Rationalizing Neural Predictions

TL;DR

This paper introduces a modular approach to extract justifications from input text for predictions, improving interpretability without using rationales during training, and demonstrates its effectiveness on sentiment analysis and question retrieval tasks.

Contribution

The paper proposes a novel generator-encoder framework for extracting rationales that are sufficient for prediction, trained with regularization instead of explicit rationale annotations.

Findings

Outperforms attention-based baselines in sentiment analysis

Effective on multi-aspect sentiment analysis and question retrieval

Rationales are coherent, short, and sufficient for accurate predictions

Abstract

Prediction without justification has limited applicability. As a remedy, we learn to extract pieces of input text as justifications -- rationales -- that are tailored to be short and coherent, yet sufficient for making the same prediction. Our approach combines two modular components, generator and encoder, which are trained to operate well together. The generator specifies a distribution over text fragments as candidate rationales and these are passed through the encoder for prediction. Rationales are never given during training. Instead, the model is regularized by desiderata for rationales. We evaluate the approach on multi-aspect sentiment analysis against manually annotated test cases. Our approach outperforms attention-based baseline by a significant margin. We also successfully illustrate the method on the question retrieval task.