In this research work the possibility of adapting image based anomaly detection into text based anomaly detection was explored. Two main approaches are being proposed, namely anomaly detection as a task of classification and unsupervised anomaly detection using text patches. Both approached explore the use of generative adversarial networks to perform anomaly detection and results presented show that such can be fruitful.

According to the Crimes Against Children Research Center, one in five U.S. teenagers who regularly use the Internet have received an unwanted sexual solicitation via the web. There is an increasing danger in online environments such as chat-rooms, where predatory behaviour is more and more frequent, creating an unsafe environment for minors. This project aims to design an approach for online communities to enhance their member's safety by detecting malicious conversations of sexual nature. This project joins the powers of computational linguistics with statistical machine learning to decipher the insight lying in conversations, then make predictions on whether or not a specific conversation should be flagged for containing sexual predatory behaviour. The contribution of this novel approach is 2-fold: firstly, the approach is able to capture the contextual details by putting an emphasis on insight that lies within the conversation, and secondly it contains a 2 stage classification system, which is highly flexible and customizable for detecting and classifying other malicious textual data.

In a world where customers can buy products with a few clicks online, future customers must consider the opinions and satisfaction levels of previous customers. In order to allow one to understand what previous customers have said, the design of an automated technique that summarizes opinions of thousands of customers is desirable. A promising technique has been developed that combines continuous vector representation models, natural language processing techniques and statistical machine learning models. This technique has been tested on labelled datasets and it extracts over 80% of opinions correctly. Future research can focus on improving the technique's limitations on edge cases.