Open Information Extraction with Tree Kernels

Open Information Extraction with Tree Kernels Ying Xu, Mi-Young Kim, Kevin Quinn, Randy Goebel and Denilson Barbosa Department of Computing Science University of Alberta {yx2,miyoung2,kfjquinn,goebel,denilson}@cs.ualberta.ca Abstract Traditional relation extraction seeks to iden- tify pre-specified semantic relations within natural language text, while open Information Extraction (Open IE) takes a more general ap- proach, and looks for a variety of relations without restriction to a fixed relation set. With this generalization comes the question, what is a relation? For example, should the more general task be restricted to relations medi- ated by verbs, nouns, or both? To help answer this question, we propose two levels of sub- tasks for Open IE. One task is to determine if a sentence potentially contains a relation be- tween two entities? The other task looks to confirm explicit relation words for two enti- ties. We propose multiple SVM models with dependency tree kernels for both tasks. For explicit relation extraction, our system can ex- tract both noun and verb relations. Our results on three datasets show that our system is su- perior when compared to state-of-the-art sys- tems like REVERB and OLLIE for both tasks. For example, in some experiments our system achieves 33% improvement on nominal rela- tion extraction over OLLIE. In addition we propose an unsupervised rule-based approach which can serve as a strong baseline for Open IE systems. 1 Introduction Relation Extraction (RE) systems are designed to discover various semantic relations (e.g. <Obama, president, the United States>) from natural lan- guage text. Traditional RE systems extract spe- cific relations for prespecified name-entity types (Bunescu and Mooney, 2005; Chan and Dan, 2011; Zhou and Zhu, 2011). To train such systems, ev- ery relation needs manually annotated training ex- amples, which supports limited scope and is diffi- cult to extend. For this reason, Banko et al. (2007) proposed Open Information Extraction (Open IE), whose goal is to extract general relations for two en- tities. The idea is to avoid the need for specific train- ing examples, and to extract a diverse range of rela- tions. This generalized form has received significant attention, e.g., (Banko et al., 2007; Akbik, 2009; Wu and Weld, 2010; Fader et al., 2011; Mausam et al., 2012). Because Open IE is not guided by or not restricted to a prespecified list of relations, the immediate chal- lenge is determining about what counts as a relation? Most recent Open IE systems have targeted verbal relations (Banko et al., 2007; Mausam et al., 2012), claiming that these are the majority. However, Chan and Dan (2011) show that only 20% of relations in the ACE programs Relation Detection and Charac- terization (RDC) are verbal. Our manually extracted relation triple set from the Penn Treebank shows that there are more nominal relations than verbal ones, 3 to 2. This difference arises because of the ambi- guity of what constitutes a relation in Open IE. It is often difficult even for humans to agree on what constitutes a relation, and which words in the sen- tence …

Content truncated to 500 words