<![CDATA[Automated business plan assessment in entrepreneurship education requires an understanding of document structure, domain-specific terminology, and logical relationships between sections. However, conventional text processing methods, including TF-IDF and flat language representations, often fail to capture structural dependencies, reducing the effectiveness of competency assessment and personalized feedback. Purpose: This study aims to develop a Section-Aware Hierarchical Transformer (SAHT) architecture that integrates hierarchical document representations, domain-adapted language modeling, and graph-based cross-sectional attention to improve automated business plan assessment within entrepreneurship education. Methods/Study Design/Approach: The proposed architecture consists of three integrated stages. First, automatic section segmentation identifies logical document components using heading recognition, typographic features, and a BERT-based fallback classifier. Second, individual sections are encoded using EntreP-BERT, a domain-adapted language model pre-trained on approximately 500,000 entrepreneurship-related texts, followed by learnable pooling to emphasize competency-relevant information. Third, cross-sectional attention and graph convolutional operations capture logical dependencies among business plan sections. The resulting representations are integrated into the CBT-DE artifact submission system for multi-label classification of twelve entrepreneurial competencies. Additionally, a conditional variational autoencoder generates personalized feedback questions to support instructional evaluation. Results/Findings: The BERT-based fallback classifier achieved 94.2% accuracy in identifying sections within non-standard business plan documents. Experimental findings indicated that SAHT improved competency prediction and feedback relevance compared with conventional flat text representation baselines. Hierarchical encoding and graph-augmented attention facilitated the integration of contextual information across related sections, enabling more structured competency evaluation and personalized feedback generation. Novelty/Originality/Value: This study introduces a section-aware hierarchical architecture combining entrepreneurship-specific language modeling, learnable section representations, and graph-augmented cross-sectional attention. Unlike conventional approaches that process business plans as unstructured text, SAHT explicitly preserves document organization and models logical dependencies between sections. Its integration with competency classification and personalized feedback generation provides a framework for improving automated assessment, instructional analytics, and competency-based entrepreneurship education.]]>
Copyrights © 2026
