Large language models for analyses in digital humanities

Currently, large language models (LLMs) are redefining methodological approaches in many scientific areas, including digital humanities in areas such as linguistics, social sciences, literary studies, historiography, journalism, political science, folkloristics, etc. Harnessing their immense potential to digest and analyse large amounts of text, LLMs can provide complex insights into complex phenomena. LLMs are pretrained on huge text corpora by predicting the next tokens. This does not make them immune to hallucinations and biases, requiring a human-in-the-loop approach, careful methodological design, and strict qualitative and quantitative evaluation. We explain the working of LLMs needed to understand their performance. In the context of complex phenomena analyses, we discuss the two most frequent adaptations of LLMs, fine-tuning and prompt engineering, on the example of social media analysis and folkloristics. We emphasise the need to establish trust in their performance when analysing complex phenomena, outlining the evaluation methodology.