Tw-StAR at SemEval-2018 Task 1: Preprocessing Impact on Multi-label Emotion Classification

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 167–171 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics Tw-StAR at SemEval-2018 Task 1: Preprocessing Impact on Multi-label Emotion Classification Hala Mulki1, Chedi Bechikh Ali2, Hatem Haddad3 and Ismail Babao˘glu1 1 Department of Computer Engineering, Selcuk University, Turkey 2 LISI laboratory, INSAT, Carthage University, Tunisia 3 Department of Computer and Decision Engineering, Universit´e Libre de Bruxelles, Belgium halamulki@selcuk.edu.tr,chedi.bechikh@gmail.com Hatem.Haddad@ulb.ac.be,ibabaoglu@selcuk.edu.tr Abstract In this paper, we describe our contribution in SemEval-2018 contest. We tackled task 1 “Af- fect in Tweets”, subtask E-c “Detecting Emo- tions (multi-label classification)”. A multi- label classification system Tw-StAR was de- veloped to recognize the emotions embedded in Arabic, English and Spanish tweets. To handle the multi-label classification problem via traditional classifiers, we employed the bi- nary relevance transformation strategy while a TF-IDF scheme was used to generate the tweets’ features. We investigated using sin- gle and combinations of several preprocess- ing tasks to further improve the performance. The results showed that specific combinations of preprocessing tasks could significantly im- prove the evaluation measures. This has been later emphasized by the official results as our system ranked 3rd for both Arabic and Spanish datasets and 14th for the English dataset. 1 Introduction Social media platforms and micro-blogging sys- tems such as Twitter have recently witnessed a high rate of accessibility (Duggan et al., 2015). Tweets usually combine multiple emotions ex- pressed by the appraisal or criticism of a spe- cific issue. Sentiment analysis represents a coarse- grained opinion classification as it detects either the subjectivity (objective/subjective) or the po- larity orientation (positive, negative or neutral) (Piryani et al., 2017). For opinionated texts which are usually rich of several emotions, a fine-grained analysis is needed. Through such analysis, specific emotions can be recognized within a tweet which is crucial for many applications. For instance, recognizing anger emotions in the tweets representing the cus- tomers’ opinions about a specific service in a hotel would definitely help to take the proper response to keep the customers satisfied (Li et al., 2016). Existing MLC systems are conducted either by problem transformation approaches or algorithm adaptation ones. Each of which combines several methods and has different merits. While problem transformation methods are simpler and easier to implement, algorithm adaptation methods have a more accurate performance but with a high com- putational cost (Zhang and Zhou, 2014). There- fore, to develop a multi-label classifier that com- bines the simplicity of the problem transformation methods along with accurate performance remains an interesting issue to investigate. Since preprocessing tasks have been found of positive impact on sentiment analysis of differ- ent languages (Haddi et al., 2013; Yıldırım et al., 2015; El-Beltagy et al., 2017), we hypothesize that the application of single or combinations of various preprocessing techniques on tweets before feeding them to the multi-label emotion classifier, can improve the classification performance with- out the need to complex methods that consider the dependencies between labels. Here, we describe the participation of our …

Content truncated to 500 words