xspring14/recurrent-convolutional-neural-network

A recurrent convolutional neural network for tweet classification using python and theano

rcnn-text-classificaiton

The model developed in this project uses combination of Recurrent and Convolution Neural Network for tweet classification. The model automatically captures important features and trains model accordingly.

Dependencies

• numpy
• theano
• twokenize
• word2vec

Dataset

Database of 10000 tweets is used for creating word embeddings using word2vec. A subset of that data of size 1532 is used for training and testing. Training is done with 1000 tweets and testing is done with 532 tweets set by randomly dividing the dataset.

Training

Training is carried using GPU, Nvdia GeForce 750M

Before training create initial word embeddings from your text corpse using word2vec.

import word2vec
#
# tweet-corpse.txt contains 10K tweets in lower case.
# Size 100 defines the vector size of the word embedding
word2vec.word2vec('data/tweet-corpse.txt', 'data/training_vec.bin', size=100, verbose=True)

We can then start training by

sudo python rcnn_theano

Manually, Once we have vector embedding, we can continue to load word2vec data with TweetPreProcess and continues training

t = TweetPreProcess() # Loads dictionary and its word embedding
x, y = t.create_set(filename="data/1700.csv", random=True) # Load dataset into x and y

Once we load data we can divide the data into various set and then start training using RCNN class.

# Sharing word embeddings to theano shared memory to sothat it also can be updated
vector_dict = theano.shared(np.asarray(t.word_vec).astype("float32"), name="vector_dict")
# Dividing into training and validation set
train_x_set = [x[i][:] for i in range(1000)]
train_x = [y[i][:] for i in range(1000)]
valid_x_set = [x[i][:] for i in range(1000, len(x))]
valid_x = [y[i][:] for i in range(1000, len(x))]
# Initializing using vector_dict
model = RCNN(vector_dict)
#
# Training 
model.process_set(train_x_set,
                  train_x,
                  num_epochs=64,
                  verbose=True,
                  verbose_interval=1,
                  file_name="model_ada.model",
                  valid_x=valid_x_set,
                  valid_y=valid_x)

Prediction

Once the model is trained you can use to predict new data

model.prediction(t.sentence2index("Flipkart has great service "), 4) # Second argument is length of string

or you can load best validation saved model and then predict

model = RCNN(file_name= "model_ada")
model.prediction(t.sentence2index("Flipkart has great service"), 4)

Results

Training this model on a 1000 tweets data and testing the best stable model on 532 tweets gives accuracy of 80%

Log files of the project is at model/report.txt along with best validation model.

Author

Moonis Javed

Version

1.0.0