Bias Analysis on Twitter

Resumen

Sentiment analysis on comments retrieved from social networks has become an area of great interest in recent years, both at academic and corporate level. This analysis, including trend analysis and control of hate posting, is required to moderate forums or even to eliminate spam, and only with the help of automation it is possible to efficiently process the enormous volume of information that is generated daily. An important application of this analysis is the detection of sexist bias in social networks, which can help to identify early behavior that could later lead to violent actions. However, there are various obstacles to the correct treatment of the messages and the detection of possible biases, among which the absence of context in the analysis of these messages and the use of tools and libraries focused on the English language stand out. In this article we study the issues associated with this automatic analysis, and we introduce the foundation of an analysis method for biases or prejudices in the messages retrieved from the Twitter network, considering the context, and using a mechanism to reduce dependencies with the analyzed language. A supervised classifier was developed whose objective is to be able to identify possible biases on any topic, taking into account context and minimizing the dependency on language. The method was tested in a case study for the detection of sexism, using idioms typical of the geographical area analyzed. In the experiments carried out, an accuracy of 73% was obtained, which can be improved by expanding the training set.