Misleading Failures of Partial-input Baselines

TL;DR

This paper investigates the limitations of partial-input baselines in dataset evaluation, showing they can be misleading and may not detect all artifacts, with implications for dataset verification and creation.

Contribution

It demonstrates that partial-input baselines can fail to identify certain dataset artifacts, challenging their reliability for dataset difficulty assessment.

Findings

Partial-input baselines can be high even when datasets contain artifacts.

Artificial datasets with hidden trivial patterns evade partial-input detection.

A hypothesis-only model can solve 15% of SNLI examples using trivial patterns.

Abstract

Recent work establishes dataset difficulty and removes annotation artifacts via partial-input baselines (e.g., hypothesis-only models for SNLI or question-only models for VQA). When a partial-input baseline gets high accuracy, a dataset is cheatable. However, the converse is not necessarily true: the failure of a partial-input baseline does not mean a dataset is free of artifacts. To illustrate this, we first design artificial datasets which contain trivial patterns in the full input that are undetectable by any partial-input model. Next, we identify such artifacts in the SNLI dataset - a hypothesis-only model augmented with trivial patterns in the premise can solve 15% of the examples that are previously considered "hard". Our work provides a caveat for the use of partial-input baselines for dataset verification and creation.