Salient Feature Extraction in Japanese Metaphor Generation

Abstract

Feature salience of concepts in metaphors plays an important role in metaphor processing. This paper presents some computational models for metaphor generation in the form of “person X is person Y in the domain Dx (domain of person X)” using feature salience as the measure. The input consists of person X, domain Dx, and domain Dy (domain of person Y). The models then extract the features of the person X, search for people in the domain Dy based on the extracted features, and output the name of a person in the domain Dy as the person Y. In the model without feature salience, the features extracted are based on the term frequency of occurrence. In the model with feature salience, the salience of features are estimated using the term frequency-inverse document frequency (tf-idf) method, and weighting for opinion/evaluation expressions. A psychological experiment was conducted to verify the validity of the method for salient feature extraction, and to examine the influence of opinion/evaluation expressions for salience in metaphor generation. It was shown that the model with feature salience based on term frequency of occurrence and opinion/evaluation expressions performed better than the model without feature salience, especially when there were only a few documents relating to the person X from which the features could be extracted. The results suggest the necessity for feature salience in metaphor generation and a certain degree of validity of the method for salient feature extraction using tf-idf and weighting for opinion/evaluation expressions.