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

Few-shot Named Entity Recognition (FS-NER) addresses the challenge of identifying entity mentions under extremely limited labeled supervision, a condition that closely reflects real-world scenarios in low-resource and morphologically rich languages such as Indonesian.Despite their strong representational capacity, transformer-based NER models exhibit substantial performance degradation under sparse supervision, severe label imbalance, and semantic ambiguity.To address these limitations, this study proposes a semantic meta-learning BIO-CRF framework for Indonesian FS-NER that integrates episodic meta-learning, semantically enriched span-level prototype learning, and BIO-constrained Conditional Random Field (CRF) decoding.The framework enhances prototype stability by fusing data-driven span representations with label-text semantic information while preserving structural consistency through sequence-level decoding.Rare-entity boosting and class-weighted learning are further incorporated to mitigate long-tail entity distributions commonly observed in Indonesian corpora.Extensive experiments conducted on heterogeneous Indonesian NER datasets demonstrate that the proposed approach consistently outperforms representative metric-based, episodic fine-tuning, and label-semantic prompting baselines across multiple K-shot settings (K = 10, 20, and 50).Under the most restrictive condition (K = 10), the framework achieves an F1-score of 82.14% and maintains stable performance gains as the support size increases, indicating effective semantic generalization rather than reliance on additional labeled data.Ablation studies and statistical significance analyses confirm the complementary contributions of semantic fusion, episodic meta-learning, and CRF decoding.These findings establish a robust benchmark for Indonesian FS-NER and demonstrate the effectiveness of combining semantic modeling with structured inference in low-resource sequence labeling.