Unsupervised Learning of Prosodic Boundaries in ASL

In both spoken and sign languages, prosodic cues signal the ends of intonational phrases. Children must somehow learn to associate these cues with phrase boundaries without explicitly being told where those boundaries are. We present two unsupervised statistical models that learn to identify the ends of intonational phrases (I-phrases) in American Sign Language (ASL) based on prosodic cues: a mixture model, and a hidden Markov model. Although neither model is presented with labeled phrase boundaries, both models recover reasonable parameters and achieve performance comparable to models that are trained with labeled boundaries. However, the between-state dependence of the hidden Markov model does not improve the performance. The success of these models sheds light on how infants might learn the prosodic system without explicit instruction