Abstract
The Event Argument Extraction (EAE) task extracts structured event records from text. Most existing approaches train models on datasets independently, ignoring "overlap knowledge"—generalizable associations between entity types and events shared across datasets. This paper defines overlap knowledge and proposes APE (Assembles Prefix and Adapter), a model that learns shared and dataset-specific knowledge in two serial phases. By formulating both as conditional generation tasks and using Stressing Entity Type Prompts, APE achieves state-of-the-art performance, particularly in low-data scenarios where it outperforms baselines by an average of 27.27% F1 with only ten training records.
Introduction
Event Argument Extraction (EAE) is the process of identifying the "who, what, where, and when" of an event described in text. Despite its importance, EAE models often struggle with data scarcity due to the high cost of annotation. While transfer learning offers a potential solution, it is complicated by varying schemas across datasets; for instance, a person might be labeled a "Victim" in one dataset and a "Target" in another.
Existing models typically treat datasets in isolation, missing the opportunity to leverage universal entity-event relationships. APE addresses this by decoupling "overlap knowledge"—the general association between entity types (e.g., Person, Location) and events—from "specific knowledge"—the dataset-specific roles (e.g., Killer, Victim).
Methodology: The APE Framework
The APE model operates under the Seek Common ground while Reserving Differences (SC-RD) framework. This framework decomposes the EAE task into two distinct learning phases designed to accumulate knowledge without catastrophic forgetting.
Two-Phase Learning
- Overlap Knowledge Learning (Phase 1): The model learns to identify entities associated with an event using a universal set of entity types (e.g., Person, Location, Time). This is framed as a Pseudo-Entity Recognition (PER) task. General knowledge is stored using Prefix-tuning, where trainable prefix vectors are added to Transformer layers.
- Specific Knowledge Learning (Phase 2): The model is tuned to the specific argument roles of a target dataset. It uses Adapters—small, trainable modules—while the previously learned "Overlap" Prefix remains frozen.
Stressing Entity Type Prompts
To bridge these phases, APE uses Stressing Entity Type Prompts. These prompts utilize identical special tokens (e.g., [person/organization]) in both phases. This consistency allows the model to "ignite" the general knowledge from Phase 1 to solve the specific role-mapping task in Phase 2.
Experimental Evaluation
APE was evaluated on three major benchmarks: ACE 2005 (sentence-level), RAMS (document-level), and WikiEvents (document-level). A manual mapping function was used to align dataset-specific roles with universal entity types.
Results and Key Insights
APE achieved new state-of-the-art results across all benchmarks, with absolute F1 gains of 2.7% on ACE05, 2.1% on RAMS, and 3.4% on WikiEvents using a BART-large backbone.
The model's most significant advantage is in few-shot learning. As shown in the table below, when only 10 training samples are available, APE dramatically outperforms the PAIE baseline.
Ablation studies confirmed that both the Prefix (for shared knowledge) and the Adapter (for specific knowledge) are critical. Removing either component resulted in significant performance degradation, validating the two-phase modular approach.
Conclusion and Future Work
APE demonstrates that explicitly modeling overlap knowledge across datasets significantly enhances EAE performance, especially when data is scarce. By isolating shared and specific knowledge, the model effectively transfers "event-sense" across different schemas. Future research directions include automating the mapping between roles and entity types and exploring cross-lingual transfer learning for EAE.
Summary updated