MinIE: Minimizing Facts in Open Information Extraction

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2630–2640 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics MinIE: Minimizing Facts in Open Information Extraction Kiril Gashteovski1 and Rainer Gemulla1 and Luciano del Corro2 1Universit¨at Mannheim, Mannheim, Germany 1{k.gashteovski,rgemulla}@uni-mannheim.de 2 Max-Planck-Institut f¨ur Informatik, Saarbr¨ucken, Germany 2corrogg@mpi-inf.mpg.de Abstract The goal of Open Information Extrac- tion (OIE) is to extract surface rela- tions and their arguments from natural- language text in an unsupervised, domain- independent manner. In this paper, we pro- pose MinIE, an OIE system that aims to provide useful, compact extractions with high precision and recall. MinIE ap- proaches these goals by (1) representing information about polarity, modality, attri- bution, and quantities with semantic anno- tations instead of in the actual extraction, and (2) identifying and removing parts that are considered overly specific. We con- ducted an experimental study with several real-world datasets and found that MinIE achieves competitive or higher precision and recall than most prior systems, while at the same time producing shorter, seman- tically enriched extractions. 1 Introduction Open Information Extraction (OIE) (Banko et al., 2007) is the task of generating a structured, machine-readable representation of information expressed in natural language text in an unsuper- vised, domain-independent manner. In contrast to traditional IE systems, OIE systems do not require an upfront specification of the target schema (e.g., target relations) or access to background knowl- edge (e.g., a knowledge base). Instead, extractions are (usually) represented in the form of surface subject-relation-object triples. OIE serves as input for deeper understanding tasks such as relation ex- traction (Riedel et al., 2013; Petroni et al., 2015), knowledge base construction (Dong et al., 2014), question answering (Fader et al., 2014), word anal- ogy (Stanovsky et al., 2015), or information re- trieval (L¨oser et al., 2012). Consider, for example, the sentence “Superman was born on Krypton.” An OIE system aims to extract the triple (Superman, was born on, Kryp- ton), which most of the available systems will cor- rectly produce. As another example, consider the more involved sentence “Pinocchio believes that the hero Superman was not actually born on beau- tiful Krypton”, and the corresponding extractions of various systems in Table 1, extractions 1–6. Al- though most of the extractions are correct, they are often overly specific in that their constituents con- tain specific modifiers or even complete clauses. Such extractions severely limit the usefulness of OIE results (e.g., they are often pruned in relation extraction tasks). The main goals of OIE should be (i) to provide useful, compact extractions and (ii) to produce extractions with high precision and recall. The key challenge in OIE is how to achieve both goals simultaneously. In fact, most of the available systems (often implicitly) focus on either compactness (e.g., ReVerb (Fader et al., 2011)) or precision/recall (e.g., ClausIE (Del Corro and Gemulla, 2013)). We propose MinIE, an OIE system that aims to address and trade-off both goals. MinIE is built on top of ClausIE, a state-of-the-art OIE system that achieves …

Content truncated to 500 words