KSU at SemEval-2019 Task 3: Hybrid Features for Emotion Recognition in Textual Conversation

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 247–250 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 247 KSU at SemEval-2019 Task 3: Hybrid Features for Emotion Recognition in Textual Conversation Nourah Alswaidan Mohamed El Bachir Menai Department of Computer Science College of Computer and Information Sciences King Saud University Saudi Arabia nourah swaidan@yahoo.com, menai@ksu.edu.sa Abstract In this paper, we present the model submitted to the SemEval-2019 Task 3 competition: con- textual emotion detection in text “EmoCon- text”. We propose a model that hybridizes au- tomatically extracted features and human en- gineered features to capture the representa- tion of a textual conversation from different perspectives. The proposed model utilizes a fast gated-recurrent-unit backed by CuDNN (CuDNNGRU), and a convolutional neural network (CNN) to automatically extract fea- tures. The human engineered features take the term frequency-inverse document frequency (TF-IDF) of semantic meaning and mood tags extracted from SinticNet. For the classifica- tion, a dense neural network (DNN) is used with a sigmoid activation function. The model achieved a micro-F1 score of 0.6717 on the test dataset. 1 Introduction Emotion recognition in text refers to the task of automatically assigning an emotion to a text se- lected from a set of predefined emotion labels. The SemEval-2019 competition (Chatterjee et al., 2019b) provides a textual dialogue and asks to classify the emotion as one of the emotion labels: happy, sad, and angry or others. Previous research shows that emotion recogni- tion has been performed on different types of text, including fairy tales1 (Alm et al., 2005), news headlines2 (Strapparava and Mihalcea, 2007), blog posts3 (Aman and Szpakowicz, 2007), and tweets4 (Mohammad et al., 2018). Whether a text ex- presses a single emotion or multiple emotions, it is challenging to recognize implicit emotions, which 1http://people.rc.rit.edu/ coagla/affectdata/index.html 2http://web.eecs.umich.edu/ mihalcea/affectivetext 3http://saimacs.github.io 4https://competitions.codalab.org/competitions/17751 requires natural language understanding (NLU). Recognizing emotions in textual conversation in- creases difficulty by adding a dialogue format. Understanding emotions in textual conversation will further boost the research on NLU. In this paper, we present an emotion recognition model that hybridizes human engineered features and automatically extracted features. For the hu- man engineered features, we opted for calculating the term frequency-inverse document frequency (TF-IDF) of semantic meaning and mood tags re- trieved from SenticNet. For the automatically extracted features, we explored two deep neural networks, a fast gated-recurrent-unit backed by CuDNN (CuDNNGRU) and convolutional neural networks (CNN). The classification is performed by a dense neural network (DNN). The remainder of this paper is organized as fol- lows. Section 2 describes the task corpus. Sec- tion 3 presents the proposed emotion recognition model. Section 4 presents the experimental re- sults, and the main conclusions and future work are presented in Section 5. 2 Corpus The organizers of the competition split the cor- pus into three datasets: a training dataset with 30160 instances, a development dataset with 2755 instances, and a test dataset with 5509 instances. The corpus was in a (.txt) format and contained five columns. The first column …

Content truncated to 500 words