Meta-CQG: A Meta-Learning Framework for Complex Question Generation over Knowledge Graph

Anonymous

Abstract: Complex question generation (CQG) aims to generate questions involving multiple Knowledge Base (KB) relations or functional constraints. Existing methods train an encoder-decoder-based model to fit all questions. However, the questions in the real world exhibit an imbalanced distribution in many dimensions, such as question type, relation class, entity class, and query structure. This results in insufficient learning for minority class samples under different dimensions. To address this problem, we propose a meta-learning framework for complex question generation. It trains a unique generator for each sample via retrieving a few most related training samples, which can deeply and quickly dive into the content features (e.g. relation and entity) and structure features (e.g. query structure) of each sample. As retrieved samples directly determine the effectiveness of each unique generator, we design a self-supervised graph retriever to learn the potential features of samples and retrieve the most related samples according to multiple dimensions. We conduct experiments on both WebQuestionSP and ComplexWebQuestion, the results on the minority class of different dimensions have been significantly improved, which demonstrates the effectiveness of the proposed framework.

Paper Type: long