NLU: Semantic parsing

Transcription

1 NLU: Semantic parsing Adam Lopez slide credits: Chris Dyer, Nathan Schneider March 30, 2018 School of Informatics University of Edinburgh

2 Recall: meaning representations Sam likes Casey likes(sam, Casey) Anna s dog Mr. PeanutButter misses her misses(mrpb, Anna) dog(mrpb) Kim likes everyone x.likes(x, Kim)

3 Recall: meaning representations Meaning representations are verifiable, unambiguous, canonical. Predicate-argument structure is a good match for FOL, as well as structures with argument-like elements (e.g. NPs) Determiners, quantifiers (e.g. everyone, anyone ), and negation can be expressed in first order logic.

4 Representing the meaning of arbitrary NL is hard Much of natural language is unverifiable, ambiguous, non-canonical. What is the finite set of predicates? When do two words map to the same predicate? When are homonyms mapped to different predicates? Can we decompose lexical meanings into a finite set of predicates (possibly composed)? What is the finite set of constants?

5 Easier: representing a closed domain Example: GEOQUERY dataset What states border Texas? λx. state(x) borders(x,texas) What is the largest state? argmax(λx. state(x) λx.size(x)) Semantic parsing is the problem of returning a logical form for an input natural language sentence.

6 Pairs of NL sentences with structured MR can be collected Example: IFTTT dataset (Quirk et al. 2015)

7 WikiTableQuestions

8 similar information powers Google s knowledge graph

9 Viewing MR as a string, semantic parsing is just conditional language modeling p(y 1,...,y y x 1,...,x x ) Model using standard sequence models

10 Viewing MR as a string, semantic parsing is just conditional language modeling p(y 1,...,y y x 1,...,x x ) Model using standard sequence models with one additional element

11 x 1 x 2 x 3 x 4 Ich möchte ein Bier

12 ! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier

13 h 1 h 2 h 3 h 4! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier

14 f i =[h i ;! h i ] h 1 h 2 h 3 h 4! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier

15 f i =[h i ;! h i ] h 1 h 2 h 3 h 4! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier

16 f i =[h i ;! h i ] h 1 h 2 h 3 h 4! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier

17 f i =[h i ;! h i ] h 1 h 2 h 3 h 4! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier

18 f i =[h i ;! h i ] h 1 h 2 h 3 h 4 F 2 R 2n f! h 1! h 2! h 3! h 4 x 1 x 2 x 3 x 4 Ich möchte ein Bier Ich möchte ein Bier

19 0 Ich möchte ein Bier

20 0 z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

21 0 z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

22 I'd 0 z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

23 I'd like 0 I'd z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

24 I'd like a 0 I'd like z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

25 I'd like a beer 0 I'd like a z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

26 I'd like a beer STOP stop 0 I'd like a beer z } { Attention history: a > 1 a > 2 a > 3 a > 4 a > 5 Ich möchte ein Bier

27 English-French English-German

28 Since logical forms are treelike, can use treelstm decoder

29 Model learns to translate words into predicates they invoke

30 Abstract meaning PENMAN notation representation (AMR) The edges (ARG0 and ARG1) are relations Each node in the graph has a variable! They are labeled with concepts! d / dog means d is an instance of dog! The dog is eating a bone (e / eat-01 :ARG0 (d / dog) :ARG1 (b / bone)) e/eat-01 ARG0 ARG1 d/dog b/bone 7

31 Abstract meaning Reentrancy representation (AMR) What if something is referenced multiple times? Notice how dog has two incoming roles now. To do this in PENMAN format, repeat the variable. We call this a reentrancy. (want-01 :ARG0 (d / dog) :ARG1 (e / eat-01 :ARG0 d! :ARG1 (b / bone))) w/want-01 The dog wants to eat the bone ARG1 ARG0 e/eat-01 ARG0 ARG1 d/dog b/bone 9

32 Coreference Bob wants Anna to give him a job. Q: who does him refer to?

33 Coreference Charles just graduated, and now Bob wants Anna to give him a job. Q: who does him refer to?

34 Metonymy Westminster decided to distribute funds throughout England, Wales, Northern Island, and Scotland

35 Metonymy Westminster decided to distribute funds throughout England, Wales, Northern Island, and Scotland decided(westminster, )

36 Metonymy Westminster decided to distribute funds throughout England, Wales, Northern Island, and Scotland decided(westminster, ) decided(parliament, )

37 Implicature That cake looks delicious

38 Implicature That cake looks delicious What Rogelio was really thinking: I would like a piece of that cake.

39 Even more phenomena Abbreviations (e.g. National Health Service=NHS) Nicknames (JLaw=Jennifer Lawrence) Metaphor (crime is a virus infecting the city) Time expressions and change of state Many others

40 Summary In many cases, meaning representation can be captured in firstorder logic. But wide-coverage meaning representation is hard; closed domains are easier, and can sometimes be harvested automatically. This leads to a proliferation of domain-specific MRs. Trainable alternative to compositional approaches: encoderdecoder neural models. The encoder and decoder can be mixed and matched: RNN, top-down tree RNN, etc. Works well on small, closed domains if we have training data, but there are many unsolved phenomena/ problems in semantics.