Appendix to "How to be a meaning holist" and "How to be a meaning atomist"<1>

Eric Lormand
University of Michigan

Draft: May, 1996

Since my proposed framework for meaning (in "Holist" and "Atomist") is neither simply a psychosemantic holism nor simply a psychosemantic atomism, but a marriage in which the two have become one, we might call it a psychosemantic holism-atomism wedlock (PSHAW). In this paper I want to:

(1) locate my discussion around that of Jerry Fodor and Ernie LePore (1992),

(2) develop further the advantages of PSHAW over Fodor’s denotational theory,

(3) develop further the advantages of PSHAW over Block’s two-factor theory, and

(4) describe some implications for public-linguistic meaning (vs. mental content).

The going will be rough (especially in section 1), because this is compressed and because I presuppose familiarity with the literature I discuss. I have tried to make the four sections independent of one another.

1 Fodor and LePore on holism and atomism

My use of the words "holism" and "atomism" differs importantly from that recently introduced by Fodor and LePore (1992). Roughly, they define "atomism" and "holism" as contraries. Since (I argue) my atomism and holism are compatible, I certainly do not want to follow them here. I use the term "nonholism" for the simple denial of holism—e.g., a meaning depends on not virtually all others—and "localism" for the extreme opposite of holism—e.g., a meaning depends on no others. Fodor and LePore sometimes use "localism" as the extreme opposite of holism (or their versions of holism—see below), as I would, but they also use "atomism" this way, as I would not. In their usage, the meaning of a representation is atomistic only if it does not depend on the meaning of other representations, so what they mean by "atomism" is incompatible with definitions as normally construed. By contrast, what I mean by "atomism" requires definitions, and a theory of referential meaning such as Fodor’s (see section 2 below) is not atomistic. Perhaps, given the influence of Fodor’s writings, I should use a different term than "atomism", but my usage reflects the traditional philosophical doctrine called "logical atomism", at least in its semantic but not metaphysical or epistemological aspects.

My "meaning holism" distinguished from their kinds

Fodor and LePore distinguish three kinds of meaning holism, which they call "content holism", "translation holism", and "anthropological holism". None of these capture the kind of dependence essential to my version of meaning holism (or weaker standard versions, for that matter). In fact, there is no entailment in either direction between my meaning holism and any of their holisms (assuming that none of these are logically necessary).

Content holism is the view that there can be no "small" systems: a mind or a language can contain one representation only if it contains many (Fodor and LePore, 1992, p. 5). To see that my meaning holism does not entail content holism, consider a (logically) possible small system with two features: the meaning of each of its representations depends on the meaning of "all" the rest (satisfying meaning holism), but the system has only a few representations (falsifying content holism). Content holism does not entail meaning holism, either, since it might be that any system must have many representations to qualify as a mind or language at all (satisfying content holism), although its particular meanings need not be interdependent (falsifying meaning holism).

Translation holism is the view that there can’t be a positive but small amount of synonymy: two minds or languages can express one meaning in common only if they express many meanings in common (Fodor and LePore, 1992, p. 6). Meaning holism does not entail translation holism, since it is logically possible for two small systems to express all the same few meanings (falsifying translation holism), though the meaning of each representation depends on all the rest in its system (satisfying meaning holism). Nor does translation holism entail meaning holism, since some specific, large stock of "core" representations might be necessary for any system to qualify as a mind or a language at all (so that all minds and languages share this core, satisfying translation holism), though particular meanings aren’t all interdependent (falsifying meaning holism).

Anthropological holism is the view that the meaning of a representation depends on nonrepresentational conditions, e.g., causal relations to represented things (Fodor and LePore, 1992, p. 6). Meaning holism without anthropological holism might approach the views of (some?) structuralists and poststructuralists, who claim that the meanings of "signs" are interdependent, but do not depend on anything besides signs. Anthropological holism without meaning holism might approach the views of (some?) Fodor, who holds that in most cases the meaning of a representation depends on causal relations, but not on other representations.

My version of meaning holism is compatible with content, translation, and anthropological holism, despite these differences. But it is not even compatible with what Putnam (1986, 1988) calls "meaning holism". As nearly as I can make out, Putnam’s holism consists of the denial of atomism (i.e., the denial of a substantial role for definitions), together with an insistence that meanings are not differentially dependent. (It isn’t clear what kind of influence or dependence relations, if any, Putnam does ascribe to meanings.)

Why my construal of "meaning holism" is more interesting

Fodor and LePore object to several alleged ramifications of meaning holism besides the ones I have mentioned, and I would like to address them briefly. First, they say that meaning holism would threaten robust intentional generalizations, where a generalization is "robust" if "individuals that fall under it are otherwise heterogeneous in lots of ways" (Fodor and LePore, 1992, pp. 13-16). However, even if meaning holism violated meaning stability, it would not violate robust generalizations. Generalizations are typically conditional—"if one thinks that p, then one will think that q". Such a conditional statement can be true of someone even if they do not even have the ideas that would be necessary for thinking that p. Second, they warn that reference holism—the view that the referent of a representation depends on the referents of all others—would prevent people at different stages of science from referring to the same things (cf. Kuhnian incommensurability). This in turn would undercut a notion of scientific progress that may be vital to standard defenses of scientific realism (Fodor and LePore, 1992, pp. 11-13). Given this possibility I want to point out that a holist can hold that meaning is holistically and differentially interdependent without holding that reference is, on the (quite standard holist) assumption that reference is not the whole of meaning. (I will defend this assumption in section 2 below.) Finally, they suggest that versions of meaning holism that rest on normativity might prevent the "naturalization" of meaning (Fodor and LePore, 1992, pp. 16-17). However, the version of holism I defend depends not on normativity but on causal relations among representations, and so leaves open the possibility that meaning can be naturalized in the direction of such causal relations.

For similar reasons, as far as I can tell, the versions of holism that Fodor and LePore focus on—content holism and translation holism—do not have these problems, either. And since they do not lead to meaning holism as a meaning-dependence claim, they don’t clearly have worrisome ramifications for atomism, stability, or compositionality. This is why the version of meaning holism I have described seems to be the version of primary philosophical interest.

