Research on named entity recognition method in cybersecurity based on soft prompt tuning and reinforcement learning

Abstract

As network technology rapidly advanced, new cybersecurity threats constantly emerged, increasing the importance of cybersecurity named entity recognition. To address the problem of poor recognition accuracy in named entity recognition methods based on large language models in the cybersecurity domain, a novel cybersecurity named entity recognition method that combined soft prompt tuning and reinforcement learning was proposed. By integrating the soft prompt tuning technique, the method precisely adjusted the recognition capabilities of large language models to handle the complexity of the cybersecurity domain, improving recognition accuracy for cybersecurity named entities while optimizing training efficiency. Additionally, a reinforcement learning-based instance filter was proposed, which effectively removed low-quality annotations from the training set, further enhancing recognition accuracy. The proposed method was evaluated on two benchmark cybersecurity NER datasets, with experimental results demonstrating superior performance in F1 score compared to state-of-the-art cybersecurity NER methods.

Keywords

• cybersecurity named entity recognition
• soft prompt tuning
• reinforcement learning
• large-scale pre-trained models