Neural-Machine-Translation

Project translates english language to hindi language in devanagari script. same script can be used for translating any one language to subsequent another language.

english to hindi example

---

actual eng text :  ['the holy dip is on november']
actual hindi text :  ['मुख्य स्नान नवंबर को है']
translated hindi text :  ['नवंबर को अंतिम रूप से ये संशोधित श्रृंखला शुरू होती है।']

well it's not perfect since it is trained on just 69201 pairs of english-hindi parallel dataset. To train good encoder-decoder translation model it requires millions of parallel corpus of dataset and good amout of training time.

Main model architecture to be trained

---

Encoder Model generated from saved Model

---

Decoder Model generated from saved Model

---

To translate using encoder-decoder model :

def translate(eng_sent):
  _, eh, ec = inf_enc_model.predict(eng_sent.reshape(1, len(eng_sent)))
  translated = []
  dec_inp_seq = np.array([2]).reshape(1,1) #2 for start token and 1 for end
  stop = False
  while not stop:
    d, [eh, ec] = dec_model_final.predict([dec_inp_seq, [eh, ec]])
    dec_inp_seq[:,] = np.argmax(d[0][0])
    translated.append(np.argmax(d[0][0]))

    if dec_inp_seq[0][0] == 1 or len(translated)>16:
      stop = True
  return translated

use this script for more info

or

open this colab notebook

License

MIT