Neural_Machine_Translation

Best model metrics

Metric	Score
Corpus BLEU	37.0347
Dev ppl	61.4084

The translation demo is available here on Streamlit Sharing.
Even is you don't know Spanish you can use the demo as there is Google Translate which will help you to convert your English sentences to Spanish.

It is a Seq2Seq Model that translates Spanish sentences into English based on Luong et al. 2015.
It consists of a bidirectional LSTM encoder and unidirectional LSTM decoder.
It also uses attention mechanism to boost its performance on the translation task.
The pipeline and the implementations is inspired by the Open-NMT package.

The model becomes more powerful as we combine character-level with word-level language modelling.
The idea is that whenever the NMT model generates a <unk> token we run a character-level language model and generate a word in the output character by character.
This hybrid word-character approach was proposed by Luong and Manning 2016 and turned out to be effective in increasing the performance of the NMT model (+1.2 BLEU).

Install from source:

git clone https://github.com/sahilkhose/Neural_Machine_Translation
cd Neural_Machine_Translation
pip3 install -r requirements.txt

To run the translation demo:

streamlit run stream_translate.py

Or just go here on Streamlit Sharing.

If you find a bug, create a GitHub issue, or even better, submit a pull request. Similarly, if you have questions, simply post them as GitHub issues.