Emotions During Covid-19: LSTM Models for Emotion Detection in Tweets

Deep learning techniques for emotion detection in micro-blogs are a relatively less explored area of research. This paper investigates the performance of Long Short-Term Memory (LSTM) networks in detecting emotions from English tweets that relate to the Covid-19 pandemic. The two proposed LSTM models viz. Simple LSTM and EmoLex Boost LSTM use a corpus of streaming tweets and train the networks to detect emotions in tweets. Simple LSTM architecture comprises two hidden layers and a fully connected layer with softmax activation. EmoLex Boost LSTM uses the NRC emotion lexicon to enhance the Simple LSTM architecture. Emotion classification experiments were conducted to test both LSTM models. While the Simple LSTM model shows an accuracy of 60.57% when trained for 30 epochs, the EmoLex Boost model shows an enhanced accuracy of 61.75% when trained for 30 epochs, and 63.09% when trained for 50 epochs. Both deep learning models identify emotions in tweets but do not compute their valence. Since a tweet can convey multiple emotions, the annotated emotion labels in the training set tend to be subjective or fuzzy. This adversely impacts the performance scores of models. The results of our experiments, however, are promising and motivate further research in deep learning models that compute the valence of emotion(s).