LexSemTm: A Semantic Dataset Based on All-words Unsupervised Sense Distribution Learning

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1513–1524, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics LEXSEMTM: A Semantic Dataset Based on All-words Unsupervised Sense Distribution Learning Andrew Bennett,♥ Timothy Baldwin,♥ Jey Han Lau,♥♦ Diana McCarthy,♣ and Francis Bond♠ ♥ Dept of Computing and Information Systems, The University of Melbourne ♦ IBM Research ♣ Dept of Theoretical and Applied Linguistics, University of Cambridge ♠ Linguistics and Multilingual Studies, Nanyang Technological University awbennett0@gmail.com, tb@ldwin.net, jeyhan.lau@gmail.com, diana@dianamccarthy.co.uk, bond@ieee.org Abstract There has recently been a lot of interest in unsupervised methods for learning sense distributions, particularly in applications where sense distinctions are needed. This paper analyses a state-of-the-art method for sense distribution learning, and op- timises it for application to the entire vocabulary of a given language. The optimised method is then used to pro- duce LEXSEMTM: a sense frequency and semantic dataset of unprecedented size, spanning approximately 88% of polyse- mous, English simplex lemmas, which is released as a public resource to the com- munity. Finally, the quality of this data is investigated, and the LEXSEMTM sense distributions are shown to be superior to those based on the WORDNET first sense for lemmas missing from SEMCOR, and at least on par with SEMCOR-based distribu- tions otherwise. 1 Introduction Word sense disambiguation (WSD), as well as more general problems involving word senses, have been of great interest to the NLP commu- nity for many years (for a detailed overview, see Agirre and Edmonds (2007) and Navigli (2009)). In particular, there has recently been a lot of work on unsupervised techniques for these problems. This includes unsupervised methods for perform- ing WSD (Postma et al., 2015; Chen et al., 2014; Boyd-Graber et al., 2007; Brody et al., 2006), as well as complementary problems dealing with word senses (Jin et al., 2009; Lau et al., 2014). One such application has been the automatic learning of sense distributions (McCarthy et al., 2004b; Lau et al., 2014). A sense distribution is a probability distribution over the senses of a given lemma. For example, if the noun crane had two senses, bird and machine, then a hypo- thetical sense distribution could indicate that the noun is expected to take the machine meaning 60% of the time and the bird meaning 40% of the time in a representative corpus. Sense distribu- tions (or simple “first sense” information) are used widely in tasks including information extraction (Tandon et al., 2015), novel word sense detection (Lau et al., 2012; Lau et al., 2014), semi-automatic dictionary construction (Cook et al., 2013), lex- ical simplification (Biran et al., 2011), and tex- tual entailment (Shnarch et al., 2011). Automat- ically acquired sense distributions themselves are also used to improve unsupervised WSD, for ex- ample by providing a most frequent sense heuris- tic (McCarthy et al., 2004b; Jin et al., 2009) or by improving unsupervised usage sampling strate- gies (Agirre and Martinez, 2004). Furthermore, the improvement due to the most frequent sense heuristic has been particularly strong …

Content truncated to 500 words