FaGoN: Fake News Detection model using Grammatic Transformation on Neural Network

Abstract

These days, most of fake news are detected and verified by people, which requires a great amount of time and effort. It is difficult to Figure out the truthfulness of the news by machine algorithm because the sentences have various forms. In this paper, we shall present a fast and efficient fake news detection model which can Figure out whether the given proposition is true or not from article by exploiting grammatical transformation based on deep learning. Our model consists of four layers: word embedding layer, context generation layer, matching layer and inference layer. In word embedding layer, the words in proposition are embedded into word vector. In context generation layer, the word vectors enter into LSTM layer and generate context vector. In matching layer, attention vector is generated from the contextual embedding vector in the previous layer computing the weighted sum. Then, the hidden state vector from LSTM layers and attention vector are compared through matching operation generating the sentences which has the same meaning but different forms. In inference layer, our model calculates the similarity between the generated sentences and the sentences in articles, and classifies the answer, true or false. We shall evaluate our model calculating the perplexity to Figure out whether the generated sentences are grammatically correct. Also, the model is tested by changing the sentence group’s size to find the optimal size of the group. By showing the Our model figured out the fake news very well with the test of CNN news dataset getting the right answer.