Automating Named Entity Recognition for Indonesian Diplomas via Template-Based Synthetic Data Generation

Abstrak

Automated information extraction from administrative documents, such as higher education diplomas, is critical to streamlining public service verification. However, the development of robust Named Entity Recognition (NER) models for this domain is hindered by the scarcity of publicly available datasets and the prohibitively high cost of manual annotation. The current study addresses this gap by proposing a novel automated pipeline for generating domain-specific NER datasets for Indonesian higher education diplomas. Constructed from 36 distinct higher education templates and 3,460 alumni records, the dataset was synthesized using Direct Label Generation, which leveraged the PDDikti higher education database and a template-rendering mechanism. The generated dataset was evaluated by benchmarking two Pre-trained Language Models (PLMs), where the Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT) outperformed the 522M baseline (F1=0.8048), achieving a superior overall F1-score of 0.8213. To validate real-world utility, a pilot evaluation on held-out, noisy real-world diplomas yielded an F1-score of 0.7491 and a recall of 0.8460. These results confirm that the synthetic generation method effectively bridges the data gap, enabling high-performance extraction even on raw, noisy documents.

Identifier

• DOI: https://doi.org/10.48084/etasr.16738
• OpenAlex: https://openalex.org/W7150138229