Generating translations with the NMT

Machine Translation with Keras

Thushan Ganegedara

Data Scientist and Author

Motivation

You have a trained model
Need to be able to assist humans on translation tasks
Test the model on unseen data
How?
- Hold-out test set to evalute the model
- You will test the model by asking it to predict translations for one sentence.

Transforming the input

English sentence

en_st = ['the united states is sometimes chilly during december , but it is sometimes freezing in june .']

Transform the encoder sentence

en_seq = sents2seqs('source', en_st, onehot=True, reverse=True)
print(np.argmax(en_seq, axis=-1)

English: ['the united states is sometimes chilly during december , 
           but it is sometimes freezing in june .']
Reversed sentence: ['june in freezing sometimes is it ...'] 
Reversed sequence: [[34   3  54       10        2  4  7 45  5 69 10  2 23 22  6]]

Generating the translation

Generating a prediction
```
fr_pred = model.predict(en_seq)
```
fr_pred.shape
- [sentences, seq len, vocab size]

Getting the predicted classes

fr_seq = np.argmax(fr_pred, axis=-1)[0]

[[ 3  7 35 34  2 ...  5  4  4  0  0]] # <= fr_seq

fr_seq.shape
- [num sentences, sequence len]

Converting the prediction to a sentence

Converting the produced word IDs to a sentence using list comprehension

fr_sentence = ' '.join([fr_id2word[i] for i in fr_seq if i != 0])

English: the united states is sometimes chilly during december , but it is sometimes freezing in june .

French:                    les états unis est parfois froid en décembre mais il est parfois le gel en

French (Google Translate): les etats-unis sont parfois froids en décembre, mais parfois gelés en juin

List comprehension in more detail

List comprehension

word_list = [fr_tok.index_word[i] for i in fr_seq if i != 0]

For loop

word_list = []
for i in fr_seq:
  if i != 0:
    word_list.append(fr_tok.index_word[i])

Let's practice!

Machine Translation with Keras