Towards an automated classifier for the moderation of Maltese news blog comments

Abstract

The issues related with offensive and insulting discourse on the web has become ubiquitous for the detriment of people of all ages, especially young people and children. News portals do not usually require logging-in to access and this exposes anyone to any offensive discourse present on the website. Unless protection is afforded with moderation with the best type of moderation being that made by people. Moderators are an added cost for any commercial organisation and a tool which could minimise these costs but most of all minimise a moderator's time by flagging potentially offensive comments would be ideal. In this research we first tried to find a suitably labelled gold-standard dataset but none were available. We therefore found and considered a number of datasets which we thought as being suitable. We also investigated ways to build our own in order to provide a means to validate our software moderator tool { T.A.C.TI.C. which uses a number of machine learning classifiers. We evaluated this tool using available datasets and validation techniques and results showed that there are aspects of our approach which have given satisfactory results with consistently low false negatives but with higher false positives. These results could be improved with the use of a gold-standard dataset and a better understanding of the data.