SERE: Structural Example Retrieval to Reduce Causal Hallucination in LLMs
Researchers propose SERE, a structural example retrieval framework to enhance Large Language Models (LLMs) in Event Causality Identification (ECI). ECI requires models to determine causal relationships between event pairs, but LLMs often overpredict causality due to biases, a phenomenon termed causal hallucination. SERE leverages few-shot learning by retrieving examples based on three structural metrics: Conceptual Path Metric using edit distance in ConceptNet, Syntactic Metric via tree edit distance on syntactic trees, and Causal Pattern Filtering. The framework aims to improve LLM accuracy by providing relevant structural examples, reducing overprediction. The paper is available on arXiv under identifier 2605.03701.
Key facts
- SERE is a structural example retrieval framework for LLMs in Event Causality Identification.
- ECI requires models to determine causal relationships between event pairs.
- LLMs often overpredict causality, causing causal hallucination.
- SERE uses three structural concepts: Conceptual Path Metric, Syntactic Metric, Causal Pattern Filtering.
- Conceptual Path Metric measures conceptual relationship via edit distance in ConceptNet.
- Syntactic Metric quantifies structural similarity through tree edit distance on syntactic trees.
- Causal Pattern Filtering filters examples based on causal patterns.
- The paper is published on arXiv with ID 2605.03701.
Entities
Institutions
- arXiv