Cross-lingual Semantic Generalization for the Detection of Metaphor

In this work, we describe a supervised cross-lingual methodology for detecting novel and conventionalized metaphors that derives generalized semantic patterns from a collection of metaphor annotations. For this purpose, we model each metaphor annotation as an abstract tuple – (source, target, relation, metaphoricity) – that packages a metaphoricity judgement with a relational grounding of the source and target lexical units in text. From these annotations, we derive a set of semantic patterns using a three-step process. First, we employ several generalized representations of the target using a variety of WordNet information and representative domain terms. Then, we generalize relations using a rule-based, pseudo-semantic role labeling. Finally, we generalize the source by partitioning a semantic hierarchy (defined by the target and the relation) into metaphoric and non-metaphoric regions so as to optimally account for the evidence in the annotated data. Experiments show that by varying the generality of the source, target, and relation representations in our derived patterns, we are able to significantly extend the impact of our annotations, detecting metaphors in a variety of domains at an F-measure of between 0.88 and 0.92 for English, Spanish, Russian, and Farsi. This generalization process both enhances our ability to jointly detect novel and conventionalized metaphors and enables us to transfer the knowledge encoded in metaphoricity annotations to novel languages.