N EU G EN Text Generation from Meaning Representations Yannis - PowerPoint PPT Presentation

N EU G EN � Text Generation from Meaning Representations Yannis Konstas � � Joint work with Mark Yatskar, Luke Zettlemoyer and Yejin Choi (UW) ~ Yonatan Bisk and Daniel Marcu (ISI)

Motivation

Motivation Machine-generated Representation

Motivation Machine-generated Representation hk source block: ms target block: W small pos RP: scale:

Motivation Machine-generated Representation hk source block: Place the heineken block west of the ms target block: mercedes block. W small pos RP: scale:

Motivation Machine-generated Representation hk source block: ups target block: W big pos RP: scale:

Motivation Machine-generated Representation hk source block: Place the heineken block in the first ups open space to the left of ups block. target block: W big pos RP: scale:

Motivation

Motivation Machine-generated Representation

Motivation Machine-generated Representation Sentence Compression

Motivation Machine-generated Representation Sentence Compression Sentence Fusion

Motivation Machine-generated Representation Sentence Compression Sentence Fusion Paraphrasing

NLG Pipeline Communicative Goal A A A A m Input Content Planning Content Selection Document Planning Sentence Planning Framework Lexicalization Reordering/Linearization Splitting/Aggregation Surface Realisation Text

NLG Pipeline Communicative Goal A A A A m Input Content Planning Content Selection Document Planning Sentence Planning Framework - Neural Encoder Lexicalization - RNN Decoder Reordering/Linearization - Graph Alignment Splitting/Aggregation Surface Realisation Text

Joint Generator ? ? Temperature Cloud Sky Cover S ? S Time Min Mean Max Time Percent (%) 06:00-21:00 9 15 21 06:00-09:00 25-50 ... Cloudy, with temperatures ... 09:00-12:00 50-75 S Training between 10 and 20 degrees. S South wind around 20 mph. Wind Speed S Wind Direction ... ... Time Mode Time Min Mean Max 06:00-21:00 S 06:00-21:00 15 20 30 ... F 0 , 1 (temp 1 ,min) S → R( start ) (1) PCFG Grammar FS 0 , 1 (temp 1 ,start) R( r i .t ) → FS( r j , start )R( r j .t ) F 0 , 1 (temp 1 ,max) R( r i .t ) → FS( r j , start ) FS( r, r.f i ) → F( r, r.f j )FS( r, r.f j ) FS( r, r.f i ) → F( r, r.f j ) F 0 , 2 (temp 1 ,min) F( r, r.f ) → W( r, r.f )F( r, r.f ) FS 0 , 2 (temp 1 ,start) (2) Hypergraph F( r, r.f ) → W( r, r.f ) F 0 , 2 (temp 1 ,max) Representation W( r, r.f ) → α W( r, r.f ) → g( f.v ) FS 1 , 2 (temp 1 ,start)   mostly cloudy the morning ? 11 am mostly cloudy after ?   (3) k-best decoding   mostly cloudy then becoming ? FS 0 , 5 (skyCover 1 .t ,start)   via integration · · ·   mostly cloudy mostly clouds     Testing cloudy ,   F 0 , 2 (skyCover 1 .t ,%) W 4 , 5 (skyCover 1 .t ,time) · · ·   morning 11 am W 0 , 1 (skyCover 1 .t ,%) W 1 , 2 (skyCover 1 .t ,%)     after       mostly mostly · · · cloudy cloudy         Konstas and Lapata, 2012, 2013 sunny sunny     · · · · · ·

More Complex Structure

More Complex Structure Input (Graph or Tree) know-01 ARG0 ARG1 I planet ARG1-of inhabit-01 ARG0 man mod lazy

More Complex Structure Input (Graph or Tree) know-01 ARG0 ARG1 I planet ARG1-of inhabit-01 ARG0 man mod lazy AMR DAG

More Complex Structure Input (Graph or Tree) know-01 ARG0 ARG1 I planet ARG1-of inhabit-01 ARG0 man mod lazy AMR DAG λ -calculus Expression Tree

More Complex Structure Input (Graph or Tree) know-01 ARG0 ARG1 I planet ARG1-of inhabit-01 ARG0 man mod lazy AMR DAG λ -calculus Expression Tree ECIs Tree

More Complex Structure Input (Graph or Tree) Output (Text) know-01 ARG0 ARG1 I know a planet that is inhabited by a lazy man . I planet ARG1-of inhabit-01 ARG0 man mod lazy AMR DAG λ -calculus Expression Tree ECIs Tree

