A coherence model based on syntactic patterns by Annie Louis and Ani Nenkova M.Sc. Seminar: Discourse Coherence Theories and Modeling Nikolina Koleva Saarland University Department of Computational Linguistics June 10, 2013 Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 1 / 32 Overview 1 Motivation 2 Coherence models based on syntax Evidence for syntactic coherence Representing syntax Local co-occurrence model Global model 3 Evaluation Prediction on reports Prediction on academic articles 4 Conclusion Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 2 / 32 Motivation Factors contributing to coherence 1 attentional structure (items under discussion) 2 organization of discourse segments 3 intentional structure (purpose of the discourse) Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 3 / 32 Motivation Factors contributing to coherence 1 2 3 " organization of discourse segments " content approaches intentional structure: purpose of the discourse % not much work attentional structure: items under discussion entity approaches Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 3 / 32 Motivation Every discourse has a purpose • explaining a concept • narrating an event • critiquing an idea • ... each sentence in a text has a communicative goal Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 4 / 32 Motivation Example 1 An aqueduct is a water supply or navigable channel constructed to convey water. 2 In modern engineering, the term is used for any system of pipes, canals, tunnels, and other structures used for this purpose. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 5 / 32 Motivation Example 1 An aqueduct is a water supply or navigable channel constructed to convey water. 2 In modern engineering, the term is used for any system of pipes, canals, tunnels, and other structures used for this purpose. 1 Cytokine receptors are receptors that bind cytokines. 2 In recent years, the cytokine receptors have come to demand more attention because their deficiency has now been directly linked to certain debilitating immunodeficiency states. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 5 / 32 Motivation Example 1 An aqueduct is a water supply or navigable channel constructed to convey water. 2 In modern engineering, the termis used for any system of pipes, canals, tunnels, and other structures used for this purpose. 1 Cytokine receptors are receptors that bind cytokines. 2 In recent years, the cytokine receptors have come to demand more attention because their deficiency has now been directly linked to certain debilitating immunodeficiency states. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 5 / 32 Motivation Example 1 An aqueduct is a water supply or navigable channel constructed to convey water. 2 In modern engineering, the term is used for any system of pipes, canals, tunnels, and other structures used for this purpose. 1 Cytokine receptors are receptors that bind cytokines. 2 In recent years, the cytokine receptors have come to demand more attention because their deficiency has now been directly linked to certain debilitating immunodeficiency states. unique syntactic structure of definitions, questions etc. syntax as proxy for the communicative goal Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 5 / 32 Coherence models based on syntax Coherence model based on syntax Underlying assumptions: 1 Sentences with similar syntax are likely to have the same communicative goal. 2 Regularities in intentional structure will be manifested in syntactic regularities between adjacent sentences. supported by recent related work Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 6 / 32 Coherence models based on syntax Evidence for syntactic coherence Pilot study for the validation of assumption No: 2 • Material: gold standard parse trees from the Penn Treebank • Unit of analysis: two adjacent sentences, a pair (S1 , S2 ) Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 7 / 32 Coherence models based on syntax Evidence for syntactic coherence Pilot study for the validation of assumption No: 2 • Material: gold standard parse trees from the Penn Treebank • Unit of analysis: two adjacent sentences, a pair (S1 , S2 ) Steps: 1 enumerate all productions = 197 unique productions • productions with frequency < 25 are removed 2 for all ordered pairs (p1 , p2 ) compute • c (p1 , p2 ) ,c (p1 , ¬p2 ), c (¬p1 , p2 ) and c (¬p1 , ¬p2 ) c (p1 , p2 ): # of sentence pairs where p1 ∈ S1 and p2 ∈ S2 3 perform chi-square test to • prove significance of the count c (p1 , p2 ) • check independence of the occurrences of p1 and p2 where, p1: production 1 and p2: production 2 Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 7 / 32 Coherence models based on syntax Evidence for syntactic coherence Outcome of the study • small fraction of repetitions (5%) p1: VP → VBD SBAR p2: VP → VBD SBAR 1 S1: Documents filed with the Securities and Exchange Commission on the pending spinoff [[disclosed]VBD [that Cray Research Inc. will withdraw the almost $ 100 million in financing it is providing the new firm if Mr. Cray leaves or if the product-design project, he heads, is scrapped]SBAR ]VP . 