Integrating segmentation and recognition in on-line cursive handwriting using error-correcting grammars

A grammar-based approach is given to recognizing online cursive handwriting that overcomes the inherent problems of variability in input size and the need to integrate segmentation and recognition. The input stream is represented as a sequence of uniform stroke descriptions that are processed by a mixture of neural-networks, each designed to recognize letters of different sizes. Words are then recognized by a best-first search over the space of all possible segmentations controlled by two grammars, one defining legal letter sequences, the other defining legal segmentations of the input stream. Results demonstrate that the method is effective at learning to recognize a variety of handwriting styles, with the one writer producing word error rates of 3.1% (with a dictionary of 1000 words) and 10.8% (with a dictionary of 10000 words)