Learning Reasoning Patterns for Relational Triple Extraction with Mutual Generation of Text and Graph

Findings of the Association for Computational Linguistics: ACL 2022, pages 1638 - 1647 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Learning Reasoning Patterns for Relational Triple Extraction with Mutual Generation of Text and Graph Yubo Chen∗, Yunqi Zhang∗, Yongfeng Huang Department of Electronic Engineering & BNRist, Tsinghua University, Beijing, China ybch14@gmail.com yq-zhang19@mails.tsinghua.edu.cn yfhuang@tsinghua.edu.cn Abstract Relational triple extraction is a critical task for constructing knowledge graphs. Exist- ing methods focused on learning text pat- terns from explicit relational mentions. How- ever, they usually suffered from ignoring re- lational reasoning patterns, thus failed to ex- tract the implicitly implied triples. Fortunately, the graph structure of a sentence’s relational triples can help find multi-hop reasoning paths. Moreover, the type inference logic through the paths can be captured with the sentence’s sup- plementary relational expressions that repre- sent the real-world conceptual meanings of the paths’ composite relations. In this paper, we propose a unified framework to learn the rela- tional reasoning patterns for this task. To iden- tify multi-hop reasoning paths, we construct a relational graph from the sentence (text-to- graph generation) and apply multi-layer graph convolutions to it. To capture the relation type inference logic of the paths, we propose to understand the unlabeled conceptual expres- sions by reconstructing the sentence from the relational graph (graph-to-text generation) in a self-supervised manner. Experimental results on several benchmark datasets demonstrate the effectiveness of our method. 1 Introduction Relational triple extraction is defined as automat- ically recognizing semantic relations with triple structures (subject, relation, object) among multi- ple entities in a sentence. It is a critical task for natural language processing, especially for Knowl- edge Graph (KG) construction from unlabeled cor- pus (Dong et al., 2014). Recent work proposed several neural network methods to extract relational triples. For example, Zheng et al. (2017) proposed a sequence tagging scheme for this task but failed to extract overlap- ping triples. Wei et al. (2020) proposed to solve the ∗Equal contribution. George is Judy’s father and David’s grandfather. George Judy David father(father(⋅)) Relational Graph Sentence grandfather(⋅) ⟺ father(father(⋅)) Relational Reasoning Pattern Figure 1: An example of the relational graph and the relational reasoning pattern. Solid arrows of the rela- tional graph are golden relational triples. The dashed arrow is a two-hop reasoning path. overlapping triple problem with a binary tagging framework. Zeng et al. (2018) proposed to address this issue by generating triple element sequences with copy mechanism. Existing methods achieved considerable success in learning text patterns of relational triples from explicit mentions. However, they usually suffered from the failure of extracting the relational triples which are implicitly implied in the text (Zhu et al., 2019). This is because they ignored relational rea- soning patterns in natural language, which usually consist of finding multi-hop paths and inferring re- lation types along these paths. For example, in Fig- ure 1, the triple (“David”, “father”, “Judy”) is not explicitly expressed in the sentence and requires relational reasoning to be extracted. Unfortunately, the ignorance of relational reasoning patterns in ex- isting …

Content truncated to 500 words