Large language models as tools for detection and mitigation of polarization on social media.

Abstract

Social networks play a central role in the lives of many people, serving as platforms for communication, social interaction, sharing and consumption of information, and entertain- ment. These platforms connect individuals on a global level and allow the rapid dissem- ination of news and cultural trends, thus becoming important tools that serve for political use. However, these networks are also associated with worrying problems that affect society, such as the sharing of fake news, echo chambers, and the increase of political polarization. Since the structure and algorithms of social networks are frequently pointed out as the main causes of these problems, most of the proposed solutions focus on them, thus lacking ap- proaches that consider the content and that can be applied to automatic moderation. Besides being scarce, the proposals directed at textual content tend to treat polarization in social networks in a one-dimensional way, limiting themselves, for example, to the detection of toxic language or sentiment analysis. In this thesis, we propose the use of Large-Scale Lan- guage Models as tools for the identification and mitigation of polarized textual content in social networks. For this, we instructed language models based on a comprehensive, opera- tionalizable definition grounded in the literature on textual polarization, applying them to the detection and mitigation of polarization characteristics in tweets previously collected in sce- narios of intense polarization, such as presidential elections. As a result, we found that the language models are consistent in their binary classifications and present good performance in the identification task. Furthermore, experiments with automatic evaluation and a survey with human participants demonstrated that these models are also effective in reducing the polarization perceived in texts.