Regarding Machine Translation

HI, I’ve been trying to implement the Hindi-to-English model same as the English to hindi model provided in the Deep learning Ciriculum. (Encoder-Decoder using Attention) and I got a good loss. However when i tried to predict the english word for hindi word its giving wrong predicition and also the predicted english word is same for the different hindi inputs. Accuracy is alos very low despite of the good loss value.
Please help me solve this issue