A Hybrid approach for Machine Translation from Ge’ez to Amharic language

Abstract

Natural Language Processing can be applied in different areas. From these areas, Machine Translation is the one and its concern is to translate one natural language in the form of text or speech into another language. Human translation has positive sides as far as language translation concerned but it has also its own limitations like slowness when translating than machines, correctness and precision of the texts or speech that are being translated, it has some delays in the Process of translation and it is time and cost consuming. To overcome the problem, many studies have been conducted. Our study, Ge’ez to Amharic machine translation using a hybrid approach, is one of these. Hybrid in this case means using the best features of statistical and rule-based machine translation approaches. Even though Ge’ez and Amharic are the Semitic language family, they have a structural difference in sentence construction. To rectify this issue, in this study we proposed a reordering approach in syntax to make the source language to have a similar sentence structure with the target language. During our research, the source and target languages are Ge’ez and Amharic respectively. There is no prior study conducted on this specific title as the researcher knowledge concerned. We start our study by collecting data from different resources. Unfortunately, our data is only from spiritual books since nowadays Ge’ez language is limited in EOTC literatures. After collecting the data and passing through the preprocessing step, we classified it into two data sets of training and testing. Reordering rules are drafted and applied on the data before classifying. Since our developed machine translation system is unidirectional and the target language is Amharic, we built our language model on it. Translation model which works on probability to generate a target language sentence from a given source language sentence also built and decoder is used to search the best sequence of translation probability. Finally, we conducted four experiments with two different approaches and evaluate the results obtained accordingly. The first and second experiments are performed by the statistical approach by changing the percent of training and testing data then we get a BLEU score of 7.36% and 7.15%. The third and fourth experiments are carried out by hybrid approach in a similar fashion and we get a BLEU score result of 18.62% and 17.38%. Thus, from these we conclude that using a hybrid approach by combining statistical with rule-based machine translation approaches provides a better result for machine translation from Ge’ez to Amharic language.