Ontology-Guided Evidence Path Inference for Multi-h... | AI Research

Ontology-Guided Evidence Path Inference for Multi-hop Knowledge Graph Question Answering
Knowledge graph question answering (KGQA) involves answering natural language questions by navigating structured facts. A common approach is to start from a topic entity and explore outward to find the answer. However, this often leads to "path explosion," where the system generates a massive number of irrelevant paths, or "semantic misalignment," where the paths found do not actually satisfy the specific requirements of the question. This paper introduces OPI, a framework designed to solve these issues by using an ontology-based guide to constrain the search and refine the reasoning process.

The Ontology Graph

The core of OPI is a "relation-centric ontology graph." Instead of looking at every individual fact in a knowledge graph, OPI creates a high-level map that tracks how different types of information connect. For example, it records that a "person" might be connected to a "country" via a "birthplace" relation. By abstracting the knowledge graph into these type-level signatures, OPI creates a compact interface that defines what kind of answer is expected, allowing the system to ignore paths that lead to the wrong category of information.

Bidirectional Retrieval

Traditional methods expand outward from a topic entity in one direction. OPI improves this by using a bidirectional retrieval mechanism. First, it predicts the likely type of the answer (e.g., a person or a location). It then uses the ontology graph to identify which relations are capable of ending at that specific type. By combining this "answer-side" constraint with the "topic-side" expansion, the system effectively prunes the search space. Instead of exploring every possible branch, it only pursues paths that are structurally capable of reaching the correct answer type, significantly reducing the number of noisy, irrelevant paths.

Iterative Refinement

Even after narrowing down the search, some paths may be technically correct in type but irrelevant to the specific question. To address this, OPI employs an iterative refinement strategy. A generator creates an answer hypothesis based on the retrieved paths, and a refiner evaluates this hypothesis against the question context. This feedback loop allows the system to reassess its findings and filter out evidence that, while compatible with the answer type, does not actually answer the user's specific query.

Performance and Results

OPI was tested on several standard benchmarks, including WebQSP, CWQ, and MetaQA. The results show that the framework significantly reduces the search space required to find answers. On the WebQSP and CWQ datasets, OPI improved performance metrics (Hit@1 and F1 scores) by notable margins over previous state-of-the-art methods. On the MetaQA benchmark, the retrieval module alone was able to achieve near-saturated accuracy, demonstrating that the ontology-guided approach is highly effective at identifying the correct reasoning evidence.