A Comparison of Word-Embeddings in Emotion Detection from Text using BiLSTM, CNN and Self-Attention

Abstract

User profiling is becoming increasingly holistic by including aspects of the user that until a few years ago seemed irrelevant. The content that users produce on the Internet and social networks is an essential source of information about their habits, preferences, and behaviors in many situations. One factor that has proved to be very important for obtaining a complete user profile that includes her psychological traits are the emotions experienced. Therefore, it is of great interest to the research community to develop approaches for identifying emotions from the text that are accurate and robust in situations of everyday writing. In this work, we propose a classification approach based on deep neural networks, Bi-LSTM, CNN, and self-attention demonstrating its effectiveness on different datasets. Moreover, we compare three pre-trained word-embeddings for words encoding. The encouraging results obtained on state-of-the-art datasets allow us to confirm the validity of the model and to discuss what are the best word embeddings to adopt for the task of emotion detection. As a consequence of the great importance of deep learning in the research community, we promote our model as a starting point for further investigations in the domain.