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

cs.CLAug 24, 2026

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

Authors: Siting Liang, Omar Adjali, Daniel Sonntag

Organizations: German Research Center for Artificial Intelligence · Carl von Ossietzky University of Oldenburg

Abstract

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.

Figures & tables

Figure 1: Illustration of cross-domain event extraction, where a single unified model generates domain-specific predictions for the same input text by conditioning on the target domain at inference time.

Figure 2: The unified framework is built on a text-to-text architecture, allowing a single model to learn multiple tasks simultaneously through task-specific prompts while training on multiple datasets. During training, domain-conditioned signals (e.g., Domain: Geneva. ) enable robust cross-domain transfer while maintaining strong domain-specific precision.

Task(Mode)	Task-Specific Input Pattern	Target Pattern
ED (P1)	[Domain: $d$ ][Task: TI] $S$	$S$ + <trigger> $w$ >
ED (P2)	[Domain: $d$ ][Task: TC] $S$ + <trigger>	$w$ = $t$
ED (E2E)	[Domain: $d$ ][Task: TI+TC] $S$	<trigger> $w$ >= $t$
EAE (P)	[Domain: $d$ ][Task: EAE] $S$ + ( $w$ , $t$ )	$a_{1}=r_{1}$ ; $a_{2}=r_{2}$
ED+EAE (E2E)	[Domain: $d$ ][Task: EE] $S$	$w=t$ \| $a_{1}=r_{1}$ ; $a_{2}=r_{2}$

Table 1: Abstract input–output patterns for multi-task event extraction under pipeline (P) and end-to-end (E2E) settings. Symbols denote:

w

(trigger word),

t

(event type),

a

(argument text), and

r

(role type).

Dataset	Example Event Type	Example Argument Roles
CASIE (Cybersecurity)	Attack_Databreach	Attacker , Target , Malware , Vulnerability , Compromised Data , Instrument
Geneva (General)	Protest	Protester , Target , Place , Time , Reason , Organizer
Genia2013 (Biomedical)	Gene_expression	Theme (Gene/Protein) , Agent (Cause) , Site (Location) , Resulting Entity
M2E2 (Geopolitical)	Conflict_attack	Location , Victims , Casualties , Time
RAMS (Newswire)	Contact_negotiate_meet	Participant , Location , Time , Topic , Outcome
WikiEvents (Wikipedia)	Personnel_EndPosition	Entity , Organization , Position , Place , Time , Reason

Table 2: Representative event-type examples and their domain-specific argument roles across diverse event-extraction datasets.

Dataset	Split	Event Types	Event Mentions	Role Types	Arguments
CASIE	Train(1047)	5	5980	26	15869
	Dev(218)	5	1221	26	3175
	Test (218)	5	1268	26	3531
Geneva	Train (2582)	115	5290	220	8618
	Dev (509)	115	1016	159	1683
	Test (593)	115	1199	171	2013

Table 3: Statistics of six event extraction datasets applied in our experiments.

Dataset	#Samples	Mean #Tokens	Mean #Annotated Tokens
CASIE	1047	295.2	46.77
Geneva	2582	26.6	14.08
Genia2013	420	225.2	20.92
M2E2	4211	28.2	0.520
RAMS	7287	132.9	5.416
WikiEvents	450	329.0	18.92

Table 4: Overview of token annotation statistics. The first column shows the number of annotated samples, the second column indicates the average number of tokens per sample, and the third column reports the average number of tokens with assigned labels.

Single-Domain Multi-Task Training
	CASIE				M2E2				Genia2013
Metrics	TI	TC	AI	AC	TI	TC	AI	AC	TI	TC	AI	AC
ED-gold	–	95.1	65.1	61.6	–	88.8	58.4	51.5	–	67.9	66.7	51.9
ED-pipeline	67.7	66.2	56.4	52.9	68.2	62.4	53.8	40.4	41.6	33.4	43.8	31.5
ED-e2e	66.2	65.3	57.7	54.2	69.1	65.5	53.2	43.6	45.6	40.5	56.7	43.9
End-to-End	62.7	61.1	39.9	37.3	64.6	59.7	28.0	23.6	41.2	36.3	21.8	19.5

Table 5: Event extraction performance (F1 %) on CASIE, M2E2, and Genia2013 under single-domain and multi-domain multi-task training. The best scores are highlighted for each dataset. Multi-domain training consistently improves performance across datasets for the fully end-to-end mode. We also include comparisons with representative prior single-domain systems.

Single-Domain Multi-Task Training
	RAMS				Geneva				WikiEvents
Metrics	TI	TC	AI	AC	TI	TC	AI	AC	TI	TC	AI	AC
ED-gold	–	37.0	53.5	46.9	–	93.4	83.1	76.8	–	72.3	64.1	58.0
ED-pipeline	79.6	30.7	46.8	34.9	84.6	81.4	72.7	63.7	45.3	38.9	39.8	31.3
ED-e2e	80.2	30.6	46.7	34.8	83.4	81.2	72.1	62.8		38.0	42.3	31.5
End-to-End	79.4	30.1	39.2	29.6	81.3	79.1	65.2	59.8	46.4	35.5	26.0	21.1

Table 6: F1 Scores for Event Extraction Tasks on RAMS, Geneva, WIKIEVENTS Datasets. Scores reported for previous systems correspond only to argument extraction tasks using gold trigger words and types.

