Fake News Classification Tool

This project aims to build a fake news classifier that can accurately distinguish fake news from genuine news. We also developed a simple UI, to enable the users to efficiently verify news articles

Figure 1: Overview of the approach

Abstract:

• The advent of technology and the development of social media platforms has made it easier to share news updates with the masses.
• The effects of fake news can be catastrophic.
• Fact checking can prevent people from reacting and taking action on fake news.
• Extremely useful for news houses to fact-check their news before they share it with the masses.

Procedure:

1. Dataset: 20,800 samples with 10387 real news,10413 fake news. Dataset is publically available here

2. Preprocessing:

   1. Removing irrelevant texts and NAN values
   2. Stopwords
   3. Numerics and special characters
   4. Lemmatization
   5. Case folding
   6. Tokenization
   7. Padding

3. Vectorisation:

   1. OneHot Encoding
   2. Count Vectoriser
   3. Hashing-Vectorizer
   4. TF-IDF
   5. GloVe Embedding
   6. Word2Vec Embedding
   7. BERT

4. Machine Learning / Deep Learning Algorithms for Fake news classification:

   1. Naive Bayes (MultinomialNB)
   2. DecionTree
   3. AdaBoost Classification
   4. Logistic Regression
   5. Passive Aggressive
   6. Multilayer Perceptron
   7. LSTM
   8. BERT

Figure 2: Pipeline of the approach used

Pipeline & Output:

Output 1: BERT Tokeniser with Bert model

Figure 3: Pipeline for BERT Tokeniser with Bert model

Figure 4: Performance of BERT Tokeniser with Bert model

2: TF-IDF Tokenizer with Passive Aggressive Classifier

Figure 5: Pipeline for TF-IDF Tokenizer + Passive Aggressive Classifier

Figure 6: Performance of TF-IDF Tokenizer + Passive Aggressive Classifier

WebApp:

https://fake-news-detection-nlp.herokuapp.com/

• An end-to-end deployed tool which allows user to verify news articles in a click
• Efficient and accurate tool for fact checking
• A simple minimalistic user interface
• It allows user to input news text along with title and author name (both optional fields)
• ‘Load sample input’ button allows users to understand and test the app

Technologies Used:

1. Flask, HTML, CSS, JS for building the webapp
2. Heroku for deploying the webapp
3. Python, along with Machine learning, deep learning frameworks.

Team Members:

This project has been completed as a course project of CSE556: Natural Language Processing.

1. Ria Gupta
2. Ruhma Mehek Khan
3. Prakriti Garg

Report & Slides

Presentation Slides

Project Report