Discovering Topical Aspects in Microblogs

We address the problem of discovering topical phrases or “aspects” from microblogging sites like Twitter, that correspond to key talking points or buzz around a particular topic or entity of interest. Inferring such topical aspects enables various applications such as trend detection and opinion mining for business analytics. However, mining high-volume microblog streams for aspects poses unique challenges due to the inherent noise, redundancy and ambiguity in users’ social posts. We address these challenges by using a probabilistic model that incorporates various global and local indicators such as “uniqueness”, “diversity” and “burstiness” of phrases, to infer relevant aspects. Our model is learned using an EM algorithm that uses automatically generated noisy labels, without requiring manual effort or domain knowledge. We present results on three months of Twitter data across different types of entities to validate our approach.