Dataset	Level 1	Level 2	Whole
M2E2	65.20	–	63.74
RAMS	77.42	64.23	31.25
WikiEvents	48.67	47.27	44.72

Table 7: Trigger Classification F1 (%) decline by increasing granularity levels.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Model	TI	TC	AI	AC
T5-base	84.6	81.4	72.7	63.7
GPT-2	58.1	48.2	44.8	27.4
BART	82.5	70.3	55.5	55.4
T5-large	85.1	81.9	70.7	59.3

Appendix

Table 8: F1 scores (%) on the Geneva dataset for trigger identification (TI), trigger classification (TC), argument identification (AI), and argument classification (AC).

F1	CASIE	M2E2	Genia2013	RAMS	Geneva	WikiEvents
TI	5.96	40.79	11.57	6.28	5.64	17.58
TC	5.63	34.21	8.76	1.96	1.98	12.68

Appendix

Table 9: Trigger identification (TI) and trigger classification (TC) F1 (%) scores for GPT-4o-mini under zero-shot ED-e2e setting across six datasets.

Model	TI	TC	EAE	Cross-Domain	Paradigm
DEGREE 7	✓	✓	✓	–	Generation
BartGEN 15	–	–	✓	–	Generation
DyGIE++ 23	✓	✓	✓	–	Classification
OneIE 17	✓	✓	✓	–	Classification
TagPrime 8	–	✓	✓	–	Classification (Seq Tagging)
Ours	✓	✓	✓	✓	Generation (Seq2Seq)

Appendix

Table 10: Coverage of event extraction tasks and paradigm for each model.

Model	Prediction
Gold Reference	trigger> pressing>: argument> Iran>=Addressee \| argument> to give up its nuclear program altogether>=Content \| argument> the US leader>=Speaker \| None=Topic
T5 Base	trigger> pressing>: argument> Iran>=Addressee \| argument> to give up its nuclear program altogether>=Content \| argument> the US leader>=Speaker \| None=Topic
BART Base	trigger> pressing>: argument> Iran>=Addressee \| argument> to give up its nuclear program altogether>=Content \| the US leader>=Speaker \| None=Topic
GPT-2	trigger> pressing>: argument> Iran>=Addressee \| argument> to give up its nuclear program altogether>=Content \| argument> Sharon>=Speaker \| None=Topic>=Topic for event Convincing of trigger> pressing>: None=Addressee \| argument> the US leader>=Content \| argument> to keep>=Speaker \| None=Topic>=Topic for event Convincing of trigger> pressing>: None=Addressee \| argument> Iran>=Content \| argument> to give up its nuclear program altogether>=Addressee \| argument

Appendix

Table 11: Event Extraction Predictions: Event with Trigger pressing . GPT-2 struggles due to being decoder-only, producing over-generation, repetition, and label errors. BART is better but still less precise than T5, especially in formatting and alignment with prompts. T5’s encoder–decoder design and text-to-text pretraining allow accurate, structured, prompt-compliant generation, making it the most suitable base model for unified event extraction.

Figure 5: Comparison of event type granularity distributions of different datasets between annotation and predictions across training Epochs.

