Using stochastic grammars for modelling and recognising cursive script

Presents a new approach to cursive script recognition which combines syntactic pattern recognition with neural networks. This has a number of advantages over previous methods. First, the ability to infer complex internal representations which are abstracted from time allows the input data to be modelled efficiently. Second, the fact that no segmentation is required should lead to robust performance, though we have not yet tested this on a large cursive script database. Third, due to the highly structured way in which the data is modelled it is quite feasible to recognise a small vocabulary with a non-stochastic temporal connectionist parser TCP. This requires only AND gates, OR gates and shift registers, and is therefore easy to implement in silicon. Finally, the TCP's ability to learn in an unsupervised fashion is interesting both in terms of machine-learning theory, and in practical terms, since it reduces the need for manual labelling in large training databases