Cognitive Systems Monographs 37. Aditya Joshi Pushpak Bhattacharyya Mark J. Carman. Investigations in Computational Sarcasm

Transcription

1 Cognitive Systems Monographs 37 Aditya Joshi Pushpak Bhattacharyya Mark J. Carman Investigations in Computational Sarcasm

2 Cognitive Systems Monographs Volume 37 Series editors Rüdiger Dillmann, University of Karlsruhe, Karlsruhe, Germany Yoshihiko Nakamura, Tokyo University, Tokyo, Japan Stefan Schaal, University of Southern California, Los Angeles, USA David Vernon, University of Skövde, Skövde, Sweden

3 The Cognitive Systems Monographs (COSMOS) publish new developments and advances in the fields of cognitive systems research, rapidly and informally but with a high quality. The intent is to bridge cognitive brain science and biology with engineering disciplines. It covers all the technical contents, applications, and multidisciplinary aspects of cognitive systems, such as Bionics, System Analysis, System Modelling, System Design, Human Motion, Understanding, Human Activity Understanding, Man-Machine Interaction, Smart and Cognitive Environments, Human and Computer Vision, Neuroinformatics, Humanoids, Biologically motivated systems and artefacts Autonomous Systems, Linguistics, Sports Engineering, Computational Intelligence, Biosignal Processing, or Cognitive Materials as well as the methodologies behind them. Within the scope of the series are monographs, lecture notes, selected contributions from specialized conferences and workshops. Advisory Board Heinrich H. Bülthoff, MPI for Biological Cybernetics, Tübingen, Germany Masayuki Inaba, The University of Tokyo, Japan J.A. Scott Kelso, Florida Atlantic University, Boca Raton, FL, USA Oussama Khatib, Stanford University, CA, USA Yasuo Kuniyoshi, The University of Tokyo, Japan Hiroshi G. Okuno, Kyoto University, Japan Helge Ritter, University of Bielefeld, Germany Giulio Sandini, University of Genova, Italy Bruno Siciliano, University of Naples, Italy Mark Steedman, University of Edinburgh, Scotland Atsuo Takanishi, Waseda University, Tokyo, Japan More information about this series at

4 Aditya Joshi Pushpak Bhattacharyya Mark J. Carman Investigations in Computational Sarcasm 123

5 Aditya Joshi IITB-Monash Research Academy Indian Institute of Technology Bombay Mumbai, Maharashtra India Mark J. Carman Faculty of Information Technology Monash University Melbourne, VIC Australia Pushpak Bhattacharyya Department of Computer Science and Engineering Indian Institute of Technology Bombay Mumbai, Maharashtra India ISSN ISSN (electronic) Cognitive Systems Monographs ISBN ISBN (ebook) Library of Congress Control Number: Springer Nature Singapore Pte Ltd This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore , Singapore

6 Preface Sarcasm is defined as verbal irony that is intended to mock or ridicule. Existing sentiment analysis systems show a degraded performance in case of sarcastic text. Hence, computational sarcasm has received attention from the sentiment analysis community. Computational sarcasm refers to computational techniques that deal with sarcastic text. This monograph presents our investigations in computational sarcasm based on the linguistic notion of incongruity. For example, the sentence I love being ignored is sarcastic because the positive word love is incongruous with the negative phrase being ignored. These investigations are divided into three parts: understanding the phenomenon of sarcasm, sarcasm detection, and sarcasm generation. To first understand the phenomenon of sarcasm, we consider two components of sarcasm: implied negative sentiment and presence of a target. To understand how implied negative sentiment plays a role in sarcasm understanding, we present an annotation study which evaluates the quality of a sarcasm-labeled dataset created by non-native annotators. Following this, in order to show how the target of sarcasm is important to understand sarcasm, we first describe an annotation study which highlights the challenges in distinguishing between sarcasm and irony (since irony does not have a target while sarcasm does) and then present a computational approach that extracts the target of a sarcastic text. We then present our approaches for sarcasm detection. To detect sarcasm, we capture incongruity in two ways: intra-textual incongruity where we look at the incongruity within the text to be classified (i.e., target text) and the context incongruity where we incorporate information outside the target text. To detect incongruity within the target text, we present four approaches: (a) a classifier that captures sentiment incongruity using sentiment-based features (as in the case of I love being ignored ), (b) a classifier that captures semantic incongruity (as in the case of A woman needs a man like a fish needs bicycle ) using word embedding-based features, (c) a topic model that captures sentiment incongruity using sentiment distributions in the text (in order to discover sarcasm-prevalent topics such as work, college), and (d) an approach that captures incongruity in the language model using sentence completion. The approaches in (a) and v

