cnn-text-classification-keras

Convolutional Neural Network for Text Classification in Keras

This is a Keras implementation of Yoon Kim's paper Convolution Neural Networks for Sentence Classification with the addition that this code also works for the Glove vectors and Fasttext vectors.

Requirements:

• numpy
• keras
• cPickle

Usage:

• Download the pre-trained Google word2vec word embedding vectors as a binary file from here

• Pre-process the text data

from text_processing_util import TextProcessing

tp = TextProcessing(texts, labels, EMBEDDING_DIM, MAX_SEQUENCE_LENGTH, MAX_NB_WORDS, VALIDATION_SPLIT)

where

- texts: a list of sentences.
- labels: a list of labels corresponding to the sentences in the list texts.
- MAX_SEQUENCE_LENGTH: maximum length of the sentence to be considered, longer sentences will be terminated at this length.(default is 50)
- MAX_NB_WORDS: maximum number of words to be used in the model (default is 10000).
- EMBEDDING_DIM: dimension of the word vectors (default is 300 for google word2vec).
- VALIDATION_SPLIT: fraction of data to be used for validation. (default is 0.2).

• Split into train and test data.

x_train, y_train, x_val, y_val, word_index = tp.preprocess()

• Build the embeddings index.

embeddings_index = tp.build_embedding_index_from_word2vec(path_to_wordvec_file, word_index)

• Serialize the data after the processing.

import cPickle

cPickle.dump([word_index, embeddings_index], open('tokenization_and_embedding.p', 'wb'))

• Get labels index.

labels_index = tp.labels_index

• Build the CNN model

from text_cnn import kimCNN

model = kimCNN(EMBEDDING_DIM, MAX_SEQUENCE_LENGTH, MAX_NB_WORDS, embeddings_index, word_index, labels_index=labels_index)

• Fit the model

model.fit(x=x_train, y=y_train, batch_size=50, epochs=25 , validation_data=(x_val, y_val))

For a detailed example see example.py. This is the same example used in Kim's paper and the original theano code.

References:

• Convolution Neural Networks for Sentence Classification
• Theano Code
• Tensorflow code