Year: 2025
Type: dissertation
Abstract: Named Entity Recognition (NER) is a key task in Natural Language Processing (NLP), enabling structured information extraction from unstructured text.It is especially important in domains such as healthcare, law, and finance.However, developing accurate NER models typically requires large annotated datasets, which are often unavailable or expensive to obtain in specialized or low-resource domains.A major challenge is how to create training data when manual annotation is impractical.While synthetic data offers a scalable alternative, generating diverse, and reliable synthetic datasets remains non-trivial.Additionally, there is a lack of standardized methods for evaluating how well such data supports downstream NER performance.In this thesis, a modular, prompt-based pipeline is developed for generating synthetic NER training data using large language models (LLMs).Five prompting strategies are explored: instruction-based, few-shot, role-based, template-based, and prompt chaining.The aim is to examine how different strategies influence data quality and model performance in low-resource settings.The approach involves generating, paraphrasing, and filtering synthetic sentences, followed by span-level annotation and model training using spaCy.Models are evaluated on held-out test sets using span-level precision, recall, and F1-score.Additional analyses explore error types and generalization to unseen entities.Results show that models trained on synthetic data can achieve strong performance, especially when using few-shot prompts with paraphrasing.This configuration reached an F1-score of 0.82 and demonstrated good generalization.However, template-only prompts, while structurally reliable, offered low lexical diversity, highlighting a broader trade-off between control and variation in prompting strategies.The contribution of this work is a transferable pipeline that reduces reliance on costly manual annotation in low-resource domains.Its main limitation is the lack of realdata validation, which restricts external validity.Nevertheless, the findings suggest that synthetic datasets, when carefully designed, can effectively support NER in lowresource domains and serve as a foundation for further research.
Author: Ocklind, Victor
Institution: Karlstad University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7111844721
