NTUA-ISLab at SemEval-2019 Task 3: Determining emotions in contextual conversations with deep learning

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 277–281 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 277 NTUA-ISLab at SemEval-2019 Task 3: Determining emotions in contextual conversations with deep learning Rolandos Alexandros Potamias, Georgios Siolas Intelligent Systems Laboratory, National Technical University of Athens, Zografou, Athens , Greece rolpotamias@gmail.com , gsiolas@islab.ntua.gr Abstract Sentiment analysis (SA) in texts is a well- studied Natural Language Processing task, which in nowadays gains popularity due to the explosion of social media, and the subse- quent accumulation of huge amounts of related data. However, capturing emotional states and the sentiment polarity of written excerpts requires knowledge on the events triggering them. Towards this goal, we present a compu- tational end-to-end context-aware SA method- ology, which was competed in the context of the SemEval-2019 / EmoContext task (Task 3). The proposed system is founded on the com- bination of two neural architectures, a deep recurrent neural network, structured by an at- tentive Bidirectional LSTM, and a deep dense network (DNN). The system achieved 0.745 micro f1-score, and ranked 26/165 (top 20%) teams among the official task submissions. 1 Introduction and Related Work One of the most challenging fields of Natural Lan- guage Processing (NLP) and Computational Lin- guistics is Sentiment Analysis (SA) that aims to- wards the automated extraction of a writer’s sen- timent or emotion as conveyed in text excerpts (Liu, 2015). Relevant efforts focus on tracking the sentiment polarity of single utterances, which in most of the cases is loaded with a lot of subjec- tivity and a degree of vagueness (Thelwall et al., 2010). Contemporary research in the field uti- lizes data from social media resources (e.g., Face- book, Twitter) as well as other short text refer- ences. However, users of social media tend to vi- olate common grammar and vocabulary rules and even use various figurative language forms to com- municate their message. In such situations, the sentiment polarity underlying the literal content of the conveyed concept may significantly differ from its figurative context, making SA tasks even more puzzling (Patra et al., 2016). Evidently, sin- gle turn text lacks in detecting sentiment polar- ity on sarcastic and ironic expressions (Potamias et al., 2019), as already indicatd in the relevant SemEval-2014 Task-9 Sentiment Analysis in Twit- ter (Sara et al., 2014). As sentiment reflects the emotion behind customer engagement, SA finds its realization in automated customer aware ser- vices (Kurniawati et al., 2013). A lot of research has already been devoted on capturing the senti- ment polarity of contextual conversations of vari- ous utterances. Most of the relevant studies utilize single turn texts from topic related sources (e.g., Twitter). Hand-crafted and sentiment-oriented features, indicative of emotion polarity, are uti- lized to represent respective excerpt cases. The formed data were used as input to various tra- ditional machine learning classifiers (e.g. SVM, Random Forests etc.) or deep learning architec- tures (e.g. recurrent neural networks, CNNs) in order to induce analytical models that are able to capture the …

Content truncated to 500 words