English to Nepali Statistical Machine Translation System

Abstract

Machine Translation (MT), perhaps the earliest NLP applications, is the method of translating one human language sentence into another, using computer or any kind of machine. The aim of this research paper is to develop an MT system for Nepali language which can translate an English sentence to its most probable Nepali sentence using Statistical Machine Translation (SMT) approach. The system is implemented using three different tools like MOSES for decoding, GIZA++ for generating translation model and IRSTLM for estimating target model probability. Also for training the system, English-Nepali parallel corpus is used and for testing, English raw corpus is used. Both these two corpora are collected from TDIL (Technology Development for Indian Languages). The system has been manually evaluated using two parameters viz. fluency and adequacy and it gives an average accuracy of 2.7 out of 4 (level no), i.e., approximately 68%. Though the implemented system achieves an accuracy of 68% but for OoV (Out of Vocabulary) words the research still continuing. A small comparison has also been made with exiting English-Nepali MT system.