Semantic Textual Similarity & more on Alignment

Transcription

1 Semantic Textual Similarity & more on Alignment CMSC 723 / LING 723 / INST 725 MARINE CARPUAT marine@cs.umd.edu

2 2 topics today P3 task: Semantic Textual Similarity Including Monolingual alignment Beyond IBM word alignment Synchronous CFGs

3 Semantic Textual Similarity Series of tasks at international workshop on semantic evaluations (SemEval), since

4 What is Semantic Textual Similarity? Hnh whdun duuhj js ijd dj iow oijd oidj dk uwhd8 yh djhdhwuih jhu h uh jhihk, jdhhii, gdytysla, yuiyduinsjsh, iodpisomkncijsi. Kjhhuduh, dhdhhd hhduhd jjhuiq Welcome to my world, trust me you will never be disappointed djijdp idiowdiw I iwfiow ifiwoufowi ioiowruo iyfi I wioiwf oid oi iwoiwy iowuouwr ujjd hihi iohoihiof uouo ou o oufois f uhdiy oioi oo ouiosufoisuf iouiouf paidp paudoi uiu fh uhhioiof 안녕하세요제가당신에게전화했지만아무소용이있을려고... 당신이시간을즐기고있었다희망 Shjkahsiunu iuhndhau dhdkhn hdhaud8 kdhikahdi dhjhd dhjh jiidh iihiiohio hihiahdiod Yo! Come over here, you will be pleasantly surprised idoasd io idjioio jidjduio iodio oi iiouio oiudoi ifuiosu fiuoi oiuiou oi io hiyuify 8iy ih iouoiu ou o ooihyiush iuh fhdfosiip upouosu oiu oi o oisyoisy oi sih oiiou ios oisuois uois oudiosu doi soiddu os oso iio oioisosuo. Semantic Similarity جدالكجد يدجياجد يجدي يج جي وغو يحيح يحسيفحس يحيحفي سف ي جي جيييدج كجساكجاس حفجحسوجح ج. كححسح حيحي حوحوس دح حدي يجدي يو جي جيحجفححكسحجسكحك حفحسوحوشيحيدويويد وي يوسحفوفوفوطبس تعالى ومالكش دعوه هتبنبسط اخر انبساط Добро пожаловать в мой мир, поверьте мне вы никогда не будете разочарованы Quantitative Graded Similarity Score Confidence Score Principled Interpretability, which semantic components/features led to results (hopefully will lead to us gaining a better understanding of semantics)

5 Why Semantic Textual Similarity? Most NLP applications need some notion of semantic similarity to overcome brittleness and sparseness Provides evaluation beyond surface text processing A hub for semantic processing as a black box in applications beyond NLP Lends itself to an extrinsic evaluation of scattered semantic components

6 What is STS? The graded process by which two snippets of text (t1 and t2) are deemed equivalent semantically, i.e. bear the same meaning An STS system will quantifiably inform us on how similar t1 and t2 are, resulting in a similarity score An STS system will tell us why t1 and t2 are similar giving a nuanced interpretation of similarity based on semantic components contributions

7 What is STS? Word similarity has been relatively well studied For example according to WN cord smile 0.02 rooster voyage 0.04 noon string 0.04 fruit furnace hill woodland 1.48 car journey 1.55 cemetery mound cemetery graveyard 3.88 automobile car 3.92 More similar

8 What is STS? Fewer datasets for similarity between sentences A forest is a large area where trees grow close together. VS. The coast is an area of land that is next to the sea. [0.25]

9 What is STS? Fewer datasets for similarity between sentences A forest is a large area where trees grow close together. VS. Woodland is land with a lot of trees. [2.51]

10 What is STS? Fewer datasets for similarity between sentences Once there was a Czar who had three lovely daughters. VS. There were three beautiful girls, whose father was a Czar. [4.3]

11 Related tasks Paraphrase detection Are 2 sentences equivalent in meaning? Textual Entailment Does premise P entail hypothesis H? STS provides graded similarity judgments

12

13 Annotation: crowd-sourcing

14 Annotation: crowd-sourcing English annotation process Pairs annotated in batches of 20 Annotators paid $1 per batch 5 annotations per pair Workers need to have Mturk master qualification Defining gold standard judgments Median value of annotations After filtering low quality annotators (<0.80 correlation with leave-on-out gold & <0.20 Kappa)

15 Diverse data sources

16 Evaluation: a shared task Subset of 2016 results (Score: Pearson correlation)

17 STS models from word to sentence vectors Can we perform STS by comparing sentence vector representation? This approach works well for word level similarity But can we capture the meaning of a sentence in a single vector?

