TY - GEN
T1 - LQRPP-KGQA
T2 - 6th International Conference on Neural Computing for Advanced Applications, NCAA 2025
AU - Wu, Tenan
AU - Liu, Hai
AU - Wang, Fu Lee
AU - Hao, Tianyong
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.
PY - 2025
Y1 - 2025
N2 - Knowledge Graph Question Answering (KGQA) aims to answer natural language questions over Knowledge Bases. Information Retrieval Based methods usually contain two stages: question-specific subgraph retrieval and answer reasoning. Most recent methods focus on the relevance between retrieved entities or relations and questions, which often results in excessively large subgraphs, making it difficult for graph neural networks to effectively reason. Since Large Language Models (LLMs) demonstrate exceptional capabilities in understanding complex natural language queries, we propose LQRPP-KGQA, a novel approach that leverages LLMs to predict question-specific reasoning paths and enhance reasoning over subgraphs. Specifically, we present a retrieval-predict-expansion-reason framework, where it first retrieves valid reasoning paths from the Knowledge Bases to conduct instruction fine-tuning. Furthermore, we define several reasoning sub-paths to prompt the fine-tuned LLMs to predict potential sub-paths, which are then expanded into question-specific subgraphs. Finally, we leverage graph neural networks to perform reasoning over these enriched subgraphs. We conduct KGQA experiments on two widely used datasets to validate its functionality. Extensive experiments on two benchmark KGQA datasets demonstrate that LQRPP-KGQA achieves competitive performance on KGQA tasks.
AB - Knowledge Graph Question Answering (KGQA) aims to answer natural language questions over Knowledge Bases. Information Retrieval Based methods usually contain two stages: question-specific subgraph retrieval and answer reasoning. Most recent methods focus on the relevance between retrieved entities or relations and questions, which often results in excessively large subgraphs, making it difficult for graph neural networks to effectively reason. Since Large Language Models (LLMs) demonstrate exceptional capabilities in understanding complex natural language queries, we propose LQRPP-KGQA, a novel approach that leverages LLMs to predict question-specific reasoning paths and enhance reasoning over subgraphs. Specifically, we present a retrieval-predict-expansion-reason framework, where it first retrieves valid reasoning paths from the Knowledge Bases to conduct instruction fine-tuning. Furthermore, we define several reasoning sub-paths to prompt the fine-tuned LLMs to predict potential sub-paths, which are then expanded into question-specific subgraphs. Finally, we leverage graph neural networks to perform reasoning over these enriched subgraphs. We conduct KGQA experiments on two widely used datasets to validate its functionality. Extensive experiments on two benchmark KGQA datasets demonstrate that LQRPP-KGQA achieves competitive performance on KGQA tasks.
KW - Knowledge Graph Question Answering
KW - Large Language Model
KW - Question-specific Subgraph
UR - https://www.scopus.com/pages/publications/105022715751
U2 - 10.1007/978-981-95-3739-6_8
DO - 10.1007/978-981-95-3739-6_8
M3 - Conference contribution
AN - SCOPUS:105022715751
SN - 9789819537389
T3 - Communications in Computer and Information Science
SP - 100
EP - 114
BT - Neural Computing for Advanced Applications - 6th International Conference, NCAA 2025, Proceedings
A2 - Zhang, Haijun
A2 - Tsang, Kim Fung
A2 - Wang, Fu Lee
A2 - Hung, Kevin
A2 - Hao, Tianyong
A2 - Wang, Zenghui
A2 - Wu, Zhou
A2 - Zhang, Zhao
Y2 - 4 July 2025 through 6 July 2025
ER -