From Detection to Discourse: Tracking Events and Communities in Breaking News

Online social networks are now an established part of our reality. People no longer rely solely on traditional media outlets to stay informed. Collectively, acts of citizen journalism have transformed news consumers into producers. Keeping up with the overwhelming volume of user-generated content from social media sources is challenging for even well-resourced news organisations. Filtering the most relevant content, however, is not trivial. Significant demand exists for editorial support systems that enable journalists to work more effectively. Social newsgathering introduces many new challenges to the tasks of detecting and tracking breaking news stories. In detection, substantial volumes of data introduce scalability challenges. When tracking developing stories, approaches developed on static collections of documents often fail to capture important changes in the content or structure of data over time. Furthermore, systems tuned on static collections can perform poorly on new, unseen data. To understand significant events, we must also consider the people and organisations who are generating content related to these events. Newsworthy sources are rarely objective and neutral, and in some cases, purposefully created for disinformation, giving rise to the "fake news" phenomenon. An individual's political ideology will inform and influence their choice of language, especially during significant political events such as elections, protests, and other polarising incidents. This thesis presents techniques developed with the intention of supporting journalists who monitor social media for breaking news. Starting with the curation of newsworthy sources, through to implementing an alert system for breaking news events, tracking the evolution of these stories over time, and finally exploring the language used by different communities to gain insights into the discourse around an event. As well as detecting and tracking significant events, it is of interest to identify the differences in language patterns between groups of people around those events. Distributional semantic language models offer a way to quantify certain aspects of discourse, allowing us to track how different communities use language, thereby revealing their stances on key issues.