AbstractThe on-going problem of child grooming online grows year on year and whilst government legislation looks to combat the issue by levying heavier penalties on perpetrators of online grooming, crime figures still increase. Government guidance directed towards digital platforms and social media providers places emphasis on child safety online. As this research shows, government initiatives have proved somewhat ineffective. Therefore, the aim of this research is to investigate the scale of the of the problem and test a variety of machine learning and deep learning techniques that could be used in a novel intelligent solution to protect children from online predation.
The heterogeneity of online platforms means that a one size fits all solution presents a complex problem that needs to be solved. The maturity of intelligent approaches to Natural Language Processing makes it possible to analyse and process text data in a wide variety of ways. Pre-processing data enables the preparation of text data in a format that machines can understand and reason about without the need for human interaction.
The on-going development of Machine Learning and Deep Learning architectures enables the construction of intelligent solutions that can classify text data in ways never imagined. This thesis presents research that tests the application of potential intelligent solutions such as Artificial Neural Networks and Machine Learning algorithms applied in Natural Language Processing. The research also tests the performance of pre-processing workflows and the impact of pre- processing of both online grooming and more general chat corpora. The storage and processing of data via a traditional relational database management system has also been tested for suitability when looking to detect grooming conversation in historical data.
The research has tested the performance of Artificial Neural Networks, Recurrent Neural Networks, and Convolutional Neural Networks in sentiment analysis tasks to establish the overall sentiment of online grooming conversation. The results of the tests conducted have proved successful and have provided opportunity for future work adapting neural networks to classification tasks related to detecting online grooming conversation.
Document similarity measures such as Cosine Similarity and Support Vector Machines have displayed positive results in identifying grooming conversation, however, a more intelligent solution may prove to have better currency in developing a smart autonomous solution given the ever-evolving lexicon used by participants in online chat conversations
|Date of Award||2023|
|Supervisor||Colin Venters (Co-Supervisor)|