Detecting variation of emotions in online activities
Introduction
Web 2.0 technologies are increasingly dominating peoples’ everyday life, such that constant and evolving digital social interactions are produced dynamically. Their impact in society is evident by the exponential rates of users and their interactions carried out in popular “mega” social networks platforms, e.g., the daily active users of Facebook overcome 1 billion (Sept. 2016).1 Such intense and large scale online presence is characterized by many behavioral norms driven by people’s emotions and views. The power of emotion is evident from recent work which documents that contagious effects in online social networks (OSNs) are due to users emotional states which are often transmitted from real to online life (Coviello et al., 2014).
Until recently, emphasis has been placed on capturing human sentiments by detecting positive and negative opinions on various text sources. However, since human emotions are more variable, not necessarily restricted to a dual emotional standing, a more challenging endeavor is to track and reveal broader emotions such as anger, joy, etc., since they are powerful elicitors and indicators of human motivational and perceptual states (Roseman, Spindel, & Jose, 1990). Ongoing discussion in the psychological science classifies emotions in terms of at least two categories:
- •
the basic or primary, i.e., a fixed number of emotions as the ones we experience instantly as a response to a pleasant (or unpleasant) stimulus. A widely recognized approach of Ekman, Friesen, and Ellsworth (1982) identifies six primary emotions, i.e., ‘anger, disgust, fear, joy, sadness, surprise’. Main characteristics of primary emotions are their automatic onset and pervasive impact on individuals’ cognitive and behavioral outcomes.
- •
the social emotions, where a person’s emotions are influenced by her fellows emotions and impact their emotions too (Parkinson, Fischer, & Manstead, 2005). Indicative social emotions are the ‘rejection’, and ‘shame’ which have been identified as quite important in social interactions (Kafetsios & Nezlek, 2012).
Up to now, efforts in emotion analysis on text sources have mostly focused on detecting emotions of individuals ignoring the social context’s influence and impact. This work is motivated by the need to deepen emotional detection by exploiting not only primary emotions, but also other facets of emotions such as social emotions. A full list of emotions from a psychological perspective is not complete without reference to emotions that are neither basic or social but that characterize general individual affective states, such as feeling anxious, calm, and interest, i.e., emotional states that demonstrate a longer duration and cause less intensive experiences (Ekman, 1992). Existing work mainly targets in detecting primary emotions without considering the socials ones, while only limited work has targeted in detecting those that characterize general affective states.
The present work proposes an extended emotions analysis approach which incorporates Ekman’s primary emotions (enabling comparisons with existing work) together with a wider spectrum of emotions at which social and more general affective states are also considered. We leverage both machine learning and lexicon-based approaches which have dominated the literature on this area so far, under a separate or hybrid scheme. Motivation for building on a hybrid scheme originates from the fact that the existing lexicon-based approaches tend to achieve high precision and low recall (Nie, Wang, He, & Sato, 2015), while machine learning approaches suffer in integrating syntactic with semantic information (El-Alfy, Thampi, Takagi, Piramuthu, & Hannen, 2015). In summary, the main contributions of this work are as follows.
- 1.
We proceed with a hybrid approach which builds upon machine learning and lexicon-based approaches, to detect Ekman’s primary emotions. To validate such approach and to compare it to existing work we utilize the SemEval-2007 Affective text competition dataset.2
- 2.
We examine distinct approaches to detect social emotions and those that characterize general affective states in addition to the primary ones. We experiment with a Twitter dataset annotated by a crowdsourcing process.
- 3.
We implement a case study in which explicit (i.e., human reports) versus implicit (i.e., automatic detection of emotions) emotional experiences are monitored to cross validate results. The participants’ native language is Greek, and so issues related to the Greek texting habits and the detection of emotions in non-English texts in general are also considered.
- 4.
We share the annotated Twitter dataset at: http://bit.ly/2bLgVUP.
The remainder of the paper is organized as follows. Section 2 reviews literature on emotional analysis. Section 3 proceeds with the data preparation for analysis, while Section 4 presents the used datasets. Section 5 overviews the used methodologies. Sections 6 and 7 proceed with the emotions detection, i.e., the primary ones or the wider spectrum, accordingly. Section 8 presents the case study, while Section 9 concludes the paper.
Section snippets
Previous work
Existing research on emotion detection out of English text sources has utilized various data sources which are summarized in Section 2.1 and heavily depended on machine learning and lexicon-based methodologies which are highlighted in Section 2.2. Also, sentiment detection out of non-English text sources is briefly outlined in Section 2.3. Table 1 shows a comparison of our work to others that are most relevant to our problem setting.
