Bachelor's Thesis: Replacing Compression Models with Large Language Models for Authorship Verificati

Background and Motivation Authorship verification (AV) addresses the question of whether two texts were written by the same author. Although it is a comparatively young field of research, many established AV methods were developed before the emergence of large language models (LLMs). In recent years, numerous new LLM-based approaches to authorship verification have been proposed. Nevertheless, depending on the task and corpus, classical methods continue to deliver competitive performance. So far, however, little research has been done on how existing classical methods can be further developed or adapted through the use of LLMs. Compression-based AV methods such as OCCAV (1), COAV (2) or NNCD (3) are particularly interesting in this respect. They rely on statistical compression models whose functionality could, in principle, be replaced or extended by modern LLMs. Goal The goal of this thesis is to investigate, for several classical AV methods, to what extent key components can be replaced or supplemented by LLMs. A particular focus lies on compression-based methods. The next-token probabilities provided by LLMs are to be used instead of conventional compression models and integrated into existing AV methods. In addition, the thesis will examine the relationship between the modeling or compression quality of an LLM and the resulting performance of the AV method. Results from related tasks suggest that this relationship is not necessarily monotonic (4). Expected Results As part of the thesis, several classical AV methods will be conceptually adapted to the use of large language models and implemented as prototypes. The resulting variants will then be evaluated on several suitable AV corpora and compared with the respective classical methods. In particular, the impact of different LLMs and their modeling quality on verification performance will be analyzed.