In the past decade, social media platforms such as Twitter and Facebook have become integral to our social interactions and decision-making processes. This has made it increasingly important to understand how people interact with information online, as evidenced by the large revenues of big internet companies and its usage during political campaigns. Understanding these interactions can help develop new strategies to prevent harmful content and behavior, such as hate speech and misinformation spreading. However, communication in social media can be very noisy, making it difficult to detect and analyze relevant conversations. Although the research community has devoted a large amount of efforts to analyze social media, the research falls short to provide an operational framework that can be used to identify conversational text online. Traditional community detection methods - most of them, unsupervised clustering AI methods, may not be enough for analyzing conversations in social media due to the complexity and dynamism of the data. One of the challenges is that social media conversations can be highly heterogeneous and often do not conform to the typical patterns of community structures - sets of actors highly tied between them and with fewer connections with other communities. In this project we are set to develop new computational methodologies to identify and disentangle opinions in online communication (e.g., social media) by levering the use of already existing large language models.