Rationalizing Transformer Predictions via End-To-End Differentiable Self-Training

TL;DR

This paper introduces an end-to-end differentiable training method for rationalized transformer classifiers, enabling simultaneous classification and token relevance scoring, resulting in improved alignment with human annotations.

Contribution

It simplifies the rationalization process by using a single model for all roles, enhancing training stability and efficiency, and extends to produce class-wise rationales with state-of-the-art alignment.

Findings

Achieves stable end-to-end training of rationalized transformers.

Produces class-wise rationales with improved human annotation alignment.

Outperforms previous methods in rationale quality and training stability.

Abstract

We propose an end-to-end differentiable training paradigm for stable training of a rationalized transformer classifier. Our approach results in a single model that simultaneously classifies a sample and scores input tokens based on their relevance to the classification. To this end, we build on the widely-used three-player-game for training rationalized models, which typically relies on training a rationale selector, a classifier and a complement classifier. We simplify this approach by making a single model fulfill all three roles, leading to a more efficient training paradigm that is not susceptible to the common training instabilities that plague existing approaches. Further, we extend this paradigm to produce class-wise rationales while incorporating recent advances in parameterizing and regularizing the resulting rationales, thus leading to substantially improved and state-of-the-art alignment with human annotations without any explicit supervision.