Return to Geoff Nunberg's home page

Transfers of Meaning

Geoffrey Nunberg

Xerox Palo Alto Research Center and Stanford University

From Journal of Semantics, 1995, 1


Abstract: "Transfers of meaning" are linguistic mechanisms that make it possible to use the same expression to refer to disjoint sorts of things. Here I discuss predicate transfer, an operation that takes names of properties into new names that denote properties to which they functionally correspond. It is this operation that is responsible for the new meaning of the predicate parked out back in the utterance "I am parked out back," as well as for the lexical alternations that figure in systematic polysemy. Predicate transfer is subject to two general conditions, which require that basic and derived property stand in a functional correspondence and that the derived property should be a "noteworthy" feature of its bearer. I argue that by appealing to predicate transfer, we can maintain a very strict definition of syntactic identity, which rules out all cases of "sortal crossing," in which a term appears to refer to things of two sorts at the same time, as in examples like Ringo squeezed himself into a tight space; in such a case, the reflexive is strictly coreferential with its antecedent.

1. Introduction

By "transfers of meaning" I mean the ensemble of productive linguistic processes that enable us to use the same expression to refer to what are intuitively distinct sorts or categories of things. Broadly speaking, transfers involve all the figures that traditional rhetoric describes as metaphors, synesthesias, metonymies, and synecdoches, in all their synchronic manifestations. The difference is that transfers are linguistic processes, whereas the rhetorical figures are defined and classified according to the independent conceptual relations that they exploit. The difference between metaphors and metonymies, for example, is that the first presupposes a resemblance and the second a contiguity. And we can go on to classify figures according to the particular conceptual schemas they rest on, either as general correspondences like "abstract for concrete," "part for whole," and "animate for inanimate," or as more specific schemas like "life is a journey" or "polities are bodies." But taken by themselves, these schemas and principles aren't sufficient to explain the linguistic phenomena of transfer. Granted that there is a salient correspondence between monarchs and crowns, for example, it still has to be explained why the word "crown" can be used to refer to monarchs - or for that matter, why this fact should have any linguistic consequences at all. For this we have to look to specifically linguistic mechanisms, which is what I will be talking about here. These mechanisms exist in the service of the expression of conceptual regularities, but they are in principle independent of them, and are constrained in ways that don't permit a purely pragmatic explanation. They are the linguistic handmaidens of figuration, but each is specialized in her offices.

2. Mechanisms of Transfer:

The easiest way to appreciate the difference between rhetorical figures and the linguistic mechanisms is to consider how we can exploit the same sorts of correspondences among things in the world to effect two different kinds of transfer. A customer hands his key to an attendant at a parking lot and says either (1) or (2):

1. This is parked out back.

2. I am parked out back.

Both these utterances involve metonymies. In (1), for example, we would be inclined to say that the subject refers not to the key that the speaker is holding, but to the car that the key goes with. And in fact all the linguistic evidence supports this analysis. For example the number of the demonstrative is determined by the intended referent, not the demonstratum. So even if the customer is holding up several keys that fit a single car, he would say "This is parked out back," whereas if he's holding up a single key that fits several cars, he would say, "These are parked out back." We can make the same point looking at languages that mark demonstratives and adjectives for grammatical gender. In Italian, for example, the word for key is feminine, la chiave, and the word for truck is masculine, il camion. And if a customer gives the attendant the key to a truck it will be the referent, not the demonstratum, that determines the gender of the demonstrative and the adjective for "parked," as in (3):

3. Holding up a key (la chiave, f) to refer to a truck (il camion, masc.) Questo (masc. sg.) è parcheggiato (masc. sg) in dietro. "This (masc.) is parked (masc.) in back."

One final example to the same effect: we can conjoin another predicate that describes the car, but not a predicate that describes the key:

4. This is parked out back and may not start.

5. ??This fits only the left front door and is parked out back.

So there's every reason for saying that the subject of sentences like these refers to the car.

But what of an utterance like (2), "I'm parked out back"? This too is plainly a metonymy of some sort, and there may be a temptation to analyze it as we would (1), saying that the subject of the sentence refers not to the speaker, but to the speaker's car. But the tests we have used to validate this analysis for the demonstrative in (1) give a different answer here. For example if the speaker has two cars, he wouldn't say:

6. We are parked out back.

(though of course this would be an appropriate utterance if there were two people who were waiting for the car). By the same token, an Italian man who was waiting for his car would express this using a masculine adjective parcheggiato for "parked," even though the word for "car" is a feminine, la macchina.

7. Io sono parcheggiato (*parcheggiata, fem. sg.) dietro.

And in this case, we can conjoin any other predicate that describes the speaker, but not always one that describes the car:

8. I am parked out back and have been waiting for 15 minutes.

9. *I am parked out back and may not start.

The conclusion is that the subject of (1) refers to the speaker, and the transfer involves the predicate. That is, the predicate parked out back contributes a property of persons, the property they possess in virtue of the locations of their cars.

