Twitter sentiment classification by Daniele Grattarola

This is a TensorFlow implementation of a convolutional neural network (CNN) to perform sentiment classification on tweets.

This code is meant to have an educational value, to train the model by yourself and play with different configurations, and was not developed to be deployed as-is (although it has been used in professional contexts). The dataset used for training is taken from here (someone reported to me that the link to the dataset appears to be dead sometimes, so dataset_downloader.py might not work. I successfully ran the script on January 20, 2018, but please report it to me if you have any problems).

NOTE: this script is for Python 2.7 only

Setup

You'll need Tensorflow >=1.1.0 and its dependecies installed in order for the script to work (see here).

Once you've installed and configured Tensorflow, download the source files and cd into the folder:

$ git clone https://gitlab.com/danielegrattarola/twitter-sentiment-cnn.git
$ cd twitter-sentiment-cnn

Before being able to use the script, some setup is needed; download the dataset from the link above by running:

$ python dataset_downloader.py

Read the dataset from the CSV into two files (.pos and .neg) with:

$ python csv_parser.py

And generate a CSV with the vocabulary (and its inverse mapping) with:

$ python vocab_builder.py

The files will be created in the twitter-sentiment-dataset/ folder. Finally, create an output/ folder that will contain all session checkpoints needed to restore the trained models:

mkdir output

Now everything is set up and you're ready to start training the model.

Usage

The simplest way to run the script is:

$ python twitter-sentiment-cnn.py

which will load the dataset in memory, create the computation graph, and quit. Try to run the script like this to see if everything is set up correctly. To run a training session on the full dataset (and save the result so that we can reuse the network later, or perform more training) run:

python twitter-sentiment-cnn.py --train --save

After training, we can test the network as follows:

$ python twitter-sentiment-cnn.py --load path/to/ckpt/folder/ --custom_input 'I love neural networks!'

which will eventually output:

...
Processing custom input: I love neural networks!
Custom input evaluation: POS
Actual output: [ 0.19249919  0.80750078]
...

By running:

$ python twitter-sentiment-cnn.py -h

the script will output a list of all customizable flags and parameters. The parameters are:

train: train the network;
save: save session checkpoints;
save_protobuf: save model as binary protobuf;
evaluate_batch: evaluate the network on a held-out batch from the dataset and print the results (for debugging/educational purposes);
load: restore a model from the given path;
custom_input: evaluate the model on the given string;
filter_sizes: comma-separated filter sizes for the convolutional layers (default: '3,4,5');
dataset_fraction: fraction of the dataset to load in memory, to reduce memory usage (default: 1.0; uses all dataset);
embedding_size: size of the word embeddings (default: 128);
num_filters: number of filters per filter size (default: 128);
batch_size: batch size (default: 128);
epochs: number of training epochs (default: 3);
valid_freq: how many times per epoch to perform validation testing (default: 1);
checkpoint_freq: how many times per epoch to save the model (default: 1);
test_data_ratio: fraction of the dataset to use for validation (default: 0.1);
device: device to use for running the model (can be either 'cpu' or 'gpu').

Model description

The network implemented in this script is a single layer CNN structured as follows:

Embedding layer: takes as input the tweets (as strings) and maps each word to an n-dimensional space so that it is represented as a sparse vector (see word2vec).
Convolution layers: a set of parallel 1D convolutional layers with the given filter sizes and 128 output channels. A filter's size is the number of embedded words that the filter covers.
Pooling layers: a set of pooling layers associated to each of the convolutional layers.
Concat layer: concatenates the output of the different pooling layers into a single tensor.
Dropout layer: performs neuron dropout (some neurons are randomly not considered during training).
Output layer: fully connected layer with a softmax activation function to perform classification.

The script will automatically log the session with Tensorboard. To visualize the computation graph and training metrics run:

$ tensorboard --logdir output/path/to/summaries/

and then navigate to localhost:6006 from your browser (you'll see the computation graph in the Graph section).

manneshiva/twitter-sentiment-cnn

Twitter sentiment classification by Daniele Grattarola

Setup

Usage

Model description