Sentiment to Signal: A Comparative Analysis of Neural Models for Emotion Recognition

Emotion recognition aims to provide emotional context to any external source. It has several applications, including analysis of audience reaction to news or customer reaction to products, mental healthcare, human-computer interaction (HCIs) and more, resulting into a highly discussed topic on blogs, forums, and other online platforms. This paper aims to analyze different models that can be used to obtain emotional context of sentences using neural networks, ranging from Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Bi-Directional Long Short-Term Memory Networks (Bi-LSTMs), and Bi-Directional Encoder Representations from Transformers (BERT). Sentiment analysis of given context provides a classification between positive and negative emotions. This paper further aims to perform multiclass classification by dividing pre-existing sentiment analysis dataset into six different emotions. Positive emotions are divided into three classes – joy, love, and surprise while negative emotions are being divided into three classes – anger, sadness, and fear. The respective dataset is hence sub-categorized into a total of six emotions. The paper attempts to understand the best fit model for the open source Emotion Dataset for Emotion Recognition Tasks from Kaggle, and visualize model statistics. Bidirectional LSTMs provide the best accuracy of 89.5% among tested models, followed by 84.5% of BERT which is a close second.