Parsing Causal Relationships in Social Science Publications

Systematically cataloging causal claims made in the social sciences and parsing them into structured cause-effect relations is vital for accumulating knowledge and testing theories across large bodies of literature. Manual review is infeasible due to the scale of the literature and linguistic ambiguity, while existing automated methods often lack domain specificity or fail to extract specific cause-effect pairs. We introduce a BERT-based multitask deep learning model trained on a novel, domain-specific benchmark dataset of 3,014 annotated sentences from social science papers. The model simultaneously (1) identifies sentences containing causal claims, (2) parses them by extracting cause and effect spans (including mediating variables via a new annotation schema), and (3) links the associated causal pairs. The model achieved a reliability level matching that of human coders (Krippendorff’s α = 0.80) and robust F1 performance across configurations. The model and dataset are available as validated, open-source resources that offer a scalable, transparent, and reproducible alternative to manual annotation. It can be used to automatically catalogue causal claims, thereby contributing to research synthesis and theory development.