This repository is made as experiments with Natural Language Processing and its techniques, demonstrating a basic application of translation from English to Mandarin(Chinese) language. The model demonstrated here is not perfect in translation between the two languages and still has multiple scope of improvement, however its performance is one of the best among the algorithms used in NLP for translations and is based upon Sequence-to-Sequence ecoder-decoder model.
The encoders tried are Seq2Seq, Seq2Seq with added Attention(Linear, Bilinear, Dot Product) and the one shown in the notebooks is a pure SelfStackedAttention encoder having Dot Product Attention The loss at the end of 10 epochs in the current model is around 1.9 ~ 2.1 approx which is still has scope of improvement in the current model script only. Next we tried for 70 epochs and 50 epochs in files train_v02.1.ipynb and train_v02.1.1.ipynb, but still accuracy did not go beyond a certain limit with the current dataset and algorithm we have.
Next thing that seems can minimize the loss is, to make the preprocessed dataset to be more efficient, as we can make all the dataset converted to simplified Chinese script before passing through training module. Soon I will make the update too.
Will also try validating them against Google translate, to see if we can move ahead. The repository will be kept updated with new content and improvised versions of Eng-Mandarin translator, as soon as some progress is made.
Datasets have been obtained from : https://tatoeba.org/eng/downloads
Credits: http://www.realworldnlpbook.com/blog/building-seq2seq-machine-translation-models-using-allennlp.html