Förbättring av proprietär tillgångsåtervinning genom språkmodellsembeddings: En metod för likhetssökning
Year: 2024
Type: dissertation
Abstract: This thesis addresses the challenge of enhancing proprietary asset retrieval through the application of language model embeddings, a cutting-edge area that merges linguistic data processing with algorithmic search techniques. The primary problem tackled is the inefficiency and inaccuracy of traditional keyword-based search mechanisms in large, complex databases. The study proposes a similarity search approach utilizing language model embeddings to improve precision and speed in retrieving digital assets.A novel area of research is about the possibility of retrieving different types of assets from a single text input.Multiple embedding models are tested to assess the impact of different architectures and pre training setups.The main challenge was the development of a framework that integrates transformer language models with similarity search algorithms to identify and retrieve assets based on contextual relevance rather than mere keyword matching.Results showed that when performing similarity search with embeddings the underlying descriptions that have to be compared to the queries are crucial in improving retrieval accuracy. Further improvement such as applying Named Entity Recognition or fine tuning are also valuable in such aspect.This thesis demonstrates that language model embeddings offer a viable and superior alternative to traditional search technologies, paving the way for more intelligent and efficient retrieval systems in the field of digital asset management.
Author: Pittaro, Francesco
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7028577216