Measuring political Bias in British media: Using recurrent neural networks for long form textual analysis

Abstract

In this thesis we aim to explore methods of determining political bias in the traditional British print media. It can be shown that much of the British public perceive there to be explicit political biases in many of the UK's most popular media outlets. It is also known that people are inherently prone to political influence from their sources of news. Due to this reason, there is motivation to seek a means to formalise political bias in British media outlets.

In our study, we took the 2016 UK referendum of EU membership as the source to identify a political bias. We sought to find a means in which to determine on a sentence level, whether a newspaper identified with a pro-leave or pro-remain philosophy. For this, we used the newspapers explicit endorsements for a certain referendum outcome that are provided by the newspapers themselves as a ground truth.

Recurrent neural networks have been shown to be useful when working over data of varying sizes, such as encoded textual data. Recurrent neural networks have also been used to perform classification tasks over short form textual data in the scope of determining political bias. However, little work has been done on processing more long-form textual data for classification tasks within a political bias domain. Here, we sought to determine if recurrent neural networks would be a viable approach for solving this problem, compared to more traditional and simple approaches, such as a Naive Bayes model.

In our study, we were able to determine that recurrent neural networks are successfully able to determine a political bias in British media outlets. Our models also indicated slight biases in supposedly unbiased outlets, such as the BBC. The generated models were also able to transfer their learnings into new domains, such as determining EU membership political bias in more recent news articles. However, due to a lack of input data, and ground truths applied in a broad manner, traditional methods such as the Naive Bayes were able to achieve similar results to the recurrent neural networks, with much less compute power required.