Now the difference between these examples clearly doesn't have anything to do with the kind of relations they exploit. In both cases we assume a correspondence between the things in one domain, the cars parked in various locations, and the things in another domain, keys or drivers as the case may be. But we can take semantic advantage of these correspondences in two different ways. Sentence (1) is a case of deferred ostension or deferred indexical reference, a process that allows a demonstrative or indexical to refer to an object that corresponds in a certain way to the contextual element picked out by a demonstration or by the semantic character of the expression. In this connection, note that we can't get this kind of deferred reading when we use a description in place of an demonstrative, as in (10):

10. *The key I'm holding is parked out back.

Whereas (2) exemplifies another kind of transfer process, which I'll call predicate transfer. The principle here is that the name of a property that applies to something in one domain can sometimes be used as the name of a property that applies to things in another domain, provided the two properties correspond in a certain way. And just to fill out the contrast here, note that unlike deferred ostension, predicate transfer is indifferent to how the bearer of this new or derived property is referred to - by an indexical or description or whatever. For example, in this situation the parking lot manager could say to the attendant:

11. The man with the cigar (Mr. McDowell, etc.) is parked out back.

What these examples show, then, is that unlike rhetorical classifications like metaphor and metonymy, the various mechanisms of transfer can't be distinguished simply by pointing at the types of correspondences they exploit. And for this reason, the description of these mechanisms is fundamentally a linguistic problem, rather than a problem of conceptual analysis. That is, there is nothing we can learn about keys, drivers, or cars that will help us to explain the differences between examples like (1) and (2).

In the rest of this paper I will be concentrating on just one of these mechanisms, predicate transfer. In this section I will schematize the conditions that license this operation. In the following sections I will spell out the role of predicate transfer in lexical polysemy, and then discuss its implications for some well-known syntactic puzzles. Finally, I will talk about some of the methodological difficulties that predicate transfer raises.

3. Conditions on Predicate Transfer

Predicate transfer is subject to two conditions. The first of these I have already mentioned: the property denoted by the derived predicate has to correspond in a certain way to the property denoted by the original predicate. With an utterance like (2), for example, we begin with a functional correspondence between the locations of cars in a lot and the properties of the owners, or more accurately, the "havers," of these cars. When two property domains correspond in an interesting or useful way - of which more in a moment - we can schematize the operation of predicate transfer as follows:

12. Condition on predicate transfer

Let P and P' be sets of properties that are related by a salient transfer function gt: P Æ P' Then if F is a predicate that denotes a property P OEP , there is also a predicate F', spelt like F, that denotes the property P', where P' = gt (P).

In the cases we have been talking about, of course, these correspondences between properties are mediated by correspondences between their bearers - e.g., the functional relation from cars to their owners - and we might want to represent this directly. So let h be a salient function from a set of things A to another (disjoint) set of things B. Then for any predicate F that denotes a property P that applies to something in A, we can represent the meaning of a derived predicate F', spelt like F, as in either (13a) or (13b):

13a. lP. ly ("x[dom h]. h(x) = y Æ P(x))

b. lP. ly ($x[dom h]. h(x) = y H P(x))

Note that this entails that utterances of this type are in fact ambiguous between "universal" and "existential" readings, depending on whether all or only some of the bearers of the original property are in the inverse image of h for a given value. And in fact both types of reading are generally available. In cases like "I am parked out back," we would normally assume that the speaker means to say that all the (relevant) cars he owns have the property in question. By contrast, when a painter says, "I am in the Whitney," she doesn't imply that all her paintings or even all her relevant paintings are in the Whitney, but only that something she painted is in the Whitney, as in (13b). And when an accountant says of his firm, "We are in Chicago," he might intend either sort of interpretation, depending on whether he is talking about the firm's headquarters or one of its offices.

Still, it is more useful to think of these two types of readings as two ways of representing the general schema given in (12), rather than as two distinct conditions that license predicate transfer. One reason is that not all of the relations between properties that figure in predicate transfers are mediated by relations between their individual bearers. This is the crucial feature that distinguishes metonymies from metaphors, where one property calls up another directly, either in virtue of some similarity between them or through the perceptual mechanisms that underlie synesthesias like a dark mood. I'll come back to metaphor at the end of this paper; for the moment we just want to bear in mind that metaphors and metonymies aren't different types of transfer, they're different conditions that can support a functional relation between properties.

The schemas in (13) do a reasonable job of representing the truth-conditions associated with utterances like "I am parked out back" and "I'm in the Whitney," but they miss some important pragmatic conditions on the use of such utterances. For example, suppose my car has the property of having once been driven by Jean Gabin. Then according to the conditions in (13), I should be able to use the name of this property to describe the property that I acquire in virtue of my relationship to my car. But it would be odd for me to say:

14. ??I was once driven by Jean Gabin.

even in a context in which it might be relevant to say "My car was once driven by Jean Gabin." By the same token, a painter might say with reference to one of her paintings, "I'm in the Whitney Museum," but not, ordinarily:

15. ??I'm in the second crate on the right.

Intuitively, the difference is this: when a painting goes into a museum its creator acquires a significant or notable property, whereas when it goes into a crate she doesn't, at least not usually.

