 ? ? G few books fell room max arg1 ? … few books labeled arcs • consistency s.t. • determinism • …

Outline ❖ Parsing semantic spans and dependencies ❖ Joint parsing ❖ Learning with latent variables ❖ Empirical results

Joint Parsing Sharing parameters: Swayamdipta et al., (2016); Hershcovich et al., (2018) top arg2 � � � � G F arg1 mwe arg1 arg1 BV A few books fell in the room . fall.v A few books fell in the room . Motion Place Theme Directional Shared LSTMs

Joint Parsing Sharing parameters: Swayamdipta et al., (2016); Hershcovich et al., (2018) top arg2 � � � � G F arg1 mwe arg1 arg1 BV A few books fell in the room . fall.v A few books fell in the room . Motion Place Theme Directional Shared LSTMs This work, joint decoding: top ⇣ ⌘ arg2 arg1 mwe arg1 arg1 BV H A few books fell in the room . fall.v Motion Place Theme Directional

Joint Parsing Sharing parameters: Swayamdipta et al., (2016); Hershcovich et al., (2018) top arg2 � � � � G F arg1 mwe arg1 arg1 BV A few books fell in the room . fall.v A few books fell in the room . Motion Place Theme Directional Shared LSTMs This work, joint decoding: Orthogonal top ⇣ ⌘ arg2 arg1 mwe arg1 arg1 BV H A few books fell in the room . fall.v Motion Place Theme Directional

Joint Parsing Input: A few books fell in the room . fall.v Score: top ⇣ arg2 ⌘ arg1 mwe arg1 arg1 BV H A few books fell in the room . fall.v Motion Place Theme Directional

Joint Parsing Input: A few books fell in the room . fall.v Score: top ⇣ arg2 ⌘ arg1 mwe arg1 arg1 BV H A few books fell in the room . fall.v Motion Place Theme Directional = top arg2 � � � � F + G A few books fell in the room . arg1 mwe arg1 arg1 BV fall.v A few books fell in the room . Motion Place Theme Directional FrameNet Score DM Score

Joint Parsing Input: A few books fell in the room . fall.v Score: top ⇣ arg2 ⌘ arg1 mwe arg1 arg1 BV H A few books fell in the room . fall.v Motion Place Theme Directional = top arg2 � � � � F + G A few books fell in the room . arg1 mwe arg1 arg1 BV fall.v A few books fell in the room . Motion Place Theme Directional � � + h joint ? FrameNet Score DM Score A ffi nities between them

Span vs. Dependencies � � ? h joint If both were dependencies Lluís et al., (2013); Peng et al., (2017) role1 � � h joint head mod role2 If both were spans Finkel and Manning, (2009) � � role1 h joint role2

Span vs. Dependencies � � ? h joint If both were dependencies Lluís et al., (2013); Peng et al., (2017) role1 � � h joint head mod role2 If both were spans Finkel and Manning, (2009) � � role1 h joint role2 Structural divergence mwe arg1 arg1 A few books fell fall.v Motion Theme Directional

Span vs. Dependencies Structural divergence mwe arg1 arg1 A few books fell fall.v Motion Theme Directional Designate a head for each span PropBank dependencies; Surdeanu et al., (2008) A few books fell fall.v Theme

Span vs. Dependencies Structural divergence mwe arg1 arg1 A few books fell fall.v Motion Theme Directional Designate a head for each span PropBank dependencies; Surdeanu et al., (2008) Head selected by syntax Collins, (2003) A few books fell fall.v Theme

Span vs. Dependencies Structural divergence mwe arg1 arg1 A few books fell fall.v Motion Theme Directional Designate a head for each span PropBank dependencies; Surdeanu et al., (2008) arg1 A few books fell fall.v Theme

Span vs. Dependencies Structural divergence mwe arg1 arg1 A few books fell fall.v Motion Theme Directional This work A few books fell A few books fell A few books fell fall.v fall.v fall.v Theme Theme Theme

Span vs. Dependencies Score: top ⇣ ⌘ arg2 arg1 mwe arg1 arg1 BV H A few books fell in the room . fall.v Motion Place Theme Directional = top arg2 � � � � F + G A few books fell in the room . arg1 mwe arg1 arg1 BV fall.v A few books fell in the room . Motion Place Theme Directional ⇣ ⌘ arg1 A few books fell + h joint fall.v Motion Theme Directional FrameNet Score DM Score A ffi nities between them Multilinear mapping

Span vs. Dependencies Decoding: ⇣ ⌘ arg1 ? arg2 ? BV ? max H A few books fell in the room . fall.v arg1? frame? arg2? arg3? frame, args labeled arcs joint parts Linear program Speed up by promoting sparsity

Outline ❖ Parsing semantic spans and dependencies ❖ Joint parsing ❖ Learning with latent variables ❖ Empirical results

Learning with Latent Variables FrameNet data DM data

Learning with Latent Variables FrameNet data DM data Supervision Supervision Theme Theme role role head mod head mod A few books fell A few books fell fall.v fall.v Theme Theme

Learning with Latent Variables Latent structured hinge Yu and Joachims, (2009) arg1 ? arg2 ? BV ? ⇣ ⌘ L = − max H A few books fell in the room . fall.v Theme Motion Place labeled arcs Directional joint parts arg1 ? arg2 ? BV ? ⇣ ⌘ + δ + max H A few books fell in the room . fall.v frame, args arg1? frame? arg2? arg3? labeled arcs joint parts FrameNet data

Learning with Latent Variables Latent structured hinge Yu and Joachims, (2009) arg1 ? arg2 ? BV ? ⇣ ⌘ L = − max H A few books fell in the room . fall.v Theme Motion Place labeled arcs Directional joint parts arg1 ? arg2 ? BV ? ⇣ ⌘ + δ + max H A few books fell in the room . fall.v frame, args arg1? frame? arg2? arg3? labeled arcs joint parts cost Prediction FrameNet data

Learning with Latent Variables Latent structured hinge Yu and Joachims, (2009) Gold FN output arg1 ? arg2 ? BV ? ⇣ ⌘ L = − max H A few books fell in the room . fall.v Theme Motion Place labeled arcs Directional joint parts arg1 ? arg2 ? BV ? ⇣ ⌘ + δ + max H A few books fell in the room . fall.v frame, args arg1? frame? arg2? arg3? labeled arcs joint parts FrameNet data

Learning with Latent Variables Latent structured hinge Yu and Joachims, (2009) arg1 ? arg2 ? BV ? ⇣ ⌘ L = − max H A few books fell in the room . fall.v Theme Motion Place labeled arcs Directional joint parts arg1 ? arg2 ? BV ? ⇣ ⌘ + δ + max H A few books fell in the room . fall.v frame, args arg1? frame? arg2? arg3? labeled arcs joint parts FrameNet data

Outline ❖ Parsing semantic spans and dependencies ❖ Joint parsing ❖ Learning with latent variables ❖ Empirical results