Crimes emerge out of complex interactions of human behaviors and situations. Linkages between crime events are highly complex. Detecting crime linkage given a set of events is a highly challenging task since we only have limited information, including text descriptions, event times, and locations; moreover, the link among incidents is subtle. In practice, there are very few labeled data about related cases. We present a statistical modeling framework for police reports based on multivariate marked spatio-temporal Hawkes processes. Inspired by the notion of modus operandi (M.O.) in crime analysis, we perform text embedding for police reports and treat embedding vectors as marks of the Hawkes process. The embedding is performed by regularized Restricted Boltzmann Machine (RBM) to promote the probabilistic sparsity of selected keywords. Numerical results using real data demonstrate the competitive performance and interpretability of our method. The proposed method can be used in other similar data in social networks, electronic health records.