CauDRE: Causal Intervention with Backdoor Adjustment for Robust Data Relation Extraction

Data Relation Extraction (DRE) is a pivotal task for knowledge graph construction and natural language processing (NLP). However, contextual biases in training data (e.g., frequent co-occurring entity types) often induce spurious correlations. Existing methods primarily rely on statistical patterns of pre-trained language models (PLMs), which fail to capture causal relationships between entities, leading to poor robustness under distribution shifts. To overcome this shortcoming, we introduce CauDRE, a causal relation extraction framework that mitigates spurious correlations via backdoor adjustment. CauDRE consists of two modules: feature extraction and causal intervention. The feature extraction module employs dual-entity context-aware attention to capture statistical correlations between entities and contexts, providing basic semantic representation, but this process does not distinguish causality from spurious associations. To address this gap, the causal intervention module introduces a re-weighted multi-sample strategy based on inverse probability weighting (IPW-MSA), effectively blocking confounding paths and removing spurious correlations. Experiments on ACE 2005 and SemEval-2010 demonstrate that CauDRE achieves state-of-the-art performance (87.84% F1 on ACE 2005 and 91.59% F1 on SemEval-2010), outperforming the RoBERTa baseline by 4.78% and 4.12%, respectively.