Motivated by the real-world application of categorizing system log messages into defined situation categories, this paper describes an interactive text categorization method, PICCIL1, that leverages supervised machine learning to reduce the burden of assigning categories to documents in large finite data sets but, by coupling human expertise to the machine learning, does so without sacrificing accuracy. PICCIL uses keywords and keyword rules both to preclassify documents and to assist in the manual process of grouping and reviewing documents. The reviewed documents, in turn, are used to refine the keyword rules iteratively to improve subsequent grouping and document review. We apply PICCIL to the problem of assigning semantic situation labels to the entries of a catalog of log events to support on-line labeling of log events.