Introduction
Event Argument Extraction (EAE) is a critical task in Information Extraction, focused on identifying participants and their roles within an event—for instance, identifying the "Buyer" in a "Transaction" event. While sentence-level extraction has seen significant progress, real-world data often presents a more complex scenario: nearly 40% of event arguments appear in different sentences than their triggers. This necessitates robust document-level EAE.
The primary challenge in document-level extraction is noise. Documents contain vast amounts of text irrelevant to a specific event. Traditional methods using syntactic dependency trees often rely on rigid rules, such as keeping only the direct path between a trigger and an argument. However, these rules frequently prune essential contextual words that lie "off-path," leading to information loss.
Methodology: The OTEAE Model
To address these challenges, researchers developed OTEAE (Optimal Transport-based Event Argument Extraction). This model introduces two key innovations to bridge the gap between sentence and document-level extraction.
1. Optimal Transport for Tree Pruning
Instead of hard-coded pruning rules, OTEAE treats pruning as an alignment problem using Optimal Transport (OT). The model uses the direct dependency path as an anchor and aligns "off-path" words to it based on two factors:
- Syntactic Importance: Determines the "mass" of a word based on its distance within the dependency tree.
- Semantic Similarity: Acts as the "transport cost," measuring the effort to move mass between words based on embedding similarity.
By introducing a "NULL" node, the model can flexibly prune words if the cheapest transport option is to the NULL node. This allows the system to retain semantically relevant words even if they are syntactically distant.
2. Contribution-Constrained Regularization
Standard contextual embeddings like BERT can "leak" information from pruned-away noisy words due to self-attention. OTEAE employs a regularization loss that ensures the model's final representations remain consistent whether calculated from the full or pruned tree, effectively "de-noising" the BERT embeddings.
Experimental Results
The model was evaluated on the RAMS dataset, a large-scale benchmark for document-level EAE, and ACE 2005.
As shown in the results, OTEAE established a new state-of-the-art (SOTA) on the RAMS dataset with an F1 score of 80.9 (Type Constrained), outperforming both sequence-based and other structure-aware models like LSR.
Handling Long-Distance Dependencies
A significant advantage of the OT-based approach is its ability to capture long-distance dependencies. In document-level extraction, the distance between a trigger and its argument can span multiple sentences.
The model demonstrates particular strength in cross-sentence scenarios (distances -2, -1, 1, 2), proving that the refined document structure effectively filters noise while preserving distant but vital context.
Conclusion
OTEAE demonstrates that combining syntactic structure with semantic similarity through Optimal Transport provides a mathematically principled way to handle the noise inherent in document-level extraction. By moving beyond rigid dependency path rules and addressing the "stickiness" of contextual embeddings, the model sets a new standard for identifying event arguments across complex documents. Future work may extend this OT-based pruning to other tasks like Relation Extraction or explore more computationally efficient implementations for even longer documents.
Summary updated