Unsupervised Keyphrase Extraction Approach for Hazard Records by Contrastive Learning

During safety inspections, analyzing hazard records can help staff quickly assess the overall risk situation and formulate corrective measures. Key Phrase Extraction (KPE) can extract crucial phrases from documents with core information, aiding staff in conducting hazard record analysis. However, general SOTA approaches have performed poorly in security domains. In this study, we introduce a KPE method called Domain-oriented Joint Scores Rank, which evaluates and ranks candidate phrases by combining scores from multiple dimensions. Further, we have developed a security domain-oriented BERT model by proposing a lightweight comparative learning fine-tuning method. The HKBERT-based DJSRank outperforms the SOTA general domain model on the KPE task for hazard records, with an average F1 improvement of 4.3.