<![CDATA[Automatically transcribed Friday sermons are generally long and contain conversational noise, making their central messages difficult to locate quickly. This study implements and evaluates extractive summarization of Indonesian Friday-sermon transcripts using IndoBERT with sentence-level pseudo-labels. Of 210 YouTube transcripts, 32 documents that produced only one segment were excluded, leaving 178 documents and 16,131 sentences. The sermons were partitioned at document level into 124 training, 27 validation, and 27 test documents without overlapping doc_id values; TF-IDF was fitted only on 11,100 training sentences. Fine-tuning stopped at epoch 4 and selected the epoch-1 checkpoint with a training loss of 0.5128, validation loss of 0.4904, and validation accuracy of 0.7654. The checkpoint generated actual summaries for 27 test documents by selecting approximately 30% of their sentences. Against author-selected and author-reviewed extractive references, IndoBERT obtained ROUGE-1, ROUGE-2, and ROUGE-L F1 scores of 0.4141, 0.2907, and 0.3206. On the same documents, references, extraction ratio, and evaluation implementation, Lead-30% obtained 0.4559, 0.3024, and 0.2912, while LexRank obtained 0.6646, 0.5648, and 0.5890. LexRank achieved the strongest overall results, whereas IndoBERT exceeded Lead-30% on ROUGE-L. The study provides a controlled evaluation and empirical evidence about the limits of weakly supervised IndoBERT in the Friday-sermon domain]]>
Copyrights © 2026
