Why Sign Language AI
A signed-language system cannot be designed as a dictionary that replaces each English word with a gesture. The languages, structures, and communication modes involved do not align one-to-one.
Kinesema begins with this mismatch. Its purpose is to explore how language analysis, intermediate representations, review, and future motion work might support more inspectable signed-language technology.
The Problem Is More Than Vocabulary
Signed communication may carry meaning through handshape, orientation, movement, location, timing, facial expression, gaze, and body position. Several of these signals can occur together and can depend on the surrounding discourse or established signing space.
This means that selecting a possible sign for each written word is not enough. A system may also need to represent:
- the intended meaning of the sentence;
- grammatical relationships;
- people, objects, locations, and references;
- information expressed through space or direction;
- non-manual information;
- timing and transitions; and
- decisions that require human review.
The exact form of these features is language- and context-dependent. Kinesema must not treat general signed-language observations as automatic rules for Singapore Sign Language.
Structure Must Precede Presentation
An avatar can make an output look complete even when the underlying language decision is uncertain. Visual fluency is therefore not evidence of linguistic correctness.
Kinesema separates the work into inspectable layers:
- analyse the natural-English input within a bounded contract;
- identify what language evidence and authority would be needed next;
- represent accepted decisions without hiding uncertainty;
- keep linguistic and community review visible;
- plan motion only from sufficiently supported information; and
- treat avatar output as presentation rather than authority.
This separation makes it possible to test one layer without claiming that the entire translation-and-motion system already works.
What Can Go Wrong
A weakly grounded system can produce output that looks plausible while losing meaning, using space inconsistently, omitting relevant non-manual information, or following English structure too closely. A polished interface can make these problems harder for visitors to notice.
The risk is not limited to an awkward animation. Incorrect or overconfident output may mislead learners, distort expectations about signed languages, or be used in situations for which it was never reviewed.
Kinesema therefore needs to show what is implemented, what is inferred, what is experimental, and what remains outside the system’s authority.
Kinesema’s Current Boundary
The language parser is now an internal research tool. The website no longer accepts sentences for analysis or returns generated candidate glosses. Private research output continues to require appropriate human review.
The public open-hand study lets visitors inspect form separately from language research. It is not an accepted sign sequence or a parser-driven signing avatar.
The Parser Research Boundary documents this separation in more detail. The Kinesema Roadmap explains how later language, review, and motion work should remain gated by evidence and explicit public contracts.
Human Review Is Part of the System
Technical checks can verify schemas, deterministic behavior, and expected test cases. They cannot by themselves establish that an SgSL expression is natural, culturally appropriate, or correct for every context.
Relevant language decisions need people with the appropriate linguistic, community, teaching, interpreting, accessibility, or lived expertise. Their review must remain attributable to a defined scope; no single reviewer or project can speak for an entire Deaf community.
Signed-language AI is difficult because it brings language, motion, culture, accessibility, and computation into the same system. Kinesema’s response is not to hide that complexity. It is to make each accepted step easier to inspect, question, and improve.