Implicit Relation Linking for Question Answering over Knowledge Graph

Findings of the Association for Computational Linguistics: ACL 2022, pages 3956 - 3968 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Implicit Relation Linking for Question Answering over Knowledge Graph Yao Zhao1, Jiacheng Huang1, Wei Hu1,2,∗, Qijin Chen3, Xiaoxia Qiu3, Chengfu Huo3, Weijun Ren3 1State Key Laboratory for Novel Software Technology, Nanjing University, China 2National Institute of Healthcare Data Science, Nanjing University, China 3Alibaba Group, China {yaozhao.nju,jchuang.nju}@gmail.com, whu@nju.edu.cn, {qijin.cqj,xiaoxia.qxx,chengfu.huocf,afei}@alibaba-inc.com Abstract Relation linking (RL) is a vital module in knowledge-based question answering (KBQA) systems. It aims to link the relations expressed in natural language (NL) to the corresponding ones in knowledge graph (KG). Existing meth- ods mainly rely on the textual similarities be- tween NL and KG to build relation links. Due to the ambiguity of NL and the incomplete- ness of KG, many relations in NL are implic- itly expressed, and may not link to a single relation in KG, which challenges the current methods. In this paper, we propose an implicit RL method called ImRL, which links relation phrases in NL to relation paths in KG. To find proper relation paths, we propose a novel path ranking model that aligns not only textual in- formation in the word embedding space but also structural information in the KG embed- ding space between relation phrases in NL and relation paths in KG. Besides, we leverage a gated mechanism with attention to inject prior knowledge from external paraphrase dictionar- ies to address the relation phrases with vague meaning. Our experiments on two benchmark and a newly-created datasets show that ImRL significantly outperforms several state-of-the- art methods, especially for implicit RL. 1 Introduction The past few years have witnessed the rapid devel- opment of knowledge-based question answering (KBQA), which aims to answer natural language (NL) questions over knowledge graph (KG), e.g., DBpedia (Auer et al., 2007) and Freebase (Bol- lacker et al., 2008). In many KBQA systems (Singh et al., 2018; Kapanipathi et al., 2021), as a funda- mental module, relation linking (RL) seeks to link a relation phrase expressed in NL to a corresponding relation in KG, which has a great impact on the ac- curacy of subsequent steps like query construction and even the entire KBQA systems. ∗Wei Hu is the corresponding author. Question: Where is the tv show Dragonaut: The Resonance from? Gold Relation Path: dbp:publisher -> dbo:country dbr:Dragonaut:_The_Resonance dbr:Shueisha dbr:Japan dbp:publisher dbo:country dbr:Television_Osaka dbr:Jump_Square dbp:magazine dbo:country Edges existing in KG Edges missing in KG dbr:TX_Network dbo:broadcastNetwork dbo:network Figure 1: An example of RL to DBpe- dia. There is no explicit relation between dbr:Dragonaut:_The_Resonance and dbr:Japan. We expect to implicitly link the phrase “from” to an indirect relation path dbp:publisher → dbo:country. Previous studies (Mulang et al., 2017; Singh et al., 2017; Naseem et al., 2021) focus on the similarities between the relation phrases and the text descriptions (e.g., local names) of relations in KG, and use the textual measures to link them. We refer to these studies as explicit RL, because they all assume that the relations in NL …

Content truncated to 500 words