section{Introduction}% Head 1section{Basic Emotions}Before discussing the methods of emotion cause detection, it is important to understand the different basic sets of emotions used in the literature under review. James-Lange Theory of Emotions was one of the probably earliest study linking an individual’s congitive consideration of an event and the corresponding response to that event. Over the time the basic set of emotions has been redefined many times, some of them being Hamburg (1968), Hinde (1972), Plutchik (1980), Emde and Goensbauer (1981), Ekman (1982), Konner (1982), and Turner (1996, 2000). The study extit{Evolutionary Explanations of Emotions} by extit{Randolph M. Nesse} is extremely detailed and explores the reason why basic set of emotions have been redefined over time and why researchers have failed to come to a consensus. In most of the emotion classification tasks, researchers have proposed the use of Plutchik’s, Ekman’s or Turner’s set of emotions.
Analyzing these three sets we see that extit{fear}, extit{anger} and extit{sadness} are present in all emotion sets. Furthermore, extit{happiness} was found to be present in two out of three lists. The following set of emotions with the modifications were found to be used in the research covered in this survey:egin{itemize} item Ekman’s six emotions: Inspired by the work done by Charles Darwin and Margaret Mead that identifies emotions as evolved traits of humans, Paul Ekman identified a set of six universal emotions based on human facial expressions. The six emotions are extit{anger}, extit{disgust}, extit{fear}, extit{happiness}, extit{sadness} and extit{surprise}. In citet{ghazi2015detecting}, the authors use Ekman’s six emotions as the primary set of emotions.
Manual annotators classify the lexical units (word/meaning pairs) into one of the six emotions. However, if a lexical unit cannot be classified into any of the six emotions then an additional set of emotions extit{shame}, extit{guilt} and extit{hope} is used for classification. Post annotation, the final emotion set consisted of Ekman’s six emotions and extit{shame}.
During the whole annotation process, it was found that some of the additional emotions were found to fit one of the basic emotions like extit{fury} can be considered to belong to extit{anger} emotion. However, emotions like extit{affront} could not be easily classified into one of the six emotions. citet{gui2014emotion} use an emotion scheme which contains 6 of Ekman’s emotion and extit{like} as the additional emotion. Since, this study involved a dataset of micro-blogs, addition of the extit{like} emotion was done. citet{gui2017question}, citet{gui2016event} and citet{li2014text} use Ekman’s six emotions or “Big6” scheme in the W3C Emotion Markup Language as the emotion set for the emotion cause detection task. item Turner’s five emotions: citet{turner2000origins} identifies five primary emotions – extit{happiness}, extit{sadness}, extit{anger}, extit{fear} and extit{surprise}. citet{lee2013detecting} used Turner’s set of emotions are the basic emotion set for emotion cause detection.end{itemize}section{Datasets}egin{enumerate}item Micro-blogging datasets (Weibo): Weibo is a Chinese micro-blogging site just like Twitter but with some additional features.
Weibo has over 340 million monthly active users in China and given the high amount of content expressed using the platform, Weibo texts/messages prove to be good for emotion analysis tasks. egin{itemize} item citet{li2014text} Construct a Weibo dataset by first selecting random posts automatically, performing some basic necessary preprocessing steps and then ICTCLAS toolkit is used to parse, segment and then perform parts of speech tagging on it. In order to detect actual micro-blogging platform patterns, a development dataset containing 1000 random emotion entries were taken. To measure the effectiveness of identifying the occurrence of emotion cause events, each word was run through a linguistic cue word list cite{lee2010emotion}. After manual examination, several cue words were removed which were less highly collocated with cause events and some other new ones were added. end{itemize} item News: egin{itemize} item SINA city news: citet{gui2016event} took Chinese city news from NEWS SINA spanning a period of 3 years.
The raw dataset contained 20,000 articles and based on a list of 10,259 Chinese emotion keywords, extracted 15,687 instances by performing keyword matching. For each matched keyword, three preceding and three following clauses were extracted. If more than three clauses were present in the sentence with the emotion cause, the entire sentence was considered. Additionally, cross paragraph context was omitted. Since, presence of emotion keywords does not necessarily imply the presence of that emotion and for reasons like negative polarity and sense ambiguity, many of the irrelevant instances were removed. Finally, 2, 105 instances were left and using two manual annotators, the emotion categories and cause(s) of each instance was annotated in the W3C Emotion Markup Language format. A third annotator was used an arbitrator when the there was conflict in the opinions of the two main annotators.
Post annotation, XML tags were removed from the file to save space. The basic analysis unit in the dataset is a clause. The tags
The same dataset was also used citet{gui2017question} also used this dataset to perform emotion cause detection. end{itemize} item Lexical Units: A lexical unit is a word/meaning pair or most commonly known as a lexeme. citet{fillmore2003framenet} defined a set of 173 emotion-directed lexemes which correspond to different emotionsend{enumerate}section{Emotion Cause Detection Task Formulation}egin{itemize}item end{itemize}section{Baselines}% Bibliographyibliographystyle{ACM-Reference-Format}ibliography{sample-bibliography}
