“Machine-learning errors in Hong Kong Sign Language handshape recognition reflect markedness patterns attested in learning” published in Sign Language & Linguistics

“Machine-learning errors in Hong Kong Sign Language handshape recognition reflect markedness patterns attested in learning” published in Sign Language & Linguistics

We report on a new publication by Frank, Aaron, Arthur, Jeff and Youngah, recently published in Sign Language & Linguistics. The paper, “Machine-learning errors in Hong Kong Sign Language handshape recognition reflect markedness patterns attested in learning”, is now available online.

This study examines how errors made by a machine-learning model trained to recognise handshapes in Hong Kong Sign Language compare with the kinds of errors produced by human learners. Handshapes are a fundamental part of sign language structure (often referred to as phonology), but they are also one of the most challenging components for learners to acquire, due to factors such as physical complexity, coordination, and underlying linguistic features.

The authors developed a handshape recognition model trained on data equivalent to what a beginner hearing adult learner might encounter over approximately three months of study. The dataset included 968 signs and 62 distinct handshapes. Unlike many existing systems that focus only on isolated moments, this model processed the sequence of visual frames surrounding the most salient part of each sign, which more closely reflects how humans perceive signing in real time. The model achieved an overall accuracy of 62 percent.

The findings suggest that machine-learning systems trained under human-like conditions can reveal patterns that align with human learning, indicating that markedness structures may also be reflected in perceptual processes, not only in production. This contributes to ongoing discussions at the intersection of sign language linguistics, learning theory, and artificial intelligence.

Tan, F. L. H., Chik, A. W. C., Thompson, A. L., Yip, J. W. T., & Do, Y. (2026). Machine-learning errors in Hong Kong Sign Language handshape recognition reflect markedness patterns attested in learning. Sign Language & Linguistics. open_in_newDOI