MTL-BERT: A Multi-Task Learning Framework for Efficient Sentiment and Topic Analysis in Engineering Applications

This study proposes MTL-BERT, a multi-task learning framework integrating BERT with hierarchical attention and dynamic weight balancing for efficient sentiment and topic analysis in engineering applications. Validated on a campus feedback dataset, it achieves 85.7 % accuracy in sentiment classification and 83.2 % in topic identification, with reduced computational overhead (12 ms sample latency, 465 MB memory). Compared to traditional methods, MTL-BERT offers improved efficiency and accuracy, enabling real-time deployment in resource-constrained engineering systems like manufacturing and infrastructure management.