Minimally Supervised Event Causality Identification Quang Do, Yee Seng, and Dan Roth University of Illinois at Urbana-Champaign EMNLP-2011 1 Event Causality The police arrested him because he killed someone. 2 Event Causality The police arrested him because he killed someone. event trigger event trigger 3 Event Causality causality The police arrested him because he killed someone. event trigger event trigger • We identify causality between event pairs, but not the direction 4 Event Causality calculate causality association: co-occurrence counts, pointwise mutual information (PMI)… The police arrested him because he killed someone. 5 Event Causality connective The police arrested him because he killed someone. contingency discourse relation 6 Event Causality distributional association score Distributional The police arrested him because he killed someone. discourse relation prediction Discourse Identify multiple cues to jointly identify event causality: Distributional association scores discourse relation predictions 7 Cause-Effect Association (CEA) and Discourse Relations We define an event e as: p(a1, a2, …, an): Distributional CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) association between event predicates association between the predicate of an event and the arguments of the other event association between event arguments Discourse [ … e … e … ] connective [ … e … e … ] • A connective is associated with two text spans • Training on the Penn Discourse Treebank (PDTB), we developed a system that predicts the discourse relations of expressed by the connectives 8 Event Definition We define an event e as: p(a1, a2, …, an): predicate p: the event trigger word a1, a2, …, an: arguments associated with e Examples: Verbs: “… he killed someone …” Nominals: “… the attack by the troops …” 9 Contributions (Event Causality) We identify causality between event pairs in context: verb-verb, verb-noun, noun-noun triggered event pairs (prior work usually focus on just verb triggers) A minimally supervised approach to detect event causality using distributional similarity methods Leverage the interactions between event causality prediction and discourse relations prediction 10 Overview (Event Causality) Event causality: Cause-Effect Association (CEA) Discourse and Causality: Discourse relations Constraints for joint inference with CEA Experiments: Interaction between event causality and discourse relations Event predicates: verbs, nominals Settings Evaluation Analysis Conclusion 11 Overview (Event Causality) Event causality: Cause-Effect Association (CEA) Discourse and Causality: Discourse relations Constraints for joint inference Experiments: Interaction between event causality and discourse relations Event predicates: verbs, nominals Settings Evaluation Analysis Conclusion 12 Cause-Effect Association (CEA) CEA: prediction of whether two events are causally related The police arrested him because he killed someone. 13 Cause-Effect Association (CEA) We define an event e as: p(a1, a2, …, an): predicate p: the event trigger word (e.g.: arrested, killed) a1, a2, …, an: arguments associated with e CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) association between event predicates association between event arguments association between the predicate of an event and the arguments of the other event 14 Predicate-Predicate Association CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) spp (ei , e j ) = PMI(pi , p j )´ IDF(pi , p j )´ Dist(pi , p j )´ max(ui , u j ) log P ( pi , p j ) P(pi )P( p j ) 15 Predicate-Predicate Association CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) spp (ei , e j ) = PMI(pi , p j )´ IDF(pi , p j )´ Dist(pi , p j )´ max(ui , u j ) idf (pi )´idf (p j )´idf (pi , p j ) D idf (p) = log 1+ N D: total number of documents in the collection N: number of documents that p occurs in 16 Predicate-Predicate Association CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) spp (ei , e j ) = PMI(pi , p j )´ IDF(pi , p j )´ Dist(pi , p j )´ max(ui , u j ) awards event pairs that are closer together in the texts (in terms of num# of sentences apart), while penalizing event pairs that are further apart 17 Predicate-Predicate Association CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) spp (ei , e j ) = PMI(pi , p j )´ IDF(pi , p j )´ Dist(pi , p j )´ max(ui , u j ) takes into account whether predicates (events) pi and pj appear most frequently with each other 18 Predicate-Predicate Association i j max(u , u ) takes into account whether predicates (events) pi and pj appear most frequently with each other i j P(p , p ) i u = max k éëP(pi , pk )ùû - P(pi , p j ) + e ui will be maximized if there is no other predicate pk (as compared to pj) having a higher co-occurrence probability with pi i j P(p , p ) j u = max k éëP(p k , p j )ùû - P(pi , p j ) + e 19 Predicate-Argument Association CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) i i 1 i 2 i n j j 1 j 2 j m p (a , a ,..., a ) p (a , a ,..., a ) Pair up the predicates and arguments across events, calculate the PMI for each link, then average them 20 Argument-Argument Association CEA(ei , e j ) = spp (ei , e j )+ s pa (ei , e j ) + saa (ei , e j ) i i 1 i 2 i n j j 1 j 2 j m p (a , a ,..., a ) p (a , a ,..., a ) calculate the PMI for each possible pairings of the arguments (across the two events), then average them 21 Cause-Effect Association (CEA) CEA score: predicts whether the two events are causally related The police arrested him because he killed someone. 