Introducing a Lexicon of Verbal Polarity Shifters for English Marc - PowerPoint PPT Presentation

Introducing a Lexicon of Verbal Polarity Shifters for English Marc Schulder Josef Ruppenhofer Michael Wiegand Stephanie Köser Spoken Language Systems Institute for German Language Saarland University, Germany Mannheim, Germany

What are Polarity Shifters? Saarland University Marc Schulder � 2

Shifters vs Negation Negation Polarity Shifters Example He did [ not have [hope] + ] - . He [ abandoned [hope] + ] - . Saarland University Marc Schulder � 3

Shifters vs Negation Negation Polarity Shifters Example He did [ not have [hope] + ] - . He [ abandoned [hope] + ] - . Word Type Closed Class Open Class Saarland University Marc Schulder � 3

Shifters vs Negation Negation Polarity Shifters Example He did [ not have [hope] + ] - . He [ abandoned [hope] + ] - . Word Type Closed Class Open Class Resource Cheap Expensive Creation Saarland University Marc Schulder � 3

Shifters vs Negation Negation Polarity Shifters Example He did [ not have [hope] + ] - . He [ abandoned [hope] + ] - . Word Type Closed Class Open Class Resource Cheap Expensive Creation Existing polarity classifiers 
 can process negation, 
 but fail to detect polarity shifters 
 due to a lack of resources . Saarland University Marc Schulder � 3

Goal Create a Lexicon 
 of 
 Verbal Polarity Shifters Saarland University Marc Schulder � 4

Goal Create a Lexicon 
 of 
 Verbal Polarity Shifters Main sentence predicate ⇒ far reaching scope Saarland University Marc Schulder � 4

Title Text Overview • Polarity Shifters • Related Work • Word Sense Ambiguity • Shifter Scope • Lexicon • Annotation Process • Lexicon Example • Conclusion Saarland University Marc Schulder � 5

Related Work Supervised Complete Bootstrapping Annotation (Schulder et al., 2017) (this work) Input Resource WordNet WordNet Lexicon Size 3,000 verbs 10,500 verbs Shifter Labels Lemma Word Sense Additional - Shifter Scope Information Saarland University Marc Schulder � 6

Word Sense Ambiguity • 50% of verbs are polysemous. • 12% of verbs are shifters in at least one word sense. • Among polysemous verbal shifters, 
 only 23% are shifters in all their word senses. Mark down: Reduce in price Shifter The agency [marked down [ their assets ] + ] - . Mark down: Write down No Shifter She [marked down [ his confession of guilt ] - ] - . Saarland University Marc Schulder � 7

Shifter Scope When a phrase contains a polarity shifter, you need to know what part of the phrase it can a ff ect. (Wiegand et al., 2017, GSCL) [ The villain ] - defeated [ the hero ] + . [ The villain ] - surrendered [ to the hero ] + . Saarland University Marc Schulder � 8

Shifter Scope When a phrase contains a polarity shifter, you need to know what part of the phrase it can a ff ect. (Wiegand et al., 2017, GSCL) dobj - [ The villain ] - defeated [ the hero ] + . [ The villain ] - surrendered [ to the hero ] + . Saarland University Marc Schulder � 8

Shifter Scope When a phrase contains a polarity shifter, you need to know what part of the phrase it can a ff ect. (Wiegand et al., 2017, GSCL) dobj - [ The villain ] - defeated [ the hero ] + . + [ The villain ] - surrendered [ to the hero ] + . subj Saarland University Marc Schulder � 8

Shifter Scope When a phrase contains a polarity shifter, you need to know what part of the phrase it can a ff ect. (Wiegand et al., 2017, GSCL) dobj - [ The villain ] - defeated [ the hero ] + . + [ The villain ] - surrendered [ to the hero ] + . subj Scope annotated for dependency relations. 
 Assumes active sentence. Saarland University Marc Schulder � 8

Shifter Scope Direct Object 
 71% Prepositional 
 Object 
 10% Subject 
 18% Clausal 
 Complement 
 1% Saarland University Marc Schulder � 9

Shifter Scope Direct Object 
 71% Prepositional 
 Object 
 10% Verbal shifters 
 a ff ect not only 
 direct objects Subject 
 18% Clausal 
 Complement 
 1% Saarland University Marc Schulder � 9

Title Text Overview • Polarity Shifters • Related Work • Word Sense Ambiguity • Shifter Scope • Lexicon • Annotation Process • Lexicon Example • Conclusion Saarland University Marc Schulder � 10

