Skip to content
SDGs Center UDINUS

Evidence record · UDINUS · SDGs

A Hybrid Semantic Meta-learning and CRF-based Framework for Improving Indonesian Few-shot Named Entity Recognition

Rektorat UDINUS 2026 Published 2026-10-30

Evidence record · UDINUS · SDGs

Unit
Rektorat UDINUS
Period
2026
Goals
Record
70c82210-3bb8-442a-b130-dcaf651e2a55
Published
2026-10-30

Universitas Dian Nuswantoro (UDINUS), through an activity involving Nurchim, Muljono, Edi Noersasongko, Ahmad Zainul Fanani, and Deshinta Arrova Dewi as participants, addressed the challenge of Few-shot Named Entity Recognition (FS-NER) for Indonesian. The activity focused on identifying entity mentions when labeled supervision is extremely limited, a condition associated in the source with low-resource and morphologically rich languages such as Indonesian. The work examined limitations of transformer-based NER models under sparse supervision, severe label imbalance, and semantic ambiguity.

The activity proposed a semantic meta-learning BIO-CRF framework for Indonesian FS-NER. Its approach integrated episodic meta-learning, semantically enriched span-level prototype learning, and BIO-constrained Conditional Random Field decoding. Prototype stability was enhanced by combining data-driven span representations with label-text semantic information, while sequence-level decoding was used to preserve structural consistency. The framework also incorporated rare-entity boosting and class-weighted learning to address long-tail entity distributions commonly observed in Indonesian corpora.

Extensive experiments were conducted on heterogeneous Indonesian NER datasets. The proposed approach was compared with representative metric-based, episodic fine-tuning, and label-semantic prompting baselines across K-shot settings of 10, 20, and 50. Under the most restrictive K = 10 condition, the framework achieved an F1-score of 82.14% and maintained stable performance gains as the support size increased. The source describes this result as indicating effective semantic generalization rather than reliance on additional labeled data.

Ablation studies and statistical significance analyses confirmed complementary contributions from semantic fusion, episodic meta-learning, and CRF decoding. The activity therefore established a benchmark for Indonesian FS-NER and demonstrated the effectiveness of combining semantic modeling with structured inference in low-resource sequence labeling. In relation to the stated goal, UDINUS frames this activity as research involving the development and evaluation of a computational framework for Indonesian language technology.

Reference https://www.scopus.com/pages/publications/105030379682?origin=resultslist