Apple researchers have introduced DiscoSign, a translation framework designed to preserve meaning across a connected passage rather than translating each sentence in isolation. The work targets text-to-gloss translation for American Sign Language, where a gloss is a written representation of signs rather than a finished signed video.
The system addresses three discourse-level problems. It keeps people and objects associated with consistent spatial locations, handles question-answer clause structures used in signed discourse, and maps the same English concept to a stable ASL sign throughout a passage. Sentence-only systems can miss those relationships even when individual lines look plausible.
DiscoSign uses a modular large-language-model pipeline grounded in linguistic rules. The researchers also created evaluation measures for each type of coherence because standard translation scores mostly compare wording and do not reveal whether references remain consistent across sentences.
Tests on sentence-level and discourse-level datasets found significant gains in spatial consistency and entity tracking while maintaining competitive quality on single-sentence gloss translation. The result is a research framework, not a consumer translator, and gloss output is only one stage of sign-language communication. Still, it provides a concrete way to evaluate whether future systems preserve the structure of a conversation instead of processing disconnected sentences.