Personal users on Twitter frequently post observations about their immediate environment as part of the 500 million tweets posted everyday. These observations and their implicitly associated time and location data are a valuable source of information for monitoring objects and events, such as earthquake, hailstorm, and shooting incidents. However, given the informal and uncertain expressions used in personal Twitter messages, and the various type of accounts existing on Twitter, capturing personal observations of objects and events is challenging. In contrast to the existing supervised approaches, which require significant efforts for annotating examples, in this paper, we propose an unsupervised approach for filtering personal observations. Our approach employs lexical analysis, user profiling and classification components to significantly improve filtering precision. To identify personal accounts, we define and compute a mean user profile for a dataset and employ distance metrics to evaluate the similarity of the user profiles under analysis to the mean. Our extensive experiments with real Twitter data show that our approach consistently improves filtering precision of personal observations by around 22 %.