2 Comparison with referential theories

A referential (or denotational) theory of meaning, as I will use the term, holds that meaning is fully determined by reference, so that any two representations have the same meaning if they have the same reference. (Such a theory focuses on the meanings of representations that in some sense purport to refer—such as names and predicates as opposed to logical particles. For brevity, I will take this restriction for granted in this section.) Referential theories differ according to their construals of "reference". I will distinguish two broad options. One might hold that sameness of meaning is insured by sameness of referent, or the existing entity, if any, that a representation is about.<2> Alternatively, one might hold that sameness of meaning is insured by sameness of reference condition, or the way the world has to be in order for the representation to be about some existing entity (i.e., to have a referent).<3> For example, on some views "intact creature with a heart" and "intact creature with a kidney" have the same referent—since all and only the (real) intact creatures with hearts have kidneys—but different reference conditions—since there are possible intact creatures with only one of these organs.

Many philosophers resist referential theories on the grounds that meanings should be individuated more finely than referents or reference conditions, if meaning is to be of service to psychological explanation. Psychologists appeal to meanings to specify groups of mental representations that (allegedly) enter into interesting psychological generalizations (e.g., "intentional laws"). However, referents and reference conditions do not appear sufficient for this purpose. On plausible assumptions, for example, [water] and [H2O] have the same referent, since water and H2O are the same phenomenon. They even have the same reference conditions, since no matter what way the world is or can be, either both refer (to the same thing) or neither refers. Similar points hold for [Cicero] and [Tully], and for isomorphic pairs of representations "involving" them, such as [Cicero was covered with water] and [Tully was covered with H2O]. Nevertheless, each of these representations enters into very different psychological generalizations than the other member of its pair. For example, although there might be some interesting psychological generalizations shared by all and only people with a [water is healthy]-belief, there might be few if any shared by all and only people with either this belief or an [H2O is healthy]-belief. (The case is the same for [Cicero was a great orator]-beliefs and [Tully was a great orator]-beliefs.) Distinctions required for psychological explanation are also required for rational assessment. It appears rational and even common for someone to deny [Cicero is Tully] or [Water is H2O], though it appears irrational and rare for someone to deny [Cicero is Cicero] or [water is water]. On any view, then, referents and reference conditions alone are not individuated finely enough to reflect these differences among mental representations. Of course, this isn’t an argument against referential theories of meaning, unless we add the premise that meanings alone are sufficient to reflect these differences. To save a referential theory, one option is to insist that the members of these pairs do have the same meaning, but differ with respect to other psychologically or rationally relevant dimensions (see Fodor, 1990). This maneuvering may be attractive because referential theories promise to provide a large amount of meaning stability. But in "Holist", section 2, I argue that PSHAW also provides at least this much stability. While the resulting situation has many earmarks of a philosophical standoff, there is some room for argument.

Even if referential theories and PSHAW are equally viable with respect to stability, PSHAW is more attractive in paving the way for the potential benefits of meaning atomism. First, PSHAW enables us to operate with a simpler metaphysics of properties. Take the mental representations [unicorn], [centaur], [phlogiston], [ether], and [caloric]. On a referential theory, their meaning should be exhausted by their reference.<4> If these terms are not all to have the same meaning, a referential theory must assume that they refer to different things. At first glance, however, they all refer to the same thing, namely, nothing (or, at best, the null set). There are no unicorns and centaurs, and there is no phlogiston, ether, and caloric. A referential theorist may respond to this sort of case by claiming that these representations refer to the different properties of being a unicorn, of being ether, etc. Of course, this presupposes that the world does contain properties that are uninstantiated, and perhaps uninstantiated as a matter of physical or logical necessity.<5> I don’t want to enter the debate about whether there are such properties. My only point is that if PSHAW pans out, we needn’t be driven to the bold metaphysical claim that they do exist, merely in order to individuate the meanings of these terms. For the terms may have psubs that display their meanings in terms of more tractable representations. For example, one person’s [unicorn] might be semantically equivalent to a syntactically complex representation [thing actually represented by "unicorn"], and/or to [thing represented in works of fiction as a white, one-horned, horse-like animal]. Even though these psubs, taken as wholes, may not differ in reference, their (nonlogical) parts do differ in reference ([works of fiction], [horned], ["unicorn"], etc.). Since PSHAW can make do with properties of being a work of fiction, of being horned, and of being "unicorn" (or perhaps with nominalistic copies of works of fiction, horned things, and "unicorn"-tokens), psychosemanticists needn’t bet their theories on the existence of the property of being a unicorn.

This benefit of PSHAW is an illustration of a wide class of benefits. Both PSHAW and referential theories must include an account of referents and reference conditions, but PSHAW only needs an account that applies directly to the atoms, while referential theories face the much more difficult task of providing an account that applies directly to all representations (see "Atomist", section 1). If we accept PSHAW, we don’t need a systematic theory of the reference of the nonatoms, and, we only need to specify the atomic meanings of each atom, rather than the meanings it has via its associations with other representations in its system.<6> Similarly, PSHAW (unlike referential theories) promises at no extra cost an account of concept acquisition according to which only atomic concepts need to be innate or otherwise "unlearned".

Impure referential theories

The basic dilemma faced by attempts to individuate meanings is that meanings are distinguished either too coarsely for a strong role in psychological explanation, or too finely for meaning stability. The fewer meaning-distinctions there are, the less psychology can do with meaning; the more meaning-distinctions there are, the less likely it is that two merely similar representations share meaning. PSHAW distinguishes meanings finely enough for psychological purposes, but avoids reducing the likelihood of shared meaning by multiplying the number of meanings per representation. Fodor (1987, 1990) has also attempted to escape this dilemma, by offering (several potential) modifications of purely referential theories. In the rest of this section, I would like to argue that none of these proposals successfully avoids both horns of the dilemma. This provides additional reasons to favor PSHAW over referential theories.