18 Composing by averaging g( shots fired at residence ) = shots fired at residence [Tai et al. 2015, Wieting et al. 2016]

19 How can we induce word vectors for composition? English paraphrases [Wieting et al. 2016] Bilingual sentence pairs [Hermann & Blunsom 2014] x 1 By our fellow members Thus in fact by our fellow members x 2 By our colleagues As que podramos nuestra colega disputado Bilingual phrase pairs by our fellow member de nuestra colega

20 STS models: monolingual alignment

21 Idea One (of many) approaches to monolingual entailment Exploit not only similarity between words But also similarity between their contexts See Sultan et al

22 2 topics today P3 task: Semantic Textual Similarity Including Monolingual alignment Beyond IBM word alignment Synchronous CFGs

23 Aligning words & constituents Alignment: mapping between spans of text in lang1 and spans of text in lang2 Sentences in document pairs Words in sentence pairs Syntactic constituents in sentence pairs Today: 2 methods for aligning constituents Parse and match biparse

24 Parse & Match

25 Parse(-Parse)-Match Idea Align spans that are consistent with existing structure Pros Builds on existing NLP tools Cons Assume availability of lots of resources Assume that representations can be matched

26 Aligning words & constituents 2 methods for aligning constituents: Parse and match assume existing parses and alignment Biparse alignment = structure

27 A straw man hypothesis: All languages have same grammar

28 A straw man hypothesis: All languages have same grammar

29 A straw man hypothesis: All languages have same grammar

30 A straw man hypothesis: All languages have same grammar

31 The biparsing hypothesis: All languages have nearly the same grammar

32 The biparsing hypothesis: All languages have nearly the same grammar

33 Example for the biparsing hypothesis: All languages have nearly the same grammar

34 The biparsing hypothesis: All languages have nearly the same grammar

35 The biparsing hypothesis: All languages have nearly the same grammar Dekai Wu and Pascale Fung, IJCNLP-2005 HKUST Human Language Technology Center

36 The biparsing hypothesis : All languages have nearly the same grammar Dekai Wu and Pascale Fung, IJCNLP-2005 HKUST Human Language Technology Center

37 The biparsing hypothesis : All languages have nearly the same grammar Dekai Wu and Pascale Fung, IJCNLP-2005 HKUST Human Language Technology Center

38 The biparsing hypothesis: All languages have nearly the same grammar Permuted SDTG/SCFG VP VV PP ; 1 2 VP VV PP ; 2 1 Indexed SDTG/SCFG notation VP VV (1) PP (2), VV (1) PP (2) VP VV (1) PP (2), PP (2) VV (1) SDTG/SCFG notation VP VV PP, VV PP VP VV PP, PP VV ITG shorthand VP [ VV PP ] VP VV PP Dekai Wu and Pascale Fung, IJCNLP-2005 HKUST Human Language Technology Center

39 Synchronous Context Free Grammars Context free grammars (CFG) Common way of representing syntax in (monolingual) NLP Synchronous context free grammars (SCFG) Generate pairs of strings Align sentences by parsing them Translate sentences by parsing them Key algorithm: how to parse with SCFGs?

40 SCFG trade off Expressiveness SCFGs cannot represent all sentence pairs in all languages Efficiency SCFGs let us view alignment as parsing & benefit from well-studied formalism

41 Synchronous parsing cannot represent all sentence pairs

42 Synchronous parsing cannot represent all sentence pairs

43 Synchronous parsing cannot represent all sentence pairs

44 A subclass of SCFGs: Inversion Transduction Grammars ITGs are the subclass of SDTGs/SCFGs: with only straight and inverted transduction rules equivalent with only transduction rules of rank < 2 with only transduction rules of rank < 3 ITGs are context-free (like SCFGs).

45 For length-4 phrases (or frames), ITGs can express 22 out of 24 permutations!

46 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

47 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

48 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

49 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

50 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

51 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

52 ITGs enable efficient DP algorithms [Wu 1995] e 0 e 1 e 2 e 3 e 4 e 5 e 6 e 7 c 0 c 1 c 2 c 3 c 4 c 5 c 6

53 Biparsing with CKY Given the following SCFG A -> fat, gordos A -> thin, delgados N -> cats, gatos VP -> eats, comen NP -> A (1) N (2),N (2) A (1) S -> NP (1) VP (2), NP (1) VP (2) Let s parse a sentence pair fat cats eat gatos gordos comen Example by Matt Post (JHU)

54 Biparsing with CKY A -> fat, gordos A -> thin, delgados N -> cats, gatos VP -> eats, comen NP -> A (1) N (2),N (2) A (1) S -> NP (1) VP (2), NP (1) VP (2) 3 comen 2 gordos 1 gatos fat cats eats Chart now enumerates pairs of spans

55 Biparsing with CKY A -> fat, gordos A -> thin, delgados N -> cats, gatos VP -> eats, comen NP -> A (1) N (2),N (2) A (1) S -> NP (1) VP (2), NP (1) VP (2) 3 comen 2 gordos 1 gatos A ((1,1), (2,2)) N ((2,2), (1,1)) VP ((3,3), (3,3)) fat cats eats Apply lexical rules

56 Biparsing with CKY A -> fat, gordos A -> thin, delgados N -> cats, gatos VP -> eats, comen NP -> A (1) N (2),N (2) A (1) S -> NP (1) VP (2), NP (1) VP (2) 3 comen 2 gordos 1 gatos A S ((1,1), ((3,3), (2,2)) NP (3,3)) ((1,2), N (1,2)) ((2,2), (1,1)) VP ((3,3), (3,3)) fat cats eats For each block, apply straight & inverted rules 1 2 3

57 Biparsing with CKY 3 comen 2 gordos 1 gatos O(GN 3 M 3 ) fat cats eats 1 2 3

58 Aligning words & constituents 2 different ways of looking at this problem: parse-parse-match assume existing parses and alignment biparse alignment = structure