Subgraph Retrieval Enhanced Model for Multi-hop Knowledge Base Question Answering

Subgraph Retrieval Enhanced Model for Multi-hop Knowledge Base Question Answering Jing Zhang1, Xiaokang Zhang1, Jifan Yu2, Jian Tang3, Jie Tang2, Cuiping Li1∗, Hong Chen1 1School of Information, Renmin University of China, Beijing, China, 2Department of Computer Science and Technology, Tsinghua University, Beijing, China 3Mila - Quebec AI Institute. {zhang-jing,zhang2718,licuiping,chong}@ruc.edu.cn, yujf18@mails.tsinghua.edu.cn, jian.tang@hec.ca, jietang@tsinghua.edu.cn Abstract Recent works on knowledge base question an- swering (KBQA) retrieve subgraphs for easier reasoning. The desired subgraph is crucial as a small one may exclude the answer but a large one might introduce more noises. However, the existing retrieval is either heuristic or inter- woven with the reasoning, causing reasoning on the partial subgraphs, which increases the reasoning bias when the intermediate supervi- sion is missing. This paper proposes a train- able subgraph retriever (SR) decoupled from the subsequent reasoning process, which en- ables a plug-and-play framework to enhance any subgraph-oriented KBQA model. Exten- sive experiments demonstrate SR achieves sig- nificantly better retrieval and QA performance than existing retrieval methods. Via weakly su- pervised pre-training as well as the end-to-end fine-tuning, SR achieves new state-of-the-art performance when combined with NSM (He et al., 2021), a subgraph-oriented reasoner, for embedding-based KBQA methods. Codes and datasets are available online1. 1 Introduction Knowledge Base Question Answering (KBQA) (Zhang et al., 2021) aims to seek answers to factoid questions from structured KBs such as Freebase, Wikidata, and DBPedia. KBQA has attracted a lot of attention, as the logically organized entities and their relations are beneficial for inferring the answer. Semantic parsing-based (SP-based) methods (Das et al., 2021; Lan and Jiang, 2020; Sun et al., 2020) and embedding-based methods (He et al., 2021; Sun et al., 2018, 2019) are two mainstream methods for addressing KBQA. The former ones heavily rely on the expensive annotation of the intermediate logic form such as SPARQL. Instead of parsing the questions, the later ones directly represent ∗Corresponding author. 1https://github.com/RUCKBReasoning/SubgraphRetrievalKBQA 1K 2K 5K 10K Subgraph size 30 40 50 60 70 80 90 100 Answer Coverage Rate (%) WebQSP CWQ (a) Answer Coverage Rate 1K 2K 5K 10K Subgraph size 30 40 50 60 70 80 90 100 Hit@1 of QA (%) WebQSP CWQ (b) Hits@1 of QA Figure 1: The impact of subgraph size on (a) answer coverage rate and (b) QA performance (Hits@1) of NSM (He et al., 2021) on WebQSP (Yih et al., 2016) and CWQ (Talmor and Berant, 2018). and rank entities based on their relevance to input questions. Among them, the models which first retrieve a question-relevant subgraph and then perform reasoning on it (He et al., 2021; Sun et al., 2018, 2019) reduce the reasoning space, showing superiority compared with reasoning on the whole KB (Chen et al., 2019a; Saxena et al., 2020; Xu et al., 2019) (Cf. Table 2 for empirical proof). Subgraph retrieval is crucial to the overall QA performance, as a small subgraph is highly likely to exclude the answer but a large one might intro- duce noises that affect the QA performance. Fig- ure 1(a) presents the …

Content truncated to 500 words