Fodor attempts to distinguish the semantic properties of proper names such as [Cicero] and [Tully] by suggesting that [Cicero] and [Tully] are "synonymous but differ in presupposition". The mental representation [Cicero was wet], he explains, "says, in effect, that he was wet and presupposes that he was called ‘Cicero’". By contrast, [Tully was wet] "says that he was wet too, but it presupposes that he was called ‘Tully’" (1987, p. 85). However, Fodor does not provide a theory explaining which things are presuppositions and which things are not. If [Cicero was called "Cicero"] is presupposed by [Cicero was wet], why isn’t everything else one believes about the referent of [Cicero] presupposed as well? Given the inferential holism described and applied in "Holist", sections 3-4—and Fodor concedes this sort of inferential holism—what (even rough) psychologically natural principle can Fodor offer to let in some associated beliefs as presuppositions, but not the rest? Pending a principle, Fodor’s theory threatens to violate meaning stability just as surely as the holistic versions of IRS that he wants to avoid. Furthermore, on independent grounds it is implausible to suppose that there can be synonymy despite differences in presuppositions, at least if these function according to a Strawsonian (1971) account. On such a view, if a statement or belief has a false presupposition, the statement or belief is neither true nor false. Given that there is no king of France, neither "the king of France is bald" nor the contrary "the king of France is not bald" is true. So if it turns out that no one is called "Tully", on Fodor’s view, it is possible for [Cicero was wet] to be true while [Tully was wet] is neither true nor false. I agree that this is possible, but this seems to imply that the two representations must not be synonymous. How can one belief be true while its "synonym" is not? It is one thing for Fodor to divorce meaning from inferential role, but if he also divorces it from truth (or reference) conditions, no eligible spouses remain.<7>

Fodor has a different proposal for the [water]-[H2O] case. He says that although [water] and [H2O] "express the same property, the second is complex, built out of formulas which themselves denote hydrogen and oxygen." He explains that he does want to "let into meaning those implications that accrue to an expression in virtue of its compositional semantics; i.e., in virtue of its relation to other expressions as occur as its grammatical constituents" (1987, pp. 86-87). Although this is a popular solution to the problem, I think its initial plausibility is a mere artifact of the particular examples usually chosen. Consider, instead, the mental representation [gold] and the mental representation [Au]. Do they have the same meaning? We have as much reason to doubt that they do as we have in the case of [water] and [H2O]. For someone can learn chemistry—the atomic theory of matter, the periodic table, etc.—and yet doubt that gold is Au, because (for instance) he doubts that gold is an element. Yet, in whatever sense [H2O] is supposed to have "grammatical constituents" or a "compositional semantics", it isn’t nearly as obvious that [Au] does.<8> So Fodor is simply lucky that the mental representation [H2O] is structured, and the mental representation [water] unstructured, if indeed they are.

To pursue Fodor’s "grammatical constituents" proposal beyond this initial worry, we need to determine what relation between representations makes for "grammatical" constituency. I think there are three salient construals to consider, since Fodor doesn’t explicitly choose among them: syntactic relations (that a representation bears to its spatiotemporal parts), functional relations (that a representation may bear to spatiotemporally separate representations), and semantic relations (whatever they may turn out to be). It will be simplest to consider these options in reverse order. Construing "grammatical" as "semantic" would at least secure Fodor’s desired relationship between having "grammatical constituents" and having a "compositional semantics". But in the context of giving a theory of meaning, this option is unavailable to Fodor, on pain of circularity. Also, without other constraints on constituency, this idea would leave it completely open whether or not every representation is a "semantic constituent" of every other. So this would not provide Fodor with a way to avoid meaning holism and meaning instability.

The situation is similar with an appeal to "functional constituents". Fodor seems to invite a functional construal of constituency, since he seems to place an inferential requirement on the possession of a complex concept:

[H]aving the concept H2O requires more than just having a concept that expresses the property of being water; it also requires that you have some beliefs about hydrogen and oxygen (viz., that water has them as constituents). (1987, p. 87)

However, if the arguments in "Holist", sections 3-4 are correct, inferential requirements threaten to lead straight to meaning holism, which Fodor’s semantic theory is intended to avoid. The question is: given inferential requirements on semantic complexity, what psychologically relevant factor would make complexity depend on only certain inferential relations? Fodor contrasts the required beliefs about hydrogen and oxygen with "irrelevant" beliefs, such as the belief that Cicero was wet: "neither the concept CICERO nor the concept WET is a constituent of the concept H2O" (1987, p. 87). But how does Fodor know which concepts are constituents—if this is to be construed functionally—and which beliefs are relevant or irrelevant to functional complexity? Consider his suggestion for a relevant belief—that H2O has hydrogen and oxygen as constituents. It at least seems possible to have a concept of H2O without having this belief. Here is one possible scenario: although hydrogen and oxygen atoms exist, H2O isn’t really made of them; rather, when two hydrogen atoms and one oxygen atom "combine", their nature changes in such a way that the resulting "molecule" is really an undifferentiated cube. (It might even be true that, through a strange quantum-mechanical process, investigating such cubes in certain ways alters them back into hydrogen and oxygen atoms.) Even if we came to believe all this, it might still be reasonable to believe that such cubes are H2O (they are, after all, the same sort of stuff as virtually all alleged-H2O-samples). Therefore, we might still believe that water is made of H2O (though not that it is made of hydrogen and oxygen), and, for that matter, that Cicero was wet. If inferential holism is true, then the belief that H2O is made of hydrogen and oxygen is simply more strongly associated (in at least some contexts) with the mental representation [H2O] than is the belief that Cicero was wet. But this would not license the principled distinction Fodor wants to draw between them.<9>

The final option that seems available to Fodor is to build into meaning distinctions based on syntactic complexity, where this is construed more strongly than mere functional complexity. [water] and [H2O] would have different meanings, on this view, because their parts have different reference—[H2O] has a part that refers to hydrogen, while [water] does not. The difficulty here is that we have no reason to believe that [H2O] has syntactic parts. Even if the linguistic representation "H2O" has parts, the mental representation [H2O] may not. Even if we have a language of thought, and even if, more particularly, some of the token mental representations that govern our use of "H2O" are syntactically structured, there is no reason to suppose that they all must be. For example, perhaps our brains can and do form syntactically complex formulae, but also form syntactically simple "abbreviations" of them (perhaps for more efficient storage). Such abbreviations may have complicated functional relations—for example, they may be psubs for the complex formulae—even though they are syntactically simple. Indeed, Fodor grants this very possibility:

