Modeling the Non-Substitutability of Multiword Expressions with Distributional Semantics and a Log-Linear Model

Abstract

Non-substitutability is a property of Multiword Expressions (MWEs) that often causes lexical rigidity and is relevant for most types of MWEs. Efficient identification of this property can result in the efficient identification of MWEs. In this work we propose using distributional semantics, in the form of word embeddings, to identify candidate substitutions for a candidate MWE and model its substitutability. We use our models to rank MWEs based on their lexical rigidity and study their performance in comparison with association measures. We also study the interaction between our models and association measures. We show that one of our models can significantly improve over the association measure baselines, identifying collocations.

Multiword Expression Processing: A Survey

“…Methods based on variant generation and/or lookup were used to discover several MWE categories, such as English verb-particle constructions (McCarthy, Keller, and Carroll 2003;Ramisch et al 2008b), English verb-noun idioms (Fazly and Stevenson 2006;Cook, Fazly, and Stevenson 2007), English noun compounds (Farahmand and Henderson 2016), and German noun-verb and noun-PP idioms (Weller and Heid 2010).…”

Idiom Type Identification with Smoothed Lexical Features and a Maximum Margin Classifier

“…Turning to the general problem of identifying multiword expressions (MWEs) with associated non-compositional meanings research recent includes Yazdani et al (2015) model for noun compounds in English and Farahmand and Henderson (2016) work on identifying of collocations. Focusing on the specific task of idiom type identification, Muzny and Zettlemoyer (2013) describes a model that classifies multi-word Wiktionary entries as idiomatic or literal.…”

Human Associations Help to Detect Conventionalized Multiword Expressions

“…Statistical idiosyncrasy presupposes that the components of a phrase co-occur more often than expected by chance. Besides, the frequency of phrases with statistical idiosynrasy is much higher than the frequency of the phrase with one component changed to its near-synonym (weather forecast vs. weather prediction), as the result of the substitutionability test (Sag et al, 2002;Farahmand and Henderson, 2016). Phrases with statistical idiosyncrasy (often called conventional-ized phrases) can be syntactically and semantically compositional.…”

“…Conventionalized phrases have statistical idiosyncrasy and usually only one approach is proposed in literature to distinguish such phrases from other compositional phrases. This is so-called substitutionability test, which shows if the phrase components can be easily substituted with their synonyms (Sag et al, 2002;Farahmand et al, 2015;Farahmand and Henderson, 2016;Pearce, 2001;Senaldi et al, 2016).…”