TASK-SPECIFIC ATTENTION LAYERS FOR THE ENHANCED FINE-TUNING OF BERT IN NLP APPLICATIONS

Leveraging BERT’s pre-trained capabilities, fine-tuning has become a standard approach for optimising model performance across various NLP tasks. However, the potential to further refine task-specific outputs remains underexplored. By incorporating external attention layers designed for individual tasks, this study introduces a method to enhance BERT’s performance on tasks such as sentiment analysis, paraphrase detection, and semantic textual similarity. The proposed approach assigns task-specific weights to different tokens during fine-tuning, allowing the model to focus more effectively on key elements in the input data. Experimental results indicate significant improvements in paraphrase detection accuracy, achieving a 29% increase, with only a slight increase in training time. This research provides a novel extension to BERT’s architecture, demonstrating that targeted attention layers can yield considerable performance gains for specialised NLP applications while maintaining efficient resource utilisation.