This paper describes the system jointly developed by the Language Technologies Lab from INAOE and the Language and Reasoning Group from UAM for the Sexual Predators Identification task at the PAN 2012. The presented system focuses on the problem of identifying sexual predators in a set of suspicious chatting. It is mainly based on the following hypotheses: (i) terms used in the process of child exploitation are categorically and psychologically different than terms used in general chatting; and (ii) predators usually apply the same course of conduct pattern when they are approaching a child. Based on these hypotheses, our participation at the PAN 2012 aimed to demonstrate that it is possible to train a classifier to learn those particular terms that turn a chat conversation into a case of online child exploitation; and, that it is also possible to learn the behavioral patters of predators during a chat conversation allowing us to accurately distinguish victims from predators.
It is well known that the World Wide Web (WWW) has vastly penetrated into social
living and allows connecting people from different geographic regions through new forms
of communication. Examples of such communication forms are instant messaging
services, chat rooms, social networks (e.g., Facebook, Twitter, etc.) and blogs. These
services have become very popular tools for personal as well as for group communication,
as they are cheap, easy to use, virtual and private in nature [
Such online services allow users to hide their personal information behind the monitor; which, on the one hand, makes this type of communication a source of fun, but on the other hand, it also represents a threat. The privacy and virtual nature of these services augment the possibilities of some heinous acts which one may not commit in the real world. Examples of such acts are the online paedophiles who “groom” children, ⋆ This work was done under partial support of CONACYT (project grants 134186 and 106013).
We also thank SNI-Mexico, INAOE and UAM for their assistance.
that is, who meet underage victims online, engage in sexually explicit text or video chat
with them, and eventually convince the children to meet them in person. According
to the National Center for Missing & Exploited Children [
Traditionally, a term that is used to describe such malicious actions with a potential
aim of sexual exploitation or emotional connection with a child is referred as “Child
Grooming” or “Grooming Attack” [
Nowadays, the usual way to catch these sexual predators is for trained law enforcement officers or volunteers to pose as children in online chat rooms, thus predators fall into the trap and are identified. However, online sexual predators always outnumber the law enforcement officers and volunteers. An organization that employs this methodology to catch sexual predators is the Perverted Justice group, located in the United States. This organization has been able to convict more than 500 predators since 20044. Nevertheless, there is a great need for software applications that can flag suspicious online chats automatically, either as a tool to aid law enforcement officials or as parental control features offered by chat service providers.
In this paper we propose a novel methodology that faces the problem of sexual predators identification as a text classification task by means of a supervised approach. The identification process is divided in two main stages: the Suspicious Conversations Identification (SCI) stage and the Victim From Predator disclosure (VFP) stage. Performed experiments showed that it is possible to train a classifier to learn those particular terms that turn a chat conversation into a case of online child exploitation; and, that it is also possible to learn the behavioral patters of predators during a chat conversation allowing us to accurately distinguishing victims from predators.
The rest of the paper is organized as follows. Section 2 presents recent work on the task of sexual predators identification. Section 3 describes the proposed methodology for identifying both suspicious chat conversations and the sexual predator within a chat conversation. Section 4 presents the experimental settings as well as the results achieved in the context of the PAN 2012 competition. Finally, Section 6 depicts our conclusions and formulates directions for future work. 2
Traditionally, the problem of identifying grooming attacks has been tackled through
text classification strategies. In [
Similarly, in [
Finally, the work proposed by RahmanMiah et al. faces the problem of grooming
attack identification from a more general point of view [
Our proposal differs from previous works in that it attacks both problems at once, i.e., we are able to identify when a chat conversation is a case of child exploitation and subsequently we are able to tell which user is the sexual predator. Thus our proposal can be used for both purposes (detecting suspect conversations and identifying predators); besides, we show that the two-step approach outperforms a single-stage method.
The proposed method for detection of misbehaving users in chats is based on two main hypotheses: (i) terms used in the process of child exploitation are categorically and psychologically different than terms used in general chatting; and (ii) predators usually apply the same course of conduct pattern when they are approaching a child. Accordingly, we propose a new methodology for solving the problem of sexual predators identification, which is divided in two main stages: the Suspicious Conversations Identification (SCI) stage and the Victim From Predator disclosure (VFP) stage. Figure 1 shows the general architecture of the proposed system.
