Introduction
Event Argument Extraction (EAE) is a critical task in information extraction, aimed at identifying the participants and their roles within an event (e.g., identifying the "Giver," "Recipient," and "Amount" in a donation event). Despite progress, current models often struggle with two key challenges: intra-class variance and inter-class separation.
Intra-class variance occurs when arguments for the same role are semantically diverse. For instance, in a "giving aid" event, the "Giver" could be an individual like "Trump" or an organization like "the United States." Standard models using a single prototype per role fail to capture these distinct sub-clusters. Conversely, inter-class separation refers to the fuzzy boundaries between different roles, such as "Beneficiary" and "Recipient," which often overlap in traditional embedding spaces.
Methodology: The HMPEAE Framework
To address these issues, we propose HMPEAE (Hyperspherical Multi-Prototype with Optimal Transport). The core innovation lies in representing each role with multiple prototypes mapped onto a high-dimensional hypersphere.
Hyperspherical Multi-Prototypes
Instead of a single point, each role is assigned $M$ prototypes. Mapping these to a hypersphere provides a geometric advantage, encouraging a large-margin separation between categories and avoiding the "crowding" effect common in Euclidean spaces.
Three-Target Prototype Training
The prototypes are optimized through a three-pronged training objective:
- Inter-class Maximization: Increasing the distance between prototypes of different roles.
- Intra-class Minimization: Keeping the multiple prototypes of a single role relatively close to maintain semantic cohesion.
- Semantic Prior Integration: Utilizing BERT-based encoders to ensure that semantically unrelated roles (e.g., "Attacker" vs. "Place") are pushed further apart than related ones.
Optimal Transport for Argument Assignment
During training, the model must assign specific argument mentions to one of the $M$ prototypes. We formulate this as an assignment problem and solve it using the Sinkhorn-Knopp algorithm for Optimal Transport. This ensures a balanced and meaningful distribution of arguments across the role's sub-clusters, preventing prototype collapse.
Experimental Results
The HMPEAE model was evaluated on the RAMS and WikiEvents benchmarks, achieving state-of-the-art results on both.
On the RAMS dataset, HMPEAE achieved an Arg-C F1 score of 53.7, outperforming previous methods. On WikiEvents, it reached 66.6, a significant 0.5-point lead over the previous state-of-the-art. Ablation studies confirmed that both the hyperspherical geometry and the multi-prototype structure are essential for these performance gains. Removing either component resulted in a significant drop in extraction accuracy.
Conclusion
The HMPEAE model demonstrates that accounting for the semantic diversity within roles is crucial for robust Event Argument Extraction. By combining hyperspherical geometry with multi-prototype representations and Optimal Transport, the model effectively captures complex role semantics while maintaining clear decision boundaries. Future work will explore adaptive prototype counts for different roles and the application of this framework to few-shot learning scenarios.
Summary updated