Förbättrade embeddings för offentlig upphandling : Nyttjandet av siamesiska neurala nätverk för förbättrad likhetsmatchning

Public procurement plays a vital role in modern economies, driving sectors from infrastructure to healthcare. However, the complex language, technical terminology, and legal structures found in procurement documents pose significant challenges for standard Natural Language Processing (NLP) models. This thesis addresses this gap by developing a domain-specific Siamese Neural Network (SNN) trained on public procurement data, aiming to improve similarity matching performance compared to OpenAI’s general-purpose text-embedding-3-small model. The study, conducted in collaboration with Tendium AB, evaluates whether a domain-specific approach can better capture the nuances of procurement language. The SNN was trained using contrastive loss to enhance its ability to distinguish between matching and non-matching pairs of procurements, with performance metrics including accuracy, precision, recall, and ranking quality. Results show that the SNN provides more precise class separation and improved recall, indicating a stronger alignment with the specific linguistic challenges of procurement texts. This highlights the potential of domain-specific embeddings to improve text matching in specialized contexts, offering practical benefits for public procurement platforms by reducing manual effort and improving retrieval accuracy.