HMEAE moves away from treating event argument roles as independent labels, instead leveraging a human-designed concept hierarchy to improve extraction. By using a modular neural network architecture where high-level concepts (like "Person" or "Time") are shared across roles (like "Buyer" or "Seller"), the model gains a powerful inductive bias that significantly boosts performance on standard benchmarks.
The Challenge of Event Argument Extraction
In Natural Language Processing, Event Argument Extraction (EAE) involves identifying participants in an event and assigning them specific roles. For instance, in the sentence "Steve Jobs sold Pixar to Disney," an EAE system identifies "Steve Jobs" as the Seller and "Disney" as the Buyer.
Existing state-of-the-art models often treat every role as an independent category. However, roles are conceptually related: a "Seller" and a "Buyer" are typically a Person or an Organization, while a "Time-within" role is strictly related to Time. Ignoring these correlations misses the "common sense" hierarchy humans use, hindering generalization, especially when data for specific roles is sparse.
The HMEAE Framework
The proposed Hierarchical Modular Event Argument Extraction (HMEAE) model introduces a modular architecture consisting of three main stages:
- Instance Encoder: Converts a sentence into hidden representations using a CNN or BERT. It employs "dynamic multi-pooling" to aggregate information about the event trigger and candidate arguments.
- Hierarchical Modular Attention: This core innovation utilizes Neural Module Networks (NMNs) based on a concept hierarchy.
- Superordinate Concept Modules (SCM): Modules trained to recognize broad concepts like "Person," "Org," or "Time."
- Logic Union Modules: These combine SCM outputs using logical operations. For example, if a "Seller" can be a Person or an Organization, the model uses a "Logic Union" of both modules to create a role-oriented embedding.
- Argument Role Classifier: Combines general sentence information with specialized role-oriented embeddings to predict the final role.
By sharing the "Person" module across multiple roles (Buyer, Seller, Attacker, etc.), the model learns general features of a person more effectively than learning them separately for every role.
Evaluation and Results
The model was evaluated on two major benchmarks: ACE 2005 (35 argument roles) and TAC KBP 2016 (testing generalization across data distributions).
HMEAE consistently outperformed previous state-of-the-art models. On the ACE 2005 dataset, HMEAE (with BERT) achieved an F1 score of 59.3%, surpassing the vanilla BERT baseline (57.2%) and the previous best, dbRNN (58.7%). On TAC KBP 2016, the model achieved an F1 of 25.1% compared to the DMCNN baseline of 16.9%, demonstrating robustness on out-of-distribution data.
Key Insights
- Inductive Bias via Hierarchy: Baking human knowledge about concept relationships into the neural architecture provides a strong inductive bias.
- Modular Sharing: Sharing high-level concepts makes the model data-efficient; it can apply knowledge learned from one role (e.g., "Origin") to another (e.g., "Destination").
- Interpretability: Modular design allows for visualization. Heatmaps show the "Time" module specifically activating on words like "Wednesday," confirming the modules learn intended concepts.
Limitations and Future Directions
Current limitations include the need for manual hierarchy design. Future work may explore integrating linguistic knowledge (like syntactic dependency trees) or expanding this modular approach to other complex NLP tasks.
Summary updated