Sexual predators have found a new manner of select victims through the use of social networks. It is relatively easy for these predators to pretend to be a child or teenager with the aim of obtain the confidence of their victim. Due to this fact, there exist diverse attempts for detecting these kind of behaviors by analyzing conversations in chat rooms.
Major research works take Pendar [1] as a reference point for this topic. Here, the authors use a dataset gathered from a website named "Perverted Justice"1 for conducting a study using automatic text categorization techniques for identifying online sexual predators. More recently, Villatoro [2] performed conversation filtering by removing the shortest conversations with unintelligible characters or those conversations in which the chat participants have a very low number of interventions. With this pre-processing step, it is possible drastically to reduce the amount of texts contained in the training set.
The methodology proposed for searching/identifying sexual predators is shown in Figure 1. This proposal is made up of two steps: 1) A classification process that allows to discriminate those conversations in which a sexual predator participate; and, 2) A classification process that allows to discriminate the predator's dialogues with respect to other participants. In both steps, we use the following classification algorithms: neural networks, random forests and decision trees. We built and used three lexical resources (dictionaries) for pre-processing step: emoticons (":-)" is normalized as "happy"), contractions("isn't" is normalized as "is not"), and SMS vocabulary ("10q" is normalized as "thank you"). Afterwards, we extracted the features using POS-tagger [3], we used every morphological feature as attribute2 . For the experiments carried out, we used the union of two different conversation sets: Perverted Justice used in [1] and the PAN 2012 Training set3 that contain conversations provided by the PAN 2012 conference committee, which are structured in XML format.
We have carried out experiments using several classification algorithms implemented in the Weka [4] tool. We selected the following classification algorithms that obtained the best results: Decision trees [5], Random forests [6], and Neural networks (BackPropagation algorithm).
For the evaluation of results, we use the Weka option "Use training set" for step 1 and "Cross-validation" with 10 folds for step 2. Subsequently, we obtain the Precision (P ), Recall (R) and F-Score (F ), and we used the best F -score model for using in the second step.
Table 1 shows the results obtained for every classifier in both steps. TRC is the total of retrieved instances, and TCP is the total of positive instances. Actually, from 2,353 positive conversations of the dataset, the models identified around 2,000 conversations. The best performance was obtained by Random Forest with, 0.983 of precision, 0.882 recall and 0.930 of F -measure, thus, identifying very well those dialogues in which a sexual predator participates.
Using the conversations detected by the Random Forest classifier, we have re-constructed the dialogues in which one sexual predator participates. Thus, the second corpus contains 480 conversations of sexual predators and 442 conversations of non-sexual predators (922 conversations). In this step, Random
We presented a two step system for detecting sexual predators on-line. The conversation representation using PoS tags allowed to identify terminology employed by sexual predators, as shown by the values obtained in the experiments.
The normalization of texts have had a high impact in the results obtained and, need to be further investigated. Additionally, we are interested on analizing new features that allow us to detect "all" the conversations in which a sexual predators participates.