Multi-Task Learning with Fusion: Framework for Handling Similar and Dissimilar Tasks

Multi-Task learning (MTL), which emerged as a powerful concept in the era of machine learning deep learning, employs a shared model trained to handle multiple tasks at simultaneously. Numerous advantages of this novel approach inspire us to instigate the insights of various tasks with similar (Identification of Sentiment, Sarcasm, Hate speech, Oensive language, etc.) and dissimilar (Identification of Sentiment, Claim, Language) genres. This paper proposes two Multi-Task Learning (MTL) framework schemes based on Bidirectional LSTM (BiLSTM) to handle both similar and dissimilar tasks. The performance of these frameworks is evaluated and compared against standalone classifiers, demonstrating their effectiveness in improving classification accuracy. In order to train our proposed MTL frameworks, different task-related publicly available datasets were collected, and each sentence was annotated with all task labels with the help of publicly available pre-trained models. Along with a simple MTL framework, this paper presents an MTL framework with a fusion technique (MTL fusion) that combines learning from task-specific layers to make predictions. Our proposed MTLfusion framework provides an F1 score of 0.76, 0.92, 0.809, 0.798, and 0.89 for sentiment, sarcasm, irony, hate speech, and offensive language classification tasks, respectively (similar tasks). It also provides an F1 score of 0.59, 0.586, and 0.707 for claim, sentiment, and language identification tasks, respectively. Our research also shows that MTL frameworks perform better than their corresponding standalone classifiers for similar tasks. On the other hand, for dissimilar tasks, the standalone classifiers perform better than MTL frameworks.