Application of Language Models to Suspect Prioritisation and Suspect Likelihood in Serial Crimes

Language models are successfully applied to the problem of analysing crime descriptions from a police database with the purpose of prioritising suspects for an unsolved crime, given details of solved crimes. The frequency of terms in each description relates to the behaviour of the offender and this can be used to link crimes to a common offender. Language modelling uses Bayes' theorem and thus require a prior probability. Such a prior can be based on each offender's past propensity to offend, derived from historic data. Language modelling yields a probability of a document being relevant, which in this case is interpreted as the probability of a suspect being the culprit. Although the absolute value of the probability does not carry any direct applied implications, the study does show that the general likelihood of identification of the actual suspect does correspond to the relative values. Thus these probabilities can be used for more than just ranking suspects.