Original title: SignSeek: Learning Transferable Representations for Sign Dictionary Retrieval
Sobhan Asasi, Ozge Mercanoglu Sincan, Richard Bowden · September 2026
Looking up a sign language video in a digital dictionary is tricky. Different people sign the same word in their own personal styles, making it hard for computers to match them. Researchers created SignSeek to fix this problem. SignSeek learns by comparing videos of the same sign performed by different people. It also identifies the single most important body part, or articulator, for that sign. The system then practices recognizing the sign using only that crucial body part. At the same time, it learns to predict that part's hidden details using the surrounding movements. SignSeek set a new record in retrieving signs without any extra training. Remarkably, it successfully recognized completely unseen British Sign Language, beating systems specifically trained on it.
Imagine looking for a specific book in a library where each librarian arranges shelves differently. Normally, you'd know the genre or author, but if the labels change, you're lost. In this paper, the system learns to spot the one consistent detail about each book, like its cover pattern, that doesn't depend on how it's shelved, so it can find the right book no matter the arrangement.
For people learning sign language, finding the right sign in a dictionary is hard because every signer moves differently. This system lets them look up any sign by showing a video, even if the signer is different or the language is one it never trained on, making learning tools more reliable and accessible.
The abstract does not specify how the system performs on continuous sign language, real-time applications, or all sign languages. It focuses on isolated sign retrieval and recognition, and the evaluation is on particular datasets.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
SignSeek: Learning Transferable Representations for Sign Dictionary Retrieval — Sobhan Asasi, Ozge Mercanoglu Sincan, Richard Bowden
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.