This thesis will apply methods from corpus linguistics to compare Norwegian texts generated by large language models (LLMs) with human-written texts. The aim is to identify possible systematic differences in vocabulary, phraseology, and grammatical patterns, and to examine whether these differences vary across models and genres. Relevant questions include whether generated texts exhibit similar lexical diversity to human-written texts, which words and expressions are disproportionately frequent, and whether models reproduce the variation characteristic of particular text types.
The project will construct comparable corpora of generated and human-written Norwegian, focusing on a small selection of models and genres, with attention to matching topic, communicative purpose, and text length. The analysis will combine quantitative methods, such as keyword analysis, collocation analysis, and measures of lexical diversity, with qualitative examination of examples in context through concordance analysis. The thesis will contribute an empirical account of the linguistic characteristics of generated Norwegian and assess the usefulness of corpus-linguistic methods for studying them.
This project can potentially be tied to the ongoing Norbruk center (Senter for norsk fagspråk i bruk), within which a large corpus of scientific and educational texts are being collected. These texts provide a suitable starting point and thematic focus for the work, and it is possible to more explicitly investigate how Norwegian terminology is treated differently in the originals and in the generated texts.