You could, presumably, have a syntactically primitive symbol that means is a round square so long as it is ‘introduced by’ a definition. Whatever, precisely, that may mean. (Fodor, 1987, p. 37)

Even if we were to have evidence that our system of mental representation in fact avoids abbreviations, a theory of meaning should not make abbreviations impossible, as would be the case if syntactic properties were built into meaning. This makes it seem plausible that what makes a (perhaps syntactically simple) representation have a compositional semantics is that it is inferentially related to other representations in a certain way, e.g., by a definition relation. Furthermore, idioms such as "kick the bucket" present some reason to doubt that syntactic complexity is even sufficient for having a compositional semantics. (I assume that a theory of meaning should not rule out the possibility of mental idioms, and also should not take advantage of their contingent absence.) The upshot of this is that syntactic simplicity and complexity are at best heuristic indicators of semantic simplicity and complexity, and that the genuine indicators are inferential (and perhaps definitional), as PSHAW would have it.

3 Comparison with two-factor theories

Among the theories of meaning currently in circulation, the ones that seem closest to securing both meaning stability and the psychological relevance of meaning are the so-called "two-factor" versions of IRS (see, especially, Block, 1986). Since these are two of the main benefits I have been claiming for PSHAW, and since PSHAW is also a version of IRS, I should make explicit my reasons for preferring PSHAW.<10> The basic idea of a two-factor theory is to identify the meaning of a mental representation R with a pair consisting of (i) R’s reference (i.e., R’s referent or reference condition), and (ii) all or part of R’s inferential role (i.e., the ways R is inferentially connected to other mental representations). The contrast with purely referential theories is clear: referential theories identify meanings with the first (referential) factor, and so take meanings to be strictly independent of the second (inferential) factor. By factoring out inferential role, the referential theorist has difficulty providing meaning with a substantial part to play in psychological explanation and rational assessment (see the previous section). By factoring in inferential role, two-factor theories seek to account for the psychological relevance of meaning.

The price of this, as one would expect, is that some maneuvering is necessary to account for meaning stability on a two-factor theory. On referential theories, two representations can be synonymous even given widespread or complete difference in inferential role. However, if there is an inferential role component to meaning, and if synonymy is sameness of meaning, then two representations can be synonymous only if they have identical inferential role components. Furthermore, if inferential roles are treated holistically—so that all or most of a representation’s role helps constitute its meaning—very few representations will ever share meaning. So there are several options regarding meaning stability available to two-factor theorists:

(a) reject meaning holism, by including only some small portion of inferential role in meaning, or

(b) reject meaning stability, perhaps in favor of mere similarity relations among meanings, or

(c) treat synonymy as sameness of the referential factor of meaning.

Option (a) faces challenges analogous to those I described in "Holist", sections 3-4, for attempts to develop a nonholistic inferential role theory of semantic equivalence. Which small part of a representation’s inferential role counts as a component of its meaning—the most frequently used part, or the strongest, or the metalinguistic part, or the natural-kind part, etc.? And why not the rest, if we want to be serious about using meaning to reflect realistic psychological complexities? Relying on the previous objections, then, I will assume that two-factor theories treat holistic inferential role as a component of meaning, ruling out option (a). Option (b) might be an acceptable price to pay for securing the psychological relevance of meaning, were it not for the fact that PSHAW does better, securing this relevance without sacrificing literal meaning stability. Option (c) can seem to be a trick, a mere redefinition of the word "synonymy" (and its "synonyms"—presumably "sameness of meaning" would have to be treated as an idiom for "sameness of the referential factor of meaning"). Alternatively, option (c) seems to entail the uncomfortable—perhaps contradictory—claim that inferential role is strictly and literally a part of meaning, but strictly and literally this is irrelevant for measuring identity of meaning. Still, if stability and psychological relevance are important enough, we should be reluctant to reject a maneuver such as (c) out of hand. Fortunately, there are independent reasons to prefer PSHAW over two-factor theories.

I can be brief about the main advantage of PSHAW, since it rests on the same point about simplifying the theory of reference that I mentioned in "Atomist", section 1. Every theory of meaning on the table—referential theories, two-factor theories, and PSHAW—needs a theory of reference. The problem is that at present it is too hard to make a theory of reference work. By securing meaning atomism—a way to define arbitrary mental representations in terms of a small stock of semantic primitives—PSHAW is designed to make this job easier. Given PSHAW, instead of trying to develop a theory of reference that can apply directly to troublesome ideas such as [God], [unicorn], and [cow], we can make do with a theory that applies only to the atoms. In my view, barring some major breakthrough, the only way we're going to get an adequate theory of reference is to accept the widespread presence of definitions, and (barring some solution to the challenges in "Holist", sections 3-4) the only way we can do that is to accept multiple meanings. While two-factor theories expect simply to add a theory of inferential role to a theory of reference (as soon as the right one drops from the sky), my hope is to use a theory of inferential role (in the form of definitions) to make it possible to develop a theory of reference.<11>

While I grant that two-factor theories account for psychological relevance and stability—at least given option (c) above—I think that PSHAW does a noticeably better job in these regards. The point about psychological relevance applies not only to holistic two-factor theories, but also to any holistic IRS without multiple meanings. I suggested that psychologists use meanings to specify mental representations, and to express generalizations about their inferential role. However, if possible, we should preserve the possibility of a stronger, explanatory role for meaning. It would be nice to be able to explain, and not merely describe, at least some aspects of a representation’s inferential role by appealing to its meaning. But if "the" meaning of a representation itself consists of all aspects of inferential role, this would not be possible. Given PSHAW, by contrast, we have the potential to explain any given bit of a representation’s role by using the representation’s "other" meanings—the ones that depend only on other parts of the representation’s role. The second advantage of PSHAW, concerning stability, applies not only to two-factor theories but also to referential theories. Simply put, PSHAW can provide more meaning stability than its rivals. While all these theories can describe two representations (in different minds, or in one mind over time) as precisely sharing referential or "coarse-grained" meaning, only PSHAW can describe two representations (with different but overlapping inferential roles) as precisely sharing "fine-grained" meanings (a.k.a., definitions, Fregean senses, or modes of presentation). Referential theories have no truck with fine-grained meanings, and holistic two-factor theories, apparently, can only have truck with fine-grained meanings that depend on entire inferential roles. For this reason, PSHAW can be more precise than two-factor theories (or referential theories) in comparing two minds (or one mind over time) for semantic identities and differences.

