<![CDATA[The extractive summarization of Indonesian scientific articles is hindered by a domain mismatch where established methodologies rely on news-corpus assumptions, whereas Indonesian scientific discourse follows rigid, IMRaD-driven structural and lexical patterns. This study aims to systematically analyze feature interaction effects and saturation behaviour in RankSum-based extractive summaries for Indonesian scientific articles. Designed as a controlled comparative experiment, this research evaluates a RankSum framework integrating variables, such as graph-based, semantic-thematic vectors, and structural heuristics. The dataset comprises 2,897 Indonesian journal articles (2021-2025) collected via web scraping from open-access university repositories. Analysis across 31 scenarios demonstrates that for Indonesian scientific articles, the assumption that increasing feature density improves performance is flawed, instead a feature saturation effect occurs. Results show that a 4-feature combination maximizes unigram lexical precision (ROUGE-1 0.3564), whereas the full 5-feature fusion is necessary to preserve global semantic integrity, structural flow, and stable (ROUGE-L 0.2018; BERTScore 0.6977). This study establishes a generalizable principle for domain-aware ATS by demonstrating that overcoming domain mismatch relies on navigating feature saturation through selection aligned with the document’s inherent logic rather than raw feature quantity.]]>
Copyrights © 2026
