PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Siting Liang, Omar Adjali, Daniel Sonntag

arXiv:2608.23235v1cs.CL

TL;DR

Event extraction models often struggle to generalize across domains with different contextual expressions and event schemas. This paper presents a unified generative model using domain indicators and task prompts to handle heterogeneous schemas across subtasks and extraction settings. It achieves competitive performance across diverse datasets while improving efficiency and cross-domain transfer, particularly for resource-limited or schema-diverse domains.

  • Problem

    Existing event extraction models show limited flexibility and generalization when applied across domains with differing contextual expressions and event schemas.

  • Method

    The framework unifies multi-domain and multi-task event extraction through a generative model conditioned on lightweight domain indicators and task-specific prompts.

  • Results

    A single unified model achieves competitive performance across diverse domains, with clear gains on resource-limited or schema-diverse datasets such as Genia2013 and WikiEvents.

  • Takeaways & Limitations

    Unified multi-domain training can improve cross-domain transfer while reducing the need to maintain separate domain-specific systems.

  • Takeaways & Limitations

    Cross-domain evaluation was limited by the availability of datasets from diverse domains, which may constrain broader generalizability.

Abstract

from arXiv · show

Event extraction aims to identify event triggers, classify event types, and extract arguments to construct structured event representations. Despite strong in-domain performance, developing models that generalize robustly across domains remains challenging due to variations in contextual expressions and event schemas. Prior unified and multi-task approaches improve in-domain accuracy but exhibit limited flexibility when applied to unseen domains. Even large language model-based methods that provide full event ontologies at inference time often underperform compared to smaller, task-specific fine-tuned models. We propose a unified multi-domain and multi-task training framework that models heterogeneous event schemas within a single model. Our approach introduces domain conditioning signals, jointly with task-specific prompts, enabling dynamic adaptation to dataset-specific schemas without requiring complete event label sets at inference time. The framework supports both pipeline and end-to-end extraction settings, facilitating efficient task- and domain-level transfer. Experiments on diverse event extraction benchmarks demonstrate that our method achieves competitive performance, strong cross-domain generalization, and practical scalability, while preserving domain-specific precision.

1 Introduction

The paper frames event extraction as detecting event triggers and extracting their arguments, then proposes a single text-to-text model with task and domain conditioning for scalable cross-domain extraction. The framework supports pipeline and end-to-end settings while adapting to domain-specific patterns without complete event ontologies at inference time.

  • Task Definition: Event extraction identifies event-triggering spans and assigns event types, then labels participants or entities that fulfill event roles.
  • Unified Framework: The framework extends unified event extraction to a fully multi-domain, multi-task setting in which one model handles multiple domains and event subtasks.
  • Unified Framework: A text-to-text encoder–decoder jointly performs event detection and event argument extraction using task-specific tokens as lightweight prompts.
  • Domain Conditioning: Domain indicators such as “Domain: Geneva.” condition generation, enabling adaptation to domain-specific event patterns without explicit schemas or full ontologies at inference time.
  • Experimental Scope: A single T5-base model supports sequential pipeline extraction and joint end-to-end prediction, with experiments demonstrating multi-task, multi-domain training and cross-domain generalization.

2 Related Work

Prior event extraction work has progressed from separately trained, domain-specific discriminative models toward joint and unified multi-task systems that improve in-domain performance. LLM prompting offers generalized extraction but remains constrained by low performance, schema dependence, and weaker F1 than fine-tuned non-LLM models.

  • Discriminative and Unified Modeling: Traditional event extraction uses discriminative sequence labeling with classifiers fine-tuned on human-annotated label sets, often producing separate domain-specific models for each dataset.Recent systems increasingly adopt joint or unified multi-task modeling to learn entities, relations, and events together and improve in-domain performance.
  • LLM-Based Event Extraction: Zero-shot LLM prompting yields consistently low performance and depends heavily on explicitly provided schemas, labels, and output constraints, with weak schema adherence.Instruction fine-tuning can improve schema adherence and reduce format inconsistency and invalid outputs.
  • LLM-Based Event Extraction: LLM-based methods focus on generation correctness rather than representation transfer and remain behind fine-tuned non-LLM models in F1 performance.Constraint-aware instructions and optimization improve output validity but do not resolve the representation-transfer limitation.

3 Approach

The approach unifies event extraction across heterogeneous domains and tasks as domain- and task-conditioned text-to-text generation. It supports both pipeline and end-to-end configurations within one framework, combining flexible operation with improved cross-domain and cross-task robustness.