Let me describe this condition by saying that predicate transfer is only possible when the property contributed by the new predicate is "noteworthy," which can be reckoned in either of two ways. The first type is exemplified in an utterance like "I'm parked out back," or "I'm in the phone book." The purposes of these utterances is to classify the members of one set of things (e.g., garage customers) relative to the immediate conversational purposes (of discharging the attendant's responsibilities, say), in virtue of their relation to the identifying properties of some other group of things (the cars). In cases like these we will say that a property is noteworthy if it offers a useful way of classifying its bearer relative to the immediate conversational interests. From the point of view of a garage attendant, for example, a customer is usefully classified in terms of the properties he acquires from the location of his car ("I'm parked out back") or its overall appearance ("I'm a red Mustang") than in terms of the properties he acquires from its provenance or mechanical condition ("I was bought from a friend," "I shimmy at high speed") - though of course some of these properties might be more useful for other conversational purposes.

The second type of noteworthiness is exemplified by utterances like "I'm in the Whitney," or "American Express is in Singapore now." In these cases the derived property has a more abiding interest or consequence for its bearer, beyond the immediate conversational purposes. The difference between the two may depend on fairly subtle features. The utterance "I'm in the phone book" most likely serves the purposes of immediate classification ("You don't have to write down my number"); the utterance "Clinton is in the phone book" is more likely offered by way of remarking Clinton's accessibility, or whatever.

Noteworthiness can be thought of as a special case of relevance, depending on how we understand the latter notion. To see this relation, we can consider examples (16) and (17), adapted from Jackendoff (Jackendoff 1992):

16. Ringo was hit in the fender by a truck when he was momentarily distracted by a motorcycle.

17. ??Ringo was hit in the fender by a truck two days after he died.

Let's assume that these utterances exemplify transfers of the meanings of the relevant relational expressions - that is, that Ringo denotes the singer rather than his car. The difference between the two cases is that when a truck hits Ringo's car while he is driving it, the event will have important consequences for him as well: he is likely to have been startled, or annoyed, or put to trouble and expense. Whereas once Ringo is dead, nothing that happens to his car is going to invest him with any properties worth mentioning. Another way of putting this would be to say that it is unlikely to be conversationally relevant to know that the deceased Ringo has the property that he acquires when his (former) car is hit on the fender, even though it may be relevant to know of the car that it has been hit. But our ability to make this distinction rests on our being able to distinguish between the relevance of a proposition (e.g., that Ringo's car was hit) and the relevance of its trivial entailments (e.g., that Ringo has the property of having had his car hit). The question is of some interest for the theory of conversation, but I won't pursue it here. For the present we can just take noteworthiness in an intuitive way: they are properties that matter.

4. Predicate Transfer on Common Nouns

The examples of predicate transfer that I've been talking about up to now have all involved the meanings of predicates or verb phrases, but I don't want to give the impression that the process is essentially a matter of function coercion, rather than argument coercion. In fact predicate transfer is available for the property names that are expressed by common nouns, as well, whether they appear in predicate position or referring position. Take (18), as uttered by a restaurant waiter:

18. Who is the ham sandwich?

The process of transfer is pretty straightforward here; from the point of view of the waiter, at least, customers acquire their most usefully distinctive properties in virtue of their relations to the dishes they order. But in this case, unlike the "parked out back" examples, the relevant property is expressed by a common noun, which can equally well be used as the content of an NP in referential position in a sentence like (19):

19. The ham sandwich is at table 7.

In earlier discussions several people, myself among them, have talked about examples like (19) as involving a kind of "reference transfer," or "deferred reference," with the implication that an actual ham sandwich must figure in the interpretation of the utterance - that is, that the transfer operates on the NP interpretation. But there are a number of reasons for concluding that the transfer here takes place on the common noun meaning - that is, that this is a case of meaning transfer, rather than reference transfer. One way to make this point is to look at examples where there is a clash between the number of the dish and the number of the patron, as in (20):

20. That (*those) french fries is (*are) getting impatient.

(though the felicity of examples like this depends as well on specific principles of English morphosyntax). A related example is They played lots of Mozart, where a transferred predicate co-occurs with a quantifier or article that is inconsistent with its literal meaning - how could Mozart used as a mass term refer in a well-formed way to the composer? Or we can consider the interpretation of the determiner in example (19), which doesn't presuppose the existence of a unique ham sandwich (think of a waiter in a fast-food restaurant who is standing in front of a table piled with ham sandwiches), but does presuppose the existence of a unique ham-sandwich orderer. Finally, we can consider some examples involving anaphors (a matter I'll take up in greater detail later on). Fauconnier [Fauconnier, 1985 #19] gives examples (21) and (22).

21. The mushroom omelet was eating with chopsticks.

22. *The mushroom omelet was eating itself/ himself with chopsticks.

The use of a reflexive in (22) would presume that the object of eat was introduced by the subject NP - that is, that the subject NP actually referred to a mushroom omelet on the route to its ultimate interpretation. But inasmuch as the transfer actually takes place at the level of the common noun, which contributes only a property of persons, the example is ill-formed. Kleiber (Kleiber 1984) raises a similar puzzle with regard to (23):

23. *Je suis en train de manger la tarte flambée qui est assise à la table 20. "*I am eating the flaming pie who~that is sitting at table 20."

