AtTGen: Attribute Tree Generation for Real-World Attribute Joint Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2139–2152 July 9-14, 2023 ©2023 Association for Computational Linguistics AtTGen: Attribute Tree Generation for Real-World Attribute Joint Extraction Yanzeng Li1,2, Bingcong Xue1, Ruoyu Zhang1, Lei Zou1,3∗ 1Wangxuan Institute of Computer Technology, Peking University. Beijing, China 2National Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China 3TopGraph.AI liyanzeng@stu.pku.edu.cn {xuebingcong, ry_zhang, zoulei}@pku.edu.cn Abstract Attribute extraction aims to identify attribute names and the corresponding values from de- scriptive texts, which is the foundation for ex- tensive downstream applications such as knowl- edge graph construction, search engines, and e-Commerce. In previous studies, attribute extraction is generally treated as a classifi- cation problem for predicting attribute types or a sequence tagging problem for labeling attribute values, where two paradigms, i.e., closed-world and open-world assumption, are involved. However, both of these paradigms have limitations in terms of real-world applica- tions. And prior studies attempting to integrate these paradigms through ensemble, pipeline, and co-training models, still face challenges like cascading errors, high computational over- head, and difficulty in training. To address these existing problems, this paper presents Attribute Tree, a unified formulation for real- world attribute extraction application, where closed-world, open-world, and semi-open at- tribute extraction tasks are modeled uniformly. Then a text-to-tree generation model, AtTGen, is proposed to learn annotations from different scenarios efficiently and consistently. Experi- ments demonstrate that our proposed paradigm well covers various scenarios for real-world applications, and the model achieves state-of- the-art, outperforming existing methods by a large margin on three datasets. Our code, pre- trained model, and datasets are available at https://github.com/lsvih/AtTGen. 1 Introduction Attribute Extraction (AE) is a practical applica- tion of the Information Extraction (IE) task, aim- ing to identify the attribute name and the corre- sponding attribute value from unstructured or semi- structured text fragments (Ghani et al., 2006; Ravi and Pasca, 2008; More, 2016). Figure 1 shows a typical product profile with extracted attribute tags. ∗Corresponding Author Figure 1: An example of attribute extraction, high- lighted with annotations in different tagging forms. As the foundation for various downstream applica- tions such as knowledge graph construction, search engines, e-Commerce and recommender systems, AE has attracted extensive research interest in re- cent years (Zheng et al., 2018; Xu et al., 2019; Zhu et al., 2020; Jain et al., 2021; Zhang et al., 2022; Li and Zou, 2022). There are two basic subtasks in the research of AE, namely, attribute name extraction and attribute value extraction. And we use the RDF-style triple1 <e, n, v> to denote the entity, attribute name, and attribute value respectively. According to whether the attribute name set is pre-defined, AE can be divided into two paradigms, i.e., the Closed-World Assumption (CWA) and the Open-World Assump- tion (OWA). For CWA AE, the attribute name n is limited to a finite set of the pre-defined schema, where attribute name extraction is typically mod- eled as a classification task (Zeng et al., 2014; Zhou et al., 2016), and attribute value extraction …

Content truncated to 500 words