Annotation Workflow Lemma 1 Lemma 2 Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Lemma 1 Lemma 2 Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Lemma 1 Sense 1 Sense 2 Lemma 2 Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Lemma 1 Sense 1 Shifter? Sense 2 Lemma 2 Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 Shifter? Definition Sense 2 Lexicon Lemma 2 Examples Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 Shifter? Definition Sense 2 Lexicon Lemma 2 Examples No :( Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 ✘ Shifter? Definition Sense 2 Lexicon Lemma 2 Examples Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 ✘ Shifter? Definition Sense 2 Lexicon Lemma 2 Examples Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 ✘ Shifter? Definition Sense 2 Lexicon Lemma 2 Examples Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 ✘ Shifter? Definition Sense 2 ✔ Lexicon Yes! Lemma 2 Examples Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 ✘ Shifter? Definition Sense 2 ✔ Lexicon Lemma 2 Examples What Lemma 3 Scope? Saarland University Marc Schulder � 11

Annotation Workflow Is Sense Lemma 1 Sense 1 ✘ Shifter? Definition Sense 2 ✔ Lexicon Lemma 2 Examples What Lemma 3 Scope? Saarland University Marc Schulder � 11

Annotation Workflow Lemma 1 Sense 1 ✘ Sense 2 ✔ Subj Lemma 2 Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Lemma 1 Sense 1 ✘ Sense 2 ✔ Subj Lemma 2 Lemma 3 Saarland University Marc Schulder � 11

Annotation Workflow Lemma 1 Sense 1 ✘ Sense 2 ✔ Subj Lemma 2 Sense Lemma 3 = Lemma-Synset pair Saarland University Marc Schulder � 11

Annotation Workflow Lemma 1 Sense 1 Observation: 
 ✘ Sense annotation 
 considerably faster than Sense 2 ✔ Subj lemma annotation Lemma 2 Sense Lemma 3 = Lemma-Synset pair Saarland University Marc Schulder � 11

Annotators Expert Annotator: Experience in linguistics and annotation work Inter-annotator Agreement: 2nd annotator labelled 400 word senses Cohen's 𝝀 = 0.73 ⇒ Substantial agreement Both annotators are authors of this paper. Saarland University Marc Schulder � 12

Lexicon Example Blow out Synset 00436247 SUBJ Shifter melt, break, or become otherwise unusable Blow out Synset 02767855 DOBJ Shifter put out, as of fires, flames, or lights Blow out Synset 02766970 - No Shifter erupt in an uncontrolled manner Saarland University Marc Schulder � 13

Conclusion Summary 
 We introduced a lexicon of English verbal shifters: • Covers all verbs in WordNet • Annotations for each word sense • Shifter labels • Shifter scope labels Data • https://github.com/uds-lsv/lrec2018 Future Work • Nouns, adjectives • Other languages Saarland University Marc Schulder � 14

Title Text Thank You Saarland University Marc Schulder

References N. Schneider, D. Hovy, A. Johannsen, and M. N. Jindal and B. Liu. 2008. Opinion Spam and Carpuat. 2016. SemEval-2016 Task 10: Detecting Analysis . In Proceedings of WSDM. Minimal Semantic Units and their Meanings S. Kiritchenko and S.M. Mohammad. 2016. The (DiMSUM) . In Proceedings of SemEval@NAACL- E ff ect of Negators, Modals, and Degree HLT . Adverbs on Sentiment Composition . In M. Schulder, M. Wiegand, J. Ruppenhofer, and B. Proceedings of WASSA@NAACL/HLT. Roth. 2017. Towards Bootstrapping a Polarity G. Miller, R. Beckwith, C. Fellbaum, D. Gross, and Shifter Lexicon using Linguistic Features . In K. Miller. 1990. Introduction to WordNet: An Proceedings of IJCNLP . Online Lexical Database . International Journal of M. Wiegand, A. Balahur, B. Roth, D. Klakow, and Lexicography , 3:235–244. A. Montoyo. 2010. A Survey on the Role of C. Paradis, and C. Willners. 2006. Antonymy and Negation in Sentiment Analysis . In Proceedings Negation: The Boundedness Hypothesis . of NeSp-NLP . Journal of Pragmatics, 38(7):1051–1080. M. Wiegand, M. Wolf, and J. Ruppenhofer. 2017. L. Polanyi, and A. Zaenen. 2006. Contextual Negation Modeling for German Polarity Valence Shifters . In J.G. Shanahan, et al., editors, Classification. In Proceedings of GSCL. Computing Attitude and A ff ect in Text: Theory and T. Wilson, J. Wiebe, and P . Ho ff mann. 2005. Applications , pages 1–10, Springer Netherlands. Recognizing Contextual Polarity in Phrase-level J. Ruppenhofer, and J. Brandes. 2015. Extending Sentiment Analysis. In Proceedings of HLT/ e ff ect annotation with lexical decomposition . In EMNLP . Proceedings of WASSA@EMNLP . Saarland University Marc Schulder � 16