On Computational Approaches to Metaphor and Analogy

Abstract

The field of Artificial Intelligence (AI), which concerns itself with designing and building machines — usually in the form of computer programs capable of exhibiting ‘intelligent’ behavior, has not been oblivious to the role of metaphor and analogy in creativity and problem solving, two of the major processes requiring intelligence. But its focus has not exactly been on similarity-creating metaphors and creative analogies. The term ‘metaphor’ in the AI literature is almost always used restrictively to refer to the instances of metaphor in language. In this narrower sense, metaphor has received rather scanty attention in AI. Moreover, the few attempts that have been made to model metaphors have all been comparison-theoretic in spirit. That is, they all assume that the basis of any metaphor is some underlying similarity between the source and the target, and then they either set out to compute the similarity, or take the similarity as explicitly given and set out to use this information to determine the meaning of a metaphorical phrase. Clearly, while such approaches address the mode of similarity-based metaphor, and to some extent suggestive metaphor, they leave similarity-creating metaphors completely out in the cold. Perhaps, this should not be surprising given the state-of-the-art of Natural Language Processing (NLP), which still remains in its infancy in spite of all the advances of the past thirty years. There are many key issues surrounding literal and conventional meanings that have not yet been fully and satisfactorily incorporated in any NLP system. Similarity-creating metaphors, having to do with novel and creative uses of words and phrases, have remained an unreachable and unrealistic goal. I present a brief overview of the NLP approaches to metaphor in Section 2.