Automatic Detection of Irony

Opinion Mining in Microblogs and Social Media

Jihen Karoui, Farah Benamara, Veronique Moriceau

• Publisher: Wiley-ISTE
• Year: 2019
• Print ISBN: 9781786303998
• eBook ISBN: 9781119671220
• Edition: 1
• Topic: Technology & Engineering
• Subtopic: Electrical Engineering & Telecommunications
• Index: Technology & Engineering

In recent years, there has been a proliferation of opinion-heavy texts on the Web: opinions of Internet users, comments on social networks, etc. Automating the synthesis of opinions has become crucial to gaining an overview on a given topic. Current automatic systems perform well on classifying the subjective or objective character of a document. However, classifications obtained from polarity analysis remain inconclusive, due to the algorithms' inability to understand the subtleties of human language. Automatic Detection of Irony presents, in three stages, a supervised learning approach to predicting whether a tweet is ironic or not. The book begins by analyzing some everyday examples of irony and presenting a reference corpus. It then develops an automatic irony detection model for French tweets that exploits semantic traits and extralinguistic context. Finally, it presents a study of portability in a multilingual framework (Italian, English, Arabic).

1
From Opinion Analysis to Figurative Language Treatment

1.1. Introduction

The first work on automatic opinion extraction (or opinion mining) dates back to the late 1990s, notably to (Hatzivassiloglou and McKeown 1997) seminal work on determining adjectival polarity in documents, i.e. identifying the positive or negative character of opinions expressed by these adjectives, and to (Pang et al. 2002) and (Littman and Turney 2002) work on classifying documents according to polarity.

Work on this subject has been in progress since the 2000s, and opinion extraction is one of the most active areas in both NLP and data mining, with over 26,000 publications identified by Google Scholar. Notable examples include (Wiebe et al. 2005) work on annotating the multi-perspective question answering (MPQA) opinion corpus, (Taboada et al. 2011) work on the effects of opinion operators, such as intensifiers, modalities and negations, and (Asher et al. 2009) and (Chardon et al. 2013) work on the use of the discursive structure in calculating the overall opinion expressed in a document. Finally, we note the emergence of a number of evaluation campaigns, such as the Text Retrieval Conference (TREC) (Ounis et al. 2008), the DEFT (Défi fouille de textes, data mining challenge) in French run for the first time in 2005 (Azé and Roche 2005), and the SemEval (Semantic Evaluation) campaign, started in 1998¹.

It is important to note that opinion analysis was already a subject of study in other domains, such as linguistics (Hunston and Thompson 2000), psychology (Davidson et al. 2009), sociology (Voas 2014) and economics (Rick and Loewenstein 2008) before it attracted the attention of computer scientists. Opinion analysis is a multidisciplinary domain that draws on a wide range of tools and techniques, as we shall see throughout this chapter.

The development of opinion analysis systems is no simple matter, and there are several different challenges that must be met: identifying portions of text that provide the opinions a user is looking for; evaluating the quality of opinions obtained in this way – positive, negative, etc.; presenting results to users in a relevant way; etc.

Most existing approaches are based on word-level lexical analysis, sometimes combined with phrase-level syntactic analysis to identify operators and calculate their effects on opinion words (Liu 2012). Evidently, this type of analysis is far from sufficient to take account of the full linguistic complexity of opinion expressions. Fine, or pragmatic, semantic analysis is therefore crucial, particularly when treating complex phenomena such as figurative language, the focus of our study.

The aim of this chapter is to provide a brief introduction to the field of opinion analysis and to establish key definitions relating to the notion of figurative language. Our overview makes no claim to be exhaustive, given the extent of the field of research in question. Readers interested in going further may wish to consult the excellent summaries found in (Liu 2015) and (Benamara et al. 2017).

This chapter begins with a presentation of the notion of opinion and of the main approaches used in the literature (section 1.2). In section 1.3, we present the main limitations of existing systems, focusing on the use of figurative language. Section 1.4 deals with this type of language, looking at four figurative phenomena: irony, sarcasm, satire and humor. Finally, we shall discuss the main challenges encountered in NLP in terms of automatic detection of figurative language.

1.2. Defining the notion of opinion

1.2.1. The many faces of opinion

In NLP, the word opinion is used as a generic term to denote a range of subjective expressions such as sentiments, attitudes, points of view, judgments and desires. The most widely used definition is as follows (Benamara 2017):

“An opinion is a subjective expression of language which an emitter (a person, institution, etc.) uses to judge or evaluate a subject (an object, person, action, event, etc.), positioning it on a polarized scale in relation to a social norm (such as an aesthetic judgment) or a moral norm (such as the distinction between good and bad)”.

Phrase (1.1) is a good illustration of this definition. The author expresses a positive opinion of the dishes served in the restaurant using a positive-polarity verb (to love). In opinion analysis, the ability to distinguish between subjective or objective expressions is key. Phrase (1.2) does not express an opinion, but a purely factual event.

(1.1)	I loved the dishes served in this restaurant.
(1.2)	The Prime Minister opened the new hospital.

The most important element of this definition is the notion of a polarized scale (positive vs. negative, good vs. bad, desirable vs. undesirable, agreement vs. disagreement, etc.). Thus, the sentiment of jealousy in phrase (1.3) expresses an emotion and may appear in isolation from an evaluative opinion of an entity. Similarly, certain predictive expressions, which relay opinions in everyday language, do not constitute evaluations. In the second section of phrase (1.4), for example, the author expresses a hypothesis regarding that evening’s weather, but this does not constitute an evaluation of the weather in question.

(1.3)	I’m jealous of my brother.
(1.4)	I won’t be able to go this evening, I think it’s going to rain.

In what follows, we shall focus exclusively on the automatic detection of opinions expressed on a polarized scale or evaluative opinions.

1.2.2. Opinion as a structured model

Within the context of automatic extraction, (Liu 2012) proposed a structured model Ω made up of five elements:

• – s is the subject of the opinion;
• – a is an aspect of s;
• – e is the emitter;
• – senti is the sentiment expressed by e toward s (and potentially a). senti is generally represented by a triplet (type, p, v) such that:
  • - type is the semantic ty...

Table of contents

• Cover
• Table of Contents
• Preface
• Introduction
• 1 From Opinion Analysis to Figurative Language Treatment
• 2 Toward Automatic Detection of Figurative Language
• 3 A Multilevel Scheme for Irony Annotation in Social Network Content
• 4 Three Models for Automatic Irony Detection
• 5 Towards a Multilingual System for Automatic Irony Detection
• Conclusion
• Appendix: Categories of Irony Studied in Linguistic Literature
• References
• Index
• End User License Agreement