DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning

Abstract: Zero-shot Event Detection (ED), the task of identifying event mentions in natural language text without any training data, is critical for document understanding in specialized domains. Understanding the complex event ontology, extracting domain-specific triggers from the passage, and structuring them appropriately overloads and limits the utility of LLMs for zero-shot ED. To this end, we propose DiCoRe, a divergent-convergent reasoning framework that decouples the task of ED using Dreamer and Grounder. Dreamer encourages divergent reasoning through open-ended event discovery, which helps to boost event coverage. Conversely, Grounder introduces convergent reasoning to align the free-form predictions with the task-specific instructions using finite-state machine guided constrained decoding. Additionally, an LLM-Judge verifies the final outputs to ensure high precision. Through extensive experiments on six datasets across five domains and nine LLMs, we demonstrate how DiCoRe consistently outperforms prior zero-shot, transfer-learning, and reasoning baselines, achieving 4-7% average F1 gains over the best baseline -- establishing DiCoRe as a strong zero-shot ED framework.

• The paper introduces a novel divergent-convergent LLM reasoning framework that splits event detection into an exploratory Dreamer phase and a precision-driven Grounder phase for enhanced zero-shot performance.
• It employs the Dreamer component to maximize recall by generating diverse event hypotheses and the Grounder component to ensure precision by enforcing task-specific structural constraints.
• Experimental validations across six datasets from domains like biomedicine and cybersecurity demonstrate 4-7% F1 gains and dramatic token usage reductions compared to existing chain-of-thought approaches.

An Overview of "DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning"

The paper "DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning" presents a novel framework for zero-shot Event Detection (ED) utilizing LLMs without training data. It introduces a two-part reasoning framework, namely Dreamer and Grounder, coupled with an LLM-Judge component to significantly enhance event detection performance. This model addresses the cognitive complexity and constraints that hamper the utility of LLMs in understanding and processing complex event ontologies in specialized domains.

Methodological Contribution

Event Detection is vital for applications ranging from news monitoring to epidemic forecasting, yet traditional systems require large amounts of annotated domain-specific data. The authors propose a divergent-convergent reasoning approach called DICORE, which seeks to decouple ED tasks into manageable components by employing a strategic framework:

1. Dreamer: This component encourages divergent reasoning through an unconstrained and exploratory approach to event detection, aiming to maximize event coverage by reducing task-specific constraints. Dreamer, thus, enhances recall by supporting the identification of diverse potential events in the text.
2. Grounder: Working as a counterbalance, Grounder introduces convergent reasoning, ensuring that predictions align with a predefined event ontology. This component uses a finite-state machine (FSM) to enforce structural constraints and task-specific alignment, supporting precision by confining generations to expected outputs.
3. LLM-Judge: A verification step that ensures high precision by checking the grounded predictions against task instructions, filtering out irrelevant predictions that do not conform to the specified ontology.

DICORE's pipeline thus effectively splits cognitive load across its different components, facilitating LLM's more robust zero-shot reasoning capabilities.

Experimental Validation

The authors conducted extensive experiments involving nine LLMs across six datasets from domains including biomedicine, epidemiology, and cybersecurity. Results indicate that DICORE exhibits significant improvements over existing zero-shot and transfer-learning baselines, achieving 4-7% average F1 score gains. Moreover, DICORE demonstrates a 1-2% F1 increment while reducing inference token usage by a factor of 15-55x when compared with other reasoning-based models such as chain-of-thought (CoT) approaches.

DICORE's superior performance without any model fine-tuning further underscores its robust framework design, offering strong zero-shot ED capabilities even for complex event extraction tasks with large and closed event ontologies.

Theoretical and Practical Implications

Theoretically, the divergent-convergent reasoning framework offers a pathway for tackling complex reasoning tasks by decoupling components to focus on exploratory breadth and aligned precision independently. Practically, DICORE encourages the development of adaptable LLM frameworks in scenarios with limited annotated data, maximizing their applicability across diverse domains.

Additionally, this work suggests future research in exploring similar divergent-convergent paradigms across other information extraction tasks and domains, possibly forming the foundational blocks for robust multi-domain LLM applications. Given the growing interest in zero-shot and few-shot learning paradigms, DICORE sets a precedent for efficient task-targeted model design that balances exploratory capabilities and strict adherence to task-specific constraints.

In summary, "DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning" contributes substantially to enhancing zero-shot LLM application in event detection by strategically leveraging the cognitive potential of models through a well-engineered divergent-convergent framework, supported by empirical evidence demonstrating its effectiveness.