Reproduce “Knowledge Base Question Answering by Case-based Reasoning over Subgraphs” on OpenResearch

About this paper

Rajarshi Das, Ameya Godbole, Ankita Naik, Elliot Tower, Robin Jia, Manzil Zaheer, Hannaneh Hajishirzi, Andrew McCallum

CBR-SUBG introduces a semiparametric framework for Knowledge Base Question Answering that combines nonparametric retrieval of similar queries with parametric graph neural network reasoning over adaptively collected subgraphs. This approach demonstrates superior generalization to complex reasoning patterns and new entities, achieving state-of-the-art or competitive results on MetaQA, WebQuestionsSP, and FreebaseQA.

Key insights

• Complex reasoning patterns, while varied, tend to recur for different entities within their respective subgraph neighborhoods, making case-based reasoning an effective approach for generalization.
• Representing entities using sparse, relation-based features and relative distance embeddings enables inductive reasoning over new and unseen entities, crucial for adaptability to evolving knowledge bases.
• Contrastive learning can effectively train GNNs to identify analogous answer positions within structurally similar subgraphs, leveraging information from retrieved cases to improve prediction accuracy.

Summary by alphaXiv.