LT3 at SemEval-2018 Task 1: A classifier chain to detect emotions in tweets

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 123–127 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics LT3 at SemEval-2018 Task 1: A classifier chain to detect emotions in tweets Luna De Bruyne, Orph´ee De Clercq and V´eronique Hoste LT3, Language and Translation Technology Team, Ghent University Groot-Brittanni¨elaan 45, 9000 Ghent, Belgium {luna.debruyne, orphee.declercq, veronique.hoste}@ugent.be Abstract This paper presents an emotion classification system for English tweets, submitted for the SemEval shared task on Affect in Tweets, sub- task 5: Detecting Emotions. The system com- bines lexicon, n-gram, style, syntactic and se- mantic features. For this multi-class multi- label problem, we created a classifier chain. This is an ensemble of eleven binary classi- fiers, one for each possible emotion category, where each model gets the predictions of the preceding models as additional features. The predicted labels are combined to get a multi- label representation of the predictions. Our system was ranked eleventh among thirty five participating teams, with a Jaccard accuracy of 52.0% and macro- and micro-average F1- scores of 49.3% and 64.0%, respectively. 1 Introduction Most research in the domain of sentiment analy- sis focuses on the automatic prediction of polar- ity or valence in text, but also the detection of emotions has attracted growing interest in the last couple of years (Mohammad and Bravo-Marquez, 2017). Although emotion detection is a rather new research focus in NLP, the study of emotions has a long history in fields like psychology and neuro- imaging. Many different frameworks exist, but the specific emotion approach, in which emotions are classified as specific discrete categories, predomi- nates. In a lot of those approaches, some emotions are considered more basic than others, with Ek- man’s theory of six basic emotions (joy, sadness, anger, fear, disgust, and surprise) (Ekman, 1992) as the most well-known. Another popular theory is Plutchik’s wheel of emotions (Plutchik, 1980), in which joy, sadness, anger, fear, disgust, surprise, trust, and anticipation are considered most basic. Emotion analysis in NLP makes use of the frameworks developed by psychologists, mostly by employing categorical models of (basic) emo- tions. In traditional emotion classification tasks, a ‘document’ or sentence is classified under one or more emotion classes (or classified as neutral/no class when no emotions are present). Such emo- tion classification systems have been developed and tested on different kinds of data, including fairy tales (Alm et al., 2005), newspaper head- lines (Strapparava and Mihalcea, 2007), chat mes- sages (e.g. Holzman and Pottenger, 2003; Brooks et al., 2013), and tweets (e.g. Mohammad, 2012; Wang et al., 2012). The big advantage of using tweet datasets is the relative ease with which twit- ter data can be collected and the possibility of us- ing hashtags as emotion labels (distant supervision approach). For this paper, we used the data that was col- lected for the SemEval shared task on Affect in Tweets (Mohammad et al., 2018), a collection of tweets annotated for eleven emotions: anger, an- ticipation, disgust, fear, joy, love, optimism, pes- simism, …

Content truncated to 500 words