Neural Feature Extraction for Contextual Emotion Detection

Proceedings of Recent Advances in Natural Language Processing, pages 785–794, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_091 785 Neural Feature Extraction for Contextual Emotion Detection Elham Mohammadi, Hessam Amini and Leila Kosseim Computational Linguistics at Concordia (CLaC) Lab Department of Computer Science and Software Engineering Concordia University, Montr´eal, Qu´ebec, Canada first.last@concordia.ca Abstract This paper describes a new approach for the task of contextual emotion detection. The approach is based on a neural feature extractor, composed of a recurrent neural network with an attention mechanism, fol- lowed by a classifier, that can be neural or SVM-based. We evaluated the model with the dataset of the task 3 of SemEval 2019 (EmoContext), which includes short 3-turn conversations, tagged with 4 emo- tion classes. The best performing setup was achieved using ELMo word embed- dings and POS tags as input, bidirec- tional GRU as hidden units, and an SVM as the final classifier. This configuration reached 69.93% in terms of micro-average F1 score on the main 3 emotion classes, a score that outperformed the baseline sys- tem by 11.25%. 1 Introduction Emotions are an intricate part of human commu- nication. Being able to interpret and react to the emotions of others allows one to better communi- cate. Emotion information can be extracted from a variety of physiological sources, such as elec- troencephalography (EEG) signals (Zhang et al., 2016), skin temperature (Li and Chen, 2006), speech signals (Trigeorgis et al., 2016) and fa- cial expressions (Mao et al., 2015), as well as text (Yassine and Hajj, 2010). The rise of social media and the availability of human-written online diaries, blog posts, and comments has lead to an increase in research on the automatic detection of sentiment and emotion from textual data. Sentiment analysis and opinion mining can range from coarse-grained binary classification of texts into positive or negative classes to finer- grained classification into a variety of emotion cat- egories, such as happy, sad, angry, and scared. Such a classification is useful in business and marketing (Medhat et al., 2014), and a variety of downstream NLP applications, such as text-to- speech, to maintain the emotion present in text, and human-computer interaction in order to take into account the emotional state of users and make responses more human-like (Hirat and Mit- tal, 2015). Fine-grained emotion detection based solely on text is a challenging task. As only linguistic cues are available, facial expressions and voice fea- tures, which are known to be discriminating (Poria et al., 2016), cannot be used. In addition, several emotions can be expressed textually by the same linguistic cues, such as emotion keywords, inter- jections, and emojis (Liew and Turtle, 2016). Fi- nally, many pieces of text, especially online com- ments, posts, and tweets are too short to allow for correct classification. These challenges highlight the importance of using contextual information in order to detect the emotion conveyed in a piece of text. The goal of this work is to investigate the effec- tiveness of different models using neural networks and Support Vector …

Content truncated to 500 words