Network operators often use a trouble ticket system to track all the steps of troubleshooting and maintenance activities. The history of trouble tickets carries valuable information for network management. Unfortunately, trouble tickets are not easy to mine, because they basically contain the human description of a problem in free text. Analyzing hundreds of trouble tickets by hand is too cumbersome; it is then important to have the support of some automatic mechanism. This paper proposes TroubleMiner, a mechanism based on document clustering techniques to automatically mine network trouble tickets. We show the utility and generality of TroubleMiner by applying it to thousands of trouble tickets from two research backbone networks. Network operators can apply TroubleMiner to trouble tickets from their network to analyze general trends in network incidents and maintenance activities. Researchers can use the insights from our characterization to decide which troubleshooting or maintenance techniques are most needed.