RelU-Net: Syntax-aware Graph U-Net for Relational Triple Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4208 - 4217 December 7-11, 2022 ©2022 Association for Computational Linguistics RelU-Net: Syntax-aware Graph U-Net for Relational Triple Extraction Yunqi Zhang1, Yubo Chen1∗, Yongfeng Huang1,2 1 Department of Electronic Engineering & BNRist, Tsinghua University, Beijing, China 2 Zhongguancun Laboratory, Beijing, China yq-zhang19@mails.tsinghua.edu.cn ybch14@gmail.com yfhuang@tsinghua.edu.cn Abstract Relational triple extraction is a critical task for natural language processing. Existing methods mainly focused on capturing semantic informa- tion, but suffered from ignoring the syntactic structure of the sentence, which is proved in the relation classification task to contain rich rela- tional information. This is due to the absence of entity locations, which is the prerequisite for pruning noisy edges from the dependency tree, when extracting relational triples. In this paper, we propose a unified framework to tackle this challenge and incorporate syntactic informa- tion for relational triple extraction. First, we propose to automatically contract the depen- dency tree into a core relational topology and eliminate redundant information with graph pooling operations. Then, we propose a sym- metrical expanding path with graph unpooling operations to fuse the contracted core syntactic interactions with the original sentence context. We also propose a bipartite graph matching objective function to capture the reflections be- tween the core topology and golden relational facts. Since our model shares similar contract- ing and expanding paths with encoder-decoder models like U-Net, we name our model as Re- lation U-Net (RelU-Net). We conduct experi- ments on several datasets and the results prove the effectiveness of our method. 1 Introduction Relational Triple Extraction (RTE) is defined as au- tomatically recognizing entity pairs and the seman- tic relations in the form of (subject, relation, object) from unstructured text. It is a critical task for natu- ral language processing, especially for constructing large-scale Knowledge Graphs (KGs) from unla- beled corpus (Dong et al., 2014). Recent work proposed several neural network methods to extract relational triples. For example, Zheng et al. (2017) formulated this task as sequence ∗Corresponding author. born John is in France and lives now in New York Relation Classification (RC) born_in (John, France) live_in (John, New York) born John is in France and lives now in New York John is born in France and now lives in New York. born John in France John France John New York born John in New lives York John New York France John, born_in, France John, live_in, New York Graph Pooling Extraction Classification Dependency Tree with Known Entities Dependency Tree with Unknown Entities Relational Triple Extraction (RTE) Entity-aware Pruning Pruned Structures Core Topology Relation Types Relational Triples Figure 1: An example of the dependency-based relation classification method and our syntax-aware relational triple extraction method. The dashed arrows indicate the location injection of the known entities when pruning the dependency tree. tagging problems but failed to extract overlapping triples. Wei et al. (2020) proposed a cascade tag- ging framework to solve the overlapping problem. Chen et al.(2021) proposed to extract the implicit relational triples which are …

Content truncated to 500 words