Semantically Structured Causal Systems Knowledge for Causal Question-Answering

ACL ARR 2025 February Submission5488 Authors

Abstract: Language models often require external knowledge for causal reasoning in QA settings and employ public knowledge sources such as ConceptNet. Causality is inherently contextual, requiring models to reason about causal relations within specific situations. However, existing knowledge sources present causal facts as isolated universal triples (e.g., $\langle$lit match; cause-effect; fire$\rangle$) with limited contextual details. As a result, these repositories often fail to capture the causal context necessary for reasoning applications. To address this gap, we introduce CASK-Schema and CASK-Db. Inspired by mechanism theory, CASK-Schema formalizes causal systems and augments causal facts with relevant temporal, influential, and quantitative relations. We then construct CASK-Db, a public causal knowledge base of $\sim$5.4K synthetically enriched causal systems. Our extensive empirical evaluation demonstrates that CASK-Db improves causal QA performance across six tasks in two knowledge augmentation settings: knowledge injection (average improvement of 14\% / 9pp) and retrieval-augmented zero-shot QA (average improvement of 13\% / 6pp).

Paper Type: Long

Research Area: Resources and Evaluation

Research Area Keywords: causal systems, knowledge bases, causal mechanism theory, causal question-answering, causal reasoning

Contribution Types: Model analysis & interpretability, Data resources

Languages Studied: English

Submission Number: 5488