Zero-Shot Event Extraction (ZSEE)

• Zero-Shot Event Extraction (ZSEE) is a task that extracts event triggers and arguments from unstructured data without prior training on new events by leveraging natural language definitions and schema information.
• Key methodologies include prompt-based question answering, code-generation with schema enforcement, and contrastive embedding alignment to ensure structural consistency.
• Practical implementations show improved performance on benchmarks like ACE05 and MAVEN, paving the way for open-world, multi-modal, and cross-lingual event extraction.

Zero-Shot Event Extraction (ZSEE) is the task of extracting event triggers and arguments from unstructured data for event types—and possibly roles—that the model has never observed during training. By leveraging natural language definitions, schema information, exemplar prompts, or event ontologies, ZSEE aims to generalize beyond closed ontologies, supporting rapid adaptation to novel scenarios without human-labeled data for the new types. This is particularly relevant with the proliferation of LLMs and instruction-driven frameworks, which enable LLMs to map open specifications to structured event outputs in diverse domains, including text, images, and multi-modal sources (Li et al., 22 Dec 2025).

1. Formal Problem Definition and Challenges

ZSEE considers an input document $x$ and a set of event schemas $\mathcal{S}_U$ describing unseen event types, each with argument roles possibly specified in natural language or code. The goal is to produce structured event records:

$$f_0(x|s) = \{(t, \tau, \{ (r_i, a_i) \}_{i=1}^k ) : s \in \mathcal{S}_U \}$$

where $t$ is the event type, $\tau$ is the trigger span, $r_i$ are argument roles, and $a_i$ are the corresponding text spans or objects (Li et al., 22 Dec 2025).

Key challenges include semantic label mismatch (brittleness to new type surface forms or definitions), schema enforcement (filling exactly the required roles), error propagation from prior subtasks, word-sense ambiguity, and computational cost when scaling to large ontologies (Zhang et al., 2022, Cai et al., 2023).

2. Core Methodological Paradigms

ZSEE techniques can be grouped along several axes:

• Prompt-based Generation and Question-Answering: Early work framed event argument extraction as reading comprehension or QA, using role-specific queries to extract arguments in an end-to-end fashion (Du et al., 2020, Feng et al., 2020). This paradigm is extended to both trigger and argument extraction via flexible query templates and zero-shot textual entailment (Sainz et al., 2022, Wang et al., 2021).
• Schema-based and Code-generation Approaches: Recent frameworks encode event schemas as executable class definitions (e.g., Pydantic/BaseModel) and treat extraction as a structured code-generation problem. Multi-agent frameworks, such as Agent-Event-Coder (AEC), decompose the task into specialized agents for retrieval, planning, code generation, and verification, iteratively refining candidate outputs until all schema constraints and structural checks pass (Guo et al., 17 Nov 2025). This paradigm enforces type correctness and the structural integrity of outputs.
• Context-Definition Embedding Alignment: Definition-driven methods project contextualized event mentions and NL definitions into a shared embedding space, optimizing contrastive losses to enable precise type discrimination. ZED aligns span representations with definition sentences, introducing hard negatives and a warming phase to distinguish fine-grained definitions (Zhang et al., 2022).
• Meta-Learning and Prompt-tuned Soft Verbalization: MetaEvent applies a MAML-style meta-learning loop with cloze prompts and trigger-aware soft verbalizers, learning to quickly adapt to tasks with unseen types and few or zero examples (Yue et al., 2023). Contrastive objectives (e.g., MMD) ensure separation between newly sampled event classes.
• Multi-Modal and Cross-Task Pipelines: Methods such as CLIP-Event align structured event graphs across text and images using contrastive and optimal transport losses (Li et al., 2022). Others employ staged generative pipelines to disentangle trigger expansion, contextual disambiguation, modality normalization, and argument role QA, reducing ambiguity and computational load (Cai et al., 2023).
• Template-Driven and Global Constraint Decoding: Systems for zero-shot argument classification build prompt-based candidate role assignments and enforce event schema, cardinality, and uniqueness constraints via integer linear programming during inference, yielding consistent improvements over local-only scoring (Lin et al., 2023).

3. Representation of Schemas, Definitions, and Prompts

Event schemas $S_e$ are typically defined as tuples $\langle e, R_e \rangle$, where $e$ is the event type and $\mathcal{S}_U$0 is a set of roles $\mathcal{S}_U$1. In code-centric paradigms, these are compiled into classes using strong type annotation to enforce argument cardinality and type consistency at runtime (Guo et al., 17 Nov 2025). Definition-based approaches supply paraphrased or ontology-connected NL descriptions, often generated with LLMs to ensure diversity and intra-class semantic coverage (Cai et al., 2024).