  • Prompt-Conditioned Generation: Each input is prefixed with a domain indicator and task-specific prompt, providing lightweight signals for domain-specific event semantics.The domain indicator is a simple string such as Geneva or RAMS.
  • Multi-domain Multi-task Learning: The framework jointly learns shared structures across multiple tasks and domains while adapting to heterogeneous event semantics.Training minimizes negative log-likelihood jointly across tasks and domains.
  • Multi-task Formulation: Structured input–target patterns unify pipeline and end-to-end configurations while augmenting training data through heterogeneous patterns.This formulation is designed to improve generalization and robustness across domains and tasks.
  • Pipeline Event Generation: Pipeline extraction sequentially performs trigger identification, type classification, and argument extraction using task-conditioned inputs.The subtasks are labeled ED–P1, ED–P2, and EAE–P.
  • End-to-End Event Generation: End-to-end extraction generates complete event structures in one pass, reducing cascading errors but requiring implicit modeling of role structures and dependencies.The end-to-end configuration is labeled ED+EAE–E2E.

4 Experiments

Experiments use six English event-extraction datasets spanning cybersecurity, FrameNet-aligned, biomedical, multimedia news, news, and Wikipedia domains. The evaluation covers dataset variation, generative-model implementation, deterministic structure recovery, and trigger and argument extraction metrics.

  • Datasets: The experiments use standardized English-only versions of six publicly accessible event-extraction datasets: CASIE, Geneva, Genia2013, M2E2, RAMS, and WikiEvents.M2E2 uses only text-based event and argument annotations.
  • Datasets: The datasets span specialized and diverse settings, including cybersecurity, FrameNet-aligned sources, biomedical literature, multimedia news, cross-sentence news, and document-level Wikipedia events.Their schemas range from limited roles to more than 200 role types, while Geneva and RAMS contain over 100 event types.
  • Datasets: Dataset scale, annotation density, and sequence length vary substantially, with long-document inputs for WikiEvents and CASIE but sentence inputs for Geneva and M2E2.Average lengths are 329.0, 295.2, 225.2, and 132.9 tokens for WikiEvents, CASIE, Genia2013, and RAMS, versus 26.6 and 28.2 for Geneva and M2E2.
  • Models: Preliminary experiments select T5-Base as the base model because it balances performance and computational efficiency, while T5-Large offers negligible improvement for higher cost.The implementation uses T5ForConditionalGeneration with a learning rate of 5e-05, gradient accumulation of 4, and maximum gradient norm 1.0.
  • Post-processing and Evaluation: A deterministic post-processing procedure parses generated sequences into structured triggers, event types, arguments, roles, and offsets for consistent evaluation.Evaluation distinguishes trigger identification and classification from argument identification and classification, requiring correct trigger associations and, for classification, labels or roles.

5 Results

Results show that unified multi-domain training provides competitive cross-dataset performance and improves scalability, recall, and transfer, especially for resource-limited or schema-diverse data. Trigger identification, fine-grained type classification, and argument-matching inconsistencies remain key limitations, while structured intermediate settings outperform fully generative extraction.

  • Misaligned in AI/AC Metrics: Trigger identification errors remain the primary bottleneck, with consistent performance drops from trigger classification to argument extraction across predicted-trigger settings.End-to-end modes also introduce additional spurious events and arguments, while argument metrics can be reduced by mismatches between predicted and gold argument spans.
  • Benefit of Multi-domain Learning: Unified multi-domain training matches or exceeds single-domain performance while using one model across datasets, improving scalability and deployment efficiency.Benefits are strongest on resource-limited or schema-diverse data through transfer from high-resource domains.
  • Benefit of Multi-domain Learning: Multi-domain training substantially improves trigger and argument metrics on Genia2013 and WikiEvents through cross-domain knowledge transfer.Argument classification gains are especially marked on Genia2013, while both trigger and argument metrics increase on WikiEvents.
  • Benefit of Multi-domain Learning: Joint training progressively improves recall while precision stabilizes early, indicating that shared representations primarily enhance coverage, especially in end-to-end mode.The most substantial gains are observed across most domains as training proceeds.
  • ED-pipeline, ED-e2e vs. fully End-to-End: ED-pipeline and ED-e2e outperform fully generative end-to-end extraction, with the largest differences occurring in argument identification and classification.ED-e2e provides comparable or slightly better trigger performance than ED-pipeline while retaining strong argument scores.
  • Fine-grained Event Type Classification Challenges: 77.42% F1 for Level 1 matching versus 31.25% for whole-level matching on RAMS demonstrates the difficulty of fine-grained event type classification.Fine-grained types suffer from data sparsity and require more nuanced semantic distinctions than coarse-grained types.

6 Conclusion

The conclusion presents a unified T5-based Seq2Seq framework that performs event-extraction subtasks across diverse domains, with multi-domain training improving transfer on resource-limited or schema-diverse datasets. It also identifies limitations of end-to-end inference and fine-grained event-type granularity.

