LLMs in a Knowledge Graph Pipeline: From Knowledge Extraction to SPARQL Querying. The Case of FashionDB
Keywords: Knowledge Graph Construction, Knowledge Extraction, Text to SPARQL, Knowledge Graphs Question Answering, Large Language Model
TL;DR: Describing a pipeline for a domain specific Knowledge Graph Construction and Interaction, exploiting LLMs. The domain of interest is fashion.
Abstract: The abundance of unstructured textual data available on the web presents both an opportunity and a challenge for knowledge extraction. In this work, we introduce FashionDB, a domain-specific knowledge graph designed to capture key information about fashion designers, fashion houses, and their collections. The main objective of this project is to bridge the gap between unstructured fashion-related texts and structured, query-able data. To achieve this, we develop a custom ontology that extends Wikidata properties to better represent the temporal and relational aspects of the fashion industry, such as career trajectories, collaborations, and fashion collection attributes. We employ large language models (LLMs) to automate knowledge extraction, leveraging in-context learning to extract facts and their temporal components from fashion-related texts. Furthermore, we present an approach to query generation by adapting LLMs for text-to-SPARQL translation, making FashionDB accessible even to users without expertise in SPARQL. Our work highlights the potential of LLMs not only for knowledge graph construction but also for facilitating natural language interactions with complex, domain-specific knowledge bases. Through this approach, we demonstrate the power of combining automated extraction with accessible querying methods to make structured fashion knowledge more readily available.
Submission Number: 19
Loading