Prompting strategies fall into several taxonomies:

• Natural-language event definitions and argument role questions (in QA and NLI-based systems) (Sainz et al., 2022, Feng et al., 2020)
• Cloze prompts with soft verbalizers for trigger-aware classification (Yue et al., 2023)
• Prefix-based or code-generation prompts for schema-driven extraction (Guo et al., 17 Nov 2025)
• Multi-modal templates to align text and image arguments (Li et al., 2022)

4. Verification, Structural Consistency, and Multi-Agent Systems

A core limitation for LLM-driven ZSEE is schema-violation: missing, spurious, or mistyped slots in outputs. Multi-agent frameworks such as AEC orchestrate distinct agent roles (retrieval, planning, coding, verification), with a dedicated verification agent applying deterministic semantic, type, and structural tests to candidate code objects. Failures in validation loop back to code-patching or hypothesis backtracking, producing extraction objects guaranteed to adhere to schema and type requirements (Guo et al., 17 Nov 2025). Code-centric representation enables formal enforcement not only of output structure but also of output serializability (e.g., valid JSON).

Reinforcement learning loops have been proposed for document-level argument extraction, introducing collaboration between generator and evaluator agents, with the latter scoring synthetic documents for semantic and structural fidelity and returning reward signals. Event-structure constraints within the reward function explicitly control for role presence, preventing degenerate outputs (Zhang et al., 3 Mar 2026).

5. Evaluation, Datasets, and Quantitative Benchmarks

ZSEE systems are evaluated predominantly on ACE05, FewEvent, MAVEN, GENIA, WikiEvents, CASIE, SPEED, and document-level datasets (RAMS, WikiEvents splits), with standard micro-averaged metrics: Trigger Identification (TI), Trigger Classification (TC), Argument Identification (AI), Argument Classification (AC), and Span-F1.

Representative results include:

Model	TI / TC (ACE05)	AI / AC	ZS F1 (MAVEN)
DirectEE	50.7 / 46.9	-	-
AEC (ZSEE)	57.0 / 54.6	38.4 / 34.7	-
ZED	-	-	59.37 (ID), 32.96 (ID+C)
MetaEvent	-	-	36.86 (ZS F1)
Clean-LaVe	-	81.2 (ACE05-E+)	-
DivED (LLaMA)	-	-	6–12 F1 pts > GPT-3.5 (Cai et al., 2024)

These methods outperform prior zero-shot baselines, often by 3–16 points depending on metric and dataset (Guo et al., 17 Nov 2025, Yue et al., 2023, Zhang et al., 2022, Cai et al., 2024).

Ablation studies reveal that exemplar retrieval, rationale generation, the verification/patch loop, schema-aware constraints, contrastive and ranking losses, and ontology signals are essential for high-fidelity zero-shot extraction. Removing any of these components results in significant drops in trigger or argument F1 (ranging from 2–8 points per component) (Guo et al., 17 Nov 2025, Lin et al., 2023).

6. Cross-Modal, Cross-Document, and Schema Induction Extensions

ZSEE has been advanced into the multi-modal domain by aligning event graphs extracted from captions and images, leveraging contrastive and graph-based optimal transport objectives. CLIP-Event achieves substantial zero-shot improvements in both event and argument F1 over both text-only and vision-language baselines (Li et al., 2022). Zero-shot schema induction has been formulated as the joint synthesis and extraction of topic-centered event graphs from LLM-generated corpora, with dedicated modules for stepwise timeline and hierarchy construction (Dror et al., 2022).

At the cross-document scale, the introduction of event stores and event-centric memory supports reasoning and aggregation beyond a single context window, addressing fragility in long-horizon LLM extraction (Li et al., 22 Dec 2025).

7. Practical Implications, Open Challenges, and Future Directions

Zero-shot event extraction is advancing from simple span-matching frameworks to structurally consistent, schema-aware, and multi-agent cognitive systems. Limitations include bottlenecks in definition quality, over-reliance on type-specific templates, error propagation in argument extraction, scalability to large ontologies, and performance in highly ambiguous or non-English settings (Zhang et al., 2022, Li et al., 22 Dec 2025).

Key open research directions:

1. Agentic Perception and Event-Centric Memory: EE as a real-time perception and retrieval module for LLM-based agents, supporting episodic memory and temporal/causal reasoning (Li et al., 22 Dec 2025).
2. Neuro-symbolic Decoding: Integration of automata-constrained decoding and verification modules, combining neural flexibility with symbolic guarantees (Guo et al., 17 Nov 2025).
3. Open-World Schema Induction: Interactive schema expansion (e.g., on-the-fly event type induction and user-in-the-loop query) (Dror et al., 2022).
4. Cross-lingual and Multi-modal Adaptation: Adapting ZSEE to multilingual and non-textual inputs, leveraging visual or other modalities (Li et al., 2022, Li et al., 22 Dec 2025).
5. Utility-aware Evaluation: Moving beyond F1 to extrinsic metrics reflecting downstream impact and confidence calibration.

The field is converging toward unified platforms capable of definition-driven, structurally robust, and computationally efficient event extraction for truly open-world and low-resource scenarios, by combining context-rich prompting, schema enforcement, multi-agent collaboration, and definition/ontology-based generalization (Guo et al., 17 Nov 2025, Cai et al., 2024, Li et al., 22 Dec 2025).