CharCNN

Our project is using the TensorFlow framework to implement the CharCNN model to classify English sentences.

Model Architecture

This model utilizes a convolutional neural network to extract features of sentences.

Team

• Team leader:
  1. Github: hoangcaobao
• Team members:
  1. Github: Nguyendat-bit
  2. Github: aestheteeism
• Advisor:
  1. Github: bangoc123

I. Set up environment

• Step 1: Clone repo to local

git clone https://github.com/protonx-tf-03-projects/CharCNN.git

• Step 2: Change directory to the folder

cd CharCNN

• Step 3: Make sure you have installed Miniconda. If not yet, see the setup document here.

• Step 4: Create an environment

conda env create -f environment.yml

• Step 5: Activate environment

conda activate CharCNN

Congratulation. Now you have the environment to use our project.

II. Set up your dataset

• We use data IMBD Dataset of 50K Movie Reviews for training that can be easily found at here
• After that, download it locally then move this CSV file to the CharCNN folder without changing the name
• Dataset (CSV file) has to follow this structure

review	sentiment
sentence	label
...	....

• References: NLP and CV

III. Training Process

Review training on colab:

Training script:

python train.py --train-file ${link_to_train_data} --epochs ${epochs} 

For example:

python train.py --train-file data.csv --epochs 3 

Due to a large number of datasets, we recommend using epochs smaller than 3

IV. Predict Process

You create a file CSV with 1 column named sentences to contain your dataset

Predict script:

python predict.py --test-file ${link_to_test_data} --result-file ${name_of_file_contain_result}

For example:

python predict.py --test-file test.csv --result-file result.csv

Now you have the sentiment of your sentences in test.csv in result.csv file

V. Result and Comparision

Small CharCNN performance:

Epoch 1/3
313/313 [==============================] - 1583s 5s/step - loss: 0.6929 - accuracy: 0.5117 - val_loss: 0.6782 - val_accuracy: 0.5392
Epoch 2/3
313/313 [==============================] - 1554s 5s/step - loss: 0.3542 - accuracy: 0.8433 - val_loss: 0.2596 - val_accuracy: 0.8948
Epoch 3/3
313/313 [==============================] - 1481s 5s/step - loss: 0.1619 - accuracy: 0.9414 - val_loss: 0.2819 - val_accuracy: 0.8902

Bidirectional LSTM + Dense performance:

Epoch 1/3
313/313 [==============================] - 5085s 16s/step - loss: 0.4762 - acc: 0.7655 - val_loss: 0.4360 - val_acc: 0.8097
Epoch 2/3
313/313 [==============================] - 4386s 14s/step - loss: 0.2837 - acc: 0.8928 - val_loss: 0.2808 - val_acc: 0.8905
Epoch 3/3
313/313 [==============================] - 2185s 7s/step - loss: 0.1841 - acc: 0.9351 - val_loss: 0.3262 - val_acc: 0.8924

Above, they are results of training of CharCNN model vs our custom Bidirectional LSTM in 3 epochs on IMDB Dataset of 50K Movie Reviews. From that results, CharCNN is not good as Bidirectional LSTM.

VI. Feedback

If you meet any issues when using this library, please let us know via the issues submission tab.