Weakly Supervised Definition Extraction

Proceedings of Recent Advances in Natural Language Processing, pages 176–185, Hissar, Bulgaria, Sep 7–9 2015. Weakly Supervised Definition Extraction∗ Luis Espinosa-Anke, Francesco Ronzano and Horacio Saggion TALN - DTIC Universitat Pompeu Fabra Carrer T`anger, 122-134 08018 Barcelona {luis.espinosa,francesco.ronzano,horacio.saggion}@upf.edu Abstract Definition Extraction (DE) is the task to ex- tract textual definitions from naturally occur- ring text. It is gaining popularity as a prior step for constructing taxonomies, ontologies, auto- matic glossaries or dictionary entries. These fields of application motivate greater interest in well-formed encyclopedic text from which to extract definitions, and therefore DE for academic or lay discourse has received less at- tention. In this paper we propose a weakly supervised bootstrapping approach for identi- fying textual definitions with higher linguis- tic variability than the classic encyclopedic genus-et-differentia definition, and take the domain of Natural Language Processing as a use case. We also introduce a novel set of fea- tures for DE and explore their relevance. Eval- uation is carried out on two datasets that re- flect opposed ways of expressing definitional knowledge. 1 Introduction Definition Extraction (DE) is the task to automat- ically extract textual definitions from text (Navigli and Velardi, 2010). It has received notorious at- tention for its potential application to glossary gen- eration (Muresan and Klavans, 2002; Park et al., 2002), terminological databases (Nakamura and Na- gao, 1988), question answering systems (Saggion ∗ This work is partially funded by the SKATER project, TIN2012-38584-C06-03, Ministerio de Econom´ıa y Compe- titividad, Secretar´ıa de Estado de Investigaci´on, Desarrollo e Innovaci´on, Espa˜na; and Dr. Inventor (FP7-ICT-2013.8.1 611383). and Gaizauskas, 2004; Cui et al., 2005), for support- ing terminological applications (Meyer, 2001; Sierra et al., 2006), e-learning (Westerhout and Monach- esi, 2007), and more recently for multilingual para- phrase extraction (Yan et al., 2013), ontology learn- ing (Velardi et al., 2013) or hypernym discovery (Flati et al., 2014). The corpora that have been used for evaluating DE systems are varied, although in general efforts have been greatly focused on academic and ency- clopedic genres. Some prominent examples include German technical texts (Storrer and Wellinghoff, 2006), the IULA Technical Corpus (in Spanish) (Alarc´on et al., 2009), the ACL Anthology (Jin et al., 2013; Reiplinger et al., 2012), the BNC corpus (Rodr´ıguez, 2004), Wikipedia (Navigli and Velardi, 2010), ensembles of domain glossaries and Web documents (Velardi et al., 2008), or technical texts in various languages (Westerhout and Monachesi, 2007; Przepi´orkowski et al., 2007; Borg et al., 2009; Deg´orski et al., 2008; Del Gaudio et al., 2013). We propose a DE approach which, from a start- ing set of encyclopedic definition seeds, self-trains iteratively and gradually fits its classification capa- bility to a target domain-specific test set. Evalua- tion is carried out on two corpora: First, a set of 50 abstracts of papers in the field of NLP1. Here, the target term is defined in the first sentence, and addi- tional information may appear in the form of “syn- tactically plausible false definitions”, i.e. sentences where the target term is also present, relevant infor- mation …

Content truncated to 500 words