4 What about public language?

Given PSHAW, should we say that token linguistic representations have all the same meanings as their governing mental representations? While I don’t think this result would be obviously wrong, there are other salient possibilities. For starters, the right theory of linguistic meaning may simply be independent of the right theory of mental meaning. Millikan, for one, has suggested that mental meaning and linguistic meaning should be explained separately (in terms of certain teleological functions), rather than explaining one in terms of the other (1984, ch. 3). If there is such independence, then PSHAW has no relevant consequences for linguistic meaning. Suppose, then, that linguistic meaning does depend (somehow) on mental meaning. Even so, it might turn out that linguistic meaning is exhausted by reference, as in purely referential theories of meaning. Referential theories of mental meaning are troublesome, on the assumption that mental meaning is beholden to constraints due to psychological explanation. But since linguistic representations do not function as psychological intermediaries between perceptual and motor organs, these constraints may not apply, and referential theories of linguistic meaning may be safe. If so, then PSHAW would be committed at most to claims about the referential meanings of linguistic representations. Finally, even if linguistic representations have fine-grained meanings as well as reference, and even if linguistic meaning depends wholly on mental meaning, it may be possible to extend PSHAW to language while minimizing linguistic ambiguity.

This depends, mainly, on how much "psycholinguistic synonymy" a theory of meaning needs to provide. Many lines of thought in the philosophy of language and mind seem to depend on claims of psycholinguistic synonymy: that mental representations have the same meaning as the linguistic representations they govern. For example, Putnam advances a thesis about public-language meaning (namely, that it is sensitive to holistic constraints derived from expert field-linguistic interpretation), and appeals to psycholinguistic synonymy to draw a mental moral:

If our thinking is ultimately conducted in an internal language of thought … the same will be true of the items in lingua mentis corresponding to the words in public language …. "Translating" our public language into lingua mentis will … not solve the problem of conceptual content, but will just move it from one language to another. (Putnam, 1988, p. 16, see also p. 22)

Similarly, philosophers such as Burge (1979) and Kripke (1979) who draw conclusions about the content of mental representations directly from facts about ordinary public-language attributions of content seem to commit themselves to psycholinguistic synonymy. Although my theory is consistent with psycholinguistic synonymy, in this section I will consider and reject some standard reasons for believing this thesis. Roughly, my claim is that even if the relevant mental and linguistic representations must share reference (or coarse-grained meaning), they need not share fine-grained meaning (Fregean sense or definition).

To what extent must a linguistic representation L have the same meaning(s) as some mental representation M that governs L, at least in the mind of someone competent or expert in the use of L? If no synonymy is required, then PSHAW does not force any specific theory of linguistic meaning. If complete synonymy is required—if M and L must share all meanings (or all referential meanings)—then PSHAW is committed to as much multiplicity of linguistic meaning as mental meaning (or mental referential meaning). The interesting possibility is the one in which partial synonymy is required, so that M and L must share at least one meaning. I will describe some natural ways to apply PSHAW to language in this eventuality, without trying to choose among them.

First consider the case of a word-type W, shared by various speakers.<12> W has a set of governing (type or token) mental representations {R1,...,Rn} in the speakers’ minds. Given PSHAW, each Ri has a set of meanings MRi. Among others, the following sets are candidates to be W’s set of meanings, and are each compatible with partial psycholinguistic synonymy:

(1) the union of the sets MR1,...,MRn,

(2) the intersection of the sets MR1,...,MRn,

(3) the union of the sets MRe, for each Re that belongs to a W-expert, and

(4) the intersection of the sets MRe, for each Re that belongs to a W-expert.

(If linguistic meaning is purely referential, of course, we can consider only the referential subsets of these sets.) None of these options seem particularly appealing, though, again, none seem obviously wrong. Option (1) holds the meanings of a word-type hostage to any speaker’s beliefs, however kooky. Option (2) seems to yield the result that W means only thing called "W", since in typical cases that will be the only meaning shared by the associated mental representations in everyone who uses W. While there is something odd about saying that "snow is white" means that the thing called "snow" has the property called "is white", I think it is merely odd. It postpones the question of what is called "snow" or "is white", but perhaps this question should be postponed. Perhaps, for example, the answer depends on nonmetalinguistic meanings of [snow] and [is white], rather than on nonmetalinguistic meanings of "snow" and "is white". Option (3) fares better than (1), but depends on some specification of the experts—in particular, an account that doesn’t depend on W’s meaning. (As a half-hearted stab in the dark, an expert might be someone deferred to, who in turn defers only to others who return the compliment.) Even so, (3) holds the meanings of W hostage to arbitrary "off-duty" beliefs of these experts. Option (4) seems to be the best of these, so long as there is a noncircular specification of the experts, and so long as the W-experts agree on more than their metalinguistic beliefs. Of course, there are other possibilities along these lines, ones that appeal to "most experts", "most widespread mental meanings", etc.

We can uncover a few other ways to secure partial psycholinguistic synonymy by turning to a token word t of type W. We might suppose that t has all the meanings of W plus all the meanings of the particular mental representation that governs it (i.e., all the meanings in one’s "ideolect"). Then [snow] would share all of its meanings with a token word "snow", even if it would share only some of them with the word-type "snow". Finally, given ideolect-meanings for tokens, there may be no need to believe that word-types even have meanings (in addition to the varying ideolect-meanings of their tokens). If there are reasonable constraints on linguistic meaning that can help us to decide among all of the possibilities I have mentioned, I don’t know what they are, and I don’t know who to ask. Needless to say, this wealth of live options is a source of discomfort to us as linguosemanticists. But it is not an embarrassment to us as psychosemanticists, and in particular it is not an embarrassment to PSHAW.

Why psycholinguistic synonymy?

