Language Models Are Poor Learners of Directional Inference

TL;DR

This paper critically evaluates language models' ability to understand directional inferences, revealing their limitations and introducing a new multilingual benchmark to better assess this aspect of natural language understanding.

Contribution

It highlights the inadequacy of current datasets for testing directional inference and introduces BoOQA, a robust benchmark for evaluating language models on this task.

Findings

Language models perform poorly on directional inference tasks.

Existing datasets are flawed and can be exploited by artifacts.

BoOQA provides a more reliable evaluation framework.

Abstract

We examine LMs' competence of directional predicate entailments by supervised fine-tuning with prompts. Our analysis shows that contrary to their apparent success on standard NLI, LMs show limited ability to learn such directional inference; moreover, existing datasets fail to test directionality, and/or are infested by artefacts that can be learnt as proxy for entailments, yielding over-optimistic results. In response, we present BoOQA (Boolean Open QA), a robust multi-lingual evaluation benchmark for directional predicate entailments, extrinsic to existing training sets. On BoOQA, we establish baselines and show evidence of existing LM-prompting models being incompetent directional entailment learners, in contrast to entailment graphs, however limited by sparsity.