Background and fundamentals
This section summarizes the fundamental concepts and processes required for some or all of the emotion analytics of this work, with an emphasis on the data preparation, the features modeling for the machine learning and hybrid processes and the emotion words specification, to then predict the texts’ underlying emotions. Fig. 1 overviews the features’ modeling process followed by both machine learning and hybrid approaches. After preprocessing the available text sources, for the machine learning
Dataset and study setups
This work exploits various datasets which are used at its different experimentation phases and setups. Next, we summarize these datasets and study setups to increase comprehension of the next section’s methodologies.
Emotion detection methodologies
As indicated in the introduction, lexicon-based approaches tend to result in high precision and low recall, while machine learning approaches do not consider the syntactic and semantic attributes, so both approaches embed emotions misinterpretation risks. Thus, a hybrid process which builds on the advantages of both approaches is proposed. Next, we outline both the machine learning and the lexicon-based approaches, concluded in the end to the hybrid one.
Study I: primary emotions detection
The first study setup serves as a baseline as we compare our methodologies with the already existing ones, by using Ekman’s primary emotions, i.e., anger, disgust, fear, joy, sadness, and surprise. Initially, we experiment with the lexicon-based approach, followed by the machine learning process, to finally conclude to the hybrid one. Similar to the SemEval-2007 Affective text task, our objective is to classify news headlines extracted from news websites.
Study II: capturing a wider emotions spectrum
In the second study setup we proceed with the detection of a wider spectrum of emotions by considering social emotions and those that characterize general affective states in addition to the primary ones. For the experiments reported in this section we use the dataset collected from Twitter.
Due to the confirmation of the hypothesis testing (see Section 6.3), here we proceed with a hybrid approach for the emotions detection process. Initially, we present the results obtained with the
Case study: correlations between implicit and explicit emotional states
Previous study setups build upon datasets where independent annotators identified the underlying emotions. Since humans perceive emotions in a subjective manner mis-annotations are highly possible, so, here, we proceed with a case study in which we monitor participants’ emotions both implicitly and explicitly. Our objective is to examine the existence of possible similarities or differences among participants explicit emotion declaration and the emotions as these are identified with an
Conclusions
This work addresses the problem of detecting a wide spectrum of emotions from online text sources. Current research efforts mainly focus on detecting specific primary emotions without considering the social ones or those that characterize general affective states. Motivated by such lack of a wider emotional spectrum analysis, we proceeded with an approach which permits the detection of 12 emotions, i.e., Ekman’s six primary emotions, three social ones, and three emotions that characterize
Acknowledgments
Part of this work (especially work in Section 8) has been funded by the Network of Excellence in Internet Science (7th EU Framework Programme, under GA number 288021).
References (63)
- et al.
Comparative experiments using supervised learning and machine translation for multilingual sentiment analysis
Computer Speech & Language
(2014) - et al.
Tom: Twitter opinion mining framework using hybrid classification scheme
Decision Support Systems
(2014) - et al.
Sentiment polarity detection in spanish reviews combining supervised and unsupervised approaches
Expert Systems with Applications
(2013) - et al.
Sentiment, emotion, purpose, and style in electoral tweets
Information Processing & Management
(2015) More than words: Social networks’ text mining for consumer brand sentiments
Expert Systems with Applications
(2013)- et al.
Sentiment analysis: A combined approach
Journal of Informetrics
(2009) - et al.
Lexicon-enhanced sentiment analysis framework using rule-based classification scheme
PloS One
(2017) - et al.
Simpler is better? Lexicon-based ensemble sentiment classification beats supervised methods
Proceedings of the 2014 international conference on advances in social networks analysis and mining
(2014) - et al.
Sentiwordnet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining
Proceedings of the seventh international conference on language resources and evaluation
(2010) - et al.
Affective computing with primary and secondary emotions in a virtual human
Autonomous Agents and Multi-Agent Systems
(2009)
Rt news: An analysis of news agency ego networks in a microblogging environment
ACM Transactions on Management Information Systems
A tutorial on support vector machines for pattern recognition
Data Mining and Knowledge Discovery
Micro-blogging content analysis via emotionally-driven clustering
Proceedings of the humaine association conference on affective computing and intelligent interaction
Multispot: Spotting sentiments with semantic aware multilevel cascaded analysis.
