The bias effect of news media sources on social media users

Abstract

Bias in media news is such an interesting topic that it increases its popularity day by day. Classification of news sources according to their political tendency is a very well studied

subject. However, we have not been able to find any research based on how these polit- ical orientations are transferred to the users on social media. Thus, in this research we

aim to detect changes in the opinion of social media users over time in comparison to the

news articles. In light of this, we have created an hybrid model, by combining Convolu- tional neural networks (CNN) with Long short-term memory (LSTM), which is initiated

by state-of-the-art BERT embeddings. Our choice of BERT embeddings was based on a long trial and error process with other word embeddings such as GloVe, word2vec

and fastText, which led us to the conclusion that the transformer and attention mecha- nism properties of BERT embeddings make it superior to the others. Domain adaptation,

which a transfer learning method is employed as the supplementary method in this pa- per, to overcome the language usage differences between news articles and user tweets.

Thanks to transfer learning, we have observed a significant improvement in the models performance. The created model has been trained on 600.000 labeled news articles and 84.000 politically leaned tweets during the transfer learning. As a result of our testing on recently published and labeled news articles, as expected our model has been successful in proving that politically biased news articles provoke Twitter users to make more biased comments, whereas objective news articles do not lead people to express more political bias. This thesis also includes an Apache NiFi implementation of the idea of monitoring bias dynamically.