Multi-task learning transformers: comparative analysis for emotion classification and intensity prediction in social media

In the realm of sentiment analysis, transformer models have revolutionised sentiment classification tasks driven by their exponential growth in pre-training parameters, significantly enhancing performance. However, this study aims to push the boundaries further by introducing a novel framework for a multi-task learning transformer model1 with a dual learning objective: conducting emotion classification on tweets while concurrently detecting their intensity. Various pre-trained transformer models, including BERT, RoBERTa, and DistillBERT are employed to evaluate the effectiveness of this framework using the WASSA 2017 EmoInt dataset. Given its distinctive approach, this multi-task learning study identifies no direct competitors pursuing a similar methodology on the same dataset, prompting a comparative analysis among these transformer models for multi-task learning on the WASSA 2017 EmoInt dataset. The results showcase an impressive F1 score of 0.864 for emotion classification, while emotion intensity prediction achieves a Pearson correlation of 0.705 (R=0.705), closely aligning with the best results in the WASSA 2017 competition. Additionally, a supplementary investigation was conducted focusing solely on single-task learning for emotion classification and emotion intensity prediction, respectively. Comparing the performance of multi-task learning to single-task learning demonstrates the effectiveness and accuracy of the proposed multi-task learning framework. Furthermore, it yields a Pearson correlation of 0.779 (R=0.779) in emotion intensity prediction, surpassing the state-of-the-art baseline method for sentiment analysis.