A Generalisation of Lexical Functions for Composition in Distributional Semantics

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 281–291, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics A Generalisation of Lexical Functions for Composition in Distributional Semantics Antoine Bride IRIT & Universit´e de Toulouse antoine.bride@irit.fr Tim Van de Cruys IRIT & CNRS, Toulouse tim.vandecruys@irit.fr Nicholas Asher IRIT & CNRS, Toulouse nicholas.asher@irit.fr Abstract Over the last two decades, numerous algo- rithms have been developed that success- fully capture something of the semantics of single words by looking at their distri- bution in text and comparing these distri- butions in a vector space model. How- ever, it is not straightforward to construct meaning representations beyond the level of individual words – i.e. the combina- tion of words into larger units – using dis- tributional methods. Our contribution is twofold. First of all, we carry out a large- scale evaluation, comparing different com- position methods within the distributional framework for the cases of both adjective- noun and noun-noun composition, making use of a newly developed dataset. Sec- ondly, we propose a novel method for composition, which generalises the ap- proach by Baroni and Zamparelli (2010). The performance of our novel method is also evaluated on our new dataset and proves competitive with the best methods. 1 Introduction In the course of the last two decades, there has been a growing interest in distributional meth- ods for lexical semantics (Landauer and Dumais, 1997; Lin, 1998; Turney and Pantel, 2010). These methods are based on the distributional hypothe- sis (Harris, 1954), according to which words that appear in the same contexts tend to be similar in meaning. Inspired by Harris’ hypothesis, numer- ous researchers have developed algorithms that try to capture the semantics of individual words by looking at their distribution in a large corpus. Compared to manual studies common to formal semantics, distributional semantics offers substan- tially larger coverage since it is able to analyze massive amounts of empirical data. However, it is not trivial to combine the algebraic objects created by distributional semantics to get a sensible distri- butional representation for more complex expres- sions, consisting of several words. On the other hand, the formalism of the λ-calculus provides us with general, advanced and efficient methods for composition that can model meaning composition not only of simple phrases, but also more com- plex phenomena such as coercion or composition with fine-grained types (Asher, 2011; Luo, 2010; Bassac et al., 2010). Despite continued efforts to find a general method for composition and various approaches for the composition of specific syntac- tic structures (e.g. adjective-noun composition, or the composition of transitive verbs and direct ob- jects (Mitchell and Lapata, 2008; Coecke et al., 2010; Baroni and Zamparelli, 2010)), the model- ing of compositionality is still an important chal- lenge for distributional semantics. Moreover, the validation of proposed methods for composition has used relatively small datasets of human sim- ilarity judgements (Mitchell and Lapata, 2008).1 Although such studies comparing …

Content truncated to 500 words