Request Correction

Lost in Expression: Diagnosing Systemic Challenges with Non-Manual Generalization in Sign Language Understanding Tasks

Incorporation of non-manual information is one of the most challenging aspects of Sign Language Understanding (SLU), as these features contribute to the semantic, syntactic, and pragmatic structure of signed communication as a critical feature of compositional meaning at sign, phrase and sentence level. Despite their key linguistic role, non-manuals are often an afterthought in SLU model and dataset design, with many recent models still neglecting to implement non-manual analysis or evaluate how articulators beyond the hands are contributing to the model prediction. In this work, we identify and analyze the challenges relating to recognition of non-manuals and generalization of their linguistic roles encountered by SLU models, offering new explanations for failures to properly model non-manual behavior. We perform a case study on the subtasks of Continuous Sign Language Recognition and Sign Language Translation by applying the Uni-Sign model to Isharah-1000, a Saudi Sign Language dataset. Using controlled partitioning and feature attribution, we further analyze model behavior and failure cases. With this work we hope to set the stage for the creation of diagnostic frameworks for generalization of non-manuals.