With the rapid development of online social network (OSN), worldwide connections between each individual turn into much stronger. A huge number of information provided by some entities around the world are well dispersed in OSN every day. Most of those are useful but not all as anomalous entities utilize anomaly users to spread malicious content (like spam or rumors to achieve their pecuniary or political aims. In this paper, we propose a mechanism to detect such anomaly users according to the user profile and tweet content of each user. Besides the features considered in the past literature, we design several new features related to near-duplicate content (including lexical similarity, semantic similarity) to enhance the precision of detecting anomaly users. Utilizing the data by public honeypot dataset, the proposed approach deals with supervised learning approach to carry out the detection task.