Causal Explanations as Regularizers: Bridging Human Rationales and Latent Causal Variables for OOD Generalization

While ER-Test (Joshi et al., 2022) shows that explanation regularization (ER) can boost OOD generalization for language models, and Wu et al. (2024) demonstrate the power of causal intervention in graphs, there’s little work integrating human rationales directly as causal regularizers. The idea here is to develop a framework where human-provided explanations are treated as evidence about likely causal mechanisms, which are then used to softly constrain the model’s learned representations via regularization. This approach contrasts with existing methods that either use ER as a “black-box” alignment or causal inference with synthetic interventions. The novelty is in operationalizing rationales as partial, noisy causal labels and dynamically weighting their influence based on their estimated reliability and informativeness. This could lead to models that are not only more robust to OOD shifts but whose internal reasoning processes are more transparent and interpretable. If successful, this could bridge the gap between explainability and causal robustness, providing actionable insights into when and how explanations truly improve generalization.

References:

1. Graph Out-of-Distribution Generalization via Causal Intervention. Qitian Wu, Fan Nie, Chenxiao Yang, Tianyi Bao, Junchi Yan (2024). The Web Conference.
2. ER-Test: Evaluating Explanation Regularization Methods for Language Models. Brihi Joshi, Aaron Chan, Ziyi Liu, Shaoliang Nie, Maziar Sanjabi, Hamed Firooz, Xiang Ren (2022). Conference on Empirical Methods in Natural Language Processing.