From Natural Language Texts to RDF Triples: A Novel Approach to Generating e-Commerce Knowledge Graphs

André Gomes Regino, Rodrigo Oliveira Caus, Victor Hochgreb, Júlio Cesar dos Reis

Abstract: The use of Knowledge Graphs (KGs) has gained traction in various software systems to improve data handling. E-commerce platforms, for example, have leveraged KGs to perform a range of tasks, including advertisement and product recommendation. However, generating accurate, trustworthy, and scalable RDF triples for populating KGs remains a challenging and error-prone task, especially given the lack of accurate and complete data. This article presents the QART framework, a natural language processing-based approach for generating RDF triples from e-commerce product Q &A. Our QART framework leverages templates to extract entities and intents and generates summarized sentences useful to populate KGs. Our experimental results demonstrate results concerning how we fine-tuned and used few-shots prompts in models such as T5, PTT5, GPT Neo and Bloom in an e-commerce dataset for the summarization task. Our evaluations identified key challenges in building the framework. Our contribution paves the way for the development of automatic mechanisms for text-to-triple transformation in e-commerce systems.