Optimering och utvärdering av språk- och visionsmodeller för klassificering av svenska myndighetsdokument
Year: 2025
Type: dissertation
Abstract: This thesis investigates document classification strategies for Swedish government documents using transformer-based models. We explore both unimodal approaches leveraging either textual or visual information and multimodal fusion architectures that combine both modalities. Through comprehensive experimentation, we demonstrate that fine-tuning pre-trained models significantly outperforms using general-purpose embeddings, indicating successful incorporation of domain-specific terminology. Our results show that the KB-SBERT model serves as an excellent foundation for Swedish document classification, while DeBERTa offers strong performance for multilingual applications. We introduce a novel complex early fusion architecture that achieves the highest classification accuracy (90.56%) by effectively integrating the strengths of both text and image modalities, outperforming all unimodal approaches. We also evaluate decoder-based models, finding that while they show promising results, their computational demands present significant deployment challenges. The study further examines the trade-off between model accuracy and computational efficiency, providing practical insights for real-world implementationin government settings. Our findings contribute to the understanding of multimodal document classification in low-resource languages and offer guidance for developing efficient classification systems that balance performance with computational constraints.
Source: DiVA (Linkoping University)
Authors: Åström, Vilgot, Norén, Albin
Institution: Linköping University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7112427374
