.

more naturally paraphrased using the conventional metaphor “to catch,” which is very commonly used with infections, rather than the strictly literal and somewhat dramatic “become infected by” or “to suffer from.” The second is perhaps best paraphrased using the modifier “mild” for “flu,” even though the idea of mildness is likely a metaphor itself in this context. Indeed, it is hard to find a concise literal rendering for the tactile metaphor “touch:” a touch suggests a glancing contact, which implies a contact of short duration that lacks force. So a “touch of flu” implies a short-lived infection that does not manifest the worst symptoms of the ailment.

      An acceptance of the occasional conventional metaphor in a supposedly literal paraphrase is especially useful in the context of statistical models of metaphor, for it is in the nature of conventional metaphor to be used habitually and to assume the status of normative, literal language over time, to the extent that the two are hard to separate on statistical grounds alone. Indeed, much of what we consider literal was once a newly minted figurative form that, as the philosopher Nietzsche put it, has become the base metal of literal coinage through oft-repeated usage. An NLP system that aims to remove all metaphor from language will find much—perhaps too much—to remove, and will be left with precious little with which to compose its paraphrases. In contrast to the divide between the figurative and the literal, the unconventional/conventional divide has an observable reality in large text corpora that is conducive to statistical analysis (e.g., Shutova et al. [2012], Bollegala and Shutova [2013]). It is a quantifiable reality that allows a paraphrasing system to learn to do as human paraphrasers do: to generate helpful conventional paraphrases from unconventional metaphors. If such paraphrases contain conventional metaphors, as they are very likely to do, this need not pose a serious challenge to the semantic representations of an NLP system. Conventional metaphors are part of the furniture of language, and can easily be accommodated—with some design forethought—in a system’s core representations. For instance, both the MIDAS system of Martin [1990] and the ATT-META system of Barnden and Lee [2002], and Barnden [2006] put a semantic representation of conventional metaphors at the heart of their metaphor interpretation systems. MIDAS and ATT-META each show, through their different uses of schematic structures, how an NLP system can support a wider range of inferences, and, thus, a fuller interpretation of a metaphor, by reserving a place for the metaphor in its semantic and conceptual representation of an utterance.

2.4 METAPHOR AND SIMILE

      Since the worldview communicated by a metaphor is often shaped by a speaker’s perception of a situation or an interlocutor, simile—the figurative device most often called upon to simultaneously communicate and explain our perceptions—can be of particular use when paraphrasing a metaphorical worldview. Consider the metaphor “marriage is slavery:” while there is little insight to be gained by paraphrasing this metaphor with the simile “marriage is like slavery,” there is much to be gained from the paraphrase “this marriage is like the relationship between a slaver and his slave: you act like you own me, and treat me like your slave.” The use of similes imbues this paraphrase with three interesting qualities. First, the similes bring a desirable emotional distance to the description of an undesirable relationship, for the assertion “you act like you own me” conveys a very different affect than “you own me” or “you are my owner.” Second, the paraphrasing similes are not freighted with the same semantic tension as the original metaphor, since similes merely assert the similarity of two ideas and openly admit—through their use of “as” or “like”—to the counterfactuality of a viewpoint. That is, every simile “A is like B” asserts not just that A is similar to B, but that A is very much not B. Finally, the similes not only explain the meaning of the metaphor, but offer a rationale for it too: the paraphrase suggests that the metaphor is the speaker’s way of making sense of the bad behavior of others.

      The implicit negation in every simile makes simile a particularly good choice for paraphrasing a negative metaphor. For, just as explicit similes contain tacit negations, explicit negations often suggest tacit similes, especially when the negation emphasizes the figurative qualities of a descriptor. Since most negative metaphors are trivially true in a literal sense, insofar as the underlying positive metaphor is literally absurd, then each negative metaphor is also its own literal paraphrase. Thus, it really is the case that no man is an island; that your wife is not your maid; that your overbearing boss does not own you; that your college fund is not an ATM; etc. Yet there is more to a negative metaphor than the negation of an obvious falsehood. A negative metaphor is not so much a disavowal of an explicit metaphor, or of the tacit conceptual pact that it implies, as it is an explicit repudiation of someone else’s implicit simile. Why else would we need to say something that is so obviously true? So, the frustrated wife who cries at her boorish spouse “I am not your maid!” is in fact saying “Do not treat me like your maid!” The moody teenager who snaps at a concerned neighbor “You are not my father” is in fact saying “Stop acting like my father.” And the man who needs to be told that “No man is an island” is no doubt acting like someone who believes the opposite. Similes allow a speaker to be abundantly clear as to the perceptual foundations of a figurative viewpoint, and clear in ways that are hard to achieve in metaphor.

      Giora et al. [2010] argue that negation, when used to convey such apparently obvious facts, is a metaphor-inducing operation. So, like metaphors and similes, negations of the false or the absurd are much more likely to activate the figurative aspects of a descriptor than any literal qualities. Metaphor involves highly selective inference (see Hobbs [1981]) and the assertion “I’m no Superman” selectively activates the qualities of the cultural icon (and very dense descriptor) Superman that are most often projected by a metaphor or a simile, such as strength, resilience, and speed, rather than any literal quality related to appearance (e.g., wearing a red cape, or red underwear on the outside) or behavior (e.g., fighting crime, working in disguise as a reporter, etc.). Precisely what these figurative qualities are for a given descriptor is something that one would have difficulty discerning from only metaphorical uses of the descriptor, as metaphors rarely make explicit the qualities that are projected onto the target. This brings us then to another aspect of similes that makes them so useful to the computational modeling of metaphor. Not only do similes explicitly mark their figurative status with “like” or “as,” and frequently indicate the perceptual roots of their viewpoint with qualifiers such as “act like,” “smell like,” “look like,” etc., they also often explicitly identify the qualities that are transferred from the source/vehicle to the target/tenor.

      The scholar Quintilian saw the difference between metaphor and simile as the difference between an implicit and an explicit comparison. Not only does a simile mark the comparison between a comparandum and its comparatum, it may also provide a third element, a tertium, to state the reasons why both are seen as similar [Roberts, 2007]. In the English simile frame “X is as Y as Z” the tertium is given by the [Y] element, although a speaker may omit the tertium to say simply that “X is like Z.” But when the tertium is explicit, as it always is in “as-as” similes, a listener can acquire from the similes of others a sense of the properties of a [Z] that are most likely to be activated in a figurative comparison. In this way, similes become an important vector for the transmission of cultural knowledge via language; as Charles Dickens puts it in the opening page of A Christmas Carol, “the wisdom of our ancestors is in the simile.” Dickens was referring to the swirl of stock similes that were common currency in the language and culture of his day, but the modern scholar, or modern NLP system, now has access to a long-tail of diverse similes—and diverse tertia—for a wide range of useful descriptors on the Web. Consider again the dense descriptor Superman: a Google search for the phrasal query “as * as Superman” (here * denotes a wildcard that can match any token in a text) returns web documents that fill the tertium * position with the following values: strong, powerful, fast, mighty, invulnerable, resilient, cool, and American. As these properties are frequently highlighted in figurative similes, it seems reasonable to assume that a metaphorical use of Superman will draw on the same properties.

      The