The philosophy of language is an extremely broad field covering a number of interesting problems. Unfortunately the Stanford Encyclopedia of Philosophy does not appear to have an umbrella article on the topic, but the Wikipedia entry is fairly decent.
The SEP does, however, have an article on Conventions of Language, especially on how convention (effectively what I’ve been calling social negotiation) generates meaning. The Lewisian and Gricean accounts in that article provide an interesting “base” to build upon, which I will do by responding to some of the objections given in that article (and modifying them as I see fit along the way).
Sentences, Words, and the Components of Meaning (Stanford Section 7.2.2)
One such objection is on the relations of sentences to words, and how previously-unspoken sentences can have meaning since there must be no convention behind them (given that they’ve never before been spoken). This objection neatly knocks down the sentence-based accounts of Lewis and Grice, and appealing to just sentences-as-composed of words has its own serious problems. The underlying problem is that the unit of meaning is itself fuzzy.
Consider idiom: there is a Polish expression which roughly translates as “not my circus, not my monkey” and means, more generally “not my problem”. Here we have a phrase whose words would seem to mean one thing, but whose actual meaning is something else. Both meanings are valid, and both are due to social negotiation; it is simply that the conventional sentence meaning overrides the conventional word meaning for those who hear it (except in the context, of course, where an actual circus and monkey are present!).
It is perfectly possible for some person to speak fluent Polish but be totally unaware of this idiom: given that it is syntactically valid and composed only of known words, they would be able to assign a coherent meaning to it the first time they heard it, but the speaker would then be required to explain the convention in order to get their point across. However, if a circus and monkey were present, then our listener would be able to accurately construct the literal and intended meaning of the sentence without any additional explanation.
The point here is that the source of meanings is not specifically words, or sentences, or anything else. Meaning can be assigned to abstractions and speech-components of any size, and when we end up with multiple conflicting meanings at various layers of abstraction, we default to the “largest” (most abstract) unit unless context indicates otherwise.
Grammar and Radical Interpretation (Stanford Section 7.2.3)
Chomsky (further developed by Schiffer and others) argues that language is an internal process related to semantic and psychological properties bearing no special ties to social interaction or convention. This is a radically different tack, and not obviously wrong. It seems feasible, for example, for some person entirely isolated from any other being to develop and think in some new language. Where is the social negotiation in that?
Strictly speaking there isn’t any, but that simply makes the language entirely arbitrary. With no social interactions to bind the meanings and force conventions for practicality’s sake, the speaker is free to change it entirely at will. This makes the actual words and grammar used effectively meaningless, since the only association they actually have with the thought processes of the speaker is that speaker’s “speaker-meaning” (in the Gricean sense) at the time of utterance. If the speaker develops conventions regardless she is free to do so, but there “conventionality” is limited to the speaker’s habit, nothing more.
(Chomsky also does this bit about language as tacitly known grammatical rules, but I feel I effectively dealt with that tack in the previous section on multiple layers of meaning).
In a related way, Davidson points out that there are cases where we deviate from any conventional meaning (at any layer) and fall back to what he calls “radical interpretation” (effectively guessing based on context). Since we are capable of this radical interpretation at any time (when dealing with malapropisms, spoonerisms, etc), then in some sense language is independent of convention. I would argue that while radical interpretation is certainly communication, it is not actually language per se.
This redefinition is in some sense just a linguistic trick, but I believe it effectively answers the spirit of the argument: we are perfectly capable of communicating in some sense without language, though the process is laborious and error-prone. Just because we sometimes fall back on contextual clues when using language does not make those clues part of language itself, any more than falling back to a lower-level programming language for one part performance-critical piece of a program makes that lower-level language “part of” the principle language in use. They are simply different methods of communication that we are free to mix for our own convenience.