Constructions of Causation and Their Slot-Fillers Jesse Dunietz * , - PowerPoint PPT Presentation

Automatically Tagging Constructions of Causation and Their Slot-Fillers Jesse Dunietz * , Lori Levin * , & Jaime Carbonell * April 6, 2017 * Carnegie Mellon University

Shallow semantic parsing tags words bearing predicates and those predicates’ argument spans. PropBank cause Even brief exposures symptoms decades later . C AUSE .01 A RG 0 A RG 1 (Palmer et al., 2005) FrameNet too … a coloured poster , large for indoor display … S UFFICIENCY (Ruppenhofer et al., I TEM S CALE E NABLED _ SIT . 2016) 2

Varied linguistic expression is challenging for most shallow semantic parsers, as evidenced by causal language. Such swelling can impede breathing. (Verbal) They moved because of the schools. (Prepositional) Our success is contingent on your support. (Adjectival) We’re running late, so let’s move quickly. (Conjunctive) This opens the way for broader regulation. (Multi-word expr.) For markets to work, banks can’t expect (Complex) bailouts. 3

Shallow semantic parsers inherit the limitations of their representation schemes. Semantic parser Annotation scheme Limitations SENNA 1 , PropBank Verb arguments only ASSERT 2 End-to-end Penn Discourse Conjunctions and discourse parsers 3 Treebank (PDTB) 5 adverbials only SEMAFOR 4 , FrameNet Triggers must be words mateplus 6 or constituent MWEs word meaning 1 Collobert et al., 2011 3 Xue et al., 2015 5 Prasad et al., 2008 2 Pradhan et al., 2004 4 Das et al., 2014 6 Roth and Lapata, 2015 4

Construction Grammar (CxG) offers a way forward. _______ offensive modifier clausal comp. Linguistic form so ______ I left complementizer that Construction Meaning E XTREME (_______)  _____ I LEFT OFFENSIVE (Fillmore et al., 1988; Goldberg, 1995) 5

Full CxG theory entails a detailed hierarchy and complex interactions: “ constructions all the way down .” (Croft, 2001) 6

The “constructions on top” approach Tagging causal language … Construction recognition POS tagging, syntactic parsing Tokenization 7

T oday’s talk: 1. The BECauSE corpus of causal language 2. Causeway-L/Causeway-S: two simple systems for tagging causal constructions 3. Experiments & error analysis 8

Causal language: a clause or phrase in which one event, state, action, or entity is explicitly presented as promoting or hindering another (Dunietz et al., 2015) 9

Connective: arbitrarily complex fixed lexical cue indicating a causal construction John killed the dog because it was threatening his chickens. For markets to work, banks can’t expect bailouts. Ice cream consumption causes drowning. Not “truly” She must have met him before, because causal she recognized him yesterday. 10

We have annotated a small corpus with this scheme. B ank of E ffects and Cau ses S tated E xplicitly ( BECauSE ): Causality Documents Sentences annotations New York Times 59 2004 529 Washington section (Sandhaus, 2014) Penn Treebank WSJ 47 1542 330 2014 NLP Unshared 1 615 240 Task in PoliInformatics (Smith et al., 2014) Total 107 4161 1099 11

T oday’s talk: 1. The BECauSE corpus of causal language 2. Causeway-L/Causeway-S: two simple systems for tagging causal constructions 3. Experiments & error analysis 12

T oday’s talk: 1. The BECauSE corpus of causal language 2. Causeway-L/Causeway-S: two simple systems for tagging causal constructions 3. Experiments & error analysis 13

Our tagging approach is rooted in information extraction patterns. Y such as X Lexical patterns such Y as X… for hypernym discovery X…and/or other Y (Hearst, 1992) Y including X Y, especially X Dependency patterns for general IE (e.g., Sudo et al. 2001) Lexico-syntactic patterns 14 | hurricane | damage | ARG1+nsubj < cause > dobj+ARG2 for causal verbs 11 | hiv | ads | ARG1+nsubj < cause (Girju, 2003) > dobj+ARG2 14

Task definition: connective discovery + argument identification Connective discovery Find lexical triggers of constructions I worry because I care. I worry because I care. worry/VBP nsubj advcl I/PRP care/VBP Argument identification mark nsubj Identify cause & effect spans because/IN I/PRP for each connective (fill slots) 15

Though simplified, this task is challenging. Long tail of causal connectives ~1 per 2-3 new documents Requires sense disambiguation of connectives e.g., “necessary for us to succeed” vs. “hard for me to do” Combinatorial connective possibilities 16

