Towards a Griceian principle of compositionality

Grand, right?

Some copying and pasting from Cohen, at

http://aardvark.ucsd.edu/language/davidson.pdf

The section Cohen entitles "Schiffer on Compositionality".

Cohen writes:

"One of the most influential critics of compositionality is Schiffer,
who has argued against the need for compositional semantics for thought
([Schiffer, 1991]) and language ([Schiffer, 1987])."

"Schiffer’s case against the compositionality of language (which will be our focus here) is composed of independent arguments concerning (i) propositional attitudes and (ii) translationism.14

"I’ll examine the first challenge first (§2.3.1), and then consider the translationist challenge (§2.3.2).

2.3.1 Propositional Attitudes

"The first reason Schiffer gives for doubting the claim that natural languages
enjoy compositional semantic theories involves his view that no compositional
semantics can give an adequate treatment of propositional attitude sentences."

"Schiffer starts by arguing that a compositional semantics must treat propositional
attitude verbs relationally in order to explain the validity of the inference
from

Sam and Sally believe that aardvarks have legs.

to

There is something that Sam and Sally believe.

15

"But if a relational account of propositional attitude
verbs is required, then we need an account of the relata to which believers
are related by these verbs."

"However, Schiffer argues that each of the candidate
theories for these relata is unacceptable, and therefore he concludes that the
this line of thought as unfaithful to Davidson, and then suggests a different way of defending Davidson against Chihara and Haack. Although I cannot enter into detailed questions of Davidson-exegesis here, it looks to me as if Lepore’s response is reflected rather directly in Davidson’s characterization of his project as asking “how we shall describe the skill or ability
of a person who has learned to speak a language” ([Davidson, 1965], 7–8, emphasis added)."

"Moreover, as I have suggested, Matthews’s interpretation of Davidson as deriving the finiteness of semantic theory directly from learnability constraints makes Davidson vulnerable to his own charges against translation theories. For these reasons, I am inclined to trust Lepore’s interpretation, and to think that it should be used to rule out Matthews’s objection as well as that of Chihara and Haack."

"Note 14. In addition to the two arguments I’ll consider, he makes some further suggestions intended to motivate the same conclusion at ([Schiffer, 1987], 182) based on the existence of natural language constructions for which no successful compositional semantic account has been advanced. But these considerations (concerning which there has been substantially more
progress than Schiffer indicates) only show that tasks still remain for compositional semantics, not that the entire research program is bankrupt."

"Note 15. A relational account explains this inference since it renders the first expression as something
like

Believes (Sam, t) & Believes (Sally, t)

and the second as roughly

(9x) (Believes (Sam, x) & Believes (Sally, x))

See ([Schiffer, 1987], 7–10) for a more detailed presentation.
8

relational theory (and therefore the prospects for a compositional semantics of
propositional attitude ascriptions) is doomed.

"While I can’t reply to all the details of this argument, I think we can see that its conclusion is too hasty. There are several places where one can attempt
to defuse Schiffer’s pessimism. First, the possibility remains that Schiffer’s case
against any one of the candidate theories of the objects of the attitudes is unsuccessful (or that a new form of a relational theory not considered by Schiffer might skirt the charges he levels at other theories). This possibility is especially important since, as Fodor remarks in pursuing just such a response [Fodor, 1988],
Schiffer’s case against compositionality only works if he can successfully block
every single candidate theory. If only because of the breadth of the territory
Schiffer surveys, it seems quite likely that at least one of his targets will find a
way to evade Schiffer’s criticisms."

"Second, the measurement-theoretic account of
propositional attitude sentences in [Matthews, 1994] (see also [Davidson, 1989])
provides a way of understanding how a compositional semantics could explain
the validity of Schiffer’s inferences without treating propositional attitudes relationally. For this reason, even if Schiffer’s criticisms of every relational treatment
of the attitudes were watertight, this by itself would not show that natural
languages lack compositional semantic theories."

2.3.2 Translationism

"Schiffer’s second argument against compositional semantics is a direct attempt
to demonstrate that unboundedness arguments cannot show the need for a compositional
meaning theory for natural languages, and therefore that the claim
that natural languages have compositional semantics is unmotivated.16 He argues
for this claim by providing a description of a possible world in which a
person, Harvey, has the capacity to understand utterances of indefinitely many
sentences of the language E1 (non-indexical, unambiguous English) on the basis
of finite exposure to E1, and then attempting to explain Harvey’s capacity
without appealing to a compositional semantics for E1. Schiffer’s description
of language understanding (henceforth, ‘translationism’), assumes that Harvey
thinks by computing over formulae in Mentalese, and holds that Harvey understands
a sentence s of E1 just by translating s into a unique Mentalese sentence
f(s). Schiffer admits that the ability to compute the Mentalese translation
of E1 sentences requires knowing their syntactic configurations, but he claims
that these translations can be carried out without knowing anything about the
semantics of E1 sentences, so it will be irrelevant to this picture of language
understanding whether or not the semantics of E1 is compositional. Of course,
translationism assumes that the semantic properties of E1 expressions must be
preserved by their Mentalese translations, but this restriction merely represents..."

"Note 16. Again, there have been other arguments in the literature (not based on unboundedness) that have attempted to show that natural languages possess compositional semantic theories. If any of these is found persuasive, then Schiffer cannot demonstrate that the hypothesis that a natural language like English has a compositional semantic theory is unmotivated merely by undermining unboundedness arguments."

