SENTINET: A DEEP SENTIMENT ANALYSIS NETWORK FOR POLITICAL MEDIA BIAS DETECTION

Abstract

India has the highest circulation of newspapers in the world, but unfortunately also has high media bias rates and one of the lowest press freedom rankings for democracies. A biased media prevents citizens from receiving information that might be essential to public wellbeing by filtering information through a lens that supports government interests first. Media bias plays an influencing role even at the voting booth as propaganda can skew voter decisions and perceptions of what is true in this era of fake news. It's vital to keep an eye on bias in the news and to provide a platform where people can get unbiased and reliable news. Researchers in sentiment analysis and bias detection have been using various techniques to achieve higher accuracy to detect media bias. This study aims to take a different technical approach to the problem of Indian political media bias detection by developing SentiNet - a graphical processing unit (GPU) accelerated modified convolution neural network (CNN) model consisting of linearly inverted depth-wise separable convolutions capable of classifying news as either ‘unbiased’ or ‘biased’ from twitter data. Because of its simple architecture and lesser number of tuning parameters, it is observed that SentiNet is a good fit in terms of accuracy and loss function and its training time reduces by 50% when using a GPU. From results, Republic TV and BBC emerged as the most biased towards ruling party and Opposition parties respectively. NDTV and News19 emerged unbiased towards ruling party with balanced reporting. India TV has emerged as unbiased towards Opposition parties. From twitter political discourse, it is found that parties discuss themselves or their opposing parties and seldom issues of national interest. The research and the proposed robust model can be extended to other social media and be analysed for a bigger social network.