Named Entity Recognition for Swedish-Language Dementia Education Data: An Evaluation of Techniques and Methodologies
Year: 2025
Type: dissertation
Abstract: Recent advancements in Large Language Models (LLMs) have been remarkable, with such models increasingly integrated into daily life, often serving as digital tutors across a wide range of subjects. However, leveraging LLMs as effective, domain-specific, AI tutors requires structuring relevant data in ways that enable reliable knowledge retrieval. This paper explores various Named Entity Recognition (NER) techniques and evaluates their effectiveness in extracting entities from a dementia-related dataset. The evaluation involves benchmarking different methods of performing NER, including multiple LLMs performing NER, on a standardized test dataset using metrics such as precision, cost efficiency, and runtime. Results demonstrates that certain methods consistently outperform others, with each approach exhibiting distinct strengths and weaknesses across the chosen evaluation criteria. In terms of LLMs, GPT-4.1 performs the best with a F1-score of 46.92%, while LinkNER and SuperICL stands out from the other NER methods with high scores relative to the other models across the board, reaching an average of 57.64% and 56.07% in F1-score respectively. These two hybrid NER approaches are, however, two of the most expensive ones to use.
Authors: Eriksson, Viktor, Palmqvist, Love
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6983594720