For (23) to have an interpretation, we would require that the phrase la tarte flambée qui est assise à la table 20 refer to a dish. But it doesn't.

5. Predicate Transfer in Systematic Polysemy

The availability of transfer for common nouns, adjectives, and the rest is what underlies the patterns of lexical alternation that go by the names of "regular polysemy," "systematic polysemy," "logical metonymy" and the like. Of course not all transfers that apply lexical categories are instances of lexical processes. Many are too much dependent on the availability of a specialized context. The correspondence between the properties of dishes and customers provides a useful means of identification only in the specialized domain of a restaurant, and then only relative to the interests of waiters - this isn't a good way of classifying people in general, or even of classifying customers when the discourse isn't directly concerned with the relations between customers and orders (I doubt whether a waiter would say, "I saw the ham sandwich at table 7 the other day driving a Mercedes"). Cases like this we can think of as "occurrent metonymies," where the relation between two domains can only be exploited in a restricted range of situations. So we would be reluctant to say that the item ham sandwich is polysemous, with lexical sense along the lines of "ham sandwich orderer"; we might better class it with Clark's "contextual expressions" (see e.g., [Clark, 1993 #34]).

But often the property correspondences that license predicate transfer hold across a wider range of situations, providing a more context-independent way of classifying the bearers of derived properties, which answers to more general interests. In these cases we may very well want to say that the transferred predicate represents a lexical sense of the item in question. To a certain extent, this is a relative matter. For example dictionaries often assign the words black and white senses like "In chess, the person playing white." I expect the idea here is that even though this is a context-dependent correspondence, the derived property is much more generally useful for classifying players than the property of having ordered a ham sandwich as a means of classifying restaurant customers, since so many other things follow from which color a player takes. And as usages become progressively more useful and less context-specific, we come to the general patterns of lexical alternation commonly described by means of the formula "x for y." For example:

24. Examples of Systematic Polysemy

transmissions for cars: 4 speed, automatic, etc.

texts for inscriptions: a Webster's Third, a Guide Bleu, etc.

painters for works: a Picasso, a Derain, etc.

containers for volumes of stuff: She drank two glasses, etc.

writer for oeuvre: fifty pages of Wordsworth

place for inhabitants: Indianapolis voted for the referendum

tree for wood: The table is made of oak.

Now from one point of view, we might want to say that there is no qualitative difference between these alternations and the cases of occurrent metonymy we were looking at a moment ago. Certainly we should bear in mind that just as with occurrent metonymies, the availability of a transferred predicate in the alternations in (24) is subject to the general conditions of noteworthiness that I discussed earlier. Take the well-known rule of "grinding," which turns count nouns into mass nouns that denotes the stuff they are made of, as in "We ate chicken," "They wear beaver," or "The table was made of oak." As it happens, we don't ordinarily apply grinding to the names of breeds of animals to derive the names of their meats or their fur or skin, as in "a plate of Rhode Island Red" or "a bag made of Hereford," for the simple reason that breeds are not usually regarded as gastronomically or sartorially natural kinds; this is not a useful way of identifying a type of meat or fur. But there are exceptions, and in just these cases breed names can be used as mass terms - for example, when we are talking about the breeds of sheep from which wool is derived, as in "His sweater was made of angora," or when the government issues a warning telling consumers to "stay away from Hereford until the insecticide danger diminishes." Even if we call these alternations "lexical," then, they are still subject to the same pragmatic constraints that determine the acceptability of occurrent metonymies.

At the same time, there are also reasons for assigning alternations like these a specifically lexical status. They tend to have a high text frequency; they often become habitual or "idiomatic"; and they may develop specialized senses that aren't entirely predictable on the basis of general principles of transfer (for example the specialized meanings of mass terms like balsam or lavender). Then too, we might want to argue that the features that figure in these particular correspondences are present the lexical meanings of the items themselves: that it is the lexicon that tells us that a tree yields wood, a painter makes paintings, a glass is a container, and so on. Finally, and I think most telling, the lexical alternations that involve predicate transfer are often subject to language-specific constraints that seem to have no general pragmatic explanation. For some reason or another, for example, the English rule of grinding seems not to apply to yield the names of liquids:

25. ??We fried the chicken in safflower (olive, corn, etc.).

26. ??I enjoy a glass of orange (pear, apple, etc.) with my breakfast.

Of course it may very well be that in some kitchens cooks routinely use safflower for "safflower oil," but this not the general practice. And analogously, there seems to be no good reason why French-speakers should use the names of fruits to refer to the brandies made from them (une prune, une poire) whereas English-speakers do not. Another example drawn from Apresjian [Apresjan, 1973 #41]: in Russian you can use the name of an organ to denote a disease of that organ, as in "She has a kidneys," whereas in English you can't, though presumably the usage would be comprehensible enough. And according to Jerrold Sadock (personal communication), Greenlandic Eskimo does not permit grinding of animal names to produce mass terms for either meat or fur, though you can apply grinding to tree names to get terms for wood. So you can't say "He eats walrus" or "He wears walrus," but you can say "His boat is made of oak."

