Multi-label emotion classification of Tweets with transformer models

Analysis and classification of emotions expressed in social media content such as tweets have been useful for numerous commercial and social purposes for tasks like hate speech detection. Emotion classification of social media content has been performed using traditional techniques such as Recurrent Neural Networks (RNN) and Multivariate Long Short Term Memory (LSTM) in the past, which can be outperformed by the new transformer models. The ‘SemEval-2018 Task 1: Affect in Tweets’ (Mohommad et al. 2018) presents a challenge on multi-label classification of emotions expressed in tweets into 11 sentiment classes. The datasets given for this challenge are used in this work to explore the accuracy of the transformer models against other techniques used by the competitors of the particular challenge. Additionally the transformer models (BERT, RoBERTa and XLM RoBERTa) were compared with each other on their performance based on accuracy and speed. The best performing BERT-large model which is trained using bert-large-uncased tokenizer has shown a multi-label accuracy (Jaccard Index) which is higher than the sixth recorded score in the SemEval-2018 Task 1 competition, F1-micro which is higher than the best F1 score recorded.