Improve Example-Based Machine Translation Quality for Low-Resource Language Using Ontology

Abstract

In this research we propose to use ontology to improve the performance of an EBMT system for low-resource language pair. The EBMT architecture use (CSTs) and unknown word translation mechanism. CSTs consist of a chunk in source-language, a string in target-language, and word alignment information. For unknown word translation, we used WordNet hypernym tree and English-Bengali dictionary. CSTs improved the wide-coverage by 57 points and quality by 48.81 points in human evaluation. Currently 64.29% of the test-set translations by the system were acceptable. The combined solutions of CSTs and unknown words generated 67.85% acceptable translations from the test-set. Unknown words mechanism improved translation quality by 3.56 points in human evaluation.