Using contextual information to improve performance of character recognition machines
Rajjan Shinghal · eScholarship@McGill (McGill) · 1977
Ten la,rge cox;pora, each written on a different subject 1 matter"were compiled.,Various statistics were estimated from the corpora and analyses comparing the corpora were carried out.A textual data-base, consi~ting of machine-• printed characters taken from Ryan' 8 data, was 'compiled.The'perform~nce of contextual and non-contextual algorithrns in 1 text recognition was compared.Different n-gram statistics and Markov assulnptions were used t,o 'study the performance of the modified Vi terbi algori thm and to find the optimum depth .of sea,rch., Word-length-,and-position-independent n-gram pro-1) babilitiés estimated from different sources were used in 'classifying texts wri tten on specific subject matters.A dictionary method ~~d a propos~d hybrid algorithm --which was a function of the depth of search and the fraction of 1 the dictionary used --were extensi vely c~mpared wi th each other and wi th the modified Viterbi algorithme -',' '1 .' .' ' .' .. ""' i / ' .