Recognising hand-written Japanese sentences

Abstract

This paper makes a case for handwriting recognition compared to other input methods for communication with machines. A comparison is made with voice recognition and keyboard input systems for both western languages and for Japanese. Both single word recognition and whole sentence recognition are considered. A case is made for handwriting recognition for a language with a large character set and many homonyms, such as the Japanese language. For such a language, a fundamental problem exists for both keyboard input and for voice recognition. Both these systems need to convert a phonetic representation into Kanji, and this requires extensive knowledge of the meaning of the text if it is to be automatic. AI research has yet to deliver fast, competent text understanding systems. Consequently, both voice and keyboard input methods need to present the user with alternative choices during recognition, and this makes these methods slow and unnatural. A system is described here which is designed for accurate, fast sentence recognition of both western scripts and Japanese. The system is designed for whole sentence recognition, with the user allowed to write in a natural way. There is considerable flexibility allowed in terms of size and shape of the writing. The distinguishing characteristic of the system, is the use of a unified recognition technique applied to character, word and sentence recognition. This technique is an adaptation of chart parsing, used extensively in natural language processing in AI. Here the technique has been developed to allow weighted multiple hypotheses during recognition. This is important for a system that allows the user to write naturally. This approach to sentence recognition, allows mistakes made during low level processing to be corrected at higher levels. Knowledge of the vocabulary and allowable sentence structures are incorporated in the system in a unified way. A useful additional result of this approach, is the ability to produce a syntactic parse of the sentence recognised. Provisional results are presented for recognition of Japanese Hiragana characters and for English capital letters. The users were given considerable freedom on the style of writing used. The results show recognition rates of over 80% at present, for a variety of users. Improvements in this performance are anticipated when lexical and syntactic modules are added. Further improvements are anticipated by incorporating learning into the system, so that the knowledge base will be tuned for each user.<>