ClimateCause: Complex and Implicit Causal Structures in Climate Reports¶
Conference: ACL 2026 arXiv: 2604.14856 Code: GitHub Area: Causal Inference / Dataset Keywords: Causal Discovery, Climate Change, Implicit Causality, Nested Causality, IPCC Report
TL;DR¶
ClimateCause constructs the first expert-annotated dataset for complex and implicit causal structures in climate reports (874 causal relations), supporting nested causality, multi-event decomposition, correlation direction, and spatiotemporal context annotation. LLM benchmarking shows causal chain reasoning remains a major challenge.
Method¶
Key Designs¶
-
Noun Phrase Reconstruction and Multi-Event Decomposition: Standardizes causes/effects into comparable canonical forms, with Belongs_to and Combined fields distinguishing exemplification from joint action.
-
Implicit and Nested Causality Annotation: Captures causality expressed through semantics rather than explicit triggers (e.g., "anthropogenic greenhouse gas emissions" implicitly contains humans → greenhouse gas emissions).
-
Causal Graph Semantic Complexity Readability Metric: Five complexity dimensions with min-max normalization, measuring the cognitive complexity of causal reasoning rather than surface-level readability.
Key Experimental Results¶
- 57.33% of statements contain semantically complex causal structures
- LLMs perform far worse on causal chain reasoning than correlation inference
- Statement length significantly correlates with causal complexity (\(r=0.590\))
Highlights & Insights¶
- Causal structure readability metric is novel and practically valuable — helps assess report comprehensibility for policymakers
- Nested causality concept transferable to other specialized domains (medical reports, legal documents)
Rating¶
- Novelty: ⭐⭐⭐⭐
- Experimental Thoroughness: ⭐⭐⭐
- Writing Quality: ⭐⭐⭐⭐
- Value: ⭐⭐⭐⭐