All of this suggests that we need a second kind of device to explain how particular languages lexicalize particular cases of predicate transfer - what are described in Nunberg and Zaenen [Nunberg, 1992 #35] as "lexical licenses." These are not the same things as lexical rules, which are defined over particular lexical classes and subject to synchronically arbitrary syntactic or morphological constraints, though of course licenses may become lexical rules diachronically. What licenses do, rather, is to index specific types of correspondences (e.g., between animals and their meat, or singers and the songs they sing) as available for exploitation to produce new lexical items. In this regard they can be thought of on the model of what Morgan [Morgan, 1978 #27] calls "conventions of use" - rules that determine which practices are "idiomatic," in the common acceptation of the term, as opposed to being merely possible.

There is much more to be said about rules like these, but for our purposes, the important thing to keep in mind is that both lexical and nonlexical transfers are subject to the same range of pragmatic constraints and exploit the same semantic mechanisms. This is one reason why it is a mistake to think of transfer itself as an intrinsically lexical process, a picture that is implicit in the tendency to talk about this phenomenon in the context of discussions of lexical polysemy, as well as in the names of many of the more-or-less formal devices that have been proposed to deal with it, such as "lexical networks" (Norvig and Lakoff (1987)), "lexical subregularities" [Wilensky, 1991 #33], and "lexical implication rules" [Ostler, 1991 #42]. Even when a rule like grinding has been explicitly lexicalized, what is conventionalized is not the mechanism involved, but only the particular semantic domains the mechanism is allowed to operate on.

6. Predicate Transfer and Semantic Composition

A related problem with this emphasis on the lexicon (and I think that with most writers this is only a question of emphasis, rather than of theoretical dispositions) is that it can obscure the fact that transfer is essentially a phrasal process, which operates on predicates of any complexity, or more generally, on categories up to the type X' in the Jackendoff framework. And as such it is subject to familiar constraints on semantic composition. Take the sentence:

27. The article contains a description of a widely-studied Peruvian virus.

Let fd be a transfer function that takes the names of agents into the names of the diseases they cause. Then we could represent the four possible interpretations of the NP a widely-studied Peruvian virus as in (27):

28. (widely-studied (Peruvian (virus))) = "a microorganism that is native to Peru and widely studied."

29. (widely-studied (Peruvian fd(virus))) = "a virally-caused disease that is endemic to Peru and widely-studied.

30. (widely-studied fd(Peruvian (virus))) = "a widely-studied disease caused by a virus native to Peru"

31. fd (widely-studied (Peruvian (virus))) = "a disease that is caused by a widely-studied virus native to Peru"

But the phrase can't have an interpretation as in (41):

32. fd ((widely-studied) (Peruvian fd(virus))) = "a disease endemic to Peru that is caused by a widely studied virus"

That is, the sentence can't have a reading where Peruvian would apply to the output of the transfer function on the composition of widely studied and virus. The point may be clearer if we note the oddness of the phrase ??a tiny incurable virus , which could only mean "an incurable disease caused by a tiny virus," with the adjectives again applying in a way that would violate the ordinary process of composition. (Note that it is not the order of adjectives itself that is responsible here: we can say a tiny incurable wart.)

There are several ways to obtain the effect of compositionality here. One approach would be to try to accomplish all of this lexically, though the kind of coercion mechanism proposed by Pustejovsky and developed, among others, by Briscoe and Copestake in this volume. They discuss the phrase corn-fed chicken, where chicken is treated as a mass noun denoting chicken meat; they suggest that we can take corn-fed here as applying to an ORIGIN feature that appears in the derived qualia structure of the mass term chicken (or in the qualia structure of underived items like veal and beef). In that case corn-fed has its normal, unextended sense, and no phrasal transfer is involved. And we could tell an analogous story about Peruvian virus under the reading "virally-caused disease endemic to Peru" in which the adjective applies in its literal sense to a feature in some element (telic?) in the qualia structure of virus.

But this approach has a number of problems. For one thing, it still leaves us with the job of explaining the restrictions on compositionality that we noticed in connection with the phrase ??a tiny incurable virus.; if both adjectives have their literal senses here, why shouldn't they compose as they do in tiny incurable wart? We could introduce constraints on the operation of coercion, of course, but these would wind up simply recapitulating the process of phrasal composition. And in any case, a lexical approach can't explain many cases of phrasal transfer, as in the example The south side of Cambridge voted Conservative, given by Briscoe and Copestake in this volume, where the content of a subject NP refers to a group of people, but where there is no obvious way to derive this interpretation lexically.

