Tekstiluokittelu kielimallien avulla : hierarkkinen moniluokkainen tekstiluokitus kompleksisessa tiedonhallintasuunnitelmassa esikoulutettujen kielimallien avulla
Year: 2025
Type: dissertation
Abstract: This thesis investigates the suitability of pre-trained language models for text classification tasks against public administration information management plans (IMP). In a hierarchical multi-class (HMC) text classification (TC) task, the ability of the models to find the correct document type from a hierarchical structure was investigated. The study analysed public sector documents related to IMP and evaluated the suitability of AI/ML and NLP techniques for the HMC-TC task. The work produced a prototype classifier based on Zero-Shot Classification (ZSC), with the aim of improving the efficiency of information classification, minimising errors and easing the cognitive load of users. The modular prototype classifier navigates the IMP hierarchy and provides classification recommendations without any separate language model training or fine-tuning. The objectives of the study were to explore the potential for producing a widely deployable solution without the need for customisation, high computational requirements, or reliance on third-party services. The key findings show that although ZSC models have potential, their performance depends on the quality of the data to be classified and the content of the documents. Multilingual language models with ZSC performance performed moderately well when using Finnish text and category options. The prediction accuracy of the first hierarchy level was around 35%, which is an improvement over the baseline accuracy (6.25%), but not sufficient for independent classification. The use of translations increases the classification time and does not provide accuracy improvements with the client material used. Further research is recommended to identify the best use cases for exploiting the technology.
Source: LUTPub (LUT University)
Author: Turunen, Joni
Institution: Lappeenranta-Lahti University of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7053011553