Overlapping Entity Relation Extraction for Cybersecurity Based on Complex Tagging and Rule-Enhanced Prompt Learning

Abstract

Cyberspace security is one of the areas that must be prioritized in the development of the Internet and artificial intelligence. The fourth paradigm of “Pre-train, Prompt, Predict” in the field of natural language processing (NLP) has significantly enhanced NLP efficiency and accuracy by shifting downstream tasks from masked to generative, marking a new leap forward. In the realm of cyberspace security, where information flows are vast and multifaceted, the focus of this paper is on extracting multi-relational information among multiple entities. Based on complex tagging and rule-enhanced prompting learning, this paper proposes a method for extracting overlapping entity relationships in cybersecurity. To address the challenges posed by “overlapping entity relationships” and “nested entities” among multiple entities, we have designed a triple extraction model that integrates complex tagging and rule-enhanced prompting learning. This model is a pipeline consisting of three parts: an entity recognition model, a rule injection module, and a relation classification model. The entity recognition model tackles the “nested entity” problem through a multi-label layer and performs entity classification. Rule injection serves two purposes: constraining possible combinations of knowledge triples and constructing prompting learning templates. These templates are input into the relation classification model, which cleverly utilizes suspended position markers to improve model efficiency and maximize the utilization of global semantic information. Finally, the effectiveness of the model is validated on the general dataset DUIE 2.0 and a self-constructed Chinese cybersecurity dataset of overlapping entity relationships.