Expert-Driven Topical Classification of Short Message Streams

We study the problem of expert-driven topical classification of short messages in time-evolving streams like Face book status updates, Twitter messages, and SMS communication. While high-level topics in these streams may be fixed (e.g., Sports, News), the content associated with these topics is typically less static, reflecting temporal change in interest as these streams evolve (e.g., tweets about the Olympics wane, while tweets about the World Cup rise in popularity). Coupled with this rapid concept drift, short messages themselves provide little contextual information and result in sparse features for effective classification. With these challenges in mind, we present an expert-driven framework for time-aware topical classification framework of short messages. Three of the salient features of the framework are (i) a novel expert-centric classifier, (ii) a sliding-window training for adaptive topical classification, and (iii) a suite of enrichment-based methods (lexical, link, collocation) for overcoming feature sparsity in short messages.