Ontology-Guided Extraction Framework

The core challenge in constructing knowledge graphs (KGs) from heterogeneous, unstructured documents is the lack of structural consistency and the prevalence of redundant or conflicting information. This paper proposes an extraction layer that leverages a predefined ontology to constrain and guide the extraction process. By anchoring the extraction in a formal schema, the system ensures that entities and relationships are mapped to a standardized taxonomy, reducing the noise typically associated with open-information extraction.

Integrated Deduplication for Data Integrity

Unlike traditional pipelines that treat extraction and deduplication as sequential, decoupled steps, this approach integrates deduplication directly into the extraction layer. By performing entity resolution and conflict detection during the extraction phase, the system prevents the propagation of duplicate nodes and conflicting edges into the final graph. This 'deduplication-aware' design improves the overall quality and reliability of the resulting knowledge graph, as it resolves ambiguities—such as different surface forms referring to the same real-world entity—before they are committed to the graph structure.

Handling Heterogeneous Data

The framework is specifically designed to handle heterogeneous document sources, which often vary in format, domain, and linguistic style. By using the ontology as a semantic bridge, the extraction layer can normalize information across these disparate sources. This ensures that even when documents use different terminology or structures, the extracted data remains interoperable within the unified knowledge graph, providing a robust solution for large-scale, multi-source knowledge integration.