Proceedings of the big data analytics and knowledge discovery
Harvesting opinions and emotions from social media textual resources
IEEE Internet Computing
“They are happier and having better lives than I am”: The impact of using facebook on perceptions of others’ lives
Cyberpsychology, Behavior, and Social Networking
Detecting emotional contagion in massive social networks
PLoS ONE
Emoticons in computer-mediated communication: Social motives and social context
Cyberpsychol and Behavior
Ensemble Methods in Machine Learning
Proceedings of the first International Workshop on Multiple Classifier Systems
An argument for basic emotions
Cognition and Emotion
What emotion categories or dimensions can observers judge from facial behavior?
Advances in intelligent informatics
Emoto: Emotionally engaging interaction
Personal and Ubiquitous Computing
How are you doing? Emotions and personality in facebook
Proceedings of the empire workshop of the twenty-second international conference on user modeling, adaptation and personalization
Measuring nominal scale agreement among many raters
Psychological Bulletin
Bayesian network classifiers
Machine Learning
Sentiment analysis leveraging emotions and word embeddings
Expert Systems with Applications
What’s social about social emotions?
Journal for the Theory of Social Behaviour
Mining and summarizing customer reviews
Proceedings of the tenth ACM SIGKDD international conference on knowledge discovery and data mining
Analysis and generation of emotion in texts
Proceedings of the international conference on knowledge engineering principles and techniques
Emotion and support perceptions in everyday social interaction: Testing the “less is more hypothesis” in two different culture
Journal of Social and Personal Relationships
Cited by (29)
Internet-based identification of anxiety in university students using text and facial emotion analysis
2023, Internet InterventionsPredicting tweet impact using a novel evidential reasoning prediction method
2021, Expert Systems with ApplicationsCitation Excerpt :Emotion analysis aims to detect moods or traits based on a specific text, such as trust, sadness, fear, anger, disgust, and disgust (Li, Wu, Zhu, & Xu , 2018). It provides decision makers with a deeper analysis of what emotion could have more impact on their target audience on social media (Chatzakou, Vakali, & Kafetsios, 2017). Emotion differs from sentiment analysis in that emotion relates to people’s mood and is determined by a multi-class classifier that includes, for instance, anger, fear, and surprise.
Borrow from rich cousin: transfer learning for emotion detection using cross lingual embedding
2020, Expert Systems with ApplicationsCitation Excerpt :They used soft classification techniques (Naïve Bayes classifier) to produce the class membership of probabilities of each tweet. Chatzakou, Vakali, and Kafetsios [6] used handcrafted features with different statistical machine learning methods to classify news and social media data into different emotion classes. They performed their experiments on English and Greek languages.
An analytical system for user emotion extraction, mental state modeling, and rating
2019, Expert Systems with ApplicationsCitation Excerpt :An examination of emotions addresses emotional behaviors by the four-dimensional model or its subsets. The presence of emotions has been studied using facial expressions and gestures (Balomenos et al., 2004; Castellano, Kessous, & Caridakis, 2008), through phonetic analysis of speech (Tawari & Trivedi, 2010), and from linguistic analysis with social media such as Twitter (Chatzakou, Vakali, & Kafetsios, 2017; Mohammad, Zhu, Kiritchenko, & Martin, 2015). These studies have different applications.
Deep learning for affective computing: Text-based emotion recognition in decision support
2018, Decision Support SystemsCitation Excerpt :Previous research has experimented with different models for inferring affect from narrative materials. Examples include methods that explicitly exploit the flexibility of machine learning, such as random forests (e.g., [66]) and support vector machines (e.g., [14,19]), both of which have commonly been deployed in literature. Studies have shown that random forests tend to compute faster, while support vector machines yield superior performance [14].
Measuring the emotional state among interacting agents: A game theory approach using reinforcement learning
2018, Expert Systems with ApplicationsCitation Excerpt :This result in a model of an adaptive emotional framework that takes into account the interaction between agents and a learning process that can simulate human attributes. Several computational models of emotions have been presented in the literature considering affective processing (Becker-Asano & Wachsmuth, 2010; Champandard, 2003; Chatzakou et al., 2017; Deng et al., 2015; Dias, Mascarenhas, & Paiva, 2014; El-Nasr, Yen, & Ioerger, 2000; Funge, 1999; Lim & Kim, 2017; Lim et al., 2012; Luo et al., 2016; Marsella & Gratch, 2009). For instance, Soar (Laird, Newell, & Rosenbloom, 1987; Newell, 1990) is a cognitive architecture for human cognition realized as a production system which consists of matching the contents of working memory to the precondition of rules.