Petunjuk Operasi Keuangan Jawa Tengah

This research aims to develop an automated text summarization system for financial Operational Guidelines documents (POK) using the Bidirectional Encoder Representations from Transformers (BERT) model. Financial POK documents, issued by government institutions, are typically long, complex, and contain dense technical terminology as well as numerical data, making them difficult and time-consuming for stakeholders to comprehend efficiently.

To address this issue, this study applies an extractive summarization approach, which focuses on selecting the most important sentences from the original document without altering their structure. This approach is chosen to preserve factual accuracy and maintain the integrity of financial information. The model implemented in this research is based on BERT, specifically adapted using the BERTSUM architecture, which enhances sentence-level representation through the use of special tokens and inter-sentence Transformer layers.

Furthermore, the study utilizes IndoBERT, a pre-trained language model for the Indonesian language, and performs fine-tuning using a dataset of financial POK documents obtained from official government sources such as data.go.id. The model is expected to learn the unique characteristics of Indonesian financial documents, including bureaucratic writing styles and domain-specific terminology.

The performance of the proposed model is evaluated using standard summarization metrics, namely ROUGE-1, ROUGE-2, and ROUGE-L, to measure the similarity between the generated summaries and human-written references. Additionally, qualitative evaluation is conducted to assess the coherence, readability, and completeness of the generated summaries.

Overall, this research contributes to the advancement of Natural Language Processing (NLP) applications in the Indonesian public sector, particularly in improving accessibility and efficiency in understanding government financial documents.