Ein Programm zur Sprachanalyse von indischen Texten mit integriertem OCR-Modul

摘要

The dissertation deals with the development and implementation of two programs for digitalizing (OCR) and tagging Sanskrit texts printed in Devanagari.Images of text are digitalized using neural networks which are trained with the backpropagation algorithm. Additionally knn-classificators based on the current typeface can be activated and trained to improve the result of the recognition process. Depending on the quality of the printed text and the typeset, recognition rates of 93-98% can be obtained, most of the errors resulting from wrong segmentation of the lines of text. Furthermore the possibility (and limitations) of a language based postprocessing of the text are discussed.The automatical analysis of Sanskrit sentences (tagging) is complicated by language immanent problems - e.g. Sandhi (euphonic rules), the size of the vocabulary used and the formation of composite words. To segment sequences of untagged text the program uses a recursive algorithm based on a number of morphological rules (mainly based on Panini), the lexical contents of a large database which had to be build for the program and statistical information from already tagged texts.The main target of the digitalization is the creation of a database of tagged Sanskrit texts which can be searched and analyzed effectively. In a final chapter some simple implementations of more advanced searching algorithms are discussed.