1. Pattern-based 2. Argument 3. Statistical classifier connective discovery identification to filter results (tentative) (tentative) from worry from worry ✗ I…died I…died your your from from hotel hotel I nearly died from worry. …called …called You could have called me me me from your hotel. Approach 1: Syntactic patterns + head expansion rules 4. Remove duplicate Approach 2: Lexical patterns + CRF sequence labeler connectives 17

1. Causeway-L/Causeway-S: two simple systems for tagging causal constructions Causeway-S: Syntax-based pipeline i. Causeway-L: Lexical pattern-based pipeline ii. 18

Syntax-based connective discovery: each construction is treated as a partially-fixed parse tree fragment. worry/VBP nsubj advcl I/PRP care/VBP mark nsubj because/IN I/PRP I worry because I care. 19

Syntax-based connective discovery: each construction is treated as a partially-fixed parse tree fragment. worry/VBP advcl nsubj I/PRP care/VBP mark nsubj because/IN I/PRP I worry because I care. 20

Syntax-based connective discovery: each construction is treated as a partially-fixed parse tree fragment. advcl mark because/IN 21

Syntax-based connective discovery: TRegex patterns are extracted in training, and matched at test time. (/^because_[0-9]+$/ advcl <2 /^IN.*/ <1 mark I worry because Training: > (/.*_[0-9]+/ I care. <1 advcl mark > (/.*_[0-9]+/))) because/IN I worry because I love you. + I worry because TRegex 1 Test: (/^because_[0-9]+$/ I love you. <2 /^IN.*/ <1 mark > (/.*_[0-9]+/ <1 advcl > (/.*_[0-9]+/))) 1 Levy and Andrew, 2006 22

Syntax-based argument ID: Argument heads are expanded to include most dependents. worry/VBP nsubj advcl care/VBP I/PRP mark nsubj because/IN I/PRP 23

Syntax-based argument ID: Argument heads are expanded to include most dependents. worry/VBP nsubj advcl care/VBP I/PRP mark nsubj because/IN I/PRP 24

1. Causeway-L/Causeway-S: two simple systems for tagging causal constructions Causeway-S: Syntax-based pipeline i. Causeway-L: Lexical pattern-based pipeline ii. 25

1. Causeway-L/Causeway-S: two simple systems for tagging causal constructions Causeway-S: Syntax-based pipeline i. Causeway-L: Lexical pattern-based pipeline ii. 26

Lexical pattern-based connective discovery: constructions are matched by regular expressions over word lemmas. I worry because (ˆ | )([ \ S]+ )+?(because/IN) Training: ([ \ S]+ )+? I care. I worry because I love you. I worry because + regex Test: I love you. (ˆ | )([ \ S]+ )+?(because/IN) ([ \ S]+ )+? 27

Lexical pattern-based argument ID: Arguments are labeled by a conditional random field. E FFECT E FFECT C AUSE … labels … featurized words Features include information about: • Word • Connective • Relationship between word & connective 28

Both approaches use a weighted soft vote of three classifiers as a filter. Example classifier features Classifier 1 Classifiers 2 & 3 ( c =cause head, e = effect head) : Global: Connective X: • POS tags of c and e Connective Y: • Number of words between c and e Connective Z: • Domination relationship between c and e … • Matching connective pattern • Pair of tense/aspect/modality modifier sets of c and e • POS 1-skip-2-grams of cause and effect spans • WordNet hypernyms 29

Our baseline is an argument-aware most-frequent-sense heuristic. Parse paths to Connective other tokens Causal / Not causal prevent from nsubj, advcl 27/ 4 prevent from nsubj, advmod 0 / 8 case, case  nmod because of 14 / 1 … … … 30

T oday’s talk: 1. The BECauSE corpus of causal language 2. Causeway-L/Causeway-S: two simple systems for tagging causal constructions 3. Experiments & error analysis 31

T oday’s talk: 1. The BECauSE corpus of causal language 2. Causeway-L/Causeway-S: two simple systems for tagging causal constructions 3. Experiments & error analysis 32

Our results show the techniques are viable, but further work is needed. Connectives Causes Effects Pipeline [stages] P R F 1 S C H C J C S E H E J E Causeway-S [1-2] 7.3 71.9 13.2 65.0 84.3 39.3 30.4 63.0 30.7 51.8 Causeway-S [1-4] 57.7 47.4 67.1 84.4 39.0 37.7 70.7 33.4 Causeway-L [1-2] 8.1 91.1 14.8 56.8 67.6 33.1 39.5 59.4 30.9 47.9 Causeway-L [1-4] 60.4 39.9 74.3 85.8 42.6 53.3 76.4 38.2 33.8 Baseline 88.4 21.4 74.1 94.7 43.7 48.4 83.3 38.4 33