A BERT-Based Multi-Task Learning Framework for Automated Scoring of Business English Writing with Domain Adaptation

With the rapid advancements in Natural Language Processing (NLP), automated scoring systems have emerged as essential tools for evaluating written texts with high accuracy, scalability, and efficiency. This study proposes a novel scoring framework for Business English writing, leveraging Bidirectional Encoder Representations from Transformers (BERT), combined with multi-task learning (MTL) and domain-specific transfer learning techniques. The system addresses the specialized challenges of Business English evaluation, including the handling of domain-specific terminology, complex sentence structures, and subjective aspects like coherence, fluency, and style. The framework unifies semantic representation, syntactic analysis, and coherence evaluation into a structured model. A dataset of $\mathbf{1 0, 0 0 0}$ annotated Business English samples, covering various business contexts such as marketing reports, executive summaries, and professional emails, was constructed for training and validation. Experimental results highlight significant improvements, with the system achieving a 58.6% reduction in Mean Squared Error (MSE) and a $\mathbf{2 4 \%}$ increase in Pearson Correlation Coefficient (PCC) compared to traditional rule-based scoring systems. In practice, this system provides accurate, consistent, and real-time feedback, offering significant benefits for educational and corporate environments. It enhances personalized learning by identifying specific areas for improvement and streamlines large-scale assessments with reliable scoring. The scalability and adaptability of the system make it a valuable asset for advancing Business English writing evaluation and improving the overall quality and efficiency of language assessments.