Automatic detection and interpretation of nominal metaphor based on the theory of meaning

Automatic processing of metaphors can be explicitly divided into two subtasks: recognition and interpretation. This paper presents an approach to recognize nominal metaphorical references and to interpret metaphors by exploiting distributional semantics word embedding techniques and calculating semantic relatedness. In terms of detection, our idea is that nominal metaphors consist of source and target domains and that domains present in metaphors will be less related than domains present in non-metaphors. We represent the meaning of the concept as a vector in high-dimensional conceptual space derived from the corpus and compute the relatedness between the vectors to complete the task of detection. Relatedness here is based on the semantics of concepts. Thus, the model we present deals with metaphors where target and source have the same direct ancestors, such as "A surgeon is a butcher".Then, using the relatedness between target and source domain, based on the properties of source domain and dynamic transfer of properties, we present an approach to interpret metaphors with dynamic transfer. Based on the view that metaphor interpretation is the cooperation of source and target domains, we divide metaphor interpretation into two subtasks: properties extraction and properties transfer. Creatively, we use annotations to express a non-binary evaluation, and we take the degree of the annotators' acceptability to evaluate our interpretation of metaphors. HighlightsWe exploit distributional semantics word embedding techniques and semantic relatedness in the metaphor detection and interpretation fields.Our method is based on the theory of meaning. We consider that the difference between source and target domains is in the semantic level, rather than that the domains belong to two different categories.Our method can be flexibly applied to Chinese and English languages. In the Chinese language, we achieve detection accuracy of 85% and interpretation accuracy of 87%. In the English language, we achieve detection accuracy of 85.2% and interpretation accuracy of 85%.