Accurate, yet inconsistent? Consistency Analysis on Language Understanding Models
Myeongjun Jang, Deuk Sin Kwon, Thomas Lukasiewicz
TL;DR
This work introduces CALUM, a framework to quantify the lower-bound consistency of pretrained language understanding models using semantics-preserving perturbations (REVERSE and SIGNAL) on NLI and STS tasks in English and Korean. It demonstrates that state-of-the-art PLMs exhibit notable inconsistencies for semantically identical inputs, with greater brittleness under input order changes than format changes. A key contribution is showing that multi-task training with semantic textual similarity data improves consistency by about 13% on average, highlighting the value of meaning-focused objectives. The findings underscore the need to evaluate language understanding beyond accuracy and point to semantic data augmentation as a practical path toward more reliable models.
Abstract
Consistency, which refers to the capability of generating the same predictions for semantically similar contexts, is a highly desirable property for a sound language understanding model. Although recent pretrained language models (PLMs) deliver outstanding performance in various downstream tasks, they should exhibit consistent behaviour provided the models truly understand language. In this paper, we propose a simple framework named consistency analysis on language understanding models (CALUM)} to evaluate the model's lower-bound consistency ability. Through experiments, we confirmed that current PLMs are prone to generate inconsistent predictions even for semantically identical inputs. We also observed that multi-task training with paraphrase identification tasks is of benefit to improve consistency, increasing the consistency by 13% on average.