I've created a gist with a simple generator that builds on top of your initial idea: it's an LSTM network wired to the pre-trained word2vec embeddings, trained to predict the next word in a sentence. The data is the list of abstracts from arXiv website.
I'll highlight the most important parts here.
Gensim Word2Vec
Your code is fine, except for the number of iterations to train it. The default iter=5 seems rather low. Besides, it's definitely not the bottleneck -- LSTM training takes much longer. iter=100 looks better.
word_model = gensim.models.Word2Vec(sentences, vector_size=100, min_count=1,
window=5, iter=100)
pretrained_weights = word_model.wv.syn0
vocab_size, emdedding_size = pretrained_weights.shape
print('Result embedding shape:', pretrained_weights.shape)
print('Checking similar words:')
for word in ['model', 'network', 'train', 'learn']:
most_similar = ', '.join('%s (%.2f)' % (similar, dist)
for similar, dist in word_model.most_similar(word)[:8])
print(' %s -> %s' % (word, most_similar))
def word2idx(word):
return word_model.wv.vocab[word].index
def idx2word(idx):
return word_model.wv.index2word[idx]
The result embedding matrix is saved into pretrained_weights array which has a shape (vocab_size, emdedding_size).
Keras model
Your code is almost correct, except for the loss function. Since the model predicts the next word, it's a classification task, hence the loss should be categorical_crossentropy or sparse_categorical_crossentropy. I've chosen the latter for efficiency reasons: this way it avoids one-hot encoding, which is pretty expensive for a big vocabulary.
model = Sequential()
model.add(Embedding(input_dim=vocab_size, output_dim=emdedding_size,
weights=[pretrained_weights]))
model.add(LSTM(units=emdedding_size))
model.add(Dense(units=vocab_size))
model.add(Activation('softmax'))
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy')
Note passing the pre-trained weights to weights.
Data preparation
In order to work with sparse_categorical_crossentropy loss, both sentences and labels must be word indices. Short sentences must be padded with zeros to the common length.
train_x = np.zeros([len(sentences), max_sentence_len], dtype=np.int32)
train_y = np.zeros([len(sentences)], dtype=np.int32)
for i, sentence in enumerate(sentences):
for t, word in enumerate(sentence[:-1]):
train_x[i, t] = word2idx(word)
train_y[i] = word2idx(sentence[-1])
Sample generation
This is pretty straight-forward: the model outputs the vector of probabilities, of which the next word is sampled and appended to the input. Note that the generated text would be better and more diverse if the next word is sampled, rather than picked as argmax. The temperature based random sampling I've used is described here.
def sample(preds, temperature=1.0):
if temperature <= 0:
return np.argmax(preds)
preds = np.asarray(preds).astype('float64')
preds = np.log(preds) / temperature
exp_preds = np.exp(preds)
preds = exp_preds / np.sum(exp_preds)
probas = np.random.multinomial(1, preds, 1)
return np.argmax(probas)
def generate_next(text, num_generated=10):
word_idxs = [word2idx(word) for word in text.lower().split()]
for i in range(num_generated):
prediction = model.predict(x=np.array(word_idxs))
idx = sample(prediction[-1], temperature=0.7)
word_idxs.append(idx)
return ' '.join(idx2word(idx) for idx in word_idxs)
Examples of generated text
deep convolutional... -> deep convolutional arithmetic initialization step unbiased effectiveness
simple and effective... -> simple and effective family of variables preventing compute automatically
a nonconvex... -> a nonconvex technique compared layer converges so independent onehidden markov
a... -> a function parameterization necessary both both intuitions with technique valpola utilizes
Doesn't make too much sense, but is able to produce sentences that look at least grammatically sound (sometimes).
The link to the complete runnable script.
