A Human-Psychology-Informed Multi-Task Learning Framework for Robust Opinion Mining

Opinion mining, a critical subfield of natural language processing, seeks to extract and analyze user-generated expressions. This paper proposes a novel Multi-Task Opinion Mining framework that integrates psychological insights with a BERT-based multi-task learning model to classify sentiment, emotion, sarcasm, and subjectivity. This study aims to develop a unified opinion Mining model that leverages psychological insights to capture nuanced human intent. The Multi-Task Opinion Mining model integrates four tasks (sentiment analysis, emotion recognition, sarcasm detection, and subjectivity detection) using a transformer-based architecture with BERT embeddings. It employs hard parameter sharing with task-specific layers to improve generalization. The framework is designed for real-world applications, such as customer feedback analysis and brand monitoring, and provides a robust tool for understanding human expressions in text. The model is tested on datasets, SST-2 (sentiment), GoEmotions (emotion), iSarcasm (sarcasm), and Cornell movie review (subjectivity), with performance metrics indicating improved accuracy through mutual learning. The multi-task opinion mining achieves the highest accuracy on GoEmotions (92%), surpassing SOTA (58%) and Individual BERT (87%), demonstrating the efficacy of multi-task learning in handling complex multi-class tasks. The integration of psychological insights with computational linguistics, which distinguishes between sentiment, emotion, sarcasm, and subjectivity, addresses the limitation of treating these tasks as interchangeable.