Original title: DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation
Vasileios Baltatzis, Mert Inan, Connor Gillis, Raja Kushalnagar +3 · September 2026
Most translation tools for sign language look at only one sentence at a time. This causes mistakes across longer conversations, where meaning depends on earlier context. To solve this, researchers built DiscoSign. It is a modular system that translates written English into sign language glosses, which are written labels representing signs. It tracks three major features: placing things in consistent physical spaces, using question-and-answer structures that organize ideas, and keeping the same signs matched to the same English concepts. Because regular tests miss these connections, the team also created new measurements to check discourse quality. Tests showed the system significantly improved spatial consistency and entity tracking across longer passages, while keeping single-sentence translation quality competitive.
Imagine you're telling a friend about a busy kitchen, where you first say 'the flour is on the left counter' and later say 'grab the sugar from the right shelf.' If you only describe each action separately without remembering the layout, your friend might get confused about what's where. In this paper, the system works like keeping a mental map of the whole kitchen while describing each step, so when translating to sign language, it consistently points to the same spots for items, making the instructions clear and connected.
Sign language users rely on spatial and contextual cues for clear communication, but current translation systems often ignore these, leading to disjointed output. This approach improves coherence, making digital content more accessible for deaf communities.
The abstract does not specify the full scope of datasets or languages tested, focusing on American Sign Language gloss translation. It addresses only three discourse phenomena, so it may not cover all aspects of discourse coherence. Additionally, while new evaluation metrics are introduced, their effectiveness or generalizability is not established in the abstract.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation — Vasileios Baltatzis, Mert Inan, Connor Gillis et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.