Detecting implicit expressions of emotion in text: A comparative analysis

Abstract

Sentiment analysis is one of the recent, highly dynamic fields in Natural Language Processing. Most existing approaches are based on word-level analysis of texts and are mostly able to detect only explicit expressions of sentiment. However, in many cases, emotions are not expressed by using words with an affective meaning (e.g. happy), but by describing real-life situations, which readers (based on their commonsense knowledge) detect as being related to a specific emotion. Given the challenges of detecting emotions from contexts in which no lexical clue is present, in this article we present a comparative analysis between the performance of well-established methods for emotion detection (supervised and lexical knowledge-based) and a method we propose and extend, which is based on commonsense knowledge stored in the EmotiNet knowledge base. Our extensive evaluations show that, in the context of this task, the approach based on EmotiNet is the most appropriate.

Introduction

Research in affect has a long established tradition in many sciences—Linguistics, Psychology, Socio-Psychology, Cognitive Science, Pragmatics, Marketing or Communication Science. Recently, many closely related subtasks were developed also in the field of Natural Language Processing (NLP), such as emotion detection, subjectivity analysis and opinion mining (also known as sentiment analysis, attitude and appraisal analysis or review mining) [30]. All these research fields can be considered as part of the wider area of research in artificial intelligence (AI) of Affective Computing [37]. Among these tasks, sentiment analysis aims at detecting the expressions of sentiment in text and subsequently classifying them, according to their polarity (semantic orientation) into different categories (usually, positive and negative). The problem is defined by Pang and Lee (2008) as “the binary classification task of labeling an opinionated document as expressing either an overall positive or an overall negative sentiment.” According to the Webster dictionary,¹ “sentiment suggests a settled opinion reflective of one's feelings”, where the term feeling is defined as “the conscious subjective experience of emotion” [52], “a single component of emotion, denoting the subjective experience process” [45]. Expressions of sentiment are thus directly related to expressions of emotions in text. As such, as in the case of the latter, sentiments can be expressed directly (e.g. “I like Nokia phones.”), indirectly (e.g. “This phone is light as a feather.”) or implicitly, by describing a situation which points the reader towards a specific sentiment (e.g. “I paid 200€ for this phone and it broke in two days.”). Most of the research performed in the field of sentiment analysis and the related task of emotion detection has aimed at detecting explicit expressions of sentiment (i.e. situations where specific words or word combinations are found in texts). Nevertheless, the expression of emotion is most of the times not achieved through the use of emotion-bearing words [34], but indirectly, by presenting situations that based on commonsense knowledge can be interpreted in an affective manner [4], [5]. A clear example of such cases can be seen in self-reported affect (textual descriptions of affect-eliciting situations), a technique that has been widely used in psychological experiments [46]. In these cases, the subjects are asked to describe events that made them experience different emotions, without necessarily mentioning the emotion itself. A corpus with such examples is ISEAR [International Survey of Emotional Antecedents and Reactions 46].² In a first effort to overcome the issue of emotion detection from texts in which little or no lexical clues exist to mark the presence of a specific emotion (i.e. presence of words such as “joy”, “happy”, “angry”, etc.), Balahur et al. [6] proposed a method to build a commonsense knowledge base (EmotiNet) storing situations that trigger emotions, based on the principles of the Appraisal Theories [42]. The main idea behind our approach, inspired by the Appraisal Theories, is that situations trigger emotions based on the result of the individual evaluation of their components, in accordance to “appraisal criteria”. In order to detect the values of such criteria, each such situation was represented in EmotiNet as a chain of actions, with their corresponding actors, objects, their properties and the associated emotion. We subsequently demonstrated that by using this resource, we are able to detect emotion from examples in ISEAR describing family-related situations in which little or no explicit mention of affect is present. In the present article, we analyze the peculiarities of the ISEAR data employed in our previous evaluation [6] and comparatively evaluate the performance of our system using established supervised and lexical knowledge-based methods for emotion detection versus the use of EmotiNet as emotion detection resource. Subsequently, we extend the knowledge contained in EmotiNet and perform additional evaluations to demonstrate the appropriateness of the created knowledge base for the automatic detection of implicit expressions of emotion.