Distribution Matching for Rationalization

TL;DR

This paper introduces a distribution matching approach for rationalization in text classification, ensuring rationales share similar feature and output distributions with input text, leading to improved performance.

Contribution

It presents a novel distribution matching method that considers both feature and output spaces, addressing limitations of previous mutual information-based approaches.

Findings

Outperforms previous rationalization methods significantly

Effective in aligning rationale and input text distributions

Enhances interpretability and prediction accuracy

Abstract

The task of rationalization aims to extract pieces of input text as rationales to justify neural network predictions on text classification tasks. By definition, rationales represent key text pieces used for prediction and thus should have similar classification feature distribution compared to the original input text. However, previous methods mainly focused on maximizing the mutual information between rationales and labels while neglecting the relationship between rationales and input text. To address this issue, we propose a novel rationalization method that matches the distributions of rationales and input text in both the feature space and output space. Empirically, the proposed distribution matching approach consistently outperforms previous methods by a large margin. Our data and code are available.