Multi-view consistency for multi-hop knowledge base question answering

Abstract

The task of Knowledge Base Question Answering (KBQA) is to answer a question in natural language over a Knowledge Base. And multi-hop KBQA aims to reason over multiple hops of facts in KB to answer a complex question. Step-wised reasoning has been an important schema to solve multi-hop KBQA. But previous approaches suffer from lacking reasoning paths, causing models may answer in an incorrect way. To address the issue, we present a novel approach to enhance the KBQA model by leveraging consistency between different views of the data, with few intermediate-relation-labeled data. Previous retrieval-based methods proceeded by utilizing the data view of (question, intermediate entities, answer entities). In our method, we introduce the data view of (question, intermediate relations) and enhance the KBQA model through the consistency of different data views. Concretely, we first implement a question-to-intermediate relations(Q2R) model to obtain intermediate relations’ distributions. By utilizing a pretrained text generation model, it performs well using a small part of relation-labeled data. Then we devise a map function to map distributions of intermediate entities to distributions of intermediate. Finally, a constraint that metrics the consistency between the intermediate path distributions obtained from the Q2R model and the original KBQA model is constructed to enhance the KBQA model. Experiments over three datasets of multi-hop KBQA are conducted, and the results demonstrate the effectiveness of our method.