A Hybrid Approach for Event Extraction in Safety Production

Event extraction aims to identify events of interest from unstructured text and represent them in a structured format. However, due to the diversity and complexity of events, the current accuracy of event extraction remains low, particularly for sentences containing multiple events. Furthermore, existing event extraction studies have primarily focused on domains such as news, finance, and biomedicine, with limited application in the field of safety production. To address these challenges, this paper divides the event extraction task into two phases tailored to the characteristics of the safety production domain: event triple extraction and event classification. This approach achieves the extraction of essential event information and the categorization of event types through a phase-based extraction method. By extracting event triples based on dependent syntactic analysis and semantic role annotation, we can successfully match 84.64% of events according to the template and structure the event representations. Subsequently, the event classification task is performed on the event triples using the ELECTRA_TextRCNN model. In comparison to the TextRCNN model, our proposed ELECTRA_TextRCNN model demonstrates a 9% improvement in accuracy and a 7% increase in F1 value. Ultimately, we have significantly enhanced the accuracy of event extraction in the safety domain through a two-stage approach.