Extractive summarization of Indonesian friday sermons using fine-tuned IndoBERT and pseudo-labeled data

Abstract Automatic text summarization of Indonesian religious content, particularly Friday sermon (Khutbah Jumat) transcripts, remains underexplored. These transcripts contain long-form spoken discourse and domain-specific Islamic terminology, making manual review time-consuming. This study aims to develop an extractive summarization framework for Indonesian Friday sermon transcripts using a pseudo-labeling pipeline combined with fine-tuning IndoBERT for sentence-level classification. A corpus of 210 Khutbah Jumat transcripts from the YouTube subset was preprocessed, producing approximately 293,000 words. Sentence-level pseudo-labels were generated by combining TF-IDF relevance, sentence position, and Islamic keyword density. The pseudo-labeled sentences were then used to train and evaluate an IndoBERT-based classifier, with performance assessed using ROUGE and BERTScore metrics. The proposed system effectively selected summary-worthy sentences, achieving ROUGE-1 = 0.5181 (95% bootstrap CI 0.460–0.575), ROUGE-2 = 0.4273, and ROUGE-L = 0.4400 against pseudo-reference summaries on a 21-document test set, the highest of all evaluated methods on these three metrics; paired bootstrap comparisons showed these gains to be consistent over TextRank and generic BERTSum but not over TF-IDF. On BERTScore F1, IndoBERT (0.7310) remained below BERTSum (0.7407), indicating that lexical-overlap gains did not consistently translate into greater semantic similarity. As no human-written gold-standard summaries currently exist for this domain, these scores should be interpreted as relative comparisons among automatic methods rather than validated measures of summary quality against human judgment. An exploratory topic-wise analysis found no statistically significant difference in ROUGE-1 across the five sermon topic categories (Kruskal–Wallis p = 0.11; n = 3–7 documents per topic). The findings suggest that pseudo-labeling provides a practical low-resource strategy for extractive summarization of Indonesian religious discourse. Fine-tuning IndoBERT with pseudo-labeled data improves sentence selection and offers a scalable framework for automatic sermon summarization that substantially reduces, though does not eliminate, the need for human-annotated reference summaries.