A HYBRID MACHINE TRANSLATION SYSTEM FOR ENGLISH TO WOLAYTTA LANGUAGE

Abstract

The process of converting different documents from one natural language to another natural language by using the language and capability of a computer is a promising point of MT. MT is one of the basic tasks under computational linguistics that helps for automatic translation of resources by using different approaches to translation. Those approaches may include rule-based, example-based, statistical, neural, and hybrid. The main concern with the concept of hybridization in MT is the integration of two or more MT approaches to implement the translation process. In this study, we used the hybrid (statistical and rule-based) approach for the translation of English to the Wolaytta language. The statistical approach works based on the probability of words in the given sentences, and the rule-based approach works with different linguistic and translation rules to produce fluent output. Quality of translation is one of the big issues in MT. Rearrangement of words in SL sentences, to make them in the form of the TL, is used as a reordering rule for the rule part of our study. The design part of the study begins with the architecture of the English to Wolaytta HMT approach. We divided the collected parallel corpora for training and testing purposes. After pre-processing of the corpus, POS tagging is applied to the collected corpus to identify word classes for the reordering of words for hybridization. We performed the implementation of local reordering for source language sentences after POS tagging. We built the translation model for the SL and the language model for the TL by using the IRSTLM language modeling tool. Researchers used Moses decoder for the decoding process to see the translation output. Finally, BLEU score is used for testing. We collected the parallel corpus from three different sources. We implemented two basic experiments to determine the final result of the study. The first experiment is to determine the quality of baseline translation for the SMT approach and recorded 2.37 of BLEU score. Whereas, the second experiment used locally reordered sentences for the part of a hybrid approach and BLEU score is 3.93. Based on the experimental result of BLEU, a translation with reordering of words as a reordering rule has better translation than the statistical approach with +1.56 of BLEU score