EVADE: LLM-Based Explanation Generation and Validation for Error Detection in NLI

EVADE represents a significant advancement in the field of natural language processing by addressing the complexities of human label variation in datasets used for training NLP models. Traditional methods, such as the VARIERR framework, require costly two-round manual annotations, which can limit the diversity of plausible labels. In contrast, EVADE leverages the capabilities of large language models (LLMs) to generate and validate explanations, effectively detecting errors in natural language inference (NLI). The study demonstrates that LLM validation not only refines the distribution of generated explanations but also aligns them more closely with human annotations. This alignment is crucial as it enhances the quality of the training data, resulting in improved fine-tuning performance for NLP models. The implications of this research are profound, as it suggests a more efficient and effective method for ensuring high-quality datasets, which are essential for developing reliable NLP s…