# **Application of Transfer Learning - Usage of Sentence Embeddings for Fake News Detection by Deep Learning.**
In this project, Google Universal Sentence Encoder is used. Using Google's Sentence Encoder Deep learning Models where trained and their performance was documented.
The task was to build a model that could classify news into two different classes, Real and Fake.
Labelled dataset was available.
This technique is supposed to work with any kind of **NLP classification task**.
Research paper - [Universal Sentence Encoder](https://arxiv.org/abs/1803.11175)
Sentence Encoder by Google - [Tensorflow hub](https://tfhub.dev/google/universal-sentence-encoder/2)
*Key Requirements: Python 3+, Keras 2+, TensorFlow 1+.
Training was performed on [Google Colab](https://colab.research.google.com/)*
## Contents
### **Python Notebooks**
- 3 python Notebooks demostrating Model performance on three different datasets.
- 1 Python Notebook showing how embeddings where generated.
## Fake or Real Dataset
3 different Datasets where used in this project. These are:
- [News Dataset from Kaggle](https://www.kaggle.com/c/fake-news/data)
- [Dataset available from OpensSources](http://www.opensources.co/)
- [News obtained from George McIntire Fake News Dataset](https://bitbucket.org/WiZar-D/fake_real_dataset)
**Dataset describtion**
Dataset consisted of labelled news under two categories "Real" and "Fake".
Below is an overview of the dataset.
| Dataset 1: Kaggle |
| Type |
Real |
Fake |
| Count |
6200 |
6200 |
| Dataset 2: OpenSources |
| Type |
Real |
Fake |
| Count |
6724 |
6162 |
| Dataset 3: Work done by George McIntire |
| Type |
Real |
Fake |
| Count |
1900 |
1992 |
**The number of news articles in the respective corpus where more than the figures shown above. Only those articles in which
number of sentences was more than 7 and less than 50.**
## **Model Description**
4 Deep Learning models where developed for classification task.
[Recurrent Neural Networks](https://keras.io/layers/recurrent/) where used.
**Reason for selecting articles which has sentences in the range(7,50)**
As LSTM takes input a 3-d Tensor comprised of (samples, timesteps, features), having this limit help us to design a network
which takes input a 3-d Tensor of shape (n,50,512), where
- "n" is number of articles.
**Google sentence encoder** returns sentence embedding which is of shape (512,).
These where then reshaped into [numpy's multidimensional, homogeneous array](https://docs.scipy.org/doc/numpy-1.15.0/reference/generated/numpy.ndarray.html)
These numpy arrays made up the input for the Neural Network.
*(The figure shown below is taken from deeplearning course on Coursera. Similar version of this network was implemented.)*
**3 different architecture where used, these are:**
- Single Layer LSTM Model - A LSTM layer is defined which is followed by a dense/ output layer.
- Double Layer LSTM (Connected) Model - 2 Layer of LSTM are defined with output from previous layer fed into next LSTM layer.
- CNN LSTM Model - First a CNN layer or layers are defined, wrapping them in a TimeDistributed layer and then LSTM and output layer is defined.
All this model where build using Keras.
## Results
Evaluation of each model is shown with the help of [confusion matrix](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.confusion_matrix.html), Loss and Accuracy Graphs.
*sample output*
```
[[1178 67]
[ 63 1172]]
precision recall f1-score support
Real 0.95 0.95 0.95 1245
Fake 0.95 0.95 0.95 1235
micro avg 0.95 0.95 0.95 2480
macro avg 0.95 0.95 0.95 2480
weighted avg 0.95 0.95 0.95 2480
```
 
From the graphs it is evident that model has overfitted the data.
Work needed to be done to make generalize the model, implementing more complex and deeper network, training on more data may improve model performance.
More feautres can be added apart from sentence embeddings, like how much the headline agrees with the content, source of the article and so on..
Neural Network trained with [ELMo](https://arxiv.org/abs/1802.05365) ([*developed by AllenNLP*](https://allennlp.org/elmo)) - A Deep contextualized word representations can also be used.