A Multi-Domain and Multi-Task Generative Framework with Explicit Task and Domain Conditioning for Cross-Domain Event Extraction

Developing event extraction models that generalize robustly across diverse domains remains challenging due to variations in contextual expressions and heterogeneous event schemas. Traditional models are typically trained separately for each dataset, limiting their flexibility on unseen domains, while large language models often underperform compared to smaller, task-specific fine-tuned models. This paper investigates whether a single unified sequence-to-sequence model can jointly handle multiple event extraction subtasks across disparate domains.

The authors propose a multi-domain and multi-task generative framework built on a T5-base architecture. The core innovation involves prepending lightweight domain indicators (e.g., Domain: Geneva) and task-specific prompts to the input text. This design enables the model to dynamically adapt to dataset-specific schemas without needing complete event label sets during inference. The framework accommodates both sequential pipeline settings and fully joint end-to-end extraction modes.

To evaluate cross-domain generalization, the authors utilize six diverse event extraction benchmarks encompassing cybersecurity, news, biomedical, and Wikipedia domains. These datasets exhibit substantial variations in scale, sequence length, annotation density, and ontology complexity. For instance, Geneva and RAMS feature complex ontologies with over 100 event types, whereas Genia2013 focuses on specialized biomedical terminology with limited roles.

The study trains six single-domain models alongside a single multi-domain joint model for 40 epochs. Evaluation is conducted across multiple subtasks including trigger identification, trigger classification, argument identification, and argument classification. A deterministic post-processing parser converts the linearized generative text outputs back into structured event and argument records.

Jointly training on all datasets yields performance comparable to, and frequently exceeding, dedicated single-domain models while utilizing only a single unified architecture. The empirical results demonstrate that knowledge acquired from high-resource datasets transfers effectively to lower-resource or structurally distinct domains.

Furthermore, performance improvements are particularly pronounced for argument-level metrics such as argument identification and classification. This indicates that argument extraction benefits significantly from shared cross-domain semantic representations. Overall, the approach provides practical scalability and eliminates the need to maintain separate models for every domain.