Constrained Entity Selection under Partial Knowledge for LLM-Based Knowledge Graph QA
A research paper introduces Constrained Entity Selection under Partial Knowledge (CES-PK), a problem formulation for LLM-based knowledge graph question answering that uses lightweight symbolic constraints to verify candidate answers without requiring executable logical forms. It employs a three-valued constraint semantics (satisfied, violated, unknown) to handle incomplete knowledge graphs under open-world assumptions. The framework is instantiated over the Hetionet knowledge graph.
The paper addresses limitations in LLM-based KGQA, where full semantic parsing is brittle and LLM reasoning lacks formal guarantees. CES-PK focuses on eliminating invalid answers and providing symbolic support for valid ones using constraints derived from the question. The three-valued semantics prevents incorrect rejections when the knowledge graph is incomplete. The approach is demonstrated on Hetionet.
The method shifts from generating executable queries to post-hoc verification of candidate answers using symbolic constraints. The three-valued logic (satisfied, violated, unknown) is key for open-world reasoning, allowing the system to avoid false negatives when data is missing. The instantiation on Hetionet suggests applicability to biomedical or heterogeneous knowledge graphs.
This approach could improve reliability of LLM-based question answering over enterprise knowledge graphs, where incomplete data is common. It offers a middle ground between brittle semantic parsing and unconstrained LLM generation, potentially reducing hallucinated answers in domains requiring factual grounding.
For organizations using knowledge graphs, this method could lower the cost of deploying LLM-based QA by reducing the need for complete schema mapping and improving answer trustworthiness. It may be particularly valuable in regulated industries where answer provenance is critical.
Next signals include empirical evaluation results on Hetionet, comparisons with existing KGQA baselines, and potential extensions to other knowledge graphs or constraint types. Adoption in production systems may depend on demonstrated accuracy improvements and computational overhead.