CNN Sentence Classification

This repository implements a Convolutional Neural Network for Sentence Classification, based on the work of Yoon Kim (original paper).

Datasets Used

Movie Review

The Movie Review dataset contains 5331 labelled positive and 5331 labelled negative reviews from IMDb. Download the sentence polarity dataset v1.0 file and extract it. Place it under the data/ folder, as shown in the file directory structure below.

Subjectivity

The Subjectivity dataset contains 5000 labelled subjective and 5000 labelled objective snippets from movie reviews from Rotten Tomatoes and IMDb. Download the Subjectivity dataset v1.0 file and extract it. Place it under the data/ folder, as shown in the file directory structure below.

Hyper-parameter Search

To improve on the baseline accuracy, an extensive incremental hyper-parameter search is done to try and find the hyper-parameter values that result in a higher accuracy. The hyper-parameter values are explored and set one by one.

Architectural Changes

To address the loss of temporal information in Kim's original model, the architecture of the CNN was modified by reshaping and performing a 2D convolution before pooling and feeding the results to a dense layer.

Packages Needed

• gensim
• nltk

Word Embeddings

This project makes use of Google's word2vec embeddings, which constructs vector representations of words through their implementation of continuous bag-of-words and skip-gram. Download the GoogleNews-vectors-negative300.bin file, and place it under the data/ folder.

How to Run the Program

1. Download the Movie Reviews dataset and the Subjectivity dataset as indicated above, and place them in the data/ folder. Ensure they follow the same directory structure as below.
2. Make sure GoogleNews-vectors-negative300.bin is located under the data/ folder.
3. Open the src/config.py file and do the following:
   • While they shouldn't require modification, double check that all filepaths are ok.
   • Set the configuration to run, such as the dataset (MR for Movie Review, SUBJ for Subjectivity), vector type to use (random, word2vec) and embedding option (static, dynamic).
   • Set the number of epochs and folds desired.
   • Set the desired model hyper-parameter values (feature maps, kernel size, regularization strength, dropout rate and optimizer)
   • If you wish to run incremental parameter tuning, set RUN_INCREMENTAL_BEST_PARAMS to True. Assign arrays of the values to explore to the hyper-parameter variables ending in _INCREMENTAL.
4. Run the main.py script.

Directory Structure

.
├── data
|   ├── GoogleNews-vectors-negative300.bin
|   |
|   ├── rt-polaritydata
|   |   ├── rt-polarity.neg
|   |   └── rt-polarity.pos
|   |
|   └── SUBJ
|       ├── plot.tok.gt9.5000
|       └── quote.tok.gt9.5000
|
├── results
|
└── src
    ├── config.py
    |
    ├── main.py
    ├── cnn.py
    |
    ├── data_processing
    |   ├── data_cleaning.py
    |   ├── data_loading.py
    |   └── word_vectorizing.py
    |
    └── util
        └── results.py

The data/ folder holds the Movie Review and Subjectivity datasets.

Any results are placed automatically in the results/ folder. These results include confusion matrices, and loss and accuracy graphs.

Files in src/:

• config.py defines the datasets and word embeddings to be run, and controls the settings for specific configurations.
• main.py is a script that trains, evaluates and generates results through k-fold cross validation.
• cnn.py contains the implementation of the convolutional neural network for sentence classification.
• util/results.py defines the functionalities needed to save results of runs to files.
• data_processing/ has the logic necessary for different data processing strategies.