Fake News Detection

MAIN OBJECTIVES

Overview

Fake news detection is an application that can be used to determine whether a news is fake or not based on a dataset and specific AI algorithms that we have developed. The application can be used by either normal users to check whether the news they face in their daily life is fake or not. It can also be used by Organizations and Journalists for business needs.

Objectives

1. Assemble a trusted dataset that can be depended on for the AI model
2. Build strong AI algorithm that can be both trustworthy and fast
3. Return accurate results for news integrity.
4. Provide several options for users when checking if a news is fake. (E.g: the ability to enter the news details manually or fetch them from a URL)
5. Each user should have a profile history with his previous news submissions
6. Provide a trending fake news map to visualizes the percentage of fake news around different cities in the world.

• Get ~63% Accuracy using Liar plus data-set, This state of the art model without using (Attentions layers) BERT model .

Fake News

Depended on 3 featuers of data to training : Statement , Subject and Justification in Liar Liar plus dataset

Liar plus :

This dataset has evidence sentences extracted automatically from the full-text verdict report written by journalists in Politifact. Our objective is to provide a benchmark for evidence retrieval and show empirically that including evidence information in any automatic fake news detection method (regardless of features or classifier) always results in superior performance to any method lacking such information.

Below is the description of the TSV file taken as is from the original LIAR dataset, which was published in this paper. We added a new column at the end that has the extracted justification.

• Column 1: the ID of the statement ([ID].json).
• Column 2: the label.
• Column 3: the statement.
• Column 4: the subject(s).
• Column 5: the speaker.
• Column 6: the speaker's job title.
• Column 7: the state info.
• Column 8: the party affiliation.
• Columns 9-13: the total credit history count, including the current statement.
• 9: barely true counts.
• 10: false counts.
• 11: half true counts.
• 12: mostly true counts.
• 13: pants on fire counts.
• Column 14: the context (venue / location of the speech or statement).
• Column 15: the extracted justification.

The dataset can be found in this commit.

Using LSTM in trained model and glove file to word embeddings .

After transfer each word to vector of numbers has size 300 dims ,ex: 'trumb' --> [1,4,.,.,.,.,.,.,.,.,.,9] , then each sentence have vectors of vectors numbers.

Using mathematics trick : get cosine similarity between mean of each two featuers.

mean(tarnsferd_statement) + cosine similarity between mean(tarnsferd_statement) and mean(tarnsferd_subject) + mean(tarnsferd_subject) + cosine similarity between mean(tarnsferd_subject) and mean(tarnsferd_justification) and mean(tarnsferd_justification) + cosine similarity between mean(tarnsferd_justification) and mean(tarnsferd_statement) .

20 epochs to training model , if you need model learn more increase the number of epochs .

Trained 2 models : first one used CNN , the second one used LSTM

Detect subjects of news

can detect the subject of news when using url scrape to test the truth of news , used multiclass machine learning alogrithm (logistic regression) with sklearn multi label binarizerto encode subject calss to 1,0s the trained model and used TF-IDF to vectorization the text and save model . subject detection.

This model Implemented by Amal Ossman Abd El-moaz @amlosma .

Quality of text

Using Blue Score model to test the dependency of sentence by togther exist one version

After get the credibility and quality values by integrate two notebooks(liarplus_version1_.ipynb,quality_text_version1_.ipynb) in one notebook to test Fake news detection then sort the credibility and quality values by using weights of mulitply numbers 80% of the credibility and 20% of th quality values.

Trending Map

Provide a trending fake news map to visualizes the percentage of fake news around different cities in the world

Implemented trending map to track the fake and real news in the world Trending Map . Implemented by Mahmoud Salah @mahmoud24598salah .

To run App

example of new:

statement="The decade that shattered trust in politics."

subject="politics."

justification="It is totally normal for ministers and officials in high pressure jobs to have quarrels and tricky conversations. Arguably, a bit of healthy tension can be a good thing for governments, to make sure that ideas are tested and policies properly thought through. It is also normal from time to time for senior officials to move quietly to different government departments if a relationship breaks down with their political boss, or sometimes, for them to retire early if the situation has become impossible. There is nothing remotely normal however about a top government official quitting their job, suing the government in the belief they were forced out, deciding to go public with the reasons, and accusing one of the most senior politicians in the country of not being straight with the truth. But that is exactly what's happened. Sir Philip Rutnam has been one of the most senior civil servants for years, in charge at the Home Office for the last few. His time there has not always been an unalloyed success - the Home Office, as one of the biggest and most complicated departments in the government, has struggled with various issues, most notably the Windrush scandal. The Home Office is often seen as a poisoned chalice given the nature of its job."

the output : The credibility of new: 52.1%