		Single-Domain		Multi-Domain
Sub-Task	Reference	Input (Single)	Prediction (Single)	Input (Multi)	Prediction (Multi)
TI	When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food trigger>acquisition>.	Domain: geneva. Identify trigger words. When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food acquisition.	When farmers trigger>became> capable of trigger>producing> food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food trigger>acquisition>.	Domain: geneva. Identify trigger words. When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food acquisition.	When farmers trigger>became> capable of trigger>producing> food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food trigger>acquisition>.
ED-pipeline	acquisition=Getting	Domain: geneva. Classify trigger types. When farmers trigger>became> capable of trigger>producing> food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food trigger>acquisition>.	became $=$ Becoming \| producing $=$ Manufacturing \| acquisition $=$ Getting	Domain: geneva. Classify trigger types. When farmers trigger>became> capable of trigger>producing> food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food trigger>acquisition>.	became $=$ Becoming \| producing $=$ Manufacturing \| acquisition $=$ Getting
ED-e2e	trigger>acquisition> =Getting	Domain: geneva. Classify trigger types end-to-end. When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food acquisition.	trigger>became> =Becoming \| trigger>producing> =Manufacturing \| trigger>freed> =Causation \| trigger>acquisition> =Getting	Domain: geneva. Classify trigger types end-to-end. When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food acquisition.	trigger>became> =Becoming \| trigger>producing> =Manufacturing \| trigger>acquisition> =Getting
EAE(ED-pipeline)	trigger>acquisition> =Getting: None=Means \| None=Recipient \| None=Source \| argument>food>=Theme	Domain: geneva. Extract arguments for a specific event. argument> Entity \| Final_category \| Final_quality for trigger> became>=Becoming; argument> Factory \| Instrument \| Producer \| Product \| Resource for trigger> producing>=Manufacturing; argument> Means \| Recipient \| Source \| Theme for trigger>acquisition> =Getting: When farmers became capable of producing food beyond the needs of their own families…	trigger> became>=Becoming: argument> farmers>=Entity \| None=Final_category \| argument> capable of producing food beyond the needs of their own families>=Final_quality; trigger> producing> =Manufacturing: None=Factory \| None=Instrument \| argument> farmers>=Producer \| argument> food beyond the needs of their own families>=Product \| None=Resource; trigger>acquisition> =Getting: None=Means \| None=Recipient \| None=Source \| argument>food>=Theme	Domain: geneva. Extract arguments for a specific event. argument> Entity \| Final_category \| Final_quality for trigger> became>=Becoming; argument> Factory \| Instrument \| Producer \| Product \| Resource for trigger> producing>= Manufacturing; argument> Means \| Recipient \| Source \| Theme for trigger>acquisition> =Getting: When farmers became capable of producing food beyond the needs of their own families…	trigger> became>=Becoming: argument> farmers>=Entity \| None=Final_category \| argument> capable of producing food beyond the needs of their own families>=Final_quality; trigger> producing>=Manufacturing: None=Factory \| None=Instrument \| argument> farmers>=Producer \| argument> food>=Product \| None=Resource; trigger> acquisition>=Getting: None=Means \| None=Recipient \| None=Source \| argument> food>=Theme
End-to-End	trigger>acquisition> =Getting \| argument>food>=Theme	Domain: geneva. Event trigger and argument extraction. When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food acquisition.	trigger>became> =Becoming \| argument>farmers> =Entity \| argument>capable of producing food beyond the needs of their own families> =Final_category $\|\|$ trigger>producing> =Manufacturing \| argument>farmers> =Producer \| argument>food> =Product $\|\|$ trigger>freed>=Causation \| argument>others in their society> =Affected \| argument>to devote themselves to projects other than food acquisition> =Effect $\|\|$ trigger>acquisition> =Getting \| argument>food>=Theme	Domain: geneva. Event trigger and argument extraction. When farmers became capable of producing food beyond the needs of their own families, others in their society were freed to devote themselves to projects other than food acquisition.	trigger>became> =Becoming\| argument>farmers> =Entity \| argument>capable of producing food beyond the needs of their own families> =Final_quality $\|\|$ trigger>producing> =Manufacturing \| argument>farmers> =Producer\| argument>food> =Product $\|\|$ trigger>acquisition> =Getting\|argument>food> =Theme

Appendix

Table 12: Comparison of prediction per sub-task between single-domain and multi-domain models on Geneva.

Event Extraction

See all 15 papers

A Scalable Cross-Domain Event Extraction System via a Unified Generative Training Framework
Aug 24, 2026Siting Liang, Omar Adjali, Omair Shahzad Bhatti +1Event Extraction Document Information Extraction
German Research Center for Artificial Intelligence · Carl von Ossietzky University of Oldenburg
EAGER: Enhancing Generative Event Extraction via Reinforcement Learning with Verifiable Rewards
Sep 24, 2026Omar Adjali, Siting Liang, Omair Shahzad Bhatti +1RL for Language Models Document Information Extraction
German Research Center for Artificial Intelligence (DFKI), Germany · Carl von Ossietzky Universität Oldenburg, Germany
A Multimodal Text- and Graph-Based Approach for Open-Domain Event Extraction from Documents
Apr 23, 2026Praval SharmaEvent Extraction Document Information Extraction
College of Information Science & Technology University of Nebraska Omaha, USA

Aug 24, 2026J/K move · Enter open · S save

A Scalable Cross-Domain Event Extraction System via a Unified Generative Training Framework

Siting Liang, Omar Adjali, Omair Shahzad Bhatti, Daniel Sonntag

German Research Center for Artificial Intelligence · Carl von Ossietzky University of Oldenburg

Event Extraction Document Information Extraction Cross-Domain Generalization

Open paper AskPDF arXiv

Event extraction is fundamental to information extraction. Prior approaches often separate event detection and argument extraction or depend on dataset-specific designs, limiting scalability and cross-domain generalization. We propose a unified generative sequence-to-sequence framework that performs event extraction subtasks jointly and supports both pipeline and end-to-end configurations. We fine-tune pretrained language models on multiple event datasets across diverse domains, enabling a single model to retain domain-specific semantics while generalizing over large and evolving label spaces. We demonstrate these capabilities through a web-based application tailored for researchers and practitioners. The platform supports document upload, schema-aware event extraction, visualization of triggers and arguments, and comparison of different extraction configurations across domains.

Abstract

Event Extraction

Explore similar work

A Scalable Cross-Domain Event Extraction System via a Unified Generative Training Framework

EAGER: Enhancing Generative Event Extraction via Reinforcement Learning with Verifiable Rewards

A Multimodal Text- and Graph-Based Approach for Open-Domain Event Extraction from Documents