SignGPT’s Visual Language Toolkit (VLTK) aims to remove fundamental barriers to large scale sign language modelling by developing data-driven, linguistically grounded methods for continuous sign language recognition. We first identify fundamental issues around the ecological validity of potential data sources (e.g. broadcast media with interpreted signing or captions, scraping of social media). We contrast these with the currently highly resource-intensive development of curated sign language corpora based on linguistic principles. The VLTK addresses this scarcity of high quality sign language data by providing semi-automated glossing and other recognition tools, driving large scale corpus expansion without sacrificing linguistic principles. Unlike prior systems that rely on sparse glossing, the project integrates dense temporal annotation, non-manual and non-lexical feature tracking, and transformer-based architectures to capture the multimodal and spatial structure of signing. By aligning machine vision innovation with linguistic insights and community-embedded evaluation, SignGPT establishes a foundation for robust and extensible sign language models.

Sign languages such as British Sign Language (BSL) are visual languages which lack standard writing systems. Annotation of sign language data, especially for the purposes of machine readability, is therefore extremely slow. Tools to help automate and thus speed up the annotation process are very much needed. Here we test the development of one such tool (VIA-SLA), which uses temporal convolutional networks (Renz et al., 2021a, b) for the purpose of segmenting continuous signing in any sign language, and is designed to integrate smoothly with ELAN, the widely used annotation software for analysis of videos of sign language. We compare automatic segmentation by machine with segmentation done by a human, both in terms of time needed and accuracy of segmentation, using samples taken from the BSL Corpus (Schembri et al., 2014). A small sample of four short video files is tested (mean duration 25 seconds). We find that mean accuracy in terms of number and location of segmentations is relatively high, at around 78%. This preliminary test suggests that VIA-SLA promises to be very useful for sign linguists.