Answers are grounded in this paper and cite their sources.
Extracting relevant entities such as malware, trending incidents, and involved organizations or individuals from cyber threat intelligence and social network content is critical to ensuring the security of network communication environment and content. Chinese Named Entity Recognition(NER) is the enabling technology for extracting such entities from unstructured Chinese text. However, due to high annotation costs and highly dynamic nature of social network content, labeled data in this domain tends to be scarce, which makes few-shot Chinese NER particularly important. Existing studies mainly address annotation scarcity through transfer learning or prompt learning, the former degrades when cross-domain discrepancies are pronounced, while the latter struggles to generate effective prompts tailored to few-shot scenarios. To alleviate these challenges, we propose PLDKT, a few-shot Chinese NER framework integrating Prompt Learning with Domain-invariant Knowledge Transfer. PLDKT first learns transferable entity span knowledge from multiple annotation-rich domains. To decouple shared entity span knowledge from domain-specific characteristics, the soft prompt is decomposed into domain-invariant and domain-specific components under an orthogonality constraint. The learned domain-invariant prompt is then transferred to target few-shot domain for entity span detection, guided by a confidence-based adaptation strategy that dynamically adjusts update strengths of both components. Finally, entity typing is performed by jointly leveraging character-level, word-level, and sentence-level representations of candidate entity spans. Experiments on three public Chinese datasets demonstrate that PLDKT outperforms ten representative baselines, achieving absolute average F1-score improvements of 1.79, 1.77, and 2.12 percentage points on APTIE, CDTier, and Weibo over the strongest baseline, confirming its effectiveness for few-shot Chinese NER. Experiments on three public Chinese datasets demonstrate that PLDKT outperforms eight representative baselines, achieving average F1-score improvements of 4.20%, 2.68% and 2.92% on APTIE, CDTier and Weibo over the strongest baseline, confirming its effectiveness for few-shot Chinese NER.
No takes yet. Share an insight, caveat, or question.
Duan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: