Abstract: Aiming at the problems of insufficient modeling of trigger word contexts, inability to utilize historical event information, and difficulty in capturing complex semantic structures in generative event extraction models in complex scenarios, we proposed a generative event extraction model enhanced by event-aware memory graph network. The model enhanced its ability to focus on trigger word related contexts through an event-aware attention mechanism, utilized dynamic event memory network to achieve cross-text historical information reuse, and displayed semantic associations among modeling event arguments through an event graph structure. Experimental results on two standard datasets show that the proposed model outperforms existing baseline models in both event argument identification and classification tasks, effectively improving the extraction accuracy and generalization capability of generative models in multi-event and complex context scenarios.

Key words: generative event extraction, event-aware memory graph, dynamic event memory network, attention mechanism