# D-Separation for Causal Self-Explanation

Source: [D-Separation for Causal Self-Explanation](<https://www.sota2.com/research/paper/68fa71141e209131cea4f213>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-01-04T04:36:42.265Z

Authors: Wei Liu, Jun Wang, Haozhao Wang, Ruixuan Li, Zhiying Deng, YuanKai Zhang, Yang Qiu

Published: 2023-09-23T14:23:19.000Z

## Abstract

Rationalization is a self-explaining framework for NLP models. Conventional work typically uses the maximum mutual information (MMI) criterion to find the rationale that is most indicative of the target label. However, this criterion can be influenced by spurious features that correlate with the causal rationale or the target label. Instead of attempting to rectify the issues of the MMI criterion, we propose a novel criterion to uncover the causal rationale, termed the Minimum Conditional Dependence (MCD) criterion, which is grounded on our finding that the non-causal features and the target label are \\emph{d-separated} by the causal rationale. By minimizing the dependence between the unselected parts of the input and the target label conditioned on the selected rationale candidate, all the causes of the label are compelled to be selected. In this study, we employ a simple and practical measure of dependence, specifically the KL-divergence, to validate our proposed MCD criterion. Empirically, we demonstrate that MCD improves the F1 score by up to $13.7\\%$ compared to previous state-of-the-art MMI-based methods. Our code is available at: \\url{https://github.com/jugechengzi/Rationalization-MCD}.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Classification on Toy Synthetic Skew (test)](<https://www.sota2.com/research/sota/classification-on-toy-synthetic-skew-test.md>): Classification; Toy Synthetic Skew (test); F1 Score: 99.93.

- [Rationalization on HotelReview](<https://www.sota2.com/research/sota/rationalization-on-hotelreview.md>): Rationalization; HotelReview; Precision: 47.

- [Classification on Toy (test)](<https://www.sota2.com/research/sota/classification-on-toy-test.md>): Classification; Toy (test); F1 Score: 99.9.

- [Rationalization on correlated BeerAdvocate Appearance](<https://www.sota2.com/research/sota/rationalization-on-correlated-beeradvocate-appearance.md>): Rationalization; correlated BeerAdvocate Appearance; S Score: 29.7.

- [Classification on HateXplain Synthetic Skew (test)](<https://www.sota2.com/research/sota/classification-on-hatexplain-synthetic-skew-test.md>): Classification; HateXplain Synthetic Skew (test); Clf-F1: 70.93.

- [Rationalization on BeerAdvocate correlated Aroma](<https://www.sota2.com/research/sota/rationalization-on-beeradvocate-correlated-aroma.md>): Rationalization; BeerAdvocate correlated Aroma; S Score: 29.6.

- [Rationalization on BeerAdvocate correlated Palate](<https://www.sota2.com/research/sota/rationalization-on-beeradvocate-correlated-palate.md>): Rationalization; BeerAdvocate correlated Palate; S Score: 29.4.

- [Rationalization on Toy (test)](<https://www.sota2.com/research/sota/rationalization-on-toy-test.md>): Rationalization; Toy (test); HI-F1: 65.7.

- [Rationalization on HateXplain (test)](<https://www.sota2.com/research/sota/rationalization-on-hatexplain-test.md>): Rationalization; HateXplain (test); HI-F1: 27.92.

- [Rationalization on HateXplain Synthetic Skew (test)](<https://www.sota2.com/research/sota/rationalization-on-hatexplain-synthetic-skew-test.md>): Rationalization; HateXplain Synthetic Skew (test); HI-F1: 0.1388.

- [Hate Speech Detection on HateXplain (test)](<https://www.sota2.com/research/sota/hate-speech-detection-on-hatexplain-test.md>): Hate Speech Detection; HateXplain (test); Macro F1 Score: 70.37.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2309.13391>)

[PDF](<https://arxiv.org/pdf/2309.13391.pdf>)

[Implementation](<https://github.com/jugechengzi/Rationalization-MCD>)
