Abstract:

The phenomenon of fake news has gained a great deal of importance in recent years. Social media and the mass media are increasingly invaded by fake news, provided under the guise of real events or events. Fake news represents an evasive concept, without offering the sources and context of the information or making references to obscure sites. Fake news tends to have shocking headlines, often written in capital letters or containing capital letters.

A non-exhaustive classification of fake news:
- news containing simply invented information;
- misleading news, based on factual data taken out of context or sensationally dressed;
- partisan news, based on interpretations of real events that are manipulated to correspond to a political or business agenda;
- click-bait news, with shocking, sensationalist or compelling headlines, which urge the reader to access them and usually do not support with real information the promises of the title (clicks are counted and contribute to the site's rating, thus causing customers to buy advertising from there).

Determining whether a news item is fake or not is traditionally a method based on the expertise of some human operators in verifying and validating the facts, but due to the large volume and diversity of news to be analyzed today, brought by the explosion of the Internet, there is a clear tendency to use software applications based on artificial intelligence (AI) and especially on the sub-fields of artificial intelligence called data mining (DM) and machine learning (ML). Another reason for switching to automatic detection of fake news is that human operators are not as good as one might think in detecting lies in a text.

The present paper presents the main research directions in the field of machine learning applied in the automatic solution of the problem of detecting fake news. The implementations of two simple but nonetheless quite efficient methods are presented in detail. The first one is based on keyword search in the text, and the second one on probability theory, using the Naive Bayes method.

The paper objective is to implement and determine the performances of these two methods of classifying the news. From the experiments performed to test the functionality, the above-mentioned methods are simple, but quite efficient. Thus, both methods are able to classify news articles, solely on the basis of their extended titles and headlines, into 4 levels of authenticity: authentic news and news that is false in a weak, moderate or secure degree. The results obtained by the two methods are very promising. The performance of the Naive Bayes method can be significantly increased if a large and good quality training database is available, obtained by extending the used training data files. The more news articles, of the order of (tens of) thousands contained in the training database, the more the accuracy of the method increases, thus giving a most accurate answer, very close to the correct one. As a result of the comparison between the two methods outlined a net advantage for the Bayes Naive method, provided by the existence of an adequate training database.

The proposed method is able to improve its performances based on the learning features of the neural networks trained on large and significant databases. This advantage can be applied in e-learning integrated platforms used by journalism students, mass-media experts, social and political analysts as well as by end-users: media wide audience.

Keywords:

e-learning, Machine Learning, Artificial Intelligence, Fake News, Adaptive Methods, Automatic Systems, Decision Criteria.