There's a third difficulty with this approach that has a greater methodological interest for us: note that not every modifier that could apply to a chicken from which meat has been derived can apply happily to the mass term chicken:

33. They served meat from corn-fed (Arkansas, happy, beheaded) chickens.

34. They serve corn-fed (Arkansas, ??happy, ??beheaded) chicken.

The restriction in these examples is roughly that the property of the animals contributed by the adjective has to be one with consequences for the constitution or quality of the meat derived from them - the meat from chickens may vary according to what they were fed, but not how they were slaughtered. This makes sense if we take the adjectives here to be transferred predicates sensitive to the requirements of noteworthiness that I mentioned earlier, but it is inexplicable if they have their literal meanings and apply to a hidden feature in the noun: why should some adjectives be better than others in combination with the mass noun chicken in (33), when all apply equally well to the count noun chickens in (34)? The examples in (35) and (36) make the same point:

35. The people in the huge (overflowing, domed, old) stadium rose to give the team a standing ovation.

36. The huge (overflowing, ??domed, ??old) stadium rose to give the team a standing ovation.

That is, we assume that the group of people who rose to give an ovation acquires a noteworthy property from the dimensions of the structure in which they were housed, but not from its physical constitution or age. But again, this is explicable only if huge and overflowing have transferred senses in (36), as opposed to applying literally to a hidden feature (say, "structure") in the headword stadium, for in the latter case we are at a loss to explain why domed is not acceptable here as well.

We have to assume, then, that transfer is a phrasal process that works to a large degree in concert with the process of semantic composition. One way of dealing with this would be to permit transfer to take place independently on any simple or complex predicate or relational term, and then filter the output via constraints charged with maintaining consistency and some measure of compositionality. This approach has the advantage of providing for some violations of strict compositionality. There is in fact some evidence that this is required. For example, consider sentences where a metaphorically used subject and verb combine with a nonmetaphorical object, either idiomatically, as in (37), or nonidiomatically, as in (38):

37. The bottom dropped out of the bonds market.

38. The Austrians managed to elude Napoleon for a week, but the eagle finally swooped down on them at a small town near the border.

If transfer were constrained to adhere to strict compositionality, we would have to say that the phrases the bottom dropped out of and the eagle swooped down are constituents in these examples. But a proponent of the existence of VP's could rightfully argue that this is slim evidence for abandoning the VP hypothesis; better to assume that the metaphorical transfers in these examples are not constrained to apply in a strictly compositional way. In any case, at this time it is largely an open question as to how figuration and semantic composition interact.

7. Syntactic consequences: "Sortal Crossings"

With this general understanding of the operation of predicate transfer, we can turn to some of its syntactic implications. Consider sentences like (39), suggested by Jackendoff, and (40):

39. Ringo squeezed himself into a narrow space.

40. Yeats did not like to hear himself read in an English accent. (Contrast: I am often read in an English accent.)

On the standard way of thinking about these phenomena, the reflexives in these examples present a problem, since they seem not to be coreferential with their antecedents. In (39), that is, the subject Ringo refers to a person, whereas the reflexive appears to refer to his car. Let me describe sentences like these as involving apparent "sortal crossings." This difficulty has been known for some time, and people have suggested a number of solutions to it. In the heyday of imperial syntax, a sentence like (39) would have been handled by a syntactic operation like "car deletion" on an underlying structure like "Ringoi squeezed Ringoi's car into a narrow space." And other people have suggested various semantic and pragmatic approaches.

But all of these approaches have the same failing, which is that they presume that it is the meaning of the reflexive that has been transferred, rather than the meaning of the verb. That is, they all try in one way or another to get himself to refer to a car. Whereas what I am suggesting here is that himself in a sentence like (39) actually refers to Ringo, so that the reflexive and its antecedent are coreferential in a strict sense. What is transferred is the sense of the expression "squeeze x into a tight space." That is, in virtue of having performed a certain operation on his car, Ringo has also performed a corresponding (and noteworthy) operation on himself, and this is what permits the predicate (or more accurately, relation) transfer. And by the same token, we assume that people who read Yeats' poetry aloud are doing something to the poet as well. But here again, the transfer is only available when the derived property or relation provides some useful or noteworthy information about its bearer, in accord with modern assumptions about authorship. So we wouldn't ordinarily say:

41. ??Yeats wrote a lot of himself in sprung rhythm.

When Yeats writes a poem in a particular metrical pattern, that is, he is not performing an operation on himself in the way he is when he reads one of his poems aloud to an audience.

All of this leads to a strong hypothesis: natural language permits no sortal crossings in any of the rules or constructions that ordinarily impose conditions of identity - not just with reflexives, but with pronouns, relative constructions, and so forth. A related hypothesis would stipulate that there need be no relaxing of the coordinate structure constraint to deal with sentences like:

42. Roth is Jewish and widely read.

Instead, we will interpret both Jewish and widely-read as predicates that contribute properties of persons.

The object in all this is to honor the intuitions that originally motivated semanticists to appeal to zeugma as a test for determining whether a word has one or more senses. Predicate transfer allows us to treat syntactic identity as sortally chaste, by imputing all the promiscuity to the predicates involved. This is what the syntax has been trying to tell us all along, I think, but we weren't paying close enough attention.

