Review of

Metaphors We Live By


George Lakoff

Mark Johnson

(reviewed by Peter Norvig, UC Berkeley)

Wayne Booth [Booth] has written that, judging from the recent jump in interest in metaphor, if we extrapolate to the year 2039, there will be more students of metaphor than people. Linguists, philosophers, and psychologists have been quick to jump on the metaphorical bandwagon, but so far AI researchers have not. Lakoff and Johnson's ``Metaphors We Live By'' (henceforth ``MWLB'') is an important contribution to the study of metaphor that presents a number of controversial points. Investigating these points provides a good backdrop for presenting the state-of-the-art of metaphor in AI work.

First of all, ``Metaphors We Live By'' is an accessible and thought-provoking source of examples demonstrating the range of metaphor in everyday language and thought. This is not a technical book; it is aimed at a general audience. There is very little terminology, nary a greek letter, and no lists of `starred' ungrammatical sentences. Instead, the arguments are stated simply, and are illustrated by examples which are usually phrases one has heard, or at least could imagine someone actually saying.

The examples show that metaphor is not just a rhetorical device of poets. It is metaphor to speak of arguments in terms of battles, as in ``I demolished his argument'' or ``his claims are indefensible.'' It is metaphor to use spatial prepositions to describe non-spatial relationships, as with ``Harry is in love'' or ``Harry is in the Elks'' or ``Harry is in trouble.'' It is metaphor to personify, as when we say ``Cancer finally caught up with him.''

After demonstrating the pervasiveness of metaphor, the second contribution of Lakoff and Johnson is in showing a small number of highly productive metaphor schemata that underly much of language understanding. As an example, one particularly pervasive and productive metaphor is Michael Reddy's conduit metaphor, which underlies the understanding of communication. The conduit metaphor has three constituent metaphors: IDEAS ARE OBJECTS, LINGUISTIC EXPRESSIONS ARE CONTAINERS, and COMMUNICATING IS SENDING. The metaphor is expressed in phrases like ``it's hard to get that idea across,'' ``it's difficult to put my ideas into words,'' or ``his words carry little meaning.'' Another example of a systematic metaphor schema is MORE IS UP, which leads to expressions like ``the deficit is soaring'' or ``his income fell.'' Such schemata are motivated, but not predicted. It is easy to see why MORE IS UP is a better metaphor than MORE IS DOWN, but one still has to learn which of the many reasonable metaphors are actually used within a culture. Once the metaphor schema is learned, it is easy to generate new instances of it. Lakoff and Johnson present about fifty basic metaphor schemata, with many examples of each.

To Lakoff and Johnson, metaphors are not just matters of language, but are used extensively in reasoning and understanding. Typically, an abstract domain is understood metaphorically in terms of a more concrete domain. To a large degree, they argue, the human conceptual system is metaphorical. This is very different from the classical model of metaphor, which claims that metaphors are artifacts of language use, and have nothing to do with meaning or understanding. It is also very different from most AI models of knowledge representation and language understanding.

The classical theory of metaphor also says that metaphors arise from objective similarity. Thus, we can speak of `digesting an idea' because the mental action of attending to the expression of an idea, reasoning about it, and coming to understand it is objectively similar to the physical action of ingesting food, breaking it into nutrients, and absorbing them into the system. Lakoff and Johnson argue against the idea of a priori objective similarity. They claim metaphors do not just point out similarities that are objectively true; they create the similarities. The notion of digesting an idea is coherent only within the context of other metaphors, such as IDEAS ARE OBJECTS and THE MIND IS A CONTAINER. These basic metaphors both create similarities of their own and allow for the creation of further similarities in the IDEAS ARE FOOD metaphor. The second half of MWLB is not really about metaphors at all; it is a comparison of the traditional objectivist theory of semantics with a new view they call the experientialist theory of meaning.

AI Models of Metaphors

While little has actually been done with metaphor in existing AI systems, it has been recognized as a problem. Wilks [Wilks] calls metaphor ``central to our language capabilities,'' Hobbs [Hobbs] states ``metaphor is pervasive in everyday discourse,'' Carbonell [Carbonell] agrees that metaphors ``pervade commonly spoken English,'' and Rumelhart [Rumelhart] says ``metaphor is natural and widespread in our speech.''

There is some existing AI work on metaphor which could be greatly improved by an understanding of MWLB. For example, Winston's [Winston] Tranfer Frame system is primarily intended as a model of analogical reasoning, but Winston has used it to try to understand metaphors. The system interprets sentences such as ``Robbie is like a fox'' to mean `Robbie is clever.' It does this by adhering to the strict objectivist position that foxes really are clever (and have almost no other characteristics). There is no provision for metaphors that create similarities, only an algorithm for finding pre-defined similarities, which consists solely of counting the common features and relations. ones. A casual glance at the range of examples in MWLB shows that Winston's approach could never be extended to cover the full range of metaphorical usage.

