Robust and Efficient Indonesian Span-Based Named Entity Recognition via Compact GLiNER
Abstrak
Efficient Named Entity Recognition (NER) integration is crucial for improving retrieval precision and fact verification in Retrieval-Augmented Generation (RAG) systems. However, the development of Indonesian NER still faces challenges, including dataset heterogeneity, inefficient tokenization, and trade-offs between NER performance and inference efficiency. Although span-based frameworks such as Generalist and Lightweight Named Entity Recognition (GLiNER) offer greater flexibility than conventional Beginning-Inside-Outside (BIO) approaches, available GLiNER models are generally multilingual and have not been optimized for Indonesian characteristics. This research proposes a compact GLiNER model specifically for Indonesian, utilizing a pruned mDeBERTa with a 30k vocabulary as the encoder backbone. We built a large-scale Indonesian GLiNER training corpus by combining heterogeneous NER datasets and augmenting them with controlled translations, resulting in 56,210 training, 6,707 validation, and 9,411 test samples. Experimental results show that the proposed model achieves an F1 score of 76.58%, surpassing the IndoBERT-based GLiNER baseline (74.70%) at max_len = 192 tokens, with consistent improvements on major entities. Deployment-oriented evaluation shows up to 8 × CPU-based inference acceleration (482 ms vs. 3,897 ms per sample). In long-context evaluation (max_len = 384), the proposed model outperforms multilingual GLiNER by more than 12 F1 points while maintaining a significantly lower memory footprint on both GPU and CPU. These advantages validate the model's potential as a reliable metadata filter component for RAG architectures.
Identifier
- DOI: https://doi.org/10.48084/etasr.18482
- OpenAlex: https://openalex.org/W7163733908