SGCC: A Knowledge Graph Construction Framework for Incomplete Schemas

This paper aims to address the common issue of “incomplete Schema” in Knowledge Graph Construction (KGC). Existing automated KGC frameworks based on Large Language Models (LLMs), such as EDC, typically require a choice between two extreme assumptions: either a fully predefined target Schema exists, or none exists at all. However, when dealing with structured documents such as national standards, we observe a binary nature of their Schema. One part describes general metadata and a stable, cross-domain Generic Schema; the other part describes core technical content and a Domain-Specific Schema that dynamically changes according to specific fields. Current KGC frameworks often force a choice between these two, and this binary nature of Schema results in a trade-off in applying existing methods. To address this challenge, we propose a new four-stage framework called “Semantically-Guided Collaborative Canonicalization (SGCC)”, which includes the processes of open information extraction, relationship definition, semantic path selection, and innovative collaborative canonicalization. After completing open information extraction and relationship definition, the process is scheduled by a prior-guided path selector. This approach not only utilizes known Generic Schema to align highly certain knowledge but also initiates a self-canonicalization process for the remaining uncertain knowledge, dynamically discovering and constructing domain-specific Schemas. We conducted experimental verification on a self-constructed dataset of Chinese national standard documents. The results show that SGCC can flexibly discover new knowledge while leveraging prior knowledge. Compared to baseline methods using a single canonicalization strategy, the knowledge graphs constructed by this framework demonstrate a significant improvement in quality.