Challenges and Opportunities of LLM Integration in Danish Educational Schools
Author
Jørgensen, Mikkel Rune
Term
4. semester
Publication year
2024
Submitted on
2024-06-02
Pages
37
Abstract
This thesis examines how large language models (LLMs), with an emphasis on ChatGPT-4, could be integrated into Danish primary schools, focusing on the Danish language subject. It transfers benefits and concerns reported in international literature to a Danish context and tests the practical generation of teaching materials. The study combines qualitative semi-structured interviews with teachers and quantitative evaluations of LLM-produced assignments. In-context prompting techniques (few-shot, chain-of-thought and ReAct) and a retrieval-augmented generation (RAG) setup tailored to Danish curricular objectives were used to improve relevance and timeliness. Teacher interviews indicate limited knowledge of what LLMs can provide and a clear need for guidelines to support responsible implementation. Two teachers reviewed generated materials and judged them broadly comparable to teacher-created resources and adaptable across difficulty levels, suggesting potential for personalization and teacher support. At the same time, the work highlights challenges around data accuracy, ethics and privacy, and teacher adoption. Overall, the findings point to both opportunities and prerequisites for using ChatGPT-4 to customize and enhance educational materials in Danish primary education.
Dette speciale undersøger, hvordan store sprogmodeller (LLM’er), med særligt fokus på ChatGPT-4, kan integreres i den danske folkeskole med udgangspunkt i faget dansk. Fordele og bekymringer fra international litteratur overføres til en dansk kontekst, og den praktiske generering af undervisningsmaterialer afprøves. Studiet kombinerer kvalitative semi-strukturerede interviews med lærere og kvantitative vurderinger af LLM-frembragte opgaver. In-context-teknikker (few-shot, chain-of-thought og ReAct) samt en retrieval-augmented generation (RAG)-opsætning tilpasset danske læringsmål blev anvendt for at øge relevans og aktualitet. Lærernes interviews peger på begrænset viden om, hvad LLM’er kan tilbyde, og et tydeligt behov for retningslinjer for en ansvarlig implementering. To lærere gennemgik de genererede materialer og vurderede dem overordnet som sammenlignelige med lærerproducerede ressourcer og egnede til at variere sværhedsgrad, hvilket peger på potentiale for personalisering og lærerstøtte. Samtidig fremhæver arbejdet udfordringer omkring datanøjagtighed, etik og privatliv samt læreradoption. Samlet peger fundene på både muligheder og forudsætninger for at bruge ChatGPT-4 til at tilpasse og forbedre undervisningsmaterialer i den danske folkeskole.
[This apstract has been generated with the help of AI directly from the project full text]