Constrained Tuple Extraction with Interaction-Aware Network

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11430–11444 July 9-14, 2023 ©2023 Association for Computational Linguistics Constrained Tuple Extraction with Interaction-Aware Network Xiaojun Xue, Chunxia Zhang∗, Tianxiang Xu, Zhendong Niu School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China {xiaojunx, cxzhang, txxu, zniu}@bit.edu.cn Abstract Tuples extraction is a fundamental task for in- formation extraction and knowledge graph con- struction. The extracted tuples are usually rep- resented as knowledge triples consisting of sub- ject, relation, and object. In practice, however, the validity of knowledge triples is associated with and changes with the spatial, temporal, or other kinds of constraints. Motivated by this observation, this paper proposes a con- strained tuple extraction (CTE) task to guar- antee the validity of knowledge tuples. For- mally, the CTE task is to extract constrained tuples from unstructured text, which adds con- straints to conventional triples. To this end, we propose an interaction-aware network. Com- binatorial interactions among context-specific external features and distinct-granularity inter- nal features are exploited to effectively mine the potential constraints. Moreover, we have built a new dataset containing totally 1,748,826 constrained tuples for training and 3656 ones for evaluation. Experiments on our dataset and the public CaRB dataset demonstrate the supe- riority of the proposed model. The constructed dataset and the codes are publicly available.1 1 Introduction Tuples extraction task aims to extract knowledge tuples from unstructured texts, which is a funda- mental task for information extraction, knowledge graph construction, and so on (Cui et al., 2018; Jiang et al., 2019; Banerjee and Baral, 2020; Li et al., 2022). The extracted knowledge tuples are mainly represented in the form of (subject, rela- tion, object) (Jiang et al., 2019), and are usually acquired through named entity recognition (NER), relation extraction (RE) (Zhong and Chen, 2021; Li et al., 2020; Yahya et al., 2014), and open in- formation extraction (Open IE) (Jiang et al., 2019; Wang et al., 2022). With the rise of large-scale pre- training methods (e.g. BERT (Devlin et al., 2019)), ∗Corresponding author. 1https://github.com/ckgconstruction/ckg. the quality of knowledge triples has improved sig- nificantly. However, current knowledge triples lack con- straints for their authenticity. In practice, con- straints are ubiquitous in numerous domains, such as spatial, temporal, conditional and environmen- tal ones. Constraints are essential supplements to knowledge triples, and play an instructional role. Our deep investigation on literature indicates that this topic has not gained enough attention. Some works utilize temporal knowledge tuples to reflect temporal dynamics (Gracious et al., 2021; Jung et al., 2021). However, only temporal con- straints are not sufficient and general enough to guarantee the validity of knowledge triples. For instance, let us consider the following sentence: “Consuming the same power, the performance of ARM CPU is better than that of Intel CPU.”, only if the conditional constraint “consuming the same power” is satisfied, the knowledge triple “(ARM CPU, better than, Intel CPU)” is true. The above observations motivate us to add constraints to conventional …

Content truncated to 500 words