Our common lore about meaning tends to be based on consideration of language, since in our everyday concerns we seldom need to distinguish mental representations sharply from linguistic representations, as cognitive scientists do. If there are to be objections to the effect that PSHAW is "counterintuitive" ("Pshaw!"), then, I expect most of them to be based on (implicit) claims about language, transported to the mind via (implicit) claims about psycholinguistic synonymy. For example, IF there are good reasons to deny that token linguistic representations have (wildly) multiple meanings, then PSHAW is in trouble IF associated mental and linguistic representations must share all of their meanings. Also, IF there are good reasons to deny that linguistic representations are definable in any of the (wildly complex) ways that PSHAW takes mental representations to be, then PSHAW is in trouble IF associated mental and linguistic representations must share even some of their fine-grained meanings (senses, definitions). Since I can’t foresee or forestall the possible "good reasons" for these linguistic claims, I would like to try to block the corresponding claims of psycholinguistic synonymy, in the rest of this paper. The following phenomena seem representative of the range of things such synonymy might be taken to explain: (a) knowing what a linguistic item means, (b) sincerely expressing one’s mental states, (c) saying what one means, and (d) believing what one says. I will discuss these in turn, with most emphasis on the first. I will argue that, at most, associated mental and linguistic representations need only share (some) referential meanings, so that no complete synonymy, and no synonymy of fine-grained meanings is required.

One very natural motivation for believing that mental representations are synonymous with linguistic representations is that this provides a good explanation of what it is to know the meaning of one’s words. This is a complicated matter to discuss, since there are several different kinds of things that may be meant by the phrase "knowing what one’s words mean". One construal is that of having the ability to use the words competently, either in normal conversation or in rigorous conversation. If this involves knowing how to do something rather than knowing that a proposition is true, it’s hard to see any consequences for the meanings of one’s mental representations. However, it might be thought that this skill does involve mental representations, so we can ask what the content of these mental representations must be. It is not enough to require that a word be governed by any old mental representation, for without the right constraints, the subject will be unable to use the words competently. One way to be incompetent in the use of "snow" is to govern it with an idea of turnips. Perhaps, as a defender of psycholinguistic synonymy might suggest, the right constraint is that the governing mental representation should share fine-grained meaning with the linguistic representation. Since I do not want to assume such synonymy, I need an equally good alternative constraint. In a slightly different context, Millikan provides some reason to resist the strong requirement of sameness of fine-grained meaning:

"Knowing the meaning of a word" … is at best a vague affair. In the case of [referring] terms it is knowing what thing or kind the word refers to. Knowing this is exactly like knowing who the bearer of a certain proper name is. Depending on the setting, "knowing who" Bernard Shaw is can involve as little as knowing that he wrote plays or as much as being able to give an analysis of his character and impact on drama. Similarly, "knowing what ‘malaria’ means" can involve as little as knowing that malaria is a disease or as much as knowing what the average physician knows about malaria. Minimally, perhaps, "knowing the meaning" of a word is knowing enough about its referent that one can tell whether a good number of sentences employing the term make sense or not. … Or perhaps it is knowing enough to be able to "follow the conversation" about a term’s referent. But whether one follows the conversation is a matter of degree, and conversations come in all degrees of difficulty. Whether or not one understands a term will then depend upon how sophisticated the people are with whom one is conversing. (Millikan, 1984, pp. 156-157)

Millikan seems to be suggesting that to use a term like "malaria" competently—and in that respect to know what "malaria" means—one must know some conversationally important things about the referent of "malaria". At best, then, the governing mental representation [malaria] must refer to something that "malaria" refers to, and must figure in conversationally important beliefs. But since there is a difference between reference and fine-grained meaning, [malaria] and "malaria" can share their referent without sharing their fine-grained meaning.

It might be suggested that the relevant sense of "knowing what one’s words mean" is something more genuinely cognitive than the ability to use the words competently. To take what would seem to be the best case for fine-grained psycholinguistic synonymy, let’s focus only on full, philosophically informed knowledge of what a linguistic item L means—perhaps of the sort that might belong to someone who happens to be both the world’s leading expert on the referent of L and the world’s leading expert on lexicography and philosophy of language. Even in this case the requirement of fine-grained synonymy appears too strong. We should ask: when one (fully) knows what L means, what proposition does one know to be true? I can think of two natural ways of formulating such a proposition:

(1) Knowing what L means is knowing that L is defined by [M], where [M] is a mental representation with the same fine-grained meaning as L.

(2) Knowing what L means is knowing that the meaning of L is the F, where "the F" refers to the meaning of L.<13>

If option (1) is correct, then to know (fully) what "water" means, the expert must have a mental fine-grained synonym for it. But this is not necessary if option (2) is correct. On option (2), the expert must have a mental representation that refers to the fine-grained meaning of "water". But such a representation need not—and, plausibly, cannot—have the fine-grained meaning of "water". For example, perhaps our expert refers to the meaning of "water" using the following definite description: [the meaning whose semantic function is to pick out a substance dominantly responsible for most English speakers’ utterances of "water", namely, clear, colorless, odorless, lake-filling, H20]. Even if this description specifies the (or one particular) fine-grained meaning of "water", neither it nor any of its parts are fine-grained synonyms of "water". ("Water" refers to a liquid, but the description refers to a meaning. And [H2O] need not have the same fine-grained meaning as "water".)

So unless option (1) is the correct account of knowing what a linguistic item means, it appears that one can know this meaning without mentally "translating" the linguistic item. But there is plenty reason to resist option (1). First, merely knowing that several representations are defined in the same (unspecified) way does not itself insure knowing their meaning; one must know what one of them means in a way other than having a synonym, e.g., in the way specified by option (2). But if one can know what a representation means without fine-grained synonymy, then option (1) cannot be the correct account of what knowledge of meaning is. Second, knowing that "water" means the same as [water], for some mental representation [water], requires "quoting" or at least referring to [water] itself. It seems odd to suppose that language understanders generally refer to their mental representations, rather than merely using them.