Notice that the goal of the first stage is to act as a filter, i.e., it helps distinguishing
between general chatting and possible cases of online child exploitation; in this way,
the set of conversations to be analyzed by the VFP module will be reduced. Hence we
can focus only on conversations that potentially include sexual predators for a more
fine grained analysis. Consequently, the goal of the second stage is to identify (to point
at) the potential predator from a possible case of child exploitation. The associated
classification problem is less complex than trying to discriminate between predators
and typical users directly, see Section 4.4.
Our proposed system does not include any module for preprocessing the texts, i.e.,
we did not remove any punctuation mark, stopwords and, neither apply a stemming
process. The main reason for not applying any preprocessing was because the text in
chat conversations had unique characteristics that distinguish them from any other type
of text [
We believe that using emoticons and intentional misspelled words may contain
valuable contextual information in a chat text. For example, in the grooming phase the
perpetrator may amend the relationship by an emphasized “soryyyyyyyyy” when the child
felt threatening by any obtrusive language. Another example may be the emoticon for
“hug (>:d<)” and “kiss (:-*)” for a soft introduction of sexual stage. Preserving such
information makes traditional language processing tools, such as stemmers and POS
taggers, unsuitable for processing the chat texts [
Although we did not apply any pre-processing stage, it is important to mention that we did apply a pre-filtering stage to all the conversations that were given for the PAN 2012 sexual predator competition. The goals of the pre-filtering stage were: i) to help us focusing only in the most important cases and, ii) to reduce the computational cost for automatically processing all the information.
This pre-filtering stage consisted in removing all the conversations that accomplish at least one of the following conditions: 1) Conversations that had only one participant, 2) Conversations that had less than 6 interventions per-user and 3) Conversations that had long sequences of unrecognised characters (apparently images). Table 1 shows information about the training data before and after applying this pre-filtering stage.
As it can be seen, by means of the pre-filtering stage we are able to reach a substantial reduction ratio (90% approximately) of conversations/users. It is important to notice that by doing this pre-filtering, we also removed a few sexual predators. Thus, even if our proposed system works perfectly we will not be able to identify the 100% of the sexual predators. Nevertheless, we think the information from interventions of removed predators was not enough to effectively recognize them as predators anyways. 3.3
As we have mentioned before, our system faces the problem of sexual predators identification as a text classification task (TC). Accordingly, for training the SCI classifier (Figure 1) we employed traditional TC techniques to construct a model that distinguish between general chatting and cases of child exploitation.
In order to properly train our SCI classifier, we labeled as suspicious conversations all the chat conversations were at least one predator appears, resulting in 5790 nonsuspicious conversations and 798 suspicious ones. During the experimentation phase, the SCI classifier represents the conversations by means of the bag of words representation (BOW) employing either a boolean or a TF-IDF weighting scheme. Since we did not apply any preprocessing stage to the chat texts, we obtained features vectors of 117015 elements. 3.4
Similarly to the SCI classifier, our VFP stage was designed using traditional TC techniques. The goal of the VFP classifier was to recognize sexual predators in suspicious chat conversations, as detected by the SCI method. For training the VFP classifier we divided all the text conversations, where a predator was involved, into interventions. This means that if a text chat involved two different users, we had two sets of interventions. Therefore, we used as examples of victims the interventions of the users that had a conversation with a predator, resulting in 194 examples of victims; and as examples of predators we used the interventions of the 136 users already labeled as predators. For the experiments performed, the VFP classifier employs a BOW representation using either a boolean or a TF-IDF weighting scheme. For the VFP classifier we obtained features vectors of 16709 elements. 4 4.1
For our experiments we used the data set provided in the context of the PAN 2012 Lab: Sexual Identification competition. As we mentioned in Section 3.2 we were given for training a total of 66928 different chat conversations, where 97690 different users are involved and only 148 are tagged as sexual predators (See Table 1). Additionally, a test data set for evaluation was provided. Such corpus contained 155129 chat texts, where 218702 different users are involved and only 250 are tagged as sexual predators. For a more detailed explanation on how the training and test corpora were constructed please visit http://pan.webis.de/. 4.2
Two classifiers from the CLOP toolbox [
During the development phase we adopted two-fold cross validation to estimate the performance of our methods using training data only. This validation technique was used for all of our experiments. For the final evaluation of our system we used the test data provided by organizers of PAN 2012, see Section 4.1. An analysis of the results using both training and test data sets is given in the following section. The evaluation of training-set results was carried out mainly by means of the classification accuracy, which indicates the overall percentage of text chats correctly classified. Additionally, due to the class imbalance, we also report results in terms of F1 measure. Regarding the final evaluation of the system on test data, we used the measures proposed by organizers, namely: F-0.5 measure, precision and recall. 4.3
As baseline we used the traditional paradigm for solving the problem of sexual predators identification. Figure 2 provides a general view of the baseline definition.