"... an adequacy condition on the translation function f, and does not amount to
the assumption that there is a compositional semantics for E1. Indeed, the
heart of the translationist challenge to compositionality is the claim that Lunderstanding does not require any knowledge about the semantics of L.17
Translationism is an interesting picture of the relationship between language
and thought, and poses a serious challenge to arguments for the compositionality
of natural language. In what follows I want to consider a few challenges to the
success of translationism. One reason for doubting the adequacy of the conception of language understanding offered by translationists comes out in [Lepore, 1996]. Lepore suggests, reasonably, that a theory of language of understanding should not only describe a transition between (1) and (2), but also should make this transition rational by making available the believer’s reasons for making the transition from heard utterances to beliefs about what is said."

Sally utters ‘Aardvarks have legs’.
Sam believes that Sally said aardvarks have legs.

"For, Lepore suggests, we should no more attribute linguistic competence to
speakers who literally lack reasons for making such transitions than we should
attribute responsibility for their actions to individuals who literally cannot provide reasons for their criminal acts. How can the translationist attempt to rationalize this transition? Since translationists understand the transition from (1) to (2) as some sort of unmediated and automatic transduction (see the analogy between translation and perception at ([Schiffer, 1987], 198)), the natural place to look for a possible translationist answer to the demand for justification is in terms of the reliability of transducing mechanisms (cf. [Dretske, 1981] or [Goldman, 1986])."

"Let it be that a belief is reliable just in case it is produced (in the normal way, under normal conditions) by a reliable belief-producing mechanism, and let a belief-producing mechanism count as reliable just in case there is a nomic connection
between its producing a belief and the truth of that belief. Significantly, a reliabilist conception of the transition between (1) and (2), like other reliabilist
proposals, confer epistemic status on their products without requiring that subjects
know that they do."

"Sam’s belief will be reliable even though (unless he
engages in very unusual kinds of investigations of his cognitive endowments) he
will not know that it is reliable."

"The problem that Lepore raises, then, is exactly analogous to the difficulty
that externalist theories in epistemology have regarding epistemological justification more generally: even if we accept an externalist account of knowledge,
the very externalism of the account means that the special epistemic status of
reliably formed beliefs is not within the ken of the knower, and so is not capable
of supplying that subject with a rationalizing justification for her knowledge.
Likewise, even if the transition between (1) and (2) is a reliable transduction..."

Note 17Translationism is also advocated as a theory of language understanding in [Fodor, 1983],
[Fodor, 1989], and [Harman, 1975].

"... in Sam’s head, Sam’s ignorance of the transduction’s reliability means that the
reliability doesn’t provide Sam with a rational justification for making the transduction."

"“That he has a certain faculty that, ceteris paribus, delivers him from
heard . . . utterances to true beliefs about what is said fails to reveal his reason”
([Lepore, 1996], 51). What would suffice to give the kind of reason we are seeking?"

"If Sam knew that his linguistic competence endowed upon him a reliable
mechanism for acquiring true beliefs about what is said from heard utterances
of English, then he could answer that he knows he has the right kind of reliable
mechanism, and that this reliable mechanism produced the belief in question.
But, of course, this is precisely the sort of explanation that translationism, like
other forms of reliabilism, makes unavailable."

"In summary, if Lepore is right to demand that a theory of language interpretation
must provide the sort of justification that makes available Sam’s reasons
for coming to hold beliefs about what Sally said on the basis of her utterances,
then translationism cannot be an adequate theory of language understanding.
And if this is right, then translationism cannot be used to undermine Davidsonian
unboundedness arguments concluding that natural languages must have a
compositional semantics."

"A second challenge to translationism concerns the question whether it can,
in fact, do without a compositional semantics for natural language. Recall
that, according to translationism, an L-user understands a sentence s of L
just by translating s into a unique Mentalese sentence f(s). The difficulty
is that, if we make a few assumptions and stipulations, we can show that L
has a compositional semantics if Mentalese has a compositional semantics. In
particular, we can identify the semantic value of an L-expression s with the
semantic value of f(s), hold that an L-expression s is simple just in case f(s)
is a simple expression of Mentalese, and stipulate that if L-expression s1 is a
constituent of L-expression s2, then f(s1) is a constituent of f(s2). Since all
parties in this dispute will agree that we can say whatever we can think, it will
be uncontroversial that there is a well-defined mapping f−1 taking sentences of
Mentalese into sentences of L such that f−1(f(s)) = s and f(f−1(m)) = m. But
now it’s clear that there will be a principled and exhaustive partition between
simple and complex expressions of L if there is such a partition among Mentalese
expressions, and that there will be only finitely many simple L-expressions if
there are only finitely many simple Mentalese expressions. Finally, the semantic
value of L expression s is the semantic value of f(s), and therefore the semantic
value of s will be determined by the constituents of f(s) plus their syntactic
configuration in f(s) ifMentalese is compositional; but since constituents of s are
taken by f to constituents of f(s), this means that if Mentalese is compositional,
then the semantic value of s is determined by the the constituents of s plus their
syntactic configuration. This shows that L will have a compositional semantics
if Mentalese does."

"Schiffer therefore needs to show that translationism doesn’t
require the compositionality of Mentalese if he is to maintain his claim that
translationism can explain language understanding without presupposing that
natural languages possess compositional semantic theories."