Similarity is Closeness: Using Distributional Semantic Spaces to model Similarity in Visual and Linguistic Metaphors

The semantic similarity that characterizes two terms aligned in a metaphor is here analysed through a corpus-based distributional semantic space. We compare and contrast two samples of metaphors, representative of visual and linguistic modality of expressions respectively. Popular theories of metaphor claim that metaphors transcend their modality to influence conceptual structures, thus suggesting that different modalities of expression would typically express the same conceptual metaphors. However, we show substantial differences in the degree of similarity captured by the distributional semantic space with regard to the modality of expression (higher similarity for linguistic metaphors than for visual ones). We argue that this is due to two possible variables: Conventionality (linguistic metaphors are typically conventional, while visual are not) and Complexity (visual metaphors have modality-specific inner complexities that penalize the degree of similarity between metaphor terms captured by a language-based model). Finally, we compare the similarity scores of our original formulations with those obtained from different possible verbalizations of the same metaphors (acquired by replacing the metaphor terms with their semantic neighbours). We show that while this operation does not affect the average similarity between metaphor terms for visual metaphors, the similarity changes significantly in linguistic metaphors. These results are discussed here.