Issues and Challenges for Machine Translation of Gujarati to English Language

Abstract

We propose an algorithmic approach to build a Machine Translation System for Gujarati language, by developing an adequate software support along with manually created Dictionaries/ Corpuses. Focus has been made on building a Paninian Grammar based partial Machine Translation system, where we have used language specific rules, classification algorithms and manually prepared corpuses for statistical analysis wherever required. This approach is based on a layered structure to fulfil a complete Machine Translation of Natural Language from Gujarati to English and vice versa. At each layer, the necessary and sufficient information, namely the word, phrase and sentence level information, is extracted from the input using the software. The software uses manually tagged Corpuses as learning databases. Keeping the primary focus on the translation phase, a given sentence at the source language is translated to target language using the word, phrase and sentence level information collected during the previous phases. Noteworthy results have been attained for different phases of Machine Translation, such as POS (Part Of Speech) Tagger with 96.46% accuracy, Shallow Parser (Chunker) with 96.22% accuracy and Morph Generator/Analyzer with 100% accuracy. Out of various categories of words, such as Noun, Adjective, Verb, Adverb etc, the Morph Analyzer has been built only for Noun words. This essentially helps prove the efficiency of the algorithm. An algorithm for the final phase, i.e. the Parser, has been shown to be effective for simple sentences as is highlighted later in the thesis. The implementation of the algorithm requires manually created verb frames and a full working Morph Analyzer and hence has been excluded in the present work. Notably, the present work is believed to be first of its kind for Machine Translation for Gujarati language. newline newline