StereoRel: Relational Triple Extraction from a Stereoscopic Perspective

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4851–4861 August 1–6, 2021. ©2021 Association for Computational Linguistics 4851 StereoRel: Relational Triple Extraction from a Stereoscopic Perspective Xuetao Tian1, Liping Jing1,∗, Lu He2, Feng Liu2 1 Technology Beijing Key Lab of Traffic Data Analysis and Mining 2 Engineering Research Center of Network Management Technology for High Speed Railway School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China, 100044 {xttian, lpjing, helu19, fliu}@bjtu.edu.cn Abstract Relational triple extraction is critical to under- standing massive text corpora and construct- ing large-scale knowledge graph, which has at- tracted increasing research interest. However, existing studies still face some challenging is- sues, including information loss, error prop- agation and ignoring the interaction between entity and relation. To intuitively explore the above issues and address them, in this paper, we provide a revealing insight into relational triple extraction from a stereoscopic perspec- tive, which rationalizes the occurrence of these issues and exposes the shortcomings of exist- ing methods. Further, a novel model is pro- posed for relational triple extraction, which maps relational triples to a three-dimension (3- D) space and leverages three decoders to ex- tract them, aimed at simultaneously handling the above issues. Extensive experiments are conducted on five public datasets, demonstrat- ing that the proposed model outperforms the recent advanced baselines. 1 Introduction Relational triple is a common structural represen- tation of semantic facts. A triple is always in form of (subject, relation, object), where subject and ob- ject are two entities connected by a type of prede- fined semantic relation. Relational triple extraction from unstructured texts is critical to understand- ing massive text corpora and constructing large- scale knowledge graph (Ren et al., 2017; Wei et al., 2020), which is widely concerned in recent years. Early researches (Zhou et al., 2005; Chan and Roth, 2011; Zhang et al., 2017) first recognize entities and predict the relations for each entity pair. Such approaches suffer from error propaga- tion problem and thus recent researches (Zheng et al., 2017; Zeng et al., 2018; Fu et al., 2019; ∗Corresponding author: Liping Jing. Nayak and Ng, 2020; Wei et al., 2020; Liu et al., 2020) try to build a jointly-decoding schema for entities and relations. However, relational triple ex- traction still faces the following challenging issues: • Information loss (I-IL). Information loss in- cludes entity incompleteness (Zeng et al., 2020) and entity overlapping (Zeng et al., 2018; Wei et al., 2020). Entity incomplete- ness (I-IL-EI) refers to that only head or tail token rather than completed entity is recog- nized, while entity overlapping (I-IL-EO) is that one entity belonging to multiple triples cannot be marked. • Error propagation (I-EP). Error propagation comes from the prediction process with strict order. For examples, pipeline models (Zhang et al., 2017; Takanobu et al., 2019) recognize entities first and predict relations based on each specific entity pair. Generative models (Zeng et al., 2018, 2019) extract subject, ob- …

Content truncated to 500 words