Masterarbeit, Otto-Friedrich-Universität Bamberg, 2022

The spreading of misinformation increasingly poses a threat, especially in the online domain. With the advances in computational text generation techniques and in the context of Wikipedia, this thesis will answer the following question: Can contemporary Natural Language Generators be used to intentionally generate Wikipedia articles with embedded, authentic misinformation? For this purpose, the state of research about misinformation and the vulnerability of Wikipedia, as well as the foundations of Natural Language Generation (NLG) are reviewed. Five different, contemporary NLG models are implemented and trained on a dump of Wikipedia articles about English writers. Subsequently, they generate a variety of texts for both a quantitative and a qualitative evaluation. The quantitative outputs are then assessed by automatic NLG evaluation metrics, and the qualitative ones are reviewed manually. Even though two of the models, GPT-2 and GPT-2 (finetuned), achieve very good results concerning word use, grammar, authenticity, and can even produce a Wikipedia-like structure, overall the five implemented models do not pose a imminent threat to the online encyclopedia as they either lack the sophistication in sentence building or the stability in their performance. However, with these already good results, the further development of new NLG models should be closely monitored.