Pose Similarity Distance: A Novel Method of Estimating Sign Similarity, Ariel Goldberg (Tufts University)
The formal similarity of items in the mental lexicon plays an important role in language processing. Psycholinguistic and phonetic research indicates that segmental similarity—quantified by measures like string edit distance—is likely the most important dimension of similarity for spoken languages.
Unfortunately, there is currently no similar consensus for what the most important dimensions of similarity are for signed languages. Recent proposals have cast sign similarity in terms of shared ‘parameters’—discrete codings of the contrastive features characterizing a sign’s location, handshape, and movement (e.g., Caselli et al. 2021, Martinez dl Rio et al. 2022). While they have met with some success, these proposals do not capture some important behavioral patterns and are enormously time-consuming to implement.
I introduce Pose Similarity Distance (PoseSD), an efficient and automated procedure that uses machine vision and signal processing techniques to estimate the pairwise similarity of signs from reference videos. Analyses indicate that PoseSD predicts many measures of sign processing—accuracy, various phonetic properties, and judgments of wellformedness—on par or better than parameter-based measures. In this talk I consider how discrete/contrastive and continuous/non-contrastive dimensions of similarity may function in theories of processing and grammar.