More Complex Structure Input (Graph or Tree) Output (Text) know-01 ARG0 ARG1 I know a planet that is inhabited by a lazy man . I planet ARG1-of inhabit-01 I knew a planet that was inhabited by a lazy man . (lpp_1943.249) ARG0 man I have known a planet that was inhabited by mod a lazy man . lazy AMR DAG λ -calculus Expression Tree ECIs Tree

More Complex Structure Input (Graph or Tree) Output (Text) know-01 ARG0 ARG1 I know a planet that is inhabited by a lazy man . I planet ARG1-of inhabit-01 I knew a planet that was inhabited by a lazy man . (lpp_1943.249) ARG0 man I have known a planet that was inhabited by mod a lazy man . lazy I know about a planet. It is inhabited by a lazy man . AMR DAG λ -calculus Expression Tree ECIs Tree

N EU G EN Framework ARG1-of

N EU G EN Framework G w Encoder Decoder ARG1-of

N EU G EN Framework G w Encoder Decoder Encoder Bag of Words: Concepts and roles - ARG1-of Attention mechanism -

N EU G EN Framework G w Encoder Decoder Encoder Decoder Bag of Words: Concepts and roles Left-to-right word LSTM - - ARG1-of Attention mechanism Greedy beam search - -

N EU G EN Framework G w Encoder Decoder Reranker Encoder Decoder Bag of Words: Concepts and roles Left-to-right word LSTM - - ARG1-of Attention mechanism Greedy beam search - -

N EU G EN Framework G w Encoder Decoder Reranker Encoder Decoder Bag of Words: Concepts and roles Left-to-right word LSTM - - ARG1-of Attention mechanism Greedy beam search - - Reranker

N EU G EN Framework G w Encoder Decoder Reranker Encoder Decoder Bag of Words: Concepts and roles Left-to-right word LSTM - - ARG1-of Attention mechanism Greedy beam search - - Reranker CCG parser -> semantics -

N EU G EN Text Generation from Meaning Representations Yannis - PowerPoint PPT Presentation

N EU G EN Text Generation from Meaning Representations Yannis Konstas Joint work with Mark Yatskar, Luke Zettlemoyer and Yejin Choi (UW) ~ Yonatan Bisk and Daniel Marcu (ISI) Motivation Motivation Machine-generated

Deep Learning for Language Understanding (at Google Scale) Anjuli Kannan Software Engineer,

S8822 OPTIMIZING NMT WITH TENSORRT Micah Villmow Senior TensorRT Software Engineer 2 100

Improved Neural Machine Translation with a Syntax-Aware Encoder and Decoder Huadong Chen ! ,

Exemplar Encoder Decoder for Neural Conversation Generation By Gaurav Pandey, Danish

F i n a n c i a l R e s u l t s P r e s e n t a t i o n Aug 3, 2015

Appendix 24 Securities Code :9438 Business Overview: docomo d-menu

Measuring security with SecQua Metricon 7.0-USENIX 2012 Constantinos Patsakis Department of

Causal Belief Decomposition for Planning with Sensing: Completeness Results and Practical

Third-Order Tensor Decompositions and Their Application in Quantum Chemistry Tyler Ueltschi April

DECOMPOSITION OF COONTAIL ( CERATOPHYLLUM DEMERSUM ) IN EUTROPHIC AND OLIGOTROPHIC ENVIRONMENTS

Inflation and Unit Labor Cost Robert G. King Mark W. Watson Prepared for Studienzentrum Gerzensee

1. Team Goals for Phase 2. Functional Decomposition 3. Benchmarking 4. Morphological Chart 5.

Decomposition of Database Preferences on the Power Set of the Domain Patrick Roocks Institute

A Distributed Algorithm for Scenario-based Model Predictive Control using Primal Decomposition

Functional Decomposition of a Medium Voltage DC Integrated Power System ASNE SYMPOSIUM 2008

DECOMPOSITION ANALYSIS OF THE EU ENERGY AND GREENHOUSE GAS GOALS Sujeetha Selvakkumaran, Ph. D

Blind/Myopic Deconvolution Julian Christou UCSC CfAO PSF Reconstruction Workshop May 2004

deconvolution of complex DNA methylation data a systematic protocol Saarland University

of charge carriers Ragnar Strandbakke, Einar Vllestad, Truls Norby Truls Norby Einar

Inverse Heat Conduction Problem using TC deconvolution and IR measurements: Application to heat

Activity and Doppler Imaging of the late-type giant star Uwe Wolter Hamburger Sternwarte,

TC / TUHS Joint Board Dual Enrollment/OTA Presentation Tony / Mary Alice / Sandy [10/16/17]

1Q 2020 EARNINGS PRESENTATION April 28, 2020 DISCLOSURE STATEMENT This presentation may contain

Veirs Mill Road CAC Meeting #4 September 21, 2015 Purpose of Tonights Meeting Review