HCG-EventPre: A Pre-Training Framework for Event Trigger and Argument Extraction

Open-world machine learning addresses a key limitation of closed-world learning: it simulates real-world uncertainty by assuming training data cannot cover all possible classes. Event trigger word and argument extraction is a core task of information extraction, which provides critical semantic support for human-centric applications such as social public opinion monitoring, emergency response decision-making, and user behavior intention analysis. However, existing pre-training models for this task generally have two key limitations: on the one hand, they lack sufficient modeling of the semantic correlation between event elements, leading to low extraction accuracy in complex scenarios; on the other hand, they have difficulty effectively capturing complex event structures, particularly in cases involving overlapping and nested events, where multiple triggers and arguments interact within the same context. To address these challenges, this paper proposes an event extraction framework termed HCG-EventPre. Specifically, the framework introduces an information fusion layer to address complex overlapping and nested event structures. This layer leverages an attention mechanism to obtain abstract event-type representations and employs gated conditional normalization to adaptively integrate event-type, semantic, and structural features. Furthermore, we propose a span–relation joint prediction layer that models trigger–argument extraction as a word-pair span and relation classification task. A convolutional neural network captures contextual dependencies between word pairs, while span and relation extractors together with a deep biaffine transformation network estimate trigger–argument association probabilities. The two prediction signals are integrated to improve the accuracy of trigger and argument extraction. Extensive experiments conducted on multiple datasets further validate the effectiveness of the proposed framework. Experimental results and case studies demonstrate that the proposed model achieves superior performance in event trigger and argument identification compared with event extraction methods based on predefined event patterns. In particular, the model shows strong effectiveness under few-shot annotation settings and cross-domain evaluation scenarios, indicating its robustness in handling complex event structures and improving the overall accuracy of trigger and argument extraction.