| dc.description.abstract | In recent years, GAN has been a fascinating topic. GAN is unsupervised monitoring used in similar fields to produce images and text. In this Thesis, GAN models generate and correct textual data. Also, it is utilized to check the reliability of news. Using the results of this project, we can help fake news detection tools to increase their performance and accuracy. The reason is that the results of this thesis are some new texts that seem to be very good text so that it looks like real news and a human-readable text of the original real news. We generated a new dataset so fake news classifiers cannot detect them as real news. This database enables us to strengthen previous algorithms that detect real and fake news. The result of this thesis is divided into three classes depending on the word count they have. For the news with less than 500 words, the failure of TextGAN algorithms is slightly much more than usual. Because the TextGANs need a dictionary to run the algorithms, and these kinds of news do not have enough words to make a good dictionary. However, for words between 500 and 1000 words, TextGAN algorithms performed better because a better quality dictionary was created, resulting in higher quality output. Also, for news with a word count of more than 1000 words, although a good dictionary is produced, due to the high number of input words, there is a need for more training to produce a suitable output. Among the TextGAN algorithms implemented in this thesis, the modified MLE-Generator and modified DGSAN algorithms provide better-quality results. Finally, two datasets are created from the output of TextGAN algorithms. A dataset with more than 60% similarity contains 36,525 records, and a dataset with more than 80% similarity has 6,950 records. The discriminator classified all the records of these datasets as fake news. Nevertheless, they are reliable news that the summary of the produced text and the input text is more than 60 or 80% similar. The created dataset can help fake and reliable news detection models create more powerful models. | |