2 S2: The documents also [[said]VBD [that although the 64-year-old Mr. Cray has been working on the project for more than six years , the Cray-3 machine is at least another year away from a fully operational prototype]SBAR ]VP . Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 8 / 32 Coherence models based on syntax Evidence for syntactic coherence Outcome of the study • finance domain-specific p1: NP → NP NP-ADV p2: QP → CD CD 1 S1: The two concerns said they entered into a definitive merger agreement under which Ratners will begin a tender offer for all of Weisfield’s common shares for [$57.50 each]NP . 2 S2: Also on the takeover front, Jaguar’s ADRs rose 1/4 to 13 7/8 on turnover of [4.4 million]QP . Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 9 / 32 Coherence models based on syntax Evidence for syntactic coherence Outcome of the study • neither repetitions nor domain dependent p1: VP → VB VP p2: NP-SBJ → NNP NNP 1 S1: "The refund pool may not [be held hostage through another round of appeals]VP , " Judge Curry said. 2 S2: [Commonwealth Edison]NP −SBJ said it is already appealing the underlying commission order and is considering appealing Judge Curry’s order. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 10 / 32 Coherence models based on syntax Evidence for syntactic coherence Outcome of the study • neither repetitions nor domain dependent p1: VP → VB VP p2: NP-SBJ → NNP NNP 1 S1: "The refund pool may not [be held hostage through another round of appeals]VP , " Judge Curry said. 2 S2: [Commonwealth Edison]NP −SBJ said it is already appealing the underlying commission order and is considering appealing Judge Curry’s order. • S1 present hypothesis or speculation • S2 introduces an entity (PERS, ORG) that gives explanation or opinion on the statement • intentional structure: SPECULATE , ENDORSE Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 10 / 32 Coherence models based on syntax Evidence for syntactic coherence Outcome of the study • neither repetitions nor domain dependent p1: NP-LOC → NNP p2: S-TPC-1 → NP-SBJ VP 1 S1: "It has to be considered as an additional risk for the investor," said Gary P. Smaby of Smaby Group Inc., [Minneapolis]NP −LOC . 2 S2: ["Cray Computer will be a concept stock,"]S −TPC −1 he said. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 11 / 32 Coherence models based on syntax Evidence for syntactic coherence Outcome of the study • neither repetitions nor domain dependent p1: NP-LOC → NNP p2: S-TPC-1 → NP-SBJ VP 1 S1: "It has to be considered as an additional risk for the investor," said Gary P. Smaby of Smaby Group Inc., [Minneapolis]NP −LOC . 2 S2: ["Cray Computer will be a concept stock,"]S −TPC −1 he said. • S1 introduces location name associated with an entity • S2 contains quote from that entity • intentional structure: INTRODUCE X , STATEMENT BY X Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 11 / 32 Coherence models based on syntax Representing syntax Representing syntax 1 productions • sentence as set of grammatical productions (LHS →RHS) • RHS could be very long and thus rather specific • available information only about nodes of the same constituent Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 12 / 32 Coherence models based on syntax Representing syntax Representing syntax 1 productions • sentence as set of grammatical productions (LHS →RHS) • RHS could be very long and thus rather specific • available information only about nodes of the same constituent 2 d-sequence • cut the parse tree at level d • sentence as sequence of leaf nodes (of the cut tree) • for each node in the sequence augmented the tag of the left most child Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 12 / 32 Coherence models based on syntax Representing syntax d-sequence example • depth-2 sequence: " S:dt , " NP:nnp VP:vbd . Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 13 / 32 Coherence models based on syntax Representing syntax d-sequence example Please, write down the depth-3 sequence. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 13 / 32 Coherence models based on syntax Representing syntax d-sequence example Please, write down the depth-3 sequence. • depth-3 sequence: " NP:dt VP:vbz , " NNP NNP VBD . Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 13 / 32 Coherence models based on syntax Local co-occurrence model Local co-occurrence model idea to test assumption No 2: Regularities in intentional structure will be manifested in syntactic regularities between adjacent sentences. Steps: 1 estimate probabilities of pairs of syntactic items (from the training set) 2 use these probabilities to compute the coherence of a new text Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 14 / 32 Coherence models based on syntax Local co-occurrence model Local co-occurrence model implementation • n: number of sentences y • Sx the y th item of the x th sentence • δC : smoothing constant • |V |: size of the vocabulary Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 15 / 32 Coherence models based on syntax Local co-occurrence model Local co-occurrence model example s2: s1: 1 S → NP VP 2 NP → DT N 3 VP → VBD NP 4 NP → DT N Nikolina Koleva (CoLi Saarland) 1 S → NP VP 2 NP → DT N 3 VP → VBD PP 4 PP → P NP 5 NP → DT N Syntactic Approach to Modeling Coherence June 10, 2013 16 / 32 Coherence models based on syntax Local co-occurrence model Local co-occurrence model example s2: s1: 1 S → NP VP 2 NP → DT N 3 VP → VBD NP 4 NP → DT N 1 S → NP VP 2 NP → DT N 3 VP → VBD PP 4 PP → P NP 5 NP → DT N • [ p (S → NP VP | S → NP VP ) + p (S → NP VP | NP → DT N ) + p (S → NP VP | VP → VBD NP) + p (S → NP VP | NP → DT N ) ] * [ p (NP → DT N | S → NP VP ) + p (NP → DT N | NP → DT N ) + p (NP → DT N | VP → VBD NP) + p (NP → DT N | NP → DT N ) ] * ... Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 16 / 32 Coherence models based on syntax Local co-occurrence model Global coherence model idea to test assumption No 1: Sentences with similar syntax are likely to have the same communicative goal. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 17 / 32 Coherence models based on syntax Local co-occurrence model Clusters from abstracts of journal articles Cluster a: VP → VBZ ADJP ; ADJP → JJ PP 1 This method [is [capable of sequence-specific detection of DNA with high accuracy]ADJP ]VP 2 The same [is [true for synthetic polyamines such as polyallylamine]ADJP ]VP Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 18 / 32 Coherence models based on syntax Local co-occurrence model Clusters from abstracts of journal articles Cluster a: VP → VBZ ADJP ; ADJP → JJ PP 1 This method [is [capable of sequence-specific detection of DNA with high accuracy]ADJP ]VP 2 The same [is [true for synthetic polyamines such as polyallylamine]ADJP ]VP Cluster b: VP → MD VP ; VP → VB VP 1 Our results for the difference in reactivity [can [be linked to experimental observations]VP ]VP 2 These phenomena taken together [can [be considered as the signature of the gelation process]VP ]VP Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 18 / 32 Coherence models based on syntax Local co-occurrence model Clusters from abstracts of journal articles Cluster a: VP → VBZ ADJP ; ADJP → JJ PP 1 This method [is [capable of sequence-specific detection of DNA with high accuracy]ADJP ]VP 2 The same [is [true for synthetic polyamines such as polyallylamine]ADJP ]VP captures descriptive sentences Cluster b: VP → MD VP ; VP → VB VP 1 Our results for the difference in reactivity [can [be linked to experimental observations]VP ]VP 2 These phenomena taken together [can [be considered as the signature of the gelation process]VP ]VP captures speculative sentences Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 18 / 32 Coherence models based on syntax Global model Global coherence model idea and to capture the common patterns in the intentional structure for the domain (by using HMM) Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 19 / 32 Coherence models based on syntax Global model Global coherence model idea and to capture the common patterns in the intentional structure for the domain (by using HMM) Steps: 1 cluster sentences of different documents by syntactic similarity features : • productions: frequency of production in a parse tree • d-sequence: n-grams of size one to four 2 estimate emission and transition probabilities (from the training set) 3 use these probabilities to compute the coherence of a new text Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 19 / 32 Coherence models based on syntax Global model Global coherence model Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 20 / 32 Coherence models based on syntax Global model Global coherence model implementation • n: number of sentences, St : the t th sentence • ht : the t th state • d (ht ): # of docs whose sentences appear in ht • d (ht , ht −1 ) # of docs where subsequent sentences in subsequent states • δM : smoothing constant, C: # of clusters Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 21 / 32 Evaluation Evaluating syntactic coherence 1 prediction on reports • use pairs of articles: (original article, random permutation) • testing on identifying the original article • compare with content and entity approaches Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 22 / 32 Evaluation Evaluating syntactic coherence 1 prediction on reports • use pairs of articles: (original article, random permutation) • testing on identifying the original article • compare with content and entity approaches 2 prediction on academic articles • original vs. permuted sections • conference vs. workshop