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

Proceedings of the 29th International Conference on Computational Linguistics, pages 6105–6114 October 12–17, 2022. 6105 Meta-CQG: A Meta-Learning Framework for Complex Question Generation over Knowledge Bases Kun Zhang1,2, Yunqi Qiu4, Yuanzhuo Wang1,2,3∗, Long Bai2, Wei Li4, Xuhui Jiang1,2, Huawei Shen1,2, Xueqi Cheng2 1Data Intelligence System Research Center, Institute of Computing Technology, Chinese Academy of Sciences; 2School of Computer Science and Technology, University of Chinese Academy of Sciences; 3Big Data Academy, Zhongke; 4Baidu Inc. {zhangkun18z, wangyuanzhuo, bailong18b, jiangxuhui19g}@ict.ac.cn {shenhuawei, cxq}@ict.ac.cn {qiuyunqi, liwei85}@baidu.com Abstract Complex question generation over knowledge bases (KB) aims to generate natural language questions involving multiple KB relations or functional constraints. Existing methods train one encoder-decoder-based model to fit all questions. However, such a one-size-fits-all strategy may not perform well since complex questions exhibit an uneven distribution in many dimensions, such as question types, in- volved KB relations, and query structures, re- sulting in insufficient learning for long-tailed samples under different dimensions. To ad- dress this problem, we propose a meta-learning framework for complex question generation. The meta-trained generator can acquire uni- versal and transferable meta-knowledge and quickly adapt to long-tailed samples through a few most related training samples. To retrieve similar samples for each input query, we de- sign a self-supervised graph retriever to learn distributed representations for samples, and contrastive learning is leveraged to improve the learned representations. We conduct ex- periments on both WebQuestionsSP and Com- plexWebQuestion, and results on long-tailed samples of different dimensions have been sig- nificantly improved, which demonstrates the effectiveness of the proposed framework. 1 Introduction Question generation (QG) over knowledge base (KB) aims to generate natural language questions with structured KB queries, which has been widely used to improve the performance of question an- swering (QA) by data augmentation for the training corpora. It can also help chatbots ask questions dur- ing human-computer interaction. Traditional methods (Jia and Liang, 2016; Seyler et al., 2017) rely on hand-crafted rules and tem- plates to convert KB queries into questions, lead- ing to poor generalization. Recently, neural ap- proaches (Kumar et al., 2019; Chen et al., 2020) ∗*Corresponding author. 0 1000 2000 3000 4000 Chain-style Tree-style Ring-style Query Structure Label Frequency Query Structure 0 100 200 1 101 201 301 401 501 601 Relation Label Frequency Involved KB Relation Id 0 1000 2000 Query Structure Label Frequency Question Type Ent_1 Chain-style Tree-style Ent_1 Ent_2 Rel_1 Rel_2 Ring-style Ent_2 Ent_3 Rel_1 Rel_2 Ent_3 Ent_4 Ent_1 Ent_2 Ent_3 Rel_1 Figure 1: The distribution of question type (Typ-Con as type constraint, Ent-Con as entity constraint, Com-con as comparative constraint), relation type (the involved KB relation in the query) and query structure in training set of WebQSP. We also give the illustration of three types of query structure leveraged one encoder-decoder-based model to fit the entire training set, then used the trained model to generate questions in the testing phase. However, such a one-size-fits-all strategy may not perform well for generating complex questions which con- tain multiple KB relations or functional constraints, such as comparison and sorting, and …

Content truncated to 500 words