Reproduce “An Empirical Comparison of Instance Attribution Methods for NLP” on OpenResearch

About this paper

Pouya Pezeshkpour, Sarthak Jain, Byron C. Wallace, Sameer Singh

An empirical comparison of instance attribution methods for Natural Language Processing (NLP) models, conducted by researchers from the University of California, Irvine and Northeastern University, reveals that computationally efficient, similarity-based approaches are surprisingly competitive with, and often more effective than, complex gradient-based methods in tasks like detecting data artifacts. The study provides practical guidance for selecting scalable interpretability tools for deep learning models.

Key insights

• The substantial computational expense of state-of-the-art gradient-based methods, like full Influence Functions, largely limits their practical application for large deep NLP models.
• Simpler, similarity-based instance attribution methods, particularly those leveraging learned representations, offer competitive and sometimes superior performance in practical interpretability tasks, such as identifying data artifacts.
• While gradient-based methods without normalization can be ineffective, incorporating normalization significantly improves their ability to surface diverse and meaningful influential training examples.

Summary by alphaXiv.