Speaker
Description
With the rise of cyber threats, automating Named Entity Recognition (NER) in open-source documents is crucial for Cyber Threat Intelligence (CTI). However, cybersecurity NER models face challenges in maintaining large annotated datasets due to the ever-evolving threat landscape. To address this, we introduce LETNER, a label-efficient NER framework that balances performance and annotation demands. LETNER features Span-CNN-Gate, a convolutional gating module that enhances span-based entity representation, and integrates metric learning to effectively capture entity-span relationships in a shared metric space, improving adaptability in low-resource settings. We also propose a systematic evaluation framework for label efficiency in supervised NER models. Experimental results demonstrate that LETNER achieves state-of-the-art label efficiency, significantly reducing annotation costs while maintaining high performance. On a complex CTI dataset with 21 fine-grained entity classes, LETNER outperforms the widely adopted Flair NER framework by 11.8\% in F1 score while using only 10\% of the training data.
