Year: 2024
Type: dissertation
Abstract: Large Language Models (LLMs) have revolutionized the field of natural language understanding and text generation, outperforming traditional models by leveraging their pre-training knowledge. However, their single reliance on data from the training phase can result in information that is outdated or inaccurate, particularly in the context of domain-specific knowledge. To address these challenges, this study suggests integrating LLMs with Knowledge Graphs (KGs), which are structured knowledge sources. We conducted two separate experiments to validate this approach. The first experiment utilized the Wikipedia page of Leonhard Euler to demonstrate the capability of LLMs to effectively transform unstructured text into structured information. This transformation allows the construction of retrievable KGs. The second experiment involved the development of three distinct strategies for retrieving relevant context based on a user?s query. The results showed that enriching LLMs with structured information obtained from KGs significantly reduces, if not completely eliminates, factual errors in domain-specific knowledge. Moreover, it enhances the recall rate achieved by LLMs that do not use KGs as content repositories. In conclusion, the results obtained within this work prove that, by combining LLMs with KGs, it is possible to achieve more reliable and accurate language models.
Author: Le Boeuf Fló, Andrés
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7111479714