  • Contributions: A single T5-based Seq2Seq model performs all event-extraction subtasks across diverse domain datasets.The model uses unified multi-domain, multi-task learning.
  • Results: Multi-domain training yields clear gains on resource-limited or schema-diverse datasets such as Genia2013 and WIKIEVENTS.The conclusion attributes these gains to cross-domain knowledge transfer for improved generalization.
  • Limitations: End-to-end inference lacks explicit structural and role constraints, while scarce supervision and fine-grained event types challenge learning, annotation consistency, and prediction.The conclusion nevertheless characterizes the unified framework as practical and scalable for multi-domain information extraction.

Limitations

The study’s cross-domain evaluation was constrained by the limited availability of event extraction datasets across domains, potentially affecting broader generalizability. It also focused exclusively on event extraction, leaving broader information extraction applications for future work.

  • Dataset and task scope: Cross-domain evaluation was limited by the availability of event extraction datasets from various domains.This constraint may affect the broader generalizability of the findings.
  • Dataset and task scope: The study deliberately restricted its focus to event extraction tasks.The unified architecture is described as extensible in principle to entity and relation extraction.
  • Dataset and task scope: Extending the framework to broader information extraction tasks remains a future research opportunity.The passage specifically identifies entity and relation extraction as potential extensions.

A Comparison of capabilities of different existing EE systems

The comparison covers representative event extraction systems across key sub-tasks and cross-domain generalization. These systems comprise classification-based and generation-based paradigms, while the proposed approach uses a unified Seq2Seq framework to jointly model all event extraction components with explicit cross-domain support.

  • Representative event extraction systems are compared across key sub-tasks and their support for cross-domain generalization.
  • Existing systems span classification-based approaches, including DyGIE++, OneIE, and TagPrime, and generation-based methods, including DEGREE and BartGEN.
  • The proposed approach extends generation-based methods with a unified sequence-to-sequence framework that jointly models all event extraction components and explicitly supports cross-domain transfer.

B Comparisions of Different Model Architectures. · C Zero-shot prompting for ED-e2e.

The paper compares model architectures for event extraction and evaluates GPT-4o-mini in zero-shot ED-e2e prompting. Encoder–decoder models provide stronger structured-output control than GPT-2, while GPT-4o-mini remains inaccurate despite constrained label prompting.

  • B Comparisions of Different Model Architectures.: Encoder–decoder models show more robust trigger detection than decoder-only GPT-2, with larger architectural differences emerging in argument extraction quality.The comparison covers T5-base, BART, and GPT-2 on event extraction.
  • B Comparisions of Different Model Architectures.: BART improves output coherence over GPT-2 but still produces minor inconsistencies in argument alignment and formatting.The paper attributes these residual issues partly to differences in pretraining objectives and less effective control.
  • C Zero-shot prompting for ED-e2e.: GPT-4o-mini is evaluated in zero-shot ED-e2e with a constrained prompt that directly generates trigger-type pairs from a predefined candidate event-label set.The required format is trigger > ⟨trigger_text⟩=> ⟨EventType⟩.
  • C Zero-shot prompting for ED-e2e.: Despite explicit label constraints, GPT-4o-mini struggles with accurate trigger localization and classification in zero-shot ED-e2e extraction.Table 9 reports trigger identification and trigger classification F1 scores across six datasets.

D Benefit of Multi-domain Learning.

Joint multi-domain training produces distinct precision and recall dynamics across datasets: precision stabilizes early, while recall continues improving, especially in challenging and heterogeneous domains.

  • D Benefit of Multi-domain Learning.: Precision stabilizes early and remains largely unchanged throughout training across datasets.This pattern is particularly evident in high-resource or simpler domains such as Geneva.
  • D Benefit of Multi-domain Learning.: Recall improves progressively with training, with the largest gains occurring in challenging and heterogeneous domains.The results indicate effective cross-domain knowledge transfer.

E Granularity Distribution of different Domains.

Figure 5 examines granularity distributions across datasets and shows that their divergence affects trigger classification performance. Greater granularity creates data sparsity for rare event types and more complex semantic decision boundaries, leading to higher error rates and performance gaps.

  • Granularity Distribution: Granularity distributions diverge across datasets, directly affecting model performance in trigger classification.Figure 5 presents each dataset’s granularity distribution.
  • Granularity Distribution: Greater granularity creates data sparsity for rare event types and more complex semantic decision boundaries.These challenges contribute to higher error rates in trigger classification.
  • Granularity Distribution: Figure 5 compares event-type granularity distributions between annotations and predictions across training epochs.The comparison covers different datasets.
Loading 2608.23235v1…