Abstract
Recent progress in Sign Language Translation (SLT) has focussed primarily onimproving the representational capacity of large language models to incorporateSign Language features. This work explores an alternative direction: enhancingthe geometric properties of skeletal representations themselves. We proposeGeo-Sign, a method that leverages the properties of hyperbolic geometry tomodel the hierarchical structure inherent in sign language kinematics. Byprojecting skeletal features derived from Spatio-Temporal Graph ConvolutionalNetworks (ST-GCNs) into the Poincar\'e ball model, we aim to create morediscriminative embeddings, particularly for fine-grained motions like fingerarticulations. We introduce a hyperbolic projection layer, a weighted Fr\'echetmean aggregation scheme, and a geometric contrastive loss operating directly inhyperbolic space. These components are integrated into an end-to-endtranslation framework as a regularisation function, to enhance therepresentations within the language model. This work demonstrates the potentialof hyperbolic geometry to improve skeletal representations for Sign LanguageTranslation, improving on SOTA RGB methods while preserving privacy andimproving computational efficiency. Code available here:https://github.com/ed-fish/geo-sign.