7 vi Preface (c) incorporate sentiment incongruity relying on sentiment-bearing words, whereas approach in (b) and (d) tackles other forms of incongruity where sentiment-bearing words may not be present. On the other hand, to detect sarcasm using contextual incongruity, we describe two approaches: (a) a rule-based approach that uses historical text by an author to detect sarcasm in the text generated by them and (b) a statistical approach that uses sequence labeling techniques for sarcasm detection in dialogue. The approach in (a) attempts to detect sarcasm that requires author-specific context, while that in (b) attempts to detect sarcasm that requires conversation-specific context. Finally, we present a technique for sarcasm generation. In this case, we use a template-based approach to synthesize incongruity and generate a sarcastic response to user input. The output of our sarcasm generation system obtains high scores on three quality parameters: coherence, grammaticality, and sarcastic nature. Also, the human evaluators are able to sufficiently identify the output of our system from that of a general purpose chatbot. Our investigations demonstrate how evidences of incongruity (such as sentiment incongruity, semantic incongruity) can be modeled using different learning techniques (such as classifiers, topic models) for sarcasm detection and sarcasm generation. In addition, our findings establish the promise of novel problems, such as sarcasm target identification and sarcasm-versus-irony classification, and provide insights for future research in sarcasm detection.

8 Contents 1 Introduction Sentiment Analysis (SA) Challenges Research Problems Applications Sarcasm and Computational Sarcasm Motivation Turing Test-Completeness Impact on Sentiment Classification Prevalence of Sarcasm In Popular Culture On the Web Sarcasm Studies in Linguistics Incongruity for Sarcasm Contribution Literature Survey Problem Definition Datasets Other Datasets (Dialogues, Syntactic Patterns, etc.) Approaches Rule-Based Approaches Statistical Approaches Deep Learning-Based Approaches Shared Tasks and Benchmark Datasets Reported Performance Trends in Sarcasm Detection Pattern Discovery Role of Context in Sarcasm Detection Issues in Sarcasm Detection vii

9 viii Contents Issues with Annotation Issues with Sentiment as a Feature Dealing with Dataset Skews Sentiment Analysis at IIT Bombay Summary Monograph Organization Understanding the Phenomenon of Sarcasm Impact on Cross-Cultural Annotation Motivation Experiment Description Analysis Sarcasm-versus-irony Classification Motivation Experiment Description Analysis An Approach for Identification of the Sarcasm Target Motivation Architecture Experiment Description Results and Analysis Summary Sarcasm Detection Using Incongruity Within Target Text Sentiment Incongruity as Features Motivation Sentiment Incongruity-Based Features Experiment Setup Results Error Analysis Semantic Incongruity as Features Motivation Word Embedding-Based Features Experiment Setup Results Error Analysis Sentiment Incongruity Using Topic Model Motivation Model Experiment Setup Results Application to Sarcasm Detection... 83

10 Contents ix 3.4 Language Model Incongruity Using Sentence Completion Motivation Approach Experiment Setup Results Discussion Error Analysis Summary Sarcasm Detection Using Contextual Incongruity Contextual Incongruity in a Monologue Motivation Architecture Experiment Setup Results Error Analysis Contextual Incongruity in Dialogue Motivation Architecture Conversational Sarcasm Dataset Paradigm 1: Traditional Models Paradigm 2: Deep Learning-Based Models Results Error Analysis Summary Sarcasm Generation Motivation Architecture Input Analyzer Generator Selector Sarcasm Generator Evaluation Experiment Details Results Summary Conclusion and Future Work Summary Conclusion Future Work References

