Multi-hop question answering over incomplete knowledge graph with abstract conceptual evidence
Qi Sun,Chunhong Zhang,Zheng Hu,Zhihong Jin,Jibin Yu,Liping Liu
DOI: https://doi.org/10.1007/s10489-023-04849-1
IF: 5.3
2023-08-12
Applied Intelligence
Abstract:Multi-hop Question Answering over Knowledge Graph (KGQA) aims to reason answers through multiple triples in Knowledge Graphs (KGs). Unfortunately, in practice due to the rapid growth of information, KGs are often incomplete with missing relations, increasing the difficulty of reasoning for multi-hop KGQA. Most prior works utilize auxiliary text corpus enhancement to tackle the challenge of incomplete KG. However, they fail to consider the fine-grained semantic structure contained in texts, which would be critical to reveal the missing relations of KGs. Meanwhile, they ignore semantic associations among multiple texts, leading to the blocking of the reasoning process along texts. To address these problems, we present the idea of extracting question-candidate evidences and parsing them into Conceptual Evidences (ACEs) to explicitly capture fine-grained semantic information. Accordingly, we propose an Conceptual Evidence Reasoning (ACER) model, exploiting the semantic associations among ACEs to form multi-hop evidence reasoning paths from the question to candidate answers. Moreover, a Statement-Evidence Reasoning method is introduced to further aggregate evidences and a joint entity scoring method for answer selection. Extensive experiments demonstrate the effectiveness of the ACER by its outperformance on different extents of KG incompleteness, as well as its interpretability of the answer reasoning process.
computer science, artificial intelligence