EMOMINER at SemEval-2019 Task 3: A Stacked BiLSTM Architecture for Contextual Emotion Detection in Text

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 205–209 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 205 EMOMINER at SemEval-2019 Task 3: A Stacked BiLSTM Architecture for Contextual Emotion Detection in Text Nikhil Chakravartula Teradata / Hyderabad nikhil.chakravartula@gmail.com Vijayasaradhi Indurthi Teradata / Hyderabad vijayasaradhi.indurthi@teradata.com Abstract This paper describes our participation in the SemEval 2019 Task 3 - Contextual Emotion Detection in Text. This task aims to identify emotions, viz. happiness, anger, sadness in the context of a text conversation. Our system is a stacked Bidirectional LSTM, equipped with attention on top of word embeddings pre- trained on a large collection of Twitter data. In this paper, apart from describing our official submission, we elucidate how different deep learning models respond to this task. 1 Introduction Sentiment analysis is an established field in NLP, but just identifying positive and negative senti- ments may not be enough. Applications require systems to go further beyond sentiment analysis and perform emotion analysis, which deals with identifying discrete emotions like anger, joy, sad- ness, etc. The task is challenging because the im- portance of context in emotion analysis cannot be overstated (Malik et al., 2017; Vanzo et al., 2014). Also, text in a conversation often contains lan- guage slangs, emoticons, emojis and other noisy data that make it difficult to identify the type of feeling expressed. Many approaches have been put forward to identify emotions in a text. Purver and Battersby (2012); Balabantaray et al. (2012) used SVM clas- sifier on twitter data to carry out emotion analysis. Potapova and Gordeev (2016) deployed a model based on Random Forests to identify aggression in texts. These approaches are often assisted by lexicons and require heavy feature engineering. In recent years, deep learning approaches have outperformed traditional algorithms in NLP tasks (Bahdanau et al., 2014). Felbo et al. (2017); Gupta et al. (2017) utilized LSTMs to achieve emotion identification in text. Abdul-Mageed and Ungar happy sad angry others 14.06% 18.11% 18.25% 49.56% Table 1: Train dataset composition. (2017) showed that GRNNs achieved a very good performance on 24 fine-grained types of emotions. Kratzwald et al. (2018) proposed sent2affect, a tai- lored form of transfer learning for affective com- puting, where the network is pre-trained for sen- timent analysis task, and subsequently the output layer is tuned to the task of emotion recognition. In this paper, we propose a stacked Bidirectional LSTM architecture to recognize emotions in text. The rest of the paper is organized as follows. In sec. 2 we give a brief explanation of the shared task. In sec. 3 we describe the process of fea- ture engineering from the text. In sec. 4 we dis- cuss system architecture. It is followed by sec. 5 which contains the various settings used in our ex- periments. In sec. 6, we analyse the results and conclude our paper in sec. 7 with future ideas and vision. 2 Shared Task Description The SemEval 2019 Task 3 is as follows: Given three turns …

Content truncated to 500 words