Multiple knowledge sources for word recognition

It is not possible to get complete disambiguation of handwriting from pattern recognition alone since a written word can be interpreted in a number of different ways. The combination of several sources of information, each of which is capable of extracting a different characteristic of cursive handwriting, is more likely to be successful than pattern recognition alone. The way to produce a machine system which is both discriminatory and robust (target found even when it is not top ranked) is to combine different, but complementary recognition methods. It is only by integrating different sources of information that a stronger, more robust, machine system can be developed. Three sources of information are considered in the present paper: character-segmentation information, word shape information and lexical information. The methods used to extract these three sources of information are, respectively, a traditional pattern recognizer, a whole word recognizer and a method which uses word level contextual cues.