On the Construction of Knowledge Graphs from Unstructured Text using Neuro-Symbolic Networks

Abstract

In the past few years, Knowledge Graphs (KGs) have emerged as crucial tools for organizing structured information and building complex queries that support applications such as Websites, search engines, recommendation systems, LLMs, and data integration platforms. For KGs to scale with these applications, domain experts must transition from the current practice of curating KGs manually to generating them automatically from the huge volumes of unstructured text generated daily by different sources on the Internet. However, the difficulty of understanding the nuances of natural language and the need to extract precise relationships between entities in unstructured text are some of the major challenges of automatic KG generation. In this work, this challenge is tackled by combining entity and relationship extraction with neural networks (NNs) and the coherent representation of knowledge using symbolic reasoning techniques. An NN combined with a symbolic reasoning module can then be used for the accurate generation of KGs from a variety of unstructured text sources. Our methodology involves designing a generic neuro-symbolic network (GNSN) for KG extraction from a given text and then implementing a variant using open-source libraries. The implemented GNSN variant is trained and evaluated on WebNLG+2020 Text-to-RDF benchmark with a measured precision of 82.2%, recall of 82.4%, and an F1 score of 82.3%, which is competitive with other neuro-symbolic networks reported in the literature. This work aims to advance the state of the art in automatic KG generation from unstructured text