TMulAMR: Type-Aware Multi-AMR Fusion for Event Argument Extraction

Accurate Event Argument Extraction (EAE) is fundamental for enabling intelligent consumer technologies—such as virtual assistants, smart home systems, and conversational agents—to understand user commands and contextual events in a natural and adaptive manner. However, existing EAE models often underperform in these real-world applications due to sparse and ambiguous language inputs, which lead to incomplete semantic interpretation and hinder user experience. While Abstract Meaning Representation (AMR) provides a deep semantic structure, its effectiveness in EAE remains constrained by the limited connectivity of individual AMR graphs and the isolation of events during processing. To address these issues, we propose Type-Aware Multi-AMR (TMulAMR), a novel generation-based framework that systematically fuses AMR graphs from both semantically related and identical event types to enrich cross-event semantic reasoning. TMulAMR constructs a unified multi-AMR representation and employs three complementary prompts—an AMR prompt, a cross-context prompt, and a type-relational prompt—to enhance contextual understanding and role alignment. Experiments on benchmark datasets,WikiEventsandRAMS, show that TMulAMR achieves state-of-the-art performance on RAMS and WikiEvents, outperforming strong generation-based and classification-based baselines such as PAIE by +3.2% and +4.7% F1, respectively. The results demonstrate TMulAMR’s robustness and its potential for deployment in consumer-oriented intelligent systems requiring reliable semantic comprehension.