DSM: Question Generation over Knowledge Base via Modeling Diverse Subgraphs with Meta-learner

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4194 - 4207 December 7-11, 2022 ©2022 Association for Computational Linguistics DSM: Question Generation over Knowledge Base via Modeling Diverse Subgraphs with Meta-learner Shasha Guo1, 2, Jing Zhang1, 2∗, Yanling Wang1, 2, Qianyi Zhang1, 2, Cuiping Li1, 2, Hong Chen1, 2 1School of Information, Renmin University of China, Beijing, China 2Key Laboratory of Data Engineering and Knowledge Engineering of Ministry of Education {guoshashaxing, zhang-jing, wangyanling, qianyizhang, licuiping,chong}@ruc.edu.cn Abstract Existing methods on knowledge base question generation (KBQG) learn a one-size-fits-all model by training together all subgraphs with- out distinguishing the diverse semantics of sub- graphs. In this work, we show that making use of the past experience on semantically simi- lar subgraphs can reduce the learning difficulty and promote the performance of KBQG mod- els. To achieve this, we propose a novel ap- proach to model diverse subgraphs with meta- learner (DSM). Specifically, we devise a graph contrastive learning-based retriever to iden- tify semantically similar subgraphs, so that we can construct the semantics-aware learning tasks for the meta-learner to learn semantics- specific and semantics-agnostic knowledge on and across these tasks. Extensive experiments on two widely-adopted benchmarks for KBQG show that DSM derives new state-of-the-art per- formance and benefits the question answering tasks as a means of data augmentation. Codes and datasets are available online1. 1 Introduction In recent years, knowledge base question gener- ation (KBQG) has attracted substantial research interests as it shows great promise to improve the quality of question answering (QA). Specifically, KBQG can augment training data for QA sys- tems (Chen et al., 2020; Indurthi et al., 2017), and it can also motivate the machines to actively ask questions in human-machine conversations (Sun et al., 2018b; Zeng and Nakano, 2020). Concretely, KBQG generates natural language questions according to a set of facts extracted from KB, where each fact is typically specified as a triplet (e, r, e′) meaning entity e has relation r with entity e′. Previous efforts on KBQG can be cate- gorized into template-based models (Seyler et al., ∗Corresponding author. 1https://github.com/RUCKBReasoning/DSM 2017) and neural network-based (NN-based) mod- els (Bi et al., 2020; Kumar et al., 2019). The former ones heavily depend on hand-crafted templates, re- sulting in low scalability as these templates are lim- ited to narrow domains. Alternatively, NN-based models address this issue via inputting the set of triplets about a certain answer into a Seq2Seq ar- chitecture to automatically generate the question. In fact, for generating a question, triplets about a certain answer can naturally form a subgraph as illustrated in Figure 1. We observe that subgraphs differ in their semantics, which is especially shown in the relations that express the triplets2 as well as the structural patterns such as chain, star, and triangle3. Existing efforts do not distinguish the se- mantics of different subgraphs but learn a one-size- fits-all model by training together all subgraphs, which increases the learning difficulty. Inspired by humans who solve a problem by searching the …

Content truncated to 500 words