A Text Summarization-Based Similarity Algorithm for Inter-Subchapter Coherence Analysis in Indonesian Academic Documents

Abstract

In academic writing, the logical relationship between subchapters is essential for maintaining document coherence; however, manual evaluation of inter-subchapter consistency is time-consuming and subjective. This study proposes a hybrid text summarization and similarity-based algorithm integrating feature-based extractive summarization with BERT-based neural summarization and semantic similarity measurement using TF-IDF, cosine similarity, and Latent Semantic Analysis (LSA) to analyze inter-subchapter coherence in Indonesian dissertation qualification proposals. The principal novelty lies in operating at the subchapter level rather than the sentence or document level, enabling structural relationship analysis not addressed by existing approaches. Experiments were conducted on 30 Indonesian dissertation qualification proposals split into 21 documents (70%) for training, 3 (10%) for validation, and 6 (20%) for testing, annotated by domain expert evaluators using six quality criteria. Similarity analysis results show that the proposed method identifies strong semantic alignment between logically connected section pairs, with cosine similarity scores reaching 1.00 for the problem formulation - objectives pair and the background -methodology pair on the test set; these perfect scores reflect the structural consistency of academic proposals rather than normalization artifacts. In quality assessment, the model achieves an average exact-match accuracy of 50% against expert evaluations, with per-proposal accuracy ranging from 17% to 83%. The lower overall accuracy is attributed to BERT's tendency to over-predict quality in poorly structured documents, and these findings are reported as exploratory results given the limited test set size (n=6). The proposed framework makes a meaningful contribution toward automated academic writing assessment tools for Indonesian higher education, providing a structured, data-driven approach to evaluating proposal coherence that can serve as a foundation for future large-scale deployment.