Machine-learning errors in Hong Kong Sign Language handshape recognition reflect markedness patterns attested in learning

Handshapes are a fundamental component of sign language phonology yet pose great difficulty for learners, who acquire handshapes according to physiological complexity, phonological features, and motor skills. Advances in machine-learning have led to handshape recognition models yielding high accuracy. Such models are designed for data processing, and it is unclear whether any errors made during training resemble those of humans. We present a handshape recognition model based on Hong Kong Sign Language, which was trained using input data equivalent to the average duration of study reported for beginner hearing adult learners (3 months, 968 signs, 62 handshapes). Unlike other models, ours was fed all frames surrounding each sign’s peak, analogous to human perception, yielding an accuracy of 62%. The error output and target output exhibit phonetic similarities. These errors resemble the production errors commonly made by human learners, which can be attributed to markedness. This indicates that markedness structure can also be inferred from sign perception.