What Will it Take to Fix Benchmarking in Natural Language Understanding?

TL;DR

This paper critiques current NLU benchmarks for unreliability and bias, proposing criteria for better evaluation datasets and emphasizing improvements in dataset design, annotation, size, and bias mitigation.

Contribution

It introduces four criteria for effective NLU benchmarks and argues that current adversarial approaches do not address core evaluation issues.

Findings

Most benchmarks fail to meet the proposed criteria.

Adversarial data collection does not solve fundamental evaluation problems.

Improving benchmarks requires better dataset design, annotation, size, and bias handling.

Abstract

Evaluation for many natural language understanding (NLU) tasks is broken: Unreliable and biased systems score so highly on standard benchmarks that there is little room for researchers who develop better systems to demonstrate their improvements. The recent trend to abandon IID benchmarks in favor of adversarially-constructed, out-of-distribution test sets ensures that current models will perform poorly, but ultimately only obscures the abilities that we want our benchmarks to measure. In this position paper, we lay out four criteria that we argue NLU benchmarks should meet. We argue most current benchmarks fail at these criteria, and that adversarial data collection does not meaningfully address the causes of these failures. Instead, restoring a healthy evaluation ecosystem will require significant progress in the design of benchmark datasets, the reliability with which they are annotated, their size, and the ways they handle social bias.