Exploiting Semantic Space to Enhance Event Detection Combined with Event Knowledge

Abstract

Event detection (ED) aims to entail the identification of triggers in text and the determination of the appropriate categories. Event detection is frequently hampered by labeled data scarcity, and most current methods ignore the correlations between events. Aiming at the issues, a novel event detection framework leveraging semantic space and event knowledge (EDSSEK) is proposed. To introduce event knowledge, the pre-trained model is extended and fine-tuned based on the elaborately constructed domain corpus. The presentations of event types are encoded through the pre-trained model and mapped into the semantic space. The feature vectors of event triggers are gained using a document-level attention mechanism and then projected into the same vector space. The document embedding networks are trained by minimizing the distances between the event triggers and the relevant types. Experiments on benchmark datasets demonstrate that EDSSEK outperforms other state-of-the-art methods, and justify the effectiveness of semantic space combined with event knowledge.