The Effectiveness of Simple Hybrid Systems for Hypernym Discovery

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3362–3367 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3362 The Effectiveness of Simple Hybrid Systems for Hypernym Discovery William Held and Nizar Habash Computational Approaches to Modeling Language Lab New York University Abu Dhabi, UAE {wbh230,nizar.habash}@nyu.edu Abstract Hypernymy modeling has largely been sep- arated according to two paradigms, pattern- based methods and distributional methods. However, recent works utilizing a mix of these strategies have yielded state-of-the-art results. This paper evaluates the contribution of both paradigms to hybrid success by evaluating the benefits of hybrid treatment of baseline mod- els from each paradigm. Even with a simple methodology for each individual system, uti- lizing a hybrid approach establishes new state- of-the-art results on two domain-specific En- glish hypernym discovery tasks and outper- forms all non-hybrid approaches in a general English hypernym discovery task. 1 Introduction Discovering word-level hierarchies has long been an important step in constructing language tax- onomies. The most important of these hierar- chical relationships is hypernymy or the ISA- relationship, i.e. ‘chihuahua’ is a ‘dog’, which forms the backbone of word-level taxonomies, most notably WordNet (Fellbaum, 1998). Early works on the modeling of this relation- ship focused on the practical task of discover- ing new instances of the hypernymy relationship given a vocabulary and an existing resource with labeled data about hypernymy, described as hy- pernym discovery by Camacho-Collados (2017). For the purposes of discovery, Hearst (1992) de- veloped a landmark set of lexico-syntactic patterns which indicated hypernymy. There have been many follow-ups on this con- cept of identifying and utilizing patterns to iden- tify hypernym pairs (Caraballo, 1999; Mann, 2002; Snow et al., 2005, 2006). However, by re- stricting the sentences of interest to only those which match patterns, even very large datasets with very loose pattern matching will often return small co-occurrence numbers, especially for more indirectly connected hypernym pairs. To tackle the sparsity of pattern-based ap- proaches, recent focus has turned to distribu- tional models of hypernymy. Distributional mod- els are attractive since they use signals drawn from every sentence of training data. Distributional approaches have focused on discovering spatial properties of embedding space which capture the hypernymy relationship (Kotlerman et al., 2010; Yamane et al., 2016; Shwartz et al., 2017; Nickel and Kiela, 2017; Vulic and Mrksic, 2017). The performance of distributional approaches in hy- pernymy detection shows promise to create a more broad picture of the hypernymy relationship space. Recently, hybrid models of hypernymy, in both discovery and detection, have surpassed the per- formance of either paradigm individually. Simi- larly, the current state-of-the-art in hypernymy de- tection was set by a classifier which integrated information from both pattern data and distribu- tional word embeddings (Shwartz et al., 2016). In hypernym discovery, where purely distributional methods have struggled, a model which utilized a hybrid approach of patterns and distributional representations far and away led the results of a recent SemEval Task (Camacho-Collados et al., 2018; …

Content truncated to 500 words