8. Two Methodological Issues

Predicate transfer raises two methodological questions for semantic analysis. First, how do we determine when an expression is being used in a transferred sense? As we saw in connection with the example "I am parked out back," this is by no means always easy to decide. In the general case, we have a sentence of the form "a is G," where a is a word that ordinarily denotes something of sort A (e.g., people), and where G is a predicate that ordinarily applies to things of some disjoint sort of things B (e.g., cars). We assume that either a or G must have a transferred meaning, but how do we tell which?

Let's take a very simple example: Billy's shoes were tied. We can analyze this in either of two ways: either shoes has a transferred reading where it refers to shoelaces, or tied has a transferred reading where it denotes the property that shoes acquire when their laces have been tied. As we saw, there is no theoretical reason for choosing between these analyses: predicate transfer can operate on both functors and arguments. Nor is intuition a sure guide (or at least people seem to be of different opinions about the matter when you put the question to them baldly). But there is support for the second analysis, where it is tied that has the transferred reading, in the observation that we can conjoin tied with another expression that ordinarily applies to shoes, but not with an expression that ordinarily applies to laces:

46. Billy's shoes were neatly tied but dirty.

47. ??Billy's shoes were neatly tied but frayed.

48. Billy tied his shoes, which were dirty (??frayed).

This seems conclusive, but why? The general assumption is that where a predicate is used in a transferred sense, there will be only limited possibilities of substitution - i.e., we can say the shoes were tied but not the shoes were frayed - since other predicates will be subject to conditions of noteworthiness. (In this case, a shoe acquires an important property from its laces being tied - it is less likely to slip off, for example - but not from their being frayed.) Recall that this is the same form of argument we used above in arguing that corn-fed must have a transferred reading in the phrase corn-fedchicken, since we could not replace it with another word like beheaded or happy, even though all of these would have to be construed as applying to the animals from which the meat was derived.

Let's see now how this procedure works in the slightly more complicated case like that in (49):

49. Yeats is still widely read.

Like Billy's shoes are tied, this example admits of two analyses, but in this case the sentence is in fact ambiguous. On one reading, yeats refers to the poet himself, and the predicate has a transferred reading in which it contributes a property of poets. On this reading we would make subsequent references to the subject using the animate pronoun he, as in (50):

50. ÷though he has been dead for more than 50 years.

The restriction on sortal crossing rules out the possibility that the pronoun could refer to the poet while its antecedent referred to his works. On the second reading,yeats is a mass term that refers to the poet's works, and the predicate has a literal reading where it contributes a property of the works. On this reading we would use the pronoun it in subsequent reference:

51. ÷even though most of it is out of print.

and here the restriction on crossing forces the analysis of the subject as a mass term. But we don't have to rely only on the hypothesis about sortal crossing to make this point; we can also appeal independently to the requirements of noteworthiness that we made use of a moment ago. If the predicate has a transferred reading in the version of (49) that is continued with the pronoun he (i.e., where the subject denotes the poet himself), then we will expect that the possibilities of substituting other predicates in this version will be constrained by requirements of noteworthiness. Whereas the choice of predicates should be unconstrained in the version that is continued with the pronoun it (i.e. where the subject denotes the works):

52. Yeats is the second book from the right on the top shelf, if we still have it (??him).

The assumption is that the property of having one's books on such-and-such a shelf isn't particularly noteworthy in a poet, the way say that having one's books widely read or on the bestseller list might be (recall the painter's utterance of "I'm in the second crate on the right.") But this restriction is explained only if we assume that the predicate has a transferred reading in the version of this sentence that is continued with the animate pronoun.

This takes me to the second methodological difficulty that predicate transfer raises. Up to here we have been assuming that predicate transfer is essentially a unidirectional process. Shoes acquire noteworthy properties from the properties of their laces, meats from the properties of the animals they are derived from, and drivers from the properties of their cars, but not, in general, vice-versa. So there are a number of modifiers that apply literally to the count noun chickens on the hoof that can also apply to the mass term chicken (corn-fed, range, Arkansas), but none that apply literally to meat that can also be applied to the count noun (??sautéed chickens). This is one reason why it is usually easy to determine which use of a given expression is basic and which are derived - e.g., that it is the mass use of chicken that is created by predicate transfer. All of this reflects a basic definitional dependency: chicken meat is defined by reference to the properties of the bird, but not vice-versa.

In other cases, though, distinct categories may be interdefined in a way that makes extensive bi-directional property transfer possible. One example is the class of words that includes newspaper, magazine, dictionary, travel guide, and so on - basically any individuated type of publication that is prepared or published by a single dedicated organization. These words can refer to a token of the publication, to an edition of the publication, or to the publishing organization itself. And predicates that contribute properties of one of these sorts can frequently be applied with transferred reading to an entity of some other sort, as in:

53. The newspaper Mary works for was featured in a Madonna video.

Example (53) involves no sortal crossing. Rather the subject denotes an organization, and the predicate in the main clause has a transferred reading where it contributes a property that the organization acquires from the fact that one of its copies was featured in a video. Of course not every property of a newspaper token imparts a noteworthy property to the organization that publishes it; alongside of (53), it would be odd to say (54), suggested by Pustejovsky:

54. ??The newspaper Mary works for fell off the table. (where newspaper = "newspaper publisher")

But note that (54) does have a possible reading if we take the subject to denote, not the company that Mary works for, but a copy of the newspaper it produces. Here, that is, we have predicate transfer operating on the content of the relative clause, rather than of the main clause, and going in the opposition direction, from the properties of publishers to the properties of their products. (Note that here the noteworthiness of the transferred predicate is reckoned by its utility as a means of identifying the newspaper copy, rather than by its intrinsic interest. The point may be clearer if we change Pustejovsky's example to something like, "She handed me an envelope containing all the newspapers she had worked for," where the object NP clearly denotes a collection of copies identified in terms of the properties of the companies that produce them.)

