Bright ideas, smart colors - metaphor detection and grading through neural networks
ILC-CNR - Aula Seminari IBF SG 5
I will present two experiments in my work on figurative language.
In the first experiment, we use a fully connected neural network to distinguish metaphorical adjective-noun expressions (i.e. "cold voice") from literal expressions (i.e. "cold water"). We show that our results outperform the state of the art and that this approach allows to distinguish metaphors from literal phrases with high accuracy, also determining a "degree of metaphoricity" for a given expression.
In the second experiment (more of a work in progress) we use a deep combination of convolutional neural networks and long-short term memory to rank possible paraphrases of a given metaphor on a scale from 1 to 5. Unlike previous approaches, we tested our method on different grammatical categories (metaphoric nouns, adjectives, verbs and multiword metaphors) and we introduced syntactic and stylistic variations in the candidate paraphrases.
The bottomline of my work is that to deal reasonably with figurativity we need to assume a nuanced approach: any text has a degree of figurativity and a number of possible interpretations.