tb0se/NLP-Fake-News-Classifier

Detect fake South African based news using classification models # COMS4045A NLP - Project: South African Fake news classifier ## About the project The aim of this project is to detect whether South African based news articles are fake or legitimate. ## Implemented models * Naive Bayes Classifier * TextRNN: A bidirectional LSTM neural network used for text classification. ## Dataset * The fake news are from South African Disiniformation website data -2020. * The real news were scraped by me from News24. I then merged the two datasets and shuffled them to create a dataset that I can use to train my classifier models. ## Getting started * The Word embedding notebook was run on Google Colab using their GPU's. * The Naive Bayes notebook can be run locally using the requirements below. ## Requirements Recommended to use Anaconda for managing your environment. 1. Create a new environment using the `environment.yml` file: ```bash conda env create -f environment.yml ``` 2. Activate the new environment ```bash conda activate ENV ``` 3. Verify new environment was installed correctly ```bash conda env list ``` ## Model Performance | Metrics | Naive Bayes | textRNN | | ------------- |:-------------:| -----:| | Valid Accuracy | - | 0.54 | | Test Accuracy | 0.89 | $0.48 | | AUC| 0.95 | 0.56 | | Precision| 0.89 | 0.24 | | Recall| 0.89 | 0.50 | | F1-score| 0.89 | 0.33 | ## References * Text Classification(tfidf vs word2vec vs bert) * Text Analysis: Feature Engineering with NLP * Fake news Detection using NLP techniques * PyTorch Text Classification Tutorial * Recurrent Convolutional Neural Networks for Text Classification * Lena Voita: NLP Course|For you ## Future Work * Improve Performance of the Recurrent Neural Network. * Collect more data. * Implement a RCNN model. * Implement a Language Model such as BERT.