Multi-modal English Verb And Adjective Metaphor Identification

Metaphor is a common linguistic phenomenon,it reflects not only the way people use language,but also the way people recognize things.Metaphor is important in natural language processing researches.The semantic contradiction is a signal of metaphor.Metaphor identification model needs to use words features to find the contradictions.The multi-modal method uses textual information and graphical information to obtain features for words,which can get richer semantic features.The words can be divided into concreteness words and words.Concrete words have an entity,it can use graphical information to obtain features.Abstract words do not have an entity,it is hard to use graphical information.This paper combines the word concreteness and multi-modal method,and select the most suitable modal information for each word type.Few word concreteness calculating methods consider the affection of word part of speech(POS).Different word POS has different meanings,it may lead to different concreteness values.This paper calculates word POS concreteness,which can support the metaphor identification model.In terms of obtaining the textual features,this paper uses word POS concreteness as features.In terms of obtaining the graphical features,this paper first uses web search engine to get pictures for words,then uses Convolutional Neural Network to calculate the features,finally this paper calculates the similarities of each words.In terms of word POS concreteness calculation,this paper uses words which have only one POS as seed words,then uses multi-sense word embedding method to get word POS embeddings,and last calculates the distances of words and seed words to get the POS concreteness.This paper implements a multi-modal metaphor identification model for English verb and adjective phrases.The model uses noun concreteness in phrase to divide the concept into concrete concept or abstract concept,then uses different identification strategy for each concept.The metaphor identification model uses textual features and graphical features for concrete concept,and uses only textual features for abstract concept.In a word,this paper uses word POS concreteness and multi-modal features to implement a metaphor identification method.The experiments show the method is effective,and the method provides a new idea for metaphor identification works.