| Abstract [eng] |
Modern social media platforms now operate at volumes far greater than their predecessors. They must bend to immense government and societal pressure to moderate and review content. Among these forms of content requiring heavy moderation by most platforms is hate speech. Most of these platforms perform some level of automated content filtering in order to meet the demands required from these aforementioned pressures. Telegram, as an exception to this rule, opts for a more ambiguous moderation strategy that awaits human or government request before content is manually reviewed. Because the platform suffers from governance issues, this thesis proposes the novel application of two methods. First, the use of machine learning models trained to detect hate speech applied in the context of a Telegram bot. Second, to analyze the performance of hate speech detection natural language models against data on which they were not trained. The results show not only that this practical implementation with a bot is possible, but also that the use of these models on novel data reveals the impact that training variation can have on performance. This enriches the academic understanding of platform native content detection strategies that would otherwise be proprietary. Structure: Introduction Literature Analysis Design Proposal Implementation Performance Results Conclusions Contents: 81 text page 34 pic. 17 table. 70 bibliography ref. |
