OIE@OIA: an Adaptable and Efficient Open Information Extraction Framework

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6213 - 6226 May 22-27, 2022 c⃝2022 Association for Computational Linguistics OIE@OIA: an Adaptable and Efficient Open Information Extraction Framework Xin Wang, Minlong Peng, Mingming Sun, Ping Li Cognitive Computing Lab Baidu Research No.10 Xibeiwang East Road, Beijing 100193, China 10900 NE 8th St. Bellevue, Washington 98004, USA {wangxin60, pengminlong, sunmingming01, liping11}@baidu.com Abstract Different Open Information Extraction (OIE) tasks require different types of information, so the OIE field requires strong adaptability of OIE algorithms to meet different task re- quirements. This paper discusses the adapt- ability problem in existing OIE systems and designs a new adaptable and efficient OIE system - OIE@OIA as a solution. OIE@OIA follows the methodology of Open Information eXpression (OIX): parsing a sentence to an Open Information Annotation (OIA) Graph and then adapting the OIA graph to different OIE tasks with simple rules. As the core of our OIE@OIA system, we implement an end- to-end OIA generator by annotating a dataset (we make it open available) and designing an efficient learning algorithm for the complex OIA graph. We easily adapt the OIE@OIA system to accomplish three popular OIE tasks. The experimental show that our OIE@OIA achieves new SOTA performances on these tasks, showing the great adaptability of our OIE@OIA system. Furthermore, compared to other end-to-end OIE baselines that need mil- lions of samples for training, our OIE@OIA needs much fewer training samples (12K), showing a significant advantage in terms of efficiency. 1 Introduction Open Information Extraction (OIE) techniques are gradually attracting more and more attention (Christensen et al., 2011; Mausam et al., 2012; Corro and Gemulla, 2013; Angeli et al., 2015; Bhutani et al., 2016; Cui et al., 2018; Roy et al., 2019; Zhan and Zhao, 2020) as they build a bridge between language to knowledge. OIE tasks are generally designed for its information extraction requirements, which vary from verbal relations between entities (Banko et al., 2007; Etzioni et al., 2004), nominal attributes (Yahya et al., 2014; Pal and Mausam, 2016; Saha et al., 2017), and adverbial components (e.g., time) (Stanovsky et al., 2018). Even for the same type of information, the required facts may still differ. Table 1 shows the required form of facts of three popular OIE tasks: OIE2016 (Stanovsky and Dagan, 2016), Re-OIE2016 (Zhan and Zhao, 2020) and CaRB (Bhardwaj et al., 2019). The diversity of requirements is an essential feature in the field of OIE, which leads to the urgent need for OIE algorithms with strong adaptability. The adaptability problem in the OIE field has not been well addressed. There are two primary methodologies to obtain an OIE system: the rule- based approach and the end-to-end learning-based approach. The rule-based approaches (Christensen et al., 2011; Corro and Gemulla, 2013; Angeli et al., 2015; Bhutani et al., 2016; Gashteovski et al., 2017) use human-written or bootstrap- learned rules to convert linguistic structures of sentences into target facts. The end-to-end learning approaches (Stanovsky et al., 2018; Sun …

Content truncated to 500 words