A word like newspaper we can think of as densely metonymous. Its various denotations are interdefined: equivalence classes of copies and editions are individuated in large part by reference to the distinguishing properties of the organizations that produce them, and vice versa. ("Which newspaper did Mary burn?," we ask, and the answer might well be, "the one John works for"; "Which newspaper does John work for?," we ask, and the answer might well be, "The one Mary burned.") And the same correspondences that license the multiple uses of the word newspaper license widespread predicate transfer from the properties of one of its denotations to another: publishers acquire a number of noteworthy properties from copies of their publications or from editions of their publications, and each of these acquires many noteworthy properties from the publishers, and so on. As a result, the phrase a newspaper that supported the regime is trebly ambiguous; depending on the context it can refer to a company, an edition, or a copy. And likewise for the phrases a newspaper that the demonstrator burned, the newspaper that hit the streets yesterday morning, and so on. We can think of dense metonymy schematically as follows:

55. Dense Metonymy:

Given several disjoint sorts of things A, B,÷ and several classes of predicates F,G÷ such that members of F literally apply only to things of sort A, members of G literally apply only to things of sort B, and so on, a word W is densely metonymous iff:

a. W has distinct uses to refer to things of sorts A, B÷ and

b. When W is applied to something of sort A, it often happens that predicates belonging to G can be applied to W under transferred readings, and when W is applied to something of sort B it often happens that predicates belonging to F can be applied to W under transferred readings, and so on.

One problem with truly dense metonymies is that we may not be able to assign one or the other use a prior place in the lexicon, or to say in which direction the transfer operates. In this sense a word like newspaper is unlike an item like yeats or chicken, where the direction of transfer is clear. This creates some difficulties for lexical description, though I suspect that these are in large measure the artifacts of our theoretical approaches - the need to say which uses of a word are "basic" and which are "derived"; the desire to distinguish sharply between lexical representation and knowledge representation in general. Informally, we might think of newspaper as listed in the lexicon in association with a constellation or ring of interrelated concepts. The central problem, it seems to me, is not just how to set this up this formally, but also how to structure the knowledge representations for the relevant entities - newspaper copies, editions, publishers, etc. - so that we can predict that we will get this pattern of transfer with newspaper and dictionary but not, say, with novel or user'smanual. But I won't have any substantive proposals to offer here.

9. Metaphor

I want to conclude with a word about metaphors. Earlier on I said that the difference beween metaphors and metonymies is that the former were predicate transfers that were licensed by a direct correspondence between properties. Take the logicians' practice of referring to the symbol written as "÷" as horseshoe. We can speak here of a relation of resemblance between the set of horseshoes and horseshoe-shaped-symbols, or between prototypical horseshoes and prototypical horseshoe-shaped-symbols, but obviously there is no existential relation between any particular horseshoe and any horseshoe-symbol, the way there is, say, between a particular newspaper copy and a newspaper publisher. As a result the particular things in the range of the transfer function don't acquire properties from anything that happens to the things in its domain - nothing that happens to any real horseshoe has any consequences (noteworthy or otherwise) for any particular horseshoe-shaped-symbol. So from a metaphorical use of a predicate there is no inference to the truth of any utterance involving the literal use of the predicate. This is the difference between an utterance like "I'm parked out back" and an utterance like "I'm cruising down the fast lane of life"; with the latter you wouldn't ask the speaker, "What are you driving?" (at least not unless drive itself is used metaphorically). Hence with metaphor we don't see the same kinds of apparent sortal crossings that we saw with metonymies: a logical symbol doesn't acquire properties from anything that happens to an instance of the equine footwear that it resembles:

56. ??The second line of the proof begins with the horseshoe that Deb pinned over her door.

But we shouldn't make too much of these differences. Both types of figuration are licensed by a repertory of conceptual schemas of greater or lesser generality. And as I've suggested here, both types are available for exploitation by linguistic processes like predicate transfer, and are subject to the same kinds of functional constraints.


Earlier versions of this paper were presented at the Conference on Lexical Universals in Dagstuhl in April of 1993, at the Annual Meeting of the Association for Computational Linguistics in Columbus, Ohio in June of 1993, and at the CREA in Paris in December of 1993. For comments on this version I am grateful to Sue Atkins, Ted Briscoe, Ann Copestake, John Lamping, Ivan Sag, Annie Zaenen, and the two anonymous reviewers for this journal.