Let me turn to the other purported explananda of psycholinguistic synonymy: sincerely expressing one’s mental states, saying what one means, and believing what one says. Sincerity in expression does not ensure synonymy, since even if one is trying to express one’s beliefs sincerely, one may mislead due to using the wrong words. Genuinely, successfully, saying what one means seems more relevant. However, even without synonymy we can account for saying what one means, if we interpret this as saying what one means to say—i.e., as the genuine, successful, execution of a certain metalinguistic intention. Similarly, even without synonymy we can account for believing what one says, if we interpret this as believing (that) what one says is true—i.e., as having a certain metalinguistic belief. Of course, we need to distinguish between (a) believing-true something one says in a language one doesn’t understand and (b) believing something one says in the normal case. But we can do this without synonymy by requiring that one believe that what one says is true, while knowing the meaning of what one says.

As a final argument on behalf of psycholinguistic synonymy, I would like to consider the claim that one’s language of thought is one’s natural language. After all, it appears on first glance that English speakers do often think "in English", and it may even be insisted that this is the only sort of mental representation we have good reason (e.g., by introspection) to postulate. But if mental representations are items of natural language, fine-grained synonymy between linguistic and mental representations follows trivially. On the contrary, it appears on second glance that we have as much reason to deny that our mental representations are items in our natural language as we have for denying that they are bells and whistles. I don’t have any bells and whistles in my head—at best, what I have are representations of bells and whistles. Similarly, it seems, I don’t have any English words in my head—at best, what I have and introspect are representations of English words (e.g., auditory or visual "images" of English words). Just as nothing in my head—e.g., brain cells and blood cells—looks like a bell or sounds like a whistle, so nothing in my head looks like a written English word or sounds like a spoken English word. Of course, it can be that mental representations are transcriptions of natural language items, akin to Morse code, Braille, the sign alphabet, or ASCII code. But this provides no independent grounds for psycholinguistic synonymy. If sameness of fine-grained meaning is necessary for transcriptions, it would beg the question to suppose that there are mental transcriptions. And if not, then even if mental representations are transcriptions, they may not be fine-grained synonyms.<14>

PSHAW is distinctive as an approach to meaning in two ways: the widespread postulation of mental ambiguity, and the widespread postulation of mental definitions. Virtually all of the theoretical benefits I have claimed for PSHAW—e.g., for the theory of reference, for meaning stability, and for psychological explanation—rest on these features. If I am right that there is no need for complete or fine-grained psycholinguistic synonymy, then I have insulated these features from a potentially tempting set of objections based on the apparent absence of widespread linguistic ambiguity and definition relations.




<1> This appendix is not meant to be read alone, but only after the other two papers.

<2> I will assume that the referents of mental representations are existing objects, properties, and facts (the last of which, I will also assume, are typically instantiations of existing properties by existing objects). An idea of Paris refers to, or is about, a certain existing city, an idea of prettiness refers to, or is about, a certain existing property, and a belief or desire that Paris is pretty refers to, or is about, the existing fact that this city has this property. There are other viable conceptions of referents, particularly for the case of predicative representations (e.g., psychological "concepts") and propositional representations (e.g., beliefs). An idea of prettiness might be taken to refer, not to a property, but to all and only the pretty things. Beliefs (and, presumably, desires) might be taken to refer not to facts but to (something like) truth values. And representations that seem not to refer at all might be taken (stipulatively) to refer to something like the null set (assuming that sets exist). The arguments below would apply with only slight modifications given these conceptions of referents; if anything, my assumptions about referents strengthen the positions I will argue against.

<3> Fodor (1987, 1990) and Stalnaker (1984) offer, or at least court, referential theories. Stalnaker defends a conception of meanings (specifically, "propositions") as functions from "possible worlds", or possible ways for the world to be, to truth values. Predicative meanings (e.g., philosophical "concepts") might be treated in a similar fashion, as functions from possible worlds to reference values (or perhaps to referents themselves). Stalnaker’s possible-world functions are equivalent, in their effect on the individuation of meanings, to what I am calling reference conditions (including truth conditions).

<4> This seems true even for Fodor’s (1987) modified referential theory that applies only to syntactically simple representations (see below). It seems reasonable to suppose that these mental representations are syntactically simple, like their English counterparts.

<5> The alleged-property of being a unicorn, for example, may be logically impossible to instantiate if unicorns are necessarily fictional, as Kripke (1972) suggests.

<6> It might be useful to step through a specific case in order to show how PSHAW can specify the reference of the nonatoms without a general theory of their reference. Suppose that there is an acceptable theory of reference (e.g., a causal theory) for atoms, but that it runs into difficulties with a nonatom such as [water]. First, of course, PSHAW must use [water]’s psubs to reach its atomic definitions. For the sake of illustration, suppose [thing made-of-the-same-stuff as the clear drink in my big white box] is a psub for someone’s [water]-representation, and that each of its parts is an atom. At a minimum, PSHAW can specify a referent of [water] "metarepresentationally", as whatever has a [made-of-the-same-stuff]-relation to whatever has (a) a [clear]-property, (b) a [drink]-property, and (c) an [in]-relation to whatever has a [my]-property, a [big]-property, a [white]-property, and a [box]-property. Given the general theory of reference for atoms, we can specify these properties, relations, and objects without metarepresentation (in principle, given scientific good fortune). Such investigations might lead us, for example, to the property of triggering certain retinal color cells in certain circumstances (i.e., one of the referents of [white]), to a certain refrigerator, to certain samples of liquid in the refrigerator, and finally to anything composed of H2O. Therefore, H2O can be a referent of [water] even if it does not stand in the systematic relation to [water] that would be required for atomic reference—e.g., even if H2O never has, never will, and never can (as a matter of physical necessity) be causally related to the person’s [water]-thoughts. (For all I’ve said, the refrigerator and all H2O-samples might be in a black hole in all possible worlds in which they exists. Of course, the facts about reference don’t depend on what we can determine—H2O can satisfy the metarepresentational description even if we have scientific bad fortune.)

Finally, the same procedure is available even for the proposal to add referential meanings to PSHAW (as mentioned two paragraphs ago in the text). According to this proposal a representation has (at least) one meaning that is determined by its reference, so that [water], [H2O], [l’eau], and [Mickey Mantle’s favorite rinse] all share a meaning. By contrast, if a referential theory is meant as an alternative to meaning atomism, it cannot avail itself of this means of simplifying the theory of reference.

