Bangla-German Language Translation Using GRU Neural Networks

Abstract

Machine translation relates to highly autonomous software which is capable of translating source sentences into different languages. Previously some work was done in this sector where the result was comparatively low. Most of the researchers worked on common languages and none of them gave satisfactory Bilingual Evaluation Understudy (BLEU) score. Depending on these factors, we build a system of Bangla-German translator. This system can be used in various areas (i.e. reliable interpreters, business conduction, e-commerce merchandising, etc.). The system is built based on Gated Recurrent Unit (GRU) which is a gating mechanism of Recurrent Neural Network (RNN). Here, total five types of different RNN algorithms were used like Simple RNN, RNN with Embedding, Encoder-Decoder RNN, Bidirectional RNN, Hybrid RNN. All of them gave good accuracy. But the best result we got from the Hybrid model which was the combination of Embedded and Bidirectional algorithm. The accuracy was 85.69%. For further evaluation, BLEU score was used. The result of BLEU score of unigram to four gram was respectively increasing from 54.40% to 85.88%. Also the comparison between machine translated sentences and Google translated sentences showed that the system works very efficiently.