DiscoSign is a computational approach for discourse-aware text-to-sign-language gloss translation, developed by researchers from Northeastern University and Gallaudet University, with work done while at Apple. The authors are Vasileios Baltatzis, Mert Inan, Connor Gillis, Raja Kushalnagar, Lorna Quandt, Leah Findlater, and Colin Lea. The system is grounded in linguistic research and built as a modular Large Language Model (LLM)-based translation framework.

Sign language processing has traditionally operated at the sentence level, which omits discourse phenomena central to comprehension. DiscoSign targets three of these phenomena. Spatial coreference resolution keeps entities in consistent spatial locations across a discourse, reflecting how ASL assigns referents to signing space. Question-Answer Clauses (QACs) are treated as pseudocleft structures serving specific discourse functions rather than flattened into ordinary sentence constructions. Concept-gloss consistency maintains stable mappings between English concepts and ASL signs so a concept is not rendered by inconsistent glosses across a text.

Standard translation metrics, the authors argue, cannot capture discourse-level quality. DiscoSign therefore introduces a suite of evaluation metrics designed to assess each discourse dimension addressed by the framework. The metrics target spatial consistency, QAC handling, and concept-gloss consistency separately, making it possible to see which dimension improves without relying only on aggregate translation scores.

Experiments on sentence-level and discourse-level datasets show that discourse-aware processing significantly improves spatial consistency and entity tracking relative to sentence-only translation, while maintaining competitive single-sentence gloss translation quality. The paper positions this as the first systematic framework for discourse-level text-to-sign-language gloss translation, together with a corresponding evaluation methodology.

A related-readings section also points to neighboring work on pseudo-annotation of partially annotated ASL datasets such as ASL STEM Wiki and FLEURS-ASL, and on sign language generation, where current systems face challenges with grammatical structures, facial cues, body language, and visual/motion fidelity.