Predicting the metaphoric potential of words from lexical features
Xiaojuan Tan
Vrije Universiteit Amsterdam
Jelke Bloem
University of Amsterdam
Some words are more often used metaphorically than others . For instance, 'war' is more often used metaphorically than 'orbit' when they are collocated with 'trade' (cf. trade 'war' vs trade 'orbit'). Why does this happen? It has been hypothesized that the metaphoric potential of words depends on multiple variables, such as concreteness. Specifically, metaphors (e.g., device) typically align an abstract meaning (e.g., method) to a concrete meaning (e.g., a piece of equipment). When a word's meaning is rated as highly concrete, it tends to be used more metaphorically than other words. For instance, 'pillar' is more often used metaphorically than 'opening' when they are collocated with 'trade', possibly because 'pillar' is more concrete than 'opening' (cf. a pillar of trade vs. a trade opening). Iconicity also increases the metaphoric potential of words. We can observe this for ambiguous words – when the concrete meaning and the abstract meaning of a word stand in a higher iconic relationship with each other than with other words, this word tends to be more metaphorically used. For example, 'hourglass' is more typically used metaphorically than 'hyperboloid' when describing one's waist.
In this study, we empirically investigate the factors that influence the likelihood that a word will be used metaphorically. We have combined the metaphor-annotated VU Metaphor Corpus with datasets of various lexical-semantic properties that have been hypothesized to play a role in the metaphoric potential of words, such as concreteness, iconicity, imageability, valency, semantic class, frequency and age of acquisition. This opens up the possibility of studying the effect of these variables on the metaphoric potential of words using a multifactorial regression model. For metaphor studies, this analysis provides empirical backing for theories of metaphor use and development using natural language data. For the field of NLP, this analysis reveals potentially useful features for the popular task of metaphor identification.