Problem we’re trying to solve: could two people say the same words in all situations but mean something different by those words?

Consider a speaker as a distribution p(sentence | context) for some set of sentences and contexts. Claim: there exists a set of “accessible” contexts and conditional distributions such that we can permute all the sentences and wind up with the same p(sentence | context).

Alterantive way of asking the question: can we write a translation model that generates appropriate utterances in all contexts but does not actually preserve meaning?

[The weaker version of this—can we generate a bad translation phrasebook that preserves p(s) in the source language?—is empirically false given the success of translation-as-decipherment work.]

What’s in a context? A whole history of observations. (But not straightforward to disentangle this from experiences that define the language rather than its use.)

A “linguist” observes only p(s | c), from which they wish to recover a mapping from every s to some s’ that is meaning-preserving. [In what sense of meaning?] [Such a mapping is in any case not guaranteed to exist even for simple extensional languages.]

So present a speaker with a stimulus, get them to respond yes or no; the set of stimuli to which the speaker assents lets you identify the corresponding word. (Q seems to imagine a more interactive process where the linguist begins with a particular hypothesis about the translation.) [Compare to truth-conditional accounts of semantics.]

“meaning, supposedly, is what a sentence shares with its translation” (32). [This seems circular—how do we identify translations if we haven’t first identified what they should share? And if we also are happy with “translations” being things we’ve elicited in the manner described above, then the elicitation procedure defines meaning, not “translation”.]

Q: exactly—call this the stimulus meaning.

And versus Carnap: Framing this as a translation criterion avoids some of the problems needed to prevent the word “unicorn” from being part of the truth conditions for unicorn. [Arguably the spelling is part of the truth conditions.] In this framework subjunctives cannot be achieved via stimulus; in Carnap those judgments can be supplied by the speaker.

Distinguish between occasion sentences and standing sentences—the second don’t require stimulus of the kind above. [Does he just mean indexicals?]

Or, a standing sentence is just an occasion sentence but about past experience.

[Didn’t we say that the relevant context for a stimulus includes a whole history of observations? In this case, why isn’t the initial stimulus followed by some generic waiting period a “stimulus” for a standing sentence? Just that it’s outside the modulus.]

Obviously for purposes of elicitation it may be nontrivial for the linguist to determine the speaker’s full history.

Or, imagine a fly that always appears in the presence of rabbits [such that natives don’t have a word for “rabbit without fly”?].

Very important point: the stimulus meaning of “gavagai” includes moduli in which a local has recently made the sound [gavagai]. The stimulus meaning of “rabbit” includes moduli in which a local has recently made the sound [ræbɪt] (but not [gavagaj]). So can never have perfect synonymy via equivalence of stimulus meaning without some complicated operator that replaces all language-like stimuli with their translated versions.

And, as noted above, we can’t meaningfully distinguish between the case where a sentence S picks out worlds but only worlds upon learning some additional piece of information , and the case where S picks out to begin with.

Or, culture-dependent obeservation histories that would stun a speaker to silence. [Not a problem with “what belief does this utterance induce?”.]

For any two speakers whose social contacts are not virtually identical, the stimulus meanings of “bachelor” will diverge far more than “rabbit”. [But the social contacts are part of the context used in computing stimulus meaning. So maybe more straightforward to say in general that we’ll always define meaning w/r/t a particular initial belief assumed shared.]

Q’s solution (basically equivalent): just do everything w/r/t stimulus meanings for a single bilingual speaker. [Great, and also avoids problems with rabbit-flies etc.]

Contextual info is still tricky—morning star vs evening star, etc.

Can stimulus meaning distinguish between a rabbit and a “temporal segment of a rabbit” or “rabbithood”? Q: anything that stimulates the specific term also stimulates the general one. [How can we ever learn anything about parts that belong to wholes but are not recongizable outside of context? At least definitionally; in the case of the bilingual speaker translating via stimuli, we can use L1 as part of the stimulus to determine the meaning of an L2 word.]

[In any case if there’s no stimulus that distinguishes “Rabbit segment.” from “Rabbit.”, we might actually want to consider those sentences to be paraphrases of each other, even if the composed sentences “He is eating a rabbit.” and “He is eating a rabbit segment.” have different stimulus meanings. Cf “alleged Maoist”.]

For lots of words like “bachelor”, only the definitional part is shared—no two speakers will actually have the smae stimulus meaning. [This is why I like the “what does this sentence make you believe” formulation rather than the “what sentences are you willing to utter given this stimulus”.]

It will be difficult to find analogues of logical operators, since:

  • speakers may be imperfect reasoners and assent to statements that are logical contradictions

  • natural languages don’t actually have pure logical operators (consider Eng speaker answering a question “Yes and no.”)

  • incomplete information on part of speakers leads to problems with extensional judgments about truth conditions

That is, we can’t identify translations of logical operators except via their extension, but extensions are unreliable.

The trouble with synonymy: with a short modulus, any pair of speakers are likely to be influenced by background knowledge in a way that makes it difficult to determine equivalence on the basis of stimulus meaning. But with a sufficiently long modulus [Q says a month but it seems like nothing short of a lifetime will do in general] “there is no telling what to expect”. [Is this just a statement that there is uncertainty in the resulting distribution over responses? Can compare distributions rather than binary judgment.]

Treatment of analyticity is substantially the same as synonymy (with sentenes of the form if p then p).

Observation sentences, truth sentences, stimulus-analytic sentences can be translated.

“Questions of intrasubjective stimulus synonymy of native occasion sentences even of non-observational kind can be settled if raised, but the sentences cannot be translated” (S15).

In general we expect a good translator to turn foreign sentences that are analytic into native sentences that are analytic. But suppose we’re working with a people that regards the sentence “all rabbits are men reincarnate” as a tautology? Problem is basically that we imagine our foreign and native speakers to be working with different notions of deductive closure: if we assumed these were the same some of our problems would be much easier.

Or even more strongly, by letting the language user be the same person for both languages.

So what do we make of all this? Certainly true that as a practical matter we can’t generate all possible stimuli necessary to identify truth conditions for every utterance in a language.

More interesting possibility is for a single speaker of both languages: he has a “private implicit system of analytical hypotheses”; another speaker could have exactly the same speech dispositions in each language but a different disposition to translation.

Question is: is the distinction between the translation model possessed by our two speakers nontrivial? In the sense that e.g. someone relying on translations will take different behaviors or form different beliefs about the state of the world on the basis of these translations? Not clear that Q has provided a real example of such a pair of languages, since the rabbit-vs-rabbit-stage distinction can clearly be made on the basis of some (perhaps just imagined stimulus), or else it isn’t a distinction at all.

Of course, if we’re only concerned with the speakers having the same disposition w/r/t some subset of possible contexts (the “real ones”), then this indeterminacy is certainly possible, but not an indeterminacy of translation so much as an indeterminacy of language-learning per se (as Q himself points out at the end).

But if we actually have access to all possible contexts, we’re back to the initial problem: what does it look like to have p(s | c) s.t. we can permute s and have everything be the same?