Most state-of-the-art sign language models are trained on interpreter or isolated vocabulary data, which overlooks the variability that characterizes natural dialogue. However, human communication dynamically adapts to contexts and interlocutors through spatiotemporal changes and articulation style. This specifically manifests itself in educational settings, where novel vocabularies are used by teachers, and students. To address this gap, we collect a motion capture dataset of American Sign Language (ASL) STEM (Science, Technology, Engineering, and Mathematics) dialogue that enables quantitative comparison between dyadic interactive signing, solo signed lecture, and interpreted articles. Using continuous kinematic features, we disentangle dialogue-specific entrainment from individual effort reduction and show spatiotemporal changes across repeated mentions of STEM terms. On average, dialogue signs are 24.6%-44.6% shorter in duration than the isolated signs, and show significant reductions absent in monologue contexts. Finally, we evaluate sign embedding models on their ability to recognize STEM signs and approximate how entrained the participants become over time. Our study bridges linguistic analysis and computational modeling to understand how pragmatics shape sign articulation and its representation in sign language technologies.

The American Sign Language (ASL) lexicon contains heterogeneous sign types, including established signed words, neologisms, ad-hoc depictions of real-world phenomena, and fingerspelled words. Productions within each type exhibit unique phonotactics and substantial variability, potentially confounding models pretrained on downstream ASL tasks. To survey the capacity for representation collapse, we isolate and annotate 14 types of ASL productions as they appear in science, technology, engineering, and math (STEM) educational content. We then probe seven pretrained models’ latent representations of the annotated clips, reporting collapse metrics with respect to domain, sign type, and meaning. Our findings show that out-of-distribution signs are more frequent in STEM content, and that pretrained models lose information about their visual and semantic contents in varing ways. Among the tested models, those which encoded the phonological properties of signs generalized to STEM data more effectively, while English- and gesture-based models can exhibit more brittleness. Future work can design interventions to mitigate the kinds of collapse documented here or extend this survey by evaluating additional models and datasets.