22 Overview (Event Causality) Event causality: Cause-Effect Association (CEA) Discourse and Causality: Discourse relations Constraints for joint inference with CEA Experiments: Interaction between event causality and discourse relations Event predicates: verbs, nominals Settings Evaluation Analysis Conclusion 23 Discourse and Causality Interaction [ … e … e … ] connective [ … e … e … ] Interaction between: • Discourse relation evoked by the connective c • Relations between ep (event pairs that crosses the two text spans) causal? not-causal? 24 Penn Discourse Treebank (PDTB) Relations Discourse relations: Comparison: Contingency: Cause, Condition, Pragmatic-cause, Pragmatic-condition Expansion: Concession, Contrast, Pragmatic-concession, Pragmatic-contrast Alternative, Conjunction, Exception, Instantiation, List, Restatement Temporal: Asynchronous, Synchronous 25 Discourse Relations Comparison: Highlights differences between the situations described in the text spans: Contrast: [According to the survey, x% of Chinese Internet users prefer Google] whereas [y% prefer Baidu]. Negative evidence for causality Contingency: The situation described in one text span causally influences the situation in the other: Cause: [The first priority is search and rescue] because [many people are trapped under the rubble]. Positive evidence 26 Discourse Relations Expansion: Providing additional information, illustrating alternative situations, etc.: Conjunction: [Over the past decade, x women were killed] and [y went missing]. Negative evidence, except for Conjunction (which connects arbitrary pieces of text spans) Temporal: Synchrony: [He was sitting at his home] when [the whole world started to shake]. Temporal precedence of the (cause) event over the (effect) event is a necessary, but not sufficient requisite for causality 27 Discourse and Causality Interaction [ … e … e … ] connective [ … e … e … ] Cause, Condition 1 ei ej At least one (crossing) ep is causal 2 ei ej If we have a (crossing) ep which is causal 3 Comparison, Concession, Contrast, Pragmatic-concession, Pragmatic-contrast, Expansion, Alternative, Exception, Instantiation, List, Restatement Cause, Condition, Temporal, Asynchronous, Synchrony, Conjunction ei ej No (crossing) ep is casual 28 Joint Inference: Discourse & Distributional Causality Objective function: Arg max LDR sc (dr ) x c ,dr LER sep (er ) y ep ,er cC drLDR epEP erLER Probability that connective c is predicted with discourse relation dr discourse relation indicator variable CEA prediction that event pair ep takes on the causal or not-causal label er event pair causality indicator variable 29 Constraints [ … e … e … ] connective [ … e … e … ] 1 Cause, Condition ei ej At least one (crossing) ep is causal x c,"Cause" £ å y ep,"causal" epÎEPc x c ,"Cause" y epi ,"causal " y ep j ,"causal " If the connective is predicted with a “Cause” discourse relation, then the CEA system should predict that at least one of the (crossing) event pair is causally related 30 Constraints [ … e … e … ] connective [ … e … e … ] 2 ei ej If we have a (crossing) ep which is causal y ep,"causal" £ å dra ÎLDRa Cause, Condition, Temporal, Asynchronous, Synchrony, Conjunction x c,dra y ep ,"causal " x c ,"Cause" x c ,"Condition" x c ,"Conjunction" If a (crossing) event pair is predicted by CEA as causally related, then the associated connective should be predicted as having discourse relation; “Cause”, “Condition”, …, “Conjunction” 31 Constraints [ … e … e … ] connective [ … e … e … ] 3 Comparison, Concession, Contrast, Pragmatic-concession, Pragmatic-contrast, Expansion, Alternative, Exception, Instantiation, List, Restatement ei ej No (crossing) ep is casual x c,drb £ y ep,"Øcausal" x c,drb Þ y ep,"Øcausal" , "drb Î{“Comparison”,”Concession”…} If the connective is predicted with discourse relation “Comparison”, “Concession”, …, “Restatement”; no (crossing) event pair is causally related 32 Overview (Event Causality) Event causality: Cause-Effect Association (CEA) Discourse and Causality: Discourse relations Constraints for joint inference Experiments: Interaction between event causality and discourse relations Event predicates: verbs, nominals Settings Evaluation Analysis Conclusion 33 Experimental Settings To collect the distributional statistics for measuring CEA: 760K documents in the English Gigaword corpus 25 CNN documents from first three months of 2010: 20 documents for evaluation 5 documents for development 34 Annotation for Causal Event Pairs Annotation guidelines: The Cause event should temporally precede the Effect event; the Effect event occurs because the Cause event occurs 35 Annotation for Causal Event Pairs … Si-1 Si Si+1 … Document R (relatedness) C (causality) R (relatedness) … Drawing links between event predicates: Event arguments are not annotated, but annotators are free to look at the entire document text Annotators are not restricted to a fixed sentence window size 36 Annotation for Causal Event Pairs # relations Eval Dev C 414 71 C+R 492 92 Annotators overlap on 10 evaluation documents. Agreement ratio: 0.67 for C+R 0.58 for C 37 Performance on Extracting Causality 70 60 50 PMI ECD CEA CEA+Discourse 40 30 20 10 0 Rec% Prec% F1% 38 Performance on Extracting Causality and Relatedness 70 60 50 PMI ECD CEA CEA+Discourse 40 30 20 10 0 Rec% Prec% F1% 39 Analysis of CEA mistakes 50 (randomly selected) false-positives (precision errors): 56%: CEA assigns a high score to event pairs that are not causal 22%: involves events containing pronouns (“he”, “it”, etc.) as arguments 50 false-negatives (recall errors): 23%: CEA assigns a low score to causal event pairs 19%: involving nominal predicates that are not in our list of event evoking noun types 17%: involving nominal predicates without any argument (less information for CEA) 15%: involves events containing pronouns as arguments 40 Conclusion (Event Causality) Developed a minimally supervised approach to identify event causality Use distributional scores and discourse relations to jointly identify event causality 41