Can You Really Reason: A Novel Framework for Assessing Natural Language Reasoning Datasets and Models

Recent studies have illuminated a pressing issue in the domain of natural language understanding (NLU) and reasoning: many of these datasets are imbued with subtle statistical cues. These cues, often unnoticed, provide sophisticated models an unintended edge, allowing them to exploit these patterns, leading to a potentially misleading overestimation of their genuine capabilities. While the existence of these cues has been noted, a precise and systematic identification has remained elusive in existing literature. Addressing this gap, our paper presents a novel lightweight framework. This framework is meticulously designed to not only detect these hidden biases in multiple-choice NLU datasets but also rigorously evaluate the robustness of models that are developed based on these datasets. By unveiling these biases and assessing model integrity, we aim to pave the way for more genuine and transparent advancements in NLU research.