Word-level Models to Distinguish Sex Offenders from Victims and Differentiate Among Offender Types

The Internet has become a dominant form of online communication and has become a method used to identify, seek out, and contact children for purposes of sexual exploitation. This research examined textual chat utterances seeking differences that would allow researchers to develop acomputer model capable of classifying a person as either a victim or child based solely on the words contained in the chats. Using text analysis methods, predictive keywords were identified that could distinguish between the text-based communications of pseudo-children (volunteers or undercover officers posing as children) and sexual offenders. A model based on these keywordswas able to classify pseudo-children with up to 96% accuracy and sexual offenders with up to 100% accuracy. A similar model also was developed to successfully distinguish two subtypes of offenders based on their motives for engaging with children online. These results represent a significant step in the categorization of textual data and show that categories of individuals can be distinguished based solely on their use of specific utterances.