The Documentary Method and Artificial Intelligence

In the field of language processing by means of artificial intelligence (AI), Natural Language Processing (NLP), great progress has been made recently. The focus here is on so-called General Pretrained Transformer (GPT) models. These are language models that are pre-trained with machine learning on the basis of extremely large amounts of data (e.g. the entire Wikipedia). These models are already capable of answering arbitrary queries from users without special training and are able to carry out very complex tasks independently.

The DTEC-funded KISOFT project at the University of Munich is investigating the extent to which such models are suitable for supporting interpretation with the documentary method. This support can be provided, for example, by the AI 'pre-interpreting' passages from an interview or a group discussion. For this purpose, the CI in the project is taught by means of so-called finetuning on the basis of a larger number of human interpretations how to interpret according to the principles of the Documentary Method (e.g. the separation of formulating and reflecting interpretation). It is planned and already very advanced to implement an AI query in DokuMet QDA, by means of which the interpreters can obtain suggestions for the analysis of their materials during their interpretation work, in a way analogous to a research workshop, except that here the suggestions do not come from colleagues but from the AI.

For further reading:

• Schäffer: Möglichkeiten und Grenzen der Optimierung von Verfahren Tiefer Interpretation durch Softwareunterstützung
• Schäffer & Lieder: Distributed interpretation – teaching reconstructive methods in the social sciences supported by artificial intelligence"
• Lieder & Schäffer: Qualitative Methodenausbildung zusammen mit generativen Sprachmodellen. Zur Verteilten Interpretation in hybriden Forschungswerkstätten