Taxonomy Construction Using Syntactic Contextual Evidence

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 810–819, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Taxonomy Construction Using Syntactic Contextual Evidence Luu Anh Tuan #1, Jung-jae Kim #2, Ng See Kiong ∗3 #School of Computer Engineering, Nanyang Technological University, Singapore 1anhtuan001@e.ntu.edu.sg, 2jungjae.kim@ntu.edu.sg ∗Institute for Infocomm Research, Agency for Science, Technology and Research, Singapore 3skng@i2r.a-star.edu.sg Abstract Taxonomies are the backbone of many structured, semantic knowledge resources. Recent works for extracting taxonomic relations from text focused on collect- ing lexical-syntactic patterns to extract the taxonomic relations by matching the pat- terns to text. These approaches, however, often show low coverage due to the lack of contextual analysis across sentences. To address this issue, we propose a novel ap- proach that collectively utilizes contextual information of terms in syntactic struc- tures such that if the set of contexts of a term includes most of contexts of an- other term, a subsumption relation be- tween the two terms is inferred. We ap- ply this method to the task of taxonomy construction from scratch, where we intro- duce another novel graph-based algorithm for taxonomic structure induction. Our ex- periment results show that the proposed method is well complementary with previ- ous methods of linguistic pattern matching and significantly improves recall and thus F-measure. 1 Introduction Taxonomies that are backbone of structured on- tology knowledge have been found to be use- ful for many areas such as question answering (Harabagiu et al., 2003), document clustering (Fodeh et al., 2011) and textual entailment (Gef- fet and Dagan, 2005). There have been an in- creasing number of hand-crafted, well-structured taxonomies publicly available, including WordNet (Miller, 1995), OpenCyc (Matuszek et al., 2006), and Freebase (Bollacker et al., 2008). However, the manual curation of those taxonomies is time- consuming and human experts may miss relevant terms. As such, there are still needs to extend ex- isting taxonomies or even to construct new tax- onomies from scratch. The previous methods for identifying taxo- nomic relations (i.e. is-a relations) from text can be generally classified into two categories: statis- tical and linguistic approaches. The former in- cludes co-occurrence analysis (Budanitsky, 1999), term subsumption (Fotzo and Gallinari, 2004) and clustering (Wong et al., 2007). The main idea be- hinds these techniques is that the terms that fre- quently co-occur may have taxonomic relation- ships. Such approaches, however, usually suffer from low accuracy, though relatively high cover- age, and heavily depend on the choice of feature types and datasets. Most previous methods of the linguistic approach, on the other hand, rely on the lexical-syntactic patterns (e.g. A is a B, A such as B) (Hearst, 1992). Those patterns can be manu- ally created (Kozareva et al., 2008; Wentao et al., 2012), chosen via automatic bootstrapping (Wid- dows and Dorow, 2002; Girju et al., 2003) or iden- tified from machine-learned classifiers (Navigli et al., 2011). The pattern matching methods gen- erally achieve high precision, but low coverage due to the lack of contextual analysis across …

Content truncated to 500 words