Fast Few-shot Debugging for NLU Test Suites

TL;DR

This paper introduces a fast few-shot debugging method for transformer-based NLU models that improves accuracy on specific phenomena with minimal impact on overall performance, outperforming existing quick-fix approaches.

Contribution

The paper proposes a novel fast debugging technique that samples in-danger examples from training data, achieving better accuracy trade-offs than existing methods.

Findings

Superior original accuracy at comparable debugging accuracy

Faster debugging process than full retraining

Effective correction of specific phenomena in NLU models

Abstract

We study few-shot debugging of transformer based natural language understanding models, using recently popularized test suites to not just diagnose but correct a problem. Given a few debugging examples of a certain phenomenon, and a held-out test set of the same phenomenon, we aim to maximize accuracy on the phenomenon at a minimal cost of accuracy on the original test set. We examine several methods that are faster than full epoch retraining. We introduce a new fast method, which samples a few in-danger examples from the original training set. Compared to fast methods using parameter distance constraints or Kullback-Leibler divergence, we achieve superior original accuracy for comparable debugging accuracy.