Modeling mobile users' behavior can lead to crucial applications in accountable mobile computing such as casual authentication and anomaly detection. We introduced a language approach to model mobile users' behavior from heterogeneous sensor data. By converting temporal and spacial features generated from WiFi RSS trace into symbols and fusing them into a 1-dimension "language'' representation, we were able to leverage algorithms developed for statistical NLP to build accountable user mobility models in an office WLAN environment. We explored the continuous n-gram and skipped n-gram models to detect anomaly of mobile user's behavior, such as device theft. We have collected data from network infrastructure in an corporate office environment over 5 days. The proposed model only needs to observe the users for 8 hours to build a reliable behavior model which can detect 86% of device theft cases. We have also evaluated the effectiveness of using the n-gram models to predict the future location of the user.