[Paper Review] Towards Trustworthy Explanation: On Causal Rationalization

Wenbo Zhang, Tong Wu+3 more|arXiv (Cornell University)|Jun 25, 2023

Topic Modeling4 citations

TL;DR

This paper proposes causal rationalization, a novel framework that leverages structural causal models to identify trustworthy, non-spurious rationales in NLP by enforcing causal desiderata—non-spuriousness and efficiency. It introduces probabilities of causation (e.g., CPNS) to distinguish true causal rationales from spurious, highly correlated alternatives, achieving superior faithfulness and generalization on review and medical datasets compared to association-based methods.

ABSTRACT

With recent advances in natural language processing, rationalization becomes an essential self-explaining diagram to disentangle the black box by selecting a subset of input texts to account for the major variation in prediction. Yet, existing association-based approaches on rationalization cannot identify true rationales when two or more snippets are highly inter-correlated and thus provide a similar contribution to prediction accuracy, so-called spuriousness. To address this limitation, we novelly leverage two causal desiderata, non-spuriousness and efficiency, into rationalization from the causal inference perspective. We formally define a series of probabilities of causation based on a newly proposed structural causal model of rationalization, with its theoretical identification established as the main component of learning necessary and sufficient rationales. The superior performance of the proposed causal rationalization is demonstrated on real-world review and medical datasets with extensive experiments compared to state-of-the-art methods.

Motivation & Objective

• To address the limitation of association-based rationalization methods that fail to distinguish true causal rationales from spurious, highly correlated snippets.
• To formalize causal rationales using two causal desiderata: non-spuriousness (causal relevance) and efficiency (non-redundancy).
• To develop a structural causal model of rationalization that enables theoretical identification of necessary and sufficient rationales via probabilities of causation.
• To improve model trustworthiness in high-stakes NLP applications like healthcare by ensuring rationales reflect true causal mechanisms, not just statistical associations.

Proposed method

• Proposes a structural causal model (SCM) for rationalization, modeling the causal relationships between input snippets, outcomes, and predictors.
• Defines key probabilities of causation, including CPNS (Causal Probability of Necessary and Sufficient causes), derived from counterfactual interventions.
• Uses the ignorability assumption to identify CPNS as the difference in outcome probabilities under intervention: P(Y=y|Zj=zj, Z−j=z−j, X=x) − P(Y=y|Zj≠zj, Z−j=z−j, X=x).
• Applies monotonicity and consistency assumptions to derive and bound the CPNS, ensuring theoretical identifiability of causal rationales.
• Integrates the causal rationale identification into a select-predict framework, replacing standard association-based selectors with causal-aware selectors.
• Employs counterfactual reasoning to evaluate the necessity and sufficiency of each snippet in predicting the outcome, ensuring faithfulness to causal mechanisms.

Experimental results

Research questions

• RQ1 Can causal desiderata—non-spuriousness and efficiency—be formally embedded into rationalization to improve the trustworthiness of explanations in NLP?
• RQ2 How can probabilities of causation be theoretically identified and estimated in a structural causal model of rationalization?
• RQ3 Can causal rationalization outperform association-based methods in identifying faithful, non-redundant rationales when input snippets are highly correlated?
• RQ4 To what extent does causal rationalization improve generalization under distributional shift compared to standard rationalization methods?
• RQ5 How do the proposed causal probabilities (e.g., CPNS) enable the selection of truly sufficient and necessary rationales over spurious alternatives?

Key findings

• The proposed causal rationalization method achieves significantly higher faithfulness and better generalization on both review and medical datasets compared to state-of-the-art association-based baselines.
• On the Yelp and IMDB review datasets, causal rationalization outperforms baselines in F1-score for rationale selection while maintaining high prediction accuracy.
• On the MIMIC-III medical dataset, the method identifies clinically meaningful rationales that align with expert-annotated causal factors in patient outcomes.
• The method demonstrates robustness under distributional shift, maintaining high performance when test data distributions differ from training data, unlike association-based models prone to spurious feature reliance.
• Theoretical analysis confirms that CPNS is identifiable under standard causal assumptions, enabling reliable estimation of causal relevance of input snippets.
• Empirical results show that causal rationalization reduces reliance on spurious, correlated features—e.g., distinguishing true causal causes like 'unpleasant aroma' from correlated but non-causal features like 'palate' in beer reviews.