As can be noticed, the problem of identifying sexual predators is performed in one single step. For the baseline experiment we employed a BOW representation using either a boolean or a TF-IDF weighting scheme. By following the same procedure established for the SCI and the VFP classifiers, under this configuration we obtained features vectors of 117015 elements. 4.4
In this section we report experimental results obtained by the components of our twostep approach, as well as the results obtained with the baseline method, using training data. Next, in Section 4.5, we report the performance of the system in the test data set. Baseline results. Table 2 shows results obtained by the baseline configuration. In order to evaluate if a dimensionality reduction strategy could be helpful for the classifier, we performed several experiments varying the size of the features vectors that were used to train the classifier. For this purpose we employed the well known information gain (IG) method to rank and preserve the most distinctive features.
Recall that the size of the features vector is 117015 elements, hence using only 10% of the features means that the classifiers employed only 11,702 features to represent the 11,038 chat conversations. The baseline configuration faces the problem of learning a model that helps to classify between normal users and sexual predators from a highly unbalanced corpora, i.e., 10,902 normal users and only 136 sexual predators.
Table 2 indicates that using a binary weighting scheme allows a better performance for the classifier. It is also possible to observe that reducing the dimensionality of the feature’s vector does not allow a significant improvement. On the contrary, it decreases the ability of the classifier to accurately distinguish between normal users and sexual predators. Because of these results we decided to not use any dimensionality reduction strategy for both SCI and VFP classifiers.
SCI results. Table 3 shows results obtained from the SCI classifier. As mentioned in previous sections, the aim of this classifier is to distinguish between general chatting and possible cases of online child exploitation. It is worth mentioning that the training data employed by the SCI classifier represented an unbalanced corpus, since there are 5790 conversation labeled as general chatting and only 798 text chats labeled as cases of online child exploitation, however, obtained results showed that it is possible to accurately distinguish suspicious conversations.
Experimental results showed that both classifiers (i.e., SVM and NN) are suitable for solving the problem of classifying suspicious conversations. Contrary to the results obtained with the baseline configuration, the SCI classifier using SVM as classification method obtained better results when chat conversations are represented by means of a BOW considering a tf-idf weighting scheme.
We concluded from these experiments that, terms used in the process of child exploitation are categorically and psychologically different than terms used in general chatting, which allows to train a classifier to learn those particular terms and accurately detect cases of online child exploitation.
VFP results. Table 4 shows results obtained from the VFP classifier. As we mentioned in Section 3.4, the aim of this particular module is, once a conversation has been tagged as a suspicious, to point at the sexual predator, i.e., to tell which user is the victim and which one is the predator.
Obtained results showed that the proposed methods are adequate for solving the problem of classifying victims and predators. Similarly to the results obtained in the SCI classifier, using SVM as classification method obtained better results when a tf-idf weighting scheme was employed. Nevertheless, for the case of the VFP classifier, the best results were obtained when using NN algorithm and a binary weighing scheme.
From the experiments performed in this section we were able to show evidence that suggest predators apply the same course of conduct pattern when they are approaching a child, and that our proposed method it is able learn these behavioral patters of predators during a chat conversation allowing us to accurately distinguish victims from predators. 4.5
Previous sections described obtained results employing the training data set, and all experiments were performed in a controlled scenario. However, the main goal of the PAN 2012 Lab was to evaluate in real scenarios the proposed method for sexual predators identification. This section reports official results obtained by our system on test data from the PAN 2012 competition.
In order to apply our proposed methodology, the first step consisted on applying the filtering stage (Section 3.2) to test data. Table 5 shows some statistics of the test data before and after applying the filtering stage. From this table we can observe similar reduction ratios as in training data. A total of 28 predators were removed by our filtering approach. We manually analyzed the interventions of removed predators and found that most of them contained a few characters only; we consider that with such little information it is not possible to effectively identify to removed predators.
After filtering the test data, we were able to apply our proposed method (Figure 1). The next step was to represent all the remaining conversations (i.e., 15,330) into the generated model for the SCI classifier. Following, the chat conversations that the SCI classified as suspicious were divided into interventions and represented accordingly to the model generated for the VFP classifier.
