INTEGRATING LINGUISTIC KNOWLEDGE INTO DNNS: APPLICATION TO ONLINE GROOMING DETECTION Anonymous

Abstract

Online grooming (OG) of children is a pervasive issue in an increasingly interconnected world. We explore various complementary methods to incorporate Corpus Linguistics (CL) knowledge into accurate and interpretable Deep Learning (DL) models. They provide an implicit text normalisation that adapts embedding spaces to the groomers' usage of language, and they focus the DNN's attention onto the expressions of OG strategies. We apply these integrations to two architecture types and improve on the state-of-the-art on a new OG corpus.

1. INTRODUCTION

Online grooming (OG) is a communicative process of entrapment in which an adult lures a minor into taking part in sexual activities online and, at times, offline (Lorenzo-Dus et al., 2016; Chiang & Grant, 2019) . Our aim is to detect instances of OG. This is achieved through binary classification of whole conversations into OG (positive class) or neutral (negative class). This classification requires the ability to capture subtleties in the language used by groomers. Corpus Linguistic (CL) analysis provides a detailed characterisation of language in large textual datasets (McEnery & Wilson, 2003; Sinclair, 1991) . We argue that, when integrated into ML models, the products of CL analysis may allow a better capture of language subtleties, while simplifying and guiding the learning task. We consider two types of CL products and explore strategies for their integration into several stages of DNNs. Moreover, we show that CL knowledge may help law enforcement in interpreting the ML decision process, towards the production of evidences for potential prosecution. Our text heavily uses slang and sms-style writing, as many real-world Natural Language Processing (NLP) tasks for chat logs. Text normalisation methods were proposed to reduce variance in word choice and/or spelling and simplify learning, e.g. (Mansfield et al., 2019) for sms-style writing. However, they do not account for the final analysis goal and may discard some informative variance, e.g. the use of certain forms of slang possibly indicative of a user category. CL analysis provides with the preferred usage of spelling variants or synonyms. We propose to use this domain knowledge to selectively normalise chat logs while preserving the informative variance for the classification task. As demonstrated by the CL analysis in (Lorenzo-Dus et al., 2016) , the theme and immediate purpose of groomer messages may vary throughout the conversation, in order to achieve the overarching goal of entrapping the victims. Groomers use a series of inter-connected "sub-goals", referred to as OG processes here, namely gaining the child's trust, planning activities, building a relationship, isolating them emotionally and physically from his/her support network, checking their level of compliance, introducing sexual content and trying to secure a meeting off-line. The language used within these processes is not always sexually explicit, which makes their detection more challenging. However, CL analysis additionally flags some contexts associated to the OG processes, in the form of word collocations (i.e. words that occur within a same window of 7 words) that tend to occur more frequently in, and therefore can be associated with, OG processes. We propose to exploit the relations between the OG processes and their overarching goal of OG to improve the final OG classification. We use the CL identified context windows to guide the learning of our DNN. Our main contributions are: 1) We explore different strategies for integrating CL knowledge into DNNs. They are applied to two architecture types and demonstrated on OG detection, but may generalise to other NLP applications that involve digital language and/or complex conversational strategies. 2) The principle and several implementations of selectively normalising text through modifying a word embedding in support to classification. 3) The decomposition of conversation analysis into identifying sub-goals. Our DNN implicitly models the relations between these sub-goals and the conversation's overarching final goal. 4) A new attention mechanism for LSTM based on the direct stimulation of its input gates, with two proposed implementations. 5) A state-of-the-art (SoTA) and interpretable OG detector. 6) A new corpus for OG detection, to be publicly released on demand, and that extends PAN2012 with more conversations and with products of CL analysis.

2. RELATED WORK

Villatoro-Tello et al. ( 2012) detected OG chat logs using a DNN to classify binary bag-of-words. This simple approach highlights the importance of commonly used words amongst groomers which we exploit for selective text normalisation. This is emphasised in (Vartapetiance & Gillam, 2014; Hidalgo & Díaz, 2012) where a set of phrases are derived from the important features of a Naïve Bayes classifier to describe common behaviours among groomers. Liu et al. ( 2017) obtained the current OG detection SoTA using a word embedding for semantic of important words and an LSTM. Integrating domain knowledge into DNNs is often done with additional losses that assist with sparse and low quality data. (Muralidhar et al., 2018) penalise a DNN's output violating logical rules w.r.t. the input features. (Hu et al., 2018) use the posterior regularisation framework of (Ganchev et al., 2010) to encode domain constraints for generative models. A teacher-student architecture in (Hu et al., 2016) incorporates first-order logic rules to create an additional loss for the student network. Other works integrated prior knowledge in the design of the DNN architecture. In BrainNetCNN (Kawahara et al., 2017) , the convolutions of a convolutional neural network (CNN) are defined based on the graph data's locality to account for the brain's connectivity. The training procedure may also integrate priors without modifying the DNN's architecture. Derakhshani et al. ( 2019) use assisted excitation of CNN neurons in the images' areas of interest, thus providing both localisation and semantic information to the DNN. An attention mechanism was used in a supervised way to focus a DNN on important words in (Nguyen & Nguyen, 2018) . We experiment with these various approaches and adapt them to our domain knowledge and DNN architectures. Linguistic knowledge was integrated to learnt word embeddings in the past. Knowledge in the form of lexicons, that carry a manual categorisation and/or ranking of words, is combined with a learnt word embedding in (Margatina et al., 2019) . Three strategies are proposed, namely concatenating the lexicon and embedding features, and using the lexicon features to conditionally select or transform the word embeddings. In our study, we are concerned with a different type of linguistic knowledge. However, our modification of word embedding (Section 4.1) may also exploit this lexicon knowledge. 3 AUGMENTED PAN2012 DATASET PAN2012 (Inches & Crestani, 2012) is a standard corpus for OG detection. It was gathered from Omegle (one-to-one conversations, IRC (technical discussions in groups), and the Perverted Justice (PJ) websitefoot_0 (chat logs from convicted groomers interacting with trained adult decoys), with 396 groomers and 5700 / 216,121 OG / non-OG conversations. Some non-OG chat logs contain sexual wording, making the OG classification more challenging. Conversations are truncated to 150 messages each, which limits both CL and ML analyses. To resolve this limitation, we augment the corpus with full OG conversations and the addition of new groomers from PJ, totalling 623 groomers in 6204 OG conversations (same negatives which could not be augmented to fuller conversations due to no access to the original data). Final OG / non-OG conversations total an average (std) of 215 (689) / 13 (23) messages and 1010 (3231) / 94 (489) words, respectively. Statistics on the dataset content are in the sup. materials. PJ data is freely available online and was largely used in previous social science and NLP studies, thus its use does not raise any peculiar ethical concern. For a debate on its usability see (Chiang & Grant, 2019; Schneevogt et al., 2018) . Our dataset also includes the results of a CL analysis of the new corpus using the method described in (Lorenzo-Dus et al., 2016) , which involves a heavy use of manual analysis by CL experts. As part of data preparation for CL analysis, word variants are identified, which are either spelling variations (mistakes or intentional e.g. 'loool'→'lol'), or the same semantic meaning behind two terms (e.g. 'not comfy'→'uncomfortable'). These variants are not specific to OG, but rather reflect digital language,



http://perverted-justice.com

