Semantic Structure Based Query Graph Prediction for Question Answering over Knowledge Graph

Proceedings of the 29th International Conference on Computational Linguistics, pages 1569–1579 October 12–17, 2022. 1569 Semantic Structure based Query Graph Prediction for Question Answering over Knowledge Graph Mingchen Li Department of Computer Science Georgia State University, Atlanta, USA mli33@student.gsu.edu Shihao Ji Department of Computer Science Georgia State University, Atlanta, USA sji@gsu.edu Abstract Building query graphs from natural language questions is an important step in complex ques- tion answering over knowledge graph (Com- plex KGQA). In general, a question can be cor- rectly answered if its query graph is built cor- rectly and the right answer is then retrieved by issuing the query graph against the KG. There- fore, this paper focuses on query graph gen- eration from natural language questions. Ex- isting approaches for query graph generation ignore the semantic structure of a question, re- sulting in a large number of noisy query graph candidates that undermine prediction accura- cies. In this paper, we define six semantic structures from common questions in KGQA and develop a novel Structure-BERT to pre- dict the semantic structure of a question. By doing so, we can first filter out noisy candi- date query graphs, and then rank the remaining candidates with a BERT-based ranking model. Extensive experiments on two popular bench- marks MetaQA and WebQuestionsSP (WSP) demonstrate the effectiveness of our method as compared to state-of-the-arts. The source code can be found at https://github.com/ ToneLi/SSKGQA. 1 Introduction Knowledge graph (KG) is a graph structured database (Miller, 1995), in which nodes represent entities (e.g., Hedgehog in the Fog, Sergei Kozlov), and edges reflect the relations between entities (e.g., Hedgehog in the Fog - written_by - Sergei Kozlov). Users can get crisp answers by querying KGs with natural language questions, and this pro- cess is called Question Answering over Knowledge Graph (KGQA). Recently, consumer market wit- nesses a widespread application of this technique in a variety of virtual assistants, such as Apple Siri, Google Home, Amazon Alexa, and Microsoft Cortana, etc. Early works (Bordes et al., 2015; Golub and He, 2016) on KGQA mainly focus on simple ques- directed_by_reverse written_by Question 1: Which person wrote the films directed by Yuriy Norshteyn? Question 2: What is Cher's son's name? Topic entity Yuriy Norshteyn Hedgehog in the Fog Answer Sergei Kozlov person.children Topic entity Cher Answer Chaz Bono Constraint Male person.gender Figure 1: (Q1) Example question involving multi-hop reasoning, and (Q2) Example question with constraints tions, such as where is the hometown of Obama? This question only involves in one relation path (e.g., hometown or birth-place) in KG, and is rela- tively easy to solve. However, many questions in daily QA sessions are often more complex, man- ifested by multi-hop reasoning or questions with multiple constraints. Therefore, recently there is a flurry of interests on complex KGQA (Shi et al., 2021; Yadati et al., 2021). There are two types of complexity when deal- ing with complex KGQA, i.e., multi-hop ques- tions and questions with multiple constraints (See Figure 1 for example). Question 1 in Figure 1 is a …

Content truncated to 500 words