Transformer-Based Adaptive System for Graded Reading Difficulty in Vocational College English

Abstract

Accurate difficulty a ssessment o f t echnical E nglish d ocuments i s e ssential f or adaptive learning and intelligent information processing in engineering education, particularly for multilingual technical resources associated with electromagnetic systems and wireless communication applications. To address the limitations of conventional Transformer models in handling long documents with complex structures and domain-specific t erminology, t his s tudy p roposes a s tructure-aware a daptive reading difficulty a ssessment f ramework i ntegrating h ierarchical d ocument e ncoding, r andom-access long-text processing, terminology graph enhancement, and multi-task learning. Structural information from chapters, paragraphs, headings, and tables is explicitly incorporated into Transformer representations, while graph-enhanced professional terminology embeddings improve semantic understanding in specialized contexts. The model jointly predicts global and paragraph-level difficulty a nd c ombines l earner a bility modeling with dynamic recommendation strategies to achieve personalized reading adaptation. Experimental results demonstrate an overall difficulty p rediction a ccuracy o f 9 2.3% a nd t erminology r ecognition accuracy approaching 90%, outperforming conventional LSTM-based methods while maintaining efficient processing for long documents. The proposed framework provides an effective solution for intelligent graded reading and offers practical support for semantic understanding and adaptive management of technical documentation in electromagnetic engineering, wireless communication, and related multidisciplinary applications.