Two Context Effects on Metaphor Usage. A Computational, Corpus-Based Approach

This paper studies the relation between metaphor use and its linguistic and extralinguistic contexts. We propose a novel computational method, using token-level distributional semantic models and cluster analysis, to identify different metaphorical “senses” of a lexical item (e.g. cascade as ‘a sequence of events, each causing the next’ versus ‘a diagonally flowing array of matter’) and validate it qualitatively on a large corpus of American English fiction and newspaper texts (approximately 100 million words each). Next, we use the identified sense clusters to study how metaphor use varies as a function of extralinguistic context, in this case: genre. In a regression analysis, we model the genre bias of the sense clusters on the basis of three factors motivated by accounts of the interaction between metaphor and genre, namely sense concreteness, fixedness, and valence. Studying the metaphorical senses of 45 lexical items from the source domains of landscape and weather, we find that newspaper-biased senses involve more abstract target domains than fiction-biased ones. These results show that understanding the relation between linguistic contexts and metaphorical senses can support insight in usage variation across extralinguistic contexts, and, more broadly, that computational methods allow for novel ways to study the context-dependent nature of metaphor use.