siddarthhari95 / NLP-Fake-News_detection

Implementation of a deep learning model that does fake news classification based on Liar-Liar Dataset

NLP-Project-Fake-News-Detection

NLP-Project-Fake-News-Detection

Dataset Used : LIAR Dataset

Paper : https://arxiv.org/abs/1705.00648
Data : https://www.cs.ucsb.edu/~william/data/liar_dataset.zip

Original Source:

The paper did not release the source code.

List of commands:

Run all the cells in the ipython notebook.

Train :

function : train(model,architecture, use_pos, use_meta, use_dep)

• params :
  
  • model : keras Sequential model
    
  • architecture : 'lstm' or 'cnn'
    
  • use_pos : boolean (tells whether the POS embeddings are to be used as input to the model for training)
    
  • use_meta : boolean (tells whether the metadata is to be used as input to the model for training)
    
  • use_dep : boolean (tells whether the Dependency parse embeddings are to be used as input to the model for training)
    

Evaluation on test:

function_1 : (false_predicted,true_predicted) = evaluate(model, use_pos, use_meta, use_dep)

• params :
  
  • model : keras Sequential model
    
  • use_pos : boolean (tells whether the POS embeddings were used as input to the model while training)
    
  • use_meta : boolean (tells whether the metadata was used as input to the model while training)
    
  • use_dep : boolean (tells whether the Dependency parse embeddings were used as input to the model while training)
    
• returns :
  
  • false_predicted : dictionary with keys as indices of test samples that were classified as "pants-on-fire" (false news) and values as the softmax probability for this class label.
    
  • true_predicted : dictionary with keys as indices of test samples that were classified as "true" (not a fake news) and values as the softmax probability for this class label.
    

function_2 : print_best_false_true_predicted(false_predicted,true_predicted)

• params :
  
  • outputs from the above mentioned evaluate() function.
    
• outputs :
  • prints top 5 sentences which where predicted as "pants-on-fire" (fake news) with highest softmax probabilities.
    
  • prints top 5 sentences which where predicted as "true" (not fake news) with highest softmax probabilities.

Software and Library Requirements:

Keras==2.2.2
matplotlib==2.1.2
numpy==1.15.4
pandas==0.22.0
pydot==1.4.0
scikit-learn==0.19.1
scipy==1.0.0
sklearn==0.0
spacy==2.0.18
tensorboard==1.9.0
tensorflow==1.9.0
spacy==2.0.18
nltk==3.4