Detecting new, informative propositions in social media

The ever growing quantity of online text produced makes it increasingly challenging to find new important or useful information. This is especially so when topics of potential interest are not known a-priori, such as in “breaking news stories”. This thesis examines techniques for detecting the emergence of new, interesting information in Social Media. It sets the investigation in the context of a hypothetical knowledge discovery and acquisition system, and addresses two objectives. The first objective addressed is the detection of new topics. The second is filtering of non-informative text from Social Media. A rolling time-slicing approach is proposed for discovery, in which daily frequencies of nouns, named entities, and multiword expressions are compared to their expected daily frequencies, as estimated from previous days using a Poisson model. Trending features, those showing a significant surge in use, in Social Media are potentially interesting. Features that have not shown a similar recent surge in News are selected as indicative of new information. It is demonstrated that surges in nouns and news entities can be detected that predict corresponding surges in mainstream news. Co-occurring trending features are used to create clusters of potentially topic-related documents. Those formed from co-occurrences of named entities are shown to be the most topically coherent. Machine learning based filtering models are proposed for finding informative text in Social Media. News/Non-News and Dialogue Act models are explored using the News annotated Redites corpus of Twitter messages. A simple 5-act Dialogue scheme, used to annotate a small sample thereof, is presented. For both News/Non-News and Informative/Non-Informative classification tasks, using non-lexical message features produces more discriminative and robust classification models than using message terms alone. The combination of all investigated features yield the most accurate models.