Použití word embeddingu pro strojový překlad

Machine Translation with Keras

Thushan Ganegedara

Data Scientist and Author

Úvod do word embeddingů

Vektory s jedničkovým kódováním

cat_vector = np.array([[1,0,0,0...,0]])
dog_vector = np.array([[0,1,0,0...,0]])
window_vector = np.array([[0,0,1,0...,0]])

Vektory slov

cat_vector = np.array([[0.393,-0.263,0.086,0.011,-0.322,...,0.388]])
dog_vector = np.array([[0.399,-0.300,0.047,-0.059,-0.111,...,0.037]])
window_vector = np.array([[0.133,0.149,-0.307,0.090,-0.143,...,0.526]])

Podobnost mezi vektory slov

from sklearn.metrics.pairwise import cosine_similarity
cat_vector = np.array([[0.393,-0.263,0.086,0.011,-0.322,...,0.388]])
dog_vector = np.array([[0.399,-0.300,0.047,-0.059,-0.111,...,0.037]])
window_vector = np.array([[0.133,0.149,-0.307,0.090,-0.143,...,0.526]])

cosine_similarity(cat_vector, dog_vector)

0.601

cosine_similarity(cat_vector, window_vector)

0.323

¹ https://nlp.stanford.edu/projects/glove/

Implementace embeddingu pro enkodér

Bez vrstvy embeddingu

en_inputs = Input(shape=(en_len, en_vocab))

Vstup časové řady

Implementace embeddingu pro enkodér

S vrstvou embeddingu
```
en_inputs = Input(shape=(en_len,))
```

Vstup word embeddingu

Implementace embeddingu pro enkodér

en_inputs = Input(shape=(en_len,))
en_emb = Embedding(en_vocab, 96, input_length=en_len)(en_inputs)

Vyhledávání embeddingu

Implementace enkodéru s embeddingem

en_inputs = Input(shape=(en_len,))
en_emb = Embedding(en_vocab, 96, input_length=en_len)(en_inputs)
en_out, en_state = GRU(hsize, return_state=True)(en_emb)

Vrstva embeddingu – enkodér

Implementace dekodéru s embeddingem

de_inputs = Input(shape=(fr_len-1,))
de_emb = Embedding(fr_vocab, 96, input_length=fr_len-1)(de_inputs)
de_out, _ = GRU(hsize, return_sequences=True, return_state=True(
  de_emb, initial_state=en_state)

Model enkodér-dekodér s embeddingem

Trénování modelu

for ei in range(3):
  for i in range(0, train_size, bsize):    
    en_x = sents2seqs('source', tr_en[i:i+bsize], onehot=False, reverse=True)

    de_xy = sents2seqs('target', tr_fr[i:i+bsize], onehot=False)    
    de_x = de_xy[:,:-1]

    de_xy_oh = sents2seqs('target', tr_fr[i:i+bsize], onehot=True)    
    de_y = de_xy_oh[:,1:,:]

    nmt_emb.train_on_batch([en_x, de_x], de_y)    
    res = nmt_emb.evaluate([en_x, de_x], de_y, batch_size=bsize, verbose=0)
    print("{} => Loss:{}, Train Acc: {}".format(ei+1,res[0], res[1]*100.0))

Pojďme si procvičit!

Machine Translation with Keras