Property Modelling for Product Ontology using Vector Embeddings driven by LLMs and OCR

Lectures Property Modelling for Product Ontology using Vector Embeddings driven by LLMS and OCR March 21, 2024 Nikhil Acharya PoolParty Summit Identifying entities and relationships from heterogeneous data sources in the context of technical documentation is an important part of building a knowledge database. Technical data consists of tables, raw texts and images for various products. We use pre-trained LLM and OCR models to identify products and product attributes from these sources. The extracted product information is now disambiguated using vector embeddings and mapped to specific entities and relationships in our PIM ontology. This use of AI tools helps us build a much more concrete knowledge base for our customers compared to standard data transformation approaches that only work with structured data and are rule-based.