A more interesting approach is that taken by Hobbs and by Rumelhart. They both argue that metaphor interpretation is not only basic to language understanding, it should be indistinguishable from literal language interpretation. This challenges the traditional view of semantics, in which meaning is derived by a simple composition of the meanings of the individual lexemes in the sentence. This literal meaning may differ from the conveyed meaning according to certain rules, such as Gricean maxims. Metaphor interpretation is treated as a secondary process that follows literal interpretation, in this view. Hobbs notes that an expression can pass from a novel metaphor to a frozen idiom to a tired cliche, but at each stage the interpretation process is much the same. Thus, he argues there is no sense having separate mechanisms for `literal' and `metaphorical' interpretations. Rumelhart considers the interpretation of sentences like ``The policeman raised his hand and stopped the car.'' This uses no metaphors, but it requires a complex interpretation process that must identify knowledge structures having to do with traffic cops, drivers, brakes, and cars. This interpretation goes well beyond a simple composition of the literal meanings present in the words, and is similar to the type of interpretation that is done in processing metaphors.

Jaime Carbonell has been the most accepting of Lakoff and Johnson's ideas of anyone in AI. He has been the only one to suggest that the existence of a small number of powerful metaphors means that a good strategy for a language understander would be to try to classify inputs as instances of one metaphor or another, rather than trying to interpret them on general principles. Carbonell presents the start of a process model for language comprehension [Carbonell], but unfortunately he retreats from Rumelhart and Hobbs' position and calls for a two-step process that does literal interpretation first, and metaphorical interpretation only if that fails. Carbonell's suggestion has not yet been implemented.

It is an open question whether the experientialist model of semantics is a good one for AI work. On the one hand, the model is grounded in bodily experiences. The metaphor schema HAPPY IS UP, according to Lakoff & Johnson, is motivated by the fact that people have more erect postures when happy. Other metaphors are based on similar perceptions, none of which can be handled directly with current AI technology. On the other hand, the model stresses a knowledge-rich approach, where much of the burden of understanding is handled by known metaphor schemata. This `strong method' approach seems more in line with current AI research, and more promising than the `weak method' of metaphor understanding based on general principles of similarity.

A weakness of MWLB in terms of AI is that they have no developed process model of understanding, and no theory that relates metaphor comprehension to other comprehension tasks. The AI researcher who is looking for a theory he can immediately implement will be disappointed. The book is useful for its examples and for its questions about the nature of truth and reality, but not for a complete set of answers to these questions. For example, MWLB will tell you that, given AN ARGUMENT IS A CONTAINER and that a container has a deepest part, one can conclude that an argument has a deepest part. We are not told why, given that AN ARGUMENT IS A CONTAINER, a container is a physical object, and a physical object has a color, it is not the case that an argument has a color.

Another problem with MWLB stems from one of its strengths: its non-technical approach makes it widely accessible, but also means that much detail is left out. For those who are skeptical of the approach, the remaining detail may be unconvincing. Similarly, those who are excited by the approach will wish for more references, an index, more formal arguments, and more detailed explication of fine points.

To address that problem, three forthcoming books cover much of the same ground as MWLB, but in more scholarly detail. Lakoff is finishing a book on the theory of categorization called ``Women, Fire, and Dangerous Ideas'' that explores the experientialist, non-objectivist position in more detail. It includes a non-technical overview in a style similar to MWLB as well as three detailed `case studies' that are more technical. The book is due out from the Chicago Press in 1985. (The unusual title refers to the Aborigine language Dyirbal, a language with four classifiers. Classifiers are used to classify noun phrases-- in many languages one cannot say `a piece of paper' but must instead say `a flat-thing of paper.' One Dyirbal classifier is used for women, fire, and dangerous things, but this does not imply any similarity between them. Another classifier is used for everything that does not fit into the other three categories. Lakoff argues that categories which behave like this are common in language and thought, but are not accounted for by the objectivist model of semantics.)

Other forthcoming books of note are Ronald Langacker's ``Foundations of Cognitive Grammar [Langacker]'' and Gilles Fauconnier's ``Mental Spaces [Fauconnier].''


(Note: this paper was resurrected from a very old version, and the references were lost. Will try to fix soon. -Peter Norvig)

Booth, Wayne. Metaphor as Rhetoric

Carbonell, Jamie and Lehnert, Wendy and Ringle, Martin.

Fauconnier, Giles. Mental Spaces

Hobbs, Jerry. Metaphor Schemata

Langacker, Ron. Cognitive Grammar

Rumelhart, David and Ortony.

Wilks, Yorick. Active Preferences

Winston, Patrick. Learning Justifying