A Fine Granular Relational Reasoning KGQA model Based on Weak Supervised Learning

Knowledge graph question-answering (KGQA) task aims to provide answers in the form of entities from the knowledge graph for natural language questions. Conventional information retrieval modeling methods focus on controlling the size of the candidate reasoning path set through effective relationship filtering. However, these methods often perform iterative relational reasoning solely based on the overall semantic information of the question, neglecting the local relevance of relationships within the question statement and the value of fine granular information for filtering. KGQA task datasets typically lack annotations for inference paths, and existing weak supervised learning strategies may not effectively eliminate the misleading influence of erroneous reference reasoning paths on model reasoning capabilities. Based on the fundamental ideas of “reducing the size of candidate path sets” and “eliminating erroneous reference reasoning paths,” this paper proposes a knowledge graph question-answering model for fine granular relational reasoning based on weak supervised learning. First, it enhances relationship and path filtering capabilities from a model structure perspective and then, on this basis, improves the model’s ability to differentiate erroneous reference reasoning paths from a supervised strategy perspective. Compared to state-of-the-art KGQA models, the weak supervised fine granular relational reasoning KGQA model proposed in this paper achieves significant performance improvement.