A novel approach to automatically extracting basic units from Chinese sign language

In sign language recognition, using subwords instead of whole signs as basic units scales well with increasing vocabulary size. However, there are no subwords defined in the signs´ lexical forms. How to automatically extract subwords is a challenging issue. In this paper, a novel approach is proposed to automatically extract these subwords from Chinese sign language (CSL). Signs can be broken down into several segments using hidden Markov models in which each state represents one segment. Temporal clustering algorithm is presented to extract subwords from these segments. The 238 subwords are automatically extracted from 5113 signs, and they can be used as the basic units for large vocabulary CSL recognition with good performance.