Our team submitted three different runs: i) Baseline: it corresponds to the configuration showed in Figure 2 employing as classification method a Neural Network and using a binary weighting scheme with no dimensionality reduction; ii) SCI(NN-B)& VFP(NN-TF-IDF): it means that the SCI module was configured for using a NN and a binary weighting scheme, whereas the VFP module used a NN with a tf-idf weighting scheme; and iii)SCI(NN-B)& VFP(NN-B): it means that both the SCI and the VFP modules were configured for using a NN and a binary weighting scheme.
Official results of submitted runs are showed in Table 6. The leading evaluation measure was F-0.5 measure, which emphasizes the importance of the precision of the system, which is particularly important for sexual predator detection as mentioned by the organizers of the PAN 2012 competition.
As it is possible to observe, the best configuration was the third one, i.e., using in both modules a binary weighting scheme. Thus showing that the hypotheses of our work were right. Indeed the configuration SCI(N N − B)&V F P (N N − B) obtained the highest performance among the 16 teams that participated in the sexual predator identification track of PAN 2012. The closest entry (snider12-run-2012-06-16-0032) achieved an F-0.5 measure of 0.9168, whereas the average was of 0.5105. The results obtained by our group are promising and motivate us to pursuing several future work directions.
An additional task to that of sexual predator detection, proposed by organizers of PAN 2012, was to search for those lines (interventions) that reveal predator’s bad behavior. Traditionally, such lines are manually identified and used as evidence to convict paedophiles. PAN 2012 participants were encouraged to propose new ideas to automatically solve this issue.
We approached the line-detection task with a language-models based approach. Our
main idea was based in the following statement; it is well known that every predator
follows three main stages when approaching a child [
Our proposed solution works as follows: first we automatically divide all the conversations where a predator appears in three sections, such division is made without considering any type of contextual frontiers, i.e., we did not identify where a child approaching stage begins or ends. Next we generated the language model (lm) of the 2nd and the 3rd parts5. Finally, for a user that is tagged as predator, we compute the perplexity against the lm of each one of its interventions, and we delivered as the most distinctive lines of bad behaving those with the minor perplexity value.
For the competition, we proceed as follows: from the set of users labeled as sexual predators by our system (SCI(NN-B)& VFP(NN-B)), we select the 50 most distinctive lines and delivered as examples of predators bad behaving. Table 7 shows examples of the lines that were delivered. Official evaluation results indicated that by following this procedure we were able to identify just 1 revealing line. We believe that our proposed idea could be very effective, although more work is necessary in order to improve its performance. Although this result seems absolutely negative we have to mention that the winner participant of this sub-task submitted 63, 290 lines (grozea12-run-2012-0614-1706b).
From the 2nd parts From the 3rd parts <text>what do you want me to be?</text> <text>what do u want me to say</text> <text>what do u want me to say</text> <text>what do u want me to wear</text> <text>what do u want me to wear</text> <text>what do you want me to be?</text> <text>do u want to talk to me too</text> <text>do u want me to do it</text>
Table 7. Examples of lines with the minor perplexity values. 6
We have proposed a new methodology for detecting sexual predators in text chats. Our proposal differs from traditional approaches in that it divides the problem in two stages: 5 We used the SML toolkit http://svr-www.eng.cam.ac.uk/ prc14/toolkit.html to this end. the Suspicious Conversations Identification (SCI) stage and the Victim From Predator disclosure (VFP) stage. The goal of the first stage is to work as a filter, i.e., it helps distinguishing between general chatting and possible cases of online child exploitation; in this way, the set of conversations to be analyzed will be reduced, hence we can focus only on conversations that potentially include sexual predators for a more fine grained analysis. Consequently, the goal of the second stage is to identify (to point at) the potential predator from a possible case of child exploitation.
Performed experiments showed that it is possible to train a classifier to learn those particular terms that turn a chat conversation into a case of online child exploitation; and, that it is also possible to learn the behavioral patters of predators during a chat conversation allowing us to accurately distinguish victims from predators. Our participation in the PAN 2012 forum showed that the proposed methodology is able to produce very good results in a realistic scenario, obtaining an F-measure (β = 0.5) of 0.8936, which was the best ranked result among all of the participants.
As future work we plan to include some linguistic features, such as proposed by [