Use of Lexical and Syntactic Techniques in Recognizing Handwritten Text

USE OF LEXICAL AND SYNTACTIC TECHNIQUES IN RECOGNIZING HANDWRITTEN TEXT Rohini K. Srihari Center for Document Analysis and Recognition (CEDAR) SUNY at Buffalo Buffalo, NY 14228-2567 ABSTRACT The output of handwritten word recognizers (Wit) tends to be very noisy due to various factors. In order to compensate for this behaviour, several choices of the WR must be ini- tially considered. In the case of handwritten sentence/phrase recognition, linguistic constraints may be applied in order to improve the results of the Wit. This paper discusses two statistical methods of applying linguistic constraints to the output of an Wit on input consisting of sentences/phrases. The first is based on collocations and can be used to prOmote lower ranked word choices or to propose new words. The second is a Markov model of syntax and is based on syn- tactic categories (tags) associated with words. In each case, we show the improvement in the word recognition rate as a result of applying these constraints. 1. INTRODUCTION This paper focuses on the use of human language models in performing handwriting recognition. Systems that recog- nize handwriting are referred to as off-line or on-line sys- tems, depending on whether ordinary handwriting on paper is scanned and digitized or a special stylus and a pressure- sensitive tablet are used. The central component of a hand- written text recognizer is a word recognizer (Wit) which takes as input, a word signal and a lexicon. Its output consists of an ordered list of the best n words in the lexicon which match the word signal. Due to wide variability in writing, WRs of- ten do not return the correct word as the top choice and get worse as the lexicon size increases. Furthermore, the correct word may not even be present in the top n choices. This is illustrated in Figure 1 which shows the output of an actual word recognizer (offiine) on isolated word images. imately 200 words on the average. In the second stage, the word-image is segmented into several components; physical features of each component lead to a set of character choices for each segment thus resulting in a set of candidate words. All candidate words which are in the lexicon are returned as the direct recognition output of the Wit. In case none of the words are found in the lexicon (,~ 62% of the time), string matching (the third stage) is performed. Since the training phase (of the language module) requires the processing of several thousand sentences, the computa- tionally expensive procedure of digitizing followed by recog- nition is avoided by employing a program which simulates the output of an actual WR. Based on the intermediate results of the actual word recognizer, we have computed statistics which model the behavi'our of the second stage 1 . These in- clude substitution, splitting and merging statistics. Given an input (ASCII) word, and the above statistics, candidate (corrupted) words are generated based on simulating and pro- pogating each of the above three types of errors at …

Content truncated to 500 words