<7> PSHAW can handle semantic differences between proper names as follows. The mental name [Cicero] may have, as one of its psubs, [thing actually represented by "Cicero"], while [Tully] may psub for [thing actually represented by "Tully"]. Since these mental representations have different psubs, they have (some) different meanings. Notice, finally, that this claim is not susceptible to a familiar Kripkean objection. Following Kripke (1972), Fodor (1987) points out that "Cicero" does not mean the same as "the thing called ‘Cicero’", because Cicero might not have been called "Cicero". What this means, presumably, is that Cicero is such that in some other possible world, he isn’t called "Cicero". Suppose that analogous points hold for the mental representations [Cicero] and [the thing called "Cicero"]—i.e., Cicero is such that, in some other possible world, he isn’t called "Cicero". This does not reveal a difference in meaning between [Cicero] and [the thing actually called "Cicero"], since the thing actually called "Cicero" is such that, in some other possible world, he isn’t called "Cicero". Of course, Cicero is not such that, in some other possible world, he is not Cicero. But equally, the thing actually called "Cicero" is not such that, in some other possible world, he is not Cicero.

<8> Nor is the difference in meaning between [gold] and [Au] simply a matter of a difference in associated English expression, as with [Cicero] and [Tully]. Even putting aside metalinguistic deference, [gold] and [Au] still do not seem to have the same meaning: under normal circumstances, [Au] seems to be little more than an abbreviation for a certain entry in the periodic table. Perhaps someone’s [Au] has [element with atomic number 79] as a psub, but his [gold] does not. By PSHAW, the two representations differ in meaning, even if they do have the same reference.

<9> Of course, Fodor might weaken his inferential requirement so that, to have "the" concept of H2O, one must have some beliefs about hydrogen and oxygen, but no specific belief such as that H2O is made of them. This proposal would suffer from the other horn of the dilemma: the resulting distinction wouldn’t be fine-grained enough for relevant psychological distinctions. Suppose that Smith has the representations [H] and [O] for hydrogen and oxygen, and that Jones has the representations [Hy] and [Ox] for hydrogen and oxygen. Like us, Smith believes that H2O is the substance composed of H, H, and O. However, Jones thinks of Hy2Ox as the substance in the actual world that his grandmother likes to drink fifteen minutes after she thinks of Hy, Ox, and Hy again. Now, suppose that Hy2Ox is, in fact, water, so that [Hy2Ox] and [H2O] have the same referent. (Furthermore, because Jones "rigidly designates" the substance, [Hy2Ox] and [H2O] have the same referent in all possible worlds, which is to say that they have the same reference conditions.) The current proposal (made on behalf of Fodor) is forced to say that the meaning of [H2O] and [Hy2Ox] is the same, despite the fact that they play very different psychological roles—roles at least as different as those of [water] and [H2O]. (Also, Fodor would not be able to appeal to a difference in syntactic constituents, since their syntactic parts—[H] and [Hy], [O] and [Ox], [2] and [2]—have the same referents and reference conditions, pairwise. Thus, the [Hy2Ox] example can be used against the proposal discussed next in the text.)

<10> I want to thank Ned Block for impressing on me the need for a comparison.

<11> A two-factor theorist can also suppose that inferential role constrains reference. For example, Block (1986, pp. 643-644) suggests that aspects of inferential role—revealed in certain reactions to thought experiments—can help to decide whether causal theories or descriptivist theories of proper names are true. I’m skeptical that this is a way of making the theory of reference easier, because I think these sorts of reactions to thought experiments mainly serve to undermine causal theories and descriptivist theories and, I suspect, whatever other theories may boldly be put forward to account for "the" reference of any representation, whether atomic or not (see Unger, 1983, for some relevant thought experiments). At a minimum, it is fair to say that the emergence of two-factor theories has not led to any breakthroughs in theories of reference, and as far as I know no one claims it has or will.

<12> Perhaps two token words belong to the same word-type if they have the same (phonetic or graphemic) parts. If so, two "bank" tokens occurring in (the most natural uses of) "river bank" and "financial bank" are of the same word-type, or are "the same word". One potential way to avoid this result is to type word-tokens together if they are "copied" or "imitated" from the same earlier tokens (see Millikan, 1984).

<13> Perhaps "the F" should necessarily refer to this meaning, or at least specify certain of the essential properties of the meaning, since it would not be enough, for example, to know that the meaning of "water" is the favorite meaning of Mickey Mantle. This does not affect the argument below.

<14> It is conceivable that someone’s language of thought is English—say, that one actually does have (barely audible) English words in one’s head. Even if we found that we were such people, however, this would only serve to cast doubt on intuitive commitments about the meaning of English words. For in that case, there would be a whole side to the workings of English words that our cultural lore, our ordinary attributions of meaning, and our "expert" lexicographers had neglected. It would be unreasonable, then, simply to rely on these commitments in giving a theory of meaning for English words, or at least for those tokens that serve as mental representations.



Block, N. 1986: "Advertisement for a Semantics for Psychology", in Midwest Studies in Philosophy, volume 10.

Burge, T. 1979: "Individualism and the mental", in Midwest Studies in Philosophy, volume 4.

Fodor, J. 1987. Psychosemantics, Cambridge: MIT Press.

——— 1990: "Substitution arguments and the individuation of beliefs", in A Theory of Content. Cambridge: MIT Press.

——— and E. LePore 1992: Holism: a Shopper’s Guide, Oxford: Blackwell.

Kripke, S. 1972: Naming and Necessity. Cambridge: Harvard University Press.

———. 1979: "A puzzle about belief", in A. Margalit, ed., Meaning and Use, Dordrecht: Reidel.

Millikan, R. 1984: Language, Thought and Other Biological Categories. Cambridge: MIT Press.

Putnam, H. 1986: "Meaning Holism", Hahn and Schilpp (eds.) The Philosophy of W.V. Quine, La Salle, IL: Open Court.

——— 1988. Representation and Reality, Cambridge: MIT Press.

Stalnaker, R. 1984: Inquiry, Cambridge: MIT Press.

Strawson, P. 1971: "Identifying Reference and Truth-Values", Logico-Linguistic Papers, London: Methuen.

Unger, P. 1983: "The causal theory of reference", in Philosophical Studies, volume 43.