Multi-Emotion Classification Evaluation via Twitter

Abstract

Recently, twitter has become an indispensable social communication platform. It contains many contrasting views and cultures on miscellaneous topics. This gigantic information bulk has endeared the attention of researchers for its interpretation and serve it in various life applications (e.g. product customer feedback, tourism, voting, product branding, etc.). However, natural languages ambiguity is one of the researchers' limitations, where implicit and diverse emotions are implied in the same context. Emotion analysis is a recent research field that digs to predict the implied emotions in different media types especially written text. Traditional approaches focused on detecting single attitude from social media texts, which isn't considered accurate. This research proposes two Multi-Emotion classification (MEC) approaches, that mine users' attitudes in tweets. These approaches have diverse classifiers' architectures, representative features, and a number of emotions sets. These diversities contribute in each classifier's performance in emotion classification. Eight experiments are applied using two feature representation vectors and two supervised machine learning algorithms on two emotion sets. The proposed systems outperform the contemporary traditional approaches. The first Binary relevance approach achieves hamming score ranging from 0.36 to 0.53, and the second Convolutional neural network approach achieves hamming score ranging from 0.39 to 0.54.