Dynamic Microcluster Chains in Microtext

Two features of microtext that challenge language processing tools are addressed in the context of linking messages in the emergency response domain. First, the effect of very short texts on several classifiers is estimated by comparing the results when classifiers are applied to the full text of news reports vs. only the headlines. These experiments demonstrate a decrease of 5 - 20% in accuracy. A second challenging feature of microtexts is their accumulation in real time, which can be massive for sources such as Twitter. A dynamic hierarchical clustering algorithm that clusters messages as they accumulate is described, and a preliminary experiment in clustering tweets is demonstrated.