This Dissertation proposes a methodology to extract knowledge from databases of police criminal records. The scope of the proposed methodology comprises the full cycle for treatment of the criminal records, from the extraction of word radicals, including the construction of specialized dictionaries to support entity extraction, up to the development of criminal scenarios shaped into a relationship matrix. The scenarios are converted into intelligence maps for the analysis of criminal connections and the discovery of knowledge aimed at investigating and clarifying crimes. The intelligence maps extracted are represented by grids which are subsequently treated as capacitated graphs. Analyses of the connections extracted are carried out using the shortest path method in graphs, self-organizing neural maps, and indicators of social relationships. The method proposed in this study helps revealing evidence that was concealed by the complexity of textual information, and discovering knowledge based on criminal connections by applying hybrid algorithms. The proposed methodology was tested using databases of criminal police records related to drug traffic organizations and crimes that caused major social disturbances in Rio de Janeiro, Brazil, from 1999 to 2003.