11 About the Authors Aditya Joshi successfully defended his Ph.D. thesis at IITB-Monash Research Academy, Mumbai, a joint Ph.D. program run by the Indian Institute of Technology Bombay (IIT Bombay) and Monash University, Australia, since January His primary research focus is computational sarcasm, and he has explored different ways in which incongruity can be captured in order to detect and generate sarcasm. In addition, he has worked on innovative applications of natural language processing (NLP) such as sentiment analysis for Indian languages, drunk-texting prediction, news headline translation, and political issue extraction. The monograph is an outcome of his Ph.D. research. Dr. Pushpak Bhattacharyya is the current president of the Association for Computational Linguistics (ACL) ( ). He is the Director of the Indian Institute of Technology Patna (IIT Patna) and Vijay and Sita Vashee Chair Professor in the Department of Computer Science and Engineering at Indian Institute of Technology Bombay (IIT Bombay). He was educated at the Indian Institute of Technology Kharagpur (IIT Kharagpur) (B.Tech), Indian Institute of Technology Kanpur (IIT Kanpur) (M.Tech.), and IIT Bombay (Ph.D.). He has been a Visiting Scholar and Faculty Member at the Massachusetts Institute of Technology (MIT), Stanford, UT-Houston, and University Joseph Fourier (France). His research areas include natural language processing, machine learning, and artificial intelligence (AI). Loved by his students for his inspiring teaching and mentorship, he has guided more than 250 students (Ph.D., masters, and bachelors). He has published over 250 research papers, is the author of the textbook Machine Translation, and has led government and industry projects of international and national importance. His significant contributions in the field include multilingual lexical knowledge bases and projection. He is a fellow of the National Academy of Engineering and recipient of the IIT Bombay's Patwardhan Award and the Indian Institute of Technology Roorkee's (IIT Roorkee) VNMM award, both for technology development. He has also received IBM, Microsoft, Yahoo, and United Nations faculty grants. xi

12 xii About the Authors Dr. Mark J. Carman is a Senior Lecturer at the Faculty of Information Technology, Monash University, Australia. He obtained a Ph.D. from the University of Trento, Italy, in His research and interests span from theoretical studies (e.g., investigating statistical properties of information retrieval measures) to practical applications (e.g., technology for assisting police during digital forensic investigations). He has authored a large number of publications in prestigious venues, including full papers at SIGIR, KDD, IJCAI, CIKM, WSDM, CoNLL, and ECIR and articles in TOIS, IR, JMLR, ML, PR, JAIR, and IP&M.

13 Chapter 1 Introduction TheriseofWeb2.0 1 enabled Internet users to generate content, which often contained emotion. Considering the value of this content, automatic prediction of sentiment, i.e., sentiment analysis, became a popular area of research in natural language processing. A recent advancement in sentiment analysis research is the focus on a challenge to sentiment analysis, namely sarcasm. Sarcasm is a peculiar form of sentiment expression where words of a certain polarity are used to imply a different polarity, with an intention to mock or ridicule. While sarcasm is often used as a device to express humor, its prevalence makes it important for sentiment analysis. In 2014, a BBC story stated that the US Secret Service was also seeking a sarcasm detection system. 2 Similar interest in sarcasm has led to the work in computational approaches to process sarcasm over the last few years. We refer to them collectively as computational sarcasm. There are several facets of computational sarcasm, analogous to natural language processing. Like natural language processing covers a broad spectrum of approaches to natural language generation and several detection problems (such as sentiment detection, part-of-speech prediction), computational sarcasm covers similar problems such as sarcasm generation and sarcasm detection. This monograph takes an in-depth look into the problem of computational sarcasm. One might argue that computational sarcasm in text alone is insufficient since sarcasm is understood through non-verbal cues. For example, rolling one s eyes is a common indicator of insincerity that often accompanies sarcasm. The importance of non-verbal cues is true without a doubt. However, social media today relies heavily on text, and sarcastic content on social media today has a high volume. Therefore, it is natural that the current focus of computational sarcasm is textual data. In fact, several indicators of non-verbal cues exist in the form of hashtags, emoticons, etc. Therefore, computational sarcasm in text is a viable task in itself. This monograph Springer Nature Singapore Pte Ltd A. Joshi et al., Investigations in Computational Sarcasm, Cognitive Systems Monographs 37, 1

