New Framework Boosts Trust in Enterprise NLP Systems

Recent research has highlighted the increasing necessity of robust explanations to instill user trust in enterprise Natural Language Processing (NLP) systems. The proposed framework tackles the common obstacles faced in black-box deployments, where traditional representation-based explainers are not applicable. This new evaluation methodology focuses on token-level explanations through leave-one-out occlusion and assesses explanation robustness across various models and datasets, demonstrating significant improvements in stability when utilizing decoder-based LLMs over encoder models.