Supervised Open Information Extraction

Proceedings of NAACL-HLT 2018, pages 885–895 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Supervised Open Information Extraction Gabriel Stanovsky∗2,3, Julian Michael2, Luke Zettlemoyer2, and Ido Dagan1 1Bar-Ilan University Computer Science Department, Ramat Gan, Israel 2Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle, WA 3Allen Institute for Artificial Intelligence, Seattle, WA {gabis,julianjm,lsz}@cs.washington.edu dagan@cs.biu.ac.il Abstract We present data and methods that enable a supervised learning approach to Open Infor- mation Extraction (Open IE). Central to the approach is a novel formulation of Open IE as a sequence tagging problem, addressing challenges such as encoding multiple extrac- tions for a predicate. We also develop a bi- LSTM transducer, extending recent deep Se- mantic Role Labeling models to extract Open IE tuples and provide confidence scores for tuning their precision-recall tradeoff. Fur- thermore, we show that the recently re- leased Question-Answer Meaning Represen- tation dataset can be automatically converted into an Open IE corpus which significantly in- creases the amount of available training data. Our supervised model, made publicly avail- able,1 outperforms the state-of-the-art in Open IE on benchmark datasets. 1 Introduction Open Information Extraction (Open IE) systems extract tuples of natural language expressions that represent the basic propositions asserted by a sen- tence (see Figure 1). They have been used for a wide variety of tasks, such as textual entail- ment (Berant et al., 2011), question answering (Fader et al., 2014), and knowledge base popu- lation (Angeli et al., 2015). However, perhaps due to limited data, existing methods use semi- supervised approaches (Banko et al., 2007; Wu and Weld, 2010), or rule-based algorithms (Fader et al., 2011; Mausam et al., 2012; Del Corro and Gemulla, 2013). In this paper, we present new data and methods for Open IE, showing that supervised learning can greatly improve performance. ∗Work performed while at Bar-Ilan University. at1Our code and models are made publicly available https://github.com/gabrielStanovsky/ supervised-oie Mercury filling, particularly prevalent in the USA, was banned in the EU, partly because it causes antibiotic resistance. (mercury filling; particularly prevalent; in the USA) (mercury filling; causes; antibiotic resistance) (mercury filling; was banned; in the EU; partly because it causes antibiotic resistance) Figure 1: Open IE extractions from an example sentence. Each proposition is composed of a tu- ple with a single predicate position (in bold), and an ordered list of arguments, separated by semi- colons. We build on recent work that studies other natural-language driven representations of predi- cate argument structure, which can be annotated by non-experts. Recently, Stanovsky and Dagan (2016) created the first labeled corpus for eval- uation of Open IE by an automatic translation from question-answer driven semantic role label- ing (QA-SRL) annotations (He et al., 2015). We extend these techniques and apply them to the QAMR corpus (Michael et al., 2018), an open variant of QA-SRL that covers a wider range of predicate-argument structures (Section 5). The combined dataset is the first corpus that is large and diverse enough to train an accurate extractor. …

Content truncated to 500 words