papers Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 22 / 32 Evaluation Prediction on reports Prediction on reports Model baseline Prod d-seq PoS Prod d-seq Egrid Content Airplane Accidents Earthquake Parameters Accuracy Parameters 50.0 Local co-occurrence model 72.8 depth MVP + 2 71.8 depth MVP + 1 61.3 HMM-syntax clus. 37 74.6 clus. 5 depth MVP + 8, clus. 8 82.2 depth MVP + 9, clus. 45 Other approaches history 1 67.6 history 1 clus.48 71.4 clus. 23 Accuracy 50.0 55.0 65.1 42.6 93.8 86.5 82.2 84.5 the articles of each corpus have the same intentional structure Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 23 / 32 Evaluation Prediction on reports Combining predictions of different models Accuracy Airplane Accidents Earthquake Model Content + Egrid Content + HMM-prod Content + HMM-d-seq Egrid + HMM-prod Egrid + HMM-d-seq Egrid + Content + HMM-prod Egrid + Content + HMM-d-seq Egrid + Content + HMM-prod + HMM-d-seq Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence 76.8 74.2 82.1 79.6 84.2 79.5 84.1 83.6 90.7 95.3 90.3 93.9 91.1 95.0 92.3 95.7 June 10, 2013 24 / 32 Evaluation Prediction on reports Combining predictions of different models Accuracy Airplane Accidents Earthquake Model Content + Egrid Content + HMM-prod Content + HMM-d-seq Egrid + HMM-prod Egrid + HMM-d-seq Egrid + Content + HMM-prod Egrid + Content + HMM-d-seq Egrid + Content + HMM-prod + HMM-d-seq 76.8 74.2 82.1 79.6 84.2 79.5 84.1 83.6 90.7 95.3 90.3 93.9 91.1 95.0 92.3 95.7 syntax supplements content and entity grid methods Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 24 / 32 Evaluation Prediction on academic articles Corpora of academic articles 1 ART Corpus • 225 Chemistry journal articles • manually annotated for intentional structure 2 ACL Anthology Network (AAN) • 500 ACL-NAACL conference articles • 500 ACL-sponsored workshop articles Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 25 / 32 Evaluation Prediction on academic articles Detected clusters vs. manually annotated zones • manually annotated zones in ART Motivation Background 3 Hypothesis 4 Objective 5 ... 1 2 • compare to detected clusters • compute c (Ci , Zj ): # of sentences that are annotated as Zj and are contained in Ci Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 26 / 32 Evaluation Prediction on academic articles Original vs. permuted sections Table : Accuracy in % Data Section Test Pairs ART Corpus Abstract Introduction Abstract Introduction Rel. work 1633 1640 8815 9966 10,000 ACL Nikolina Koleva (CoLi Saarland) Local-prod Local-d-seq HMM-prod HMM-d-seq Oracle zones 57.0 44.5 44.0 54.5 54.6 52.9 54.6 47.2 53.0 54.4 64.1 58.1 58.2 64.4 57.3 55.0 64.6 63.7 74.0 67.3 80.8 94.0 Syntactic Approach to Modeling Coherence June 10, 2013 27 / 32 Evaluation Prediction on academic articles Distinguish conference and workshop articles • conference articles: complete work information presentation differ in abstracts and introductions • workshop articles: preliminary studies • used features • indicating perplexity of the local and global models • fine-grained taken from the local model • most significant 30 pairs Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 28 / 32 Evaluation Prediction on academic articles Distinguish conference and workshop articles • conference articles: complete work information presentation differ in abstracts and introductions • workshop articles: preliminary studies • used features • indicating perplexity of the local and global models • fine-grained taken from the local model • most significant 30 pairs Table : Accuracy above confidence level Conf Abstract Introduction Rel. work >= 0.5 59.3 50.3 55.4 Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 28 / 32 Conclusion Summary Syntactic patterns are reliable clues for intentional structure detection • Possible syntactic representations • • productions • d-sequence • Local coherence model: exploring pairs of adjacent sentences • Global coherence model: clustering sentences based on syntactic similarity • High accuracy on distinguishing coherent and incoherent news articles Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 29 / 32 Conclusion Thank you for your attention! Any questions? Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 30 / 32 Conclusion Discussion • Would this approach work for languages with free word order? Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 31 / 32 Conclusion References Annie Louis and Ani Nenkova. A coherence model based on syntactic patterns. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL ’12, pages 1157–1168, Stroudsburg, PA, USA, 2012. Association for Computational Linguistics. URL http: //dl.acm.org/citation.cfm?id=2390948.2391078. Manfred Stede. Discourse Processing. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers, 2011. Nikolina Koleva (CoLi Saarland) Syntactic Approach to Modeling Coherence June 10, 2013 32 / 32
© Copyright 2026 Paperzz