14 2 1 Introduction describes our investigations in computational sarcasm of text. Our investigations keep in focus prior work in sarcasm detection, while building upon it. This chapter builds the foundation of this monograph and is organized as follows. We first introduce sentiment analysis (SA) in Sect. 1.1 followed by computational sarcasm in Sect We then give the motivation behind computational sarcasm in Sect We discuss prevalence of sarcasm in popular culture and social media in Sect We describe linguistic theories of sarcasm in Sect. 1.5 and, specifically, the notion of incongruity in Sect Incongruity forms the foundation of our work. In Sect. 1.7, we specify our contribution. Section 1.8 describes prior work in computational sarcasm, specifically sarcasm detection. Finally, the monograph organization is in Sect Sentiment Analysis (SA) Sentiment analysis (SA) refers to the research area of analyzing sentiment in text. Opinion Mining (OM) has also been used as a synonym to sentiment analysis, in past literature Pang and Lee (2008). SA is the task of automatically predicting polarity in text. For example, the sentence The pizza is delicious should be labeled as positive, while the sentence The pizza tastes awful should be labeled as negative. The value of SA arises from the opportunity to understand preferences of individuals and communities, using user-generated content on the Web. To put computational sarcasm in perspective of sentiment analysis, we now highlight the challenges, research problems, and applications of SA Challenges Several challenges to SA are well-known (Pang and Lee 2008). The first challenge is negation. A negation marker can make sentiment prediction difficult. For example, the word not in the sentence I do not like this phone negates the sentiment of the verb like making the sentence negative. However, in the sentence I do not like this phone but it s still the best in its category, the negation word not negates only the portion before the word but. Scope of a negation marker has been studied in the past (Harabagiu et al. 2006). The second challenge to SA is domain dependence. Sentiment of words may differ depending on the domain. For example, the word unpredictable is positive for a movie review (for example, The plot of the movie is unpredictable ) but negative for an automobile review (e.g., The steering of a car is unpredictable ). Domain-specific sentiment is a long-studied subarea of SA (Fahrni and Klenner 2008). Similarly, polysemous words may carry different sentiment in different contexts. The word deadly may occur in the positive sentence Shane Warne is a deadly spinner and also in the negative sentence There are deadly snakes in the Amazon forest. Learning classifiers that incorporate polysemous nature of

15 1.1 Sentiment Analysis (SA) 3 words have also been reported in the past (Balamurali et al. 2011). Another challenge is thwarted expectations. An example of thwarting is: This city is polluted, has really bad traffic problems, and the weather sucks. However, I grew up here and I love the city. The second sentence reverses the sentiment expressed by the first, although, in terms of word count, negative words ( polluted, bad, and sucks ) outnumber the positive words ( love ). Thus, it can be seen that, in addition to approaches for sentiment analysis (in general), there have been explorations that focus on specific challenging aspects (in particular) such as polysemy, domain adaptation. Computational sarcasm is a similar endeavor that focuses on a specific challenge to sentiment analysis Research Problems Research in SA spans several decades now and has spawned multiple problems within the umbrella of SA. Each of these research problems has a large volume of work and several innovations. Some of these research problems are: 1. Sentiment detection on its own deals with prediction of positive or negative polarity. Both rule-based and statistical approaches have been reported for sentiment detection. Joshi et al. (2011), for example, is a rule-based sentiment detector that relies on a sentiment lexicon and a set of rules to account for words of a certain polarity and constructs such as negations and conjunctions. Similarly, many statistical sentiment detection approaches use features based on unigrams, bigrams, etc. 2. Subjectivity detection deals with prediction of a text as subjective or objective. In other words, subjectivity detection is concerned with distinguishing between text containing sentiment and the one not containing sentiment. This means that we wish to distinguish between fact and opinion in case of subjectivity detection. 3. Since a long document may contain some portions containing sentiment and some without, it is useful to separate the two. This resulted in subjectivity extraction as an area of research. Subjectivity extraction deals with identification of subset of sentences in a document that carry sentiment. Such sentences are referred to as subjective sentences, while the ones without sentiment are referred to as objective sentences. Pang and Lee (2004a) is a fundamental work in the area of subjectivity extraction where a minimum-cut algorithm is used to identify subjective extracts: subset of subjective sentences. For example, in case of the text The film is about a dog who befriends a boy. The story is very absurd but the actors do a great job. The lead is played by a new actor. He is a true find!, the goal of subjectivity extraction would be to extract the following sentences The story is very absurd but the actors do a great job. He is a true find! while discarding the other sentences. This may be useful because objective sentences do not contribute to the sentiment of the entity.