The holistic option is bound to reject the traditional view of communication if it wants to avoid the criticism of Dummett. Radical holism has bee expressed clearly in Davidson's theory of meaning based on the concept of interpretation and on Tarski's theory of truth. Not much has been said about communication in the early papers by Davidson. And radical or global holism is the typical target of Dummett's criticism. But if we change the picture of communication, maybe holism could be saved. This stance is well expressed in Davidson's "Nice derangement of epitaphs", in Stalnaker's Context and Content and in Brandom's Making it Explicit. Recently Hinzen 1999 has give further support to the idea that "shared linguistic knowledge is no precondition for communication".
The main point in Davidson is exactly to give a different view of what communication is: not a sharing of the same contents (meanings) but a work of convergence between two different idiolects, and the formation of a "passing" theory during the dialogue, provoked by the utterances of the other speaker. This passing theory is what a speaker is able to build on the ground of what he expects the speaker is thinking and of what she actually says.
A similar stance, in a different framework, has been developed by Robert Stalnaker in Content and Context. While Davidson keeps Quine's attitude of using canonical extensional first order language, Stalnaker's framework is given by model theoretic semantics in the tradition of Carnap's and Montague's modal languages. In Stalnaker the expectations of a speaker are framed in terms of "presuppositions"; the context is what is given by the (probably different) presuppositions of the speakers, by what they assume for granted: «context dependence means dependence on certain facts, but the facts must be available, or presumed to be available, to the participants in the conversation. So I propose to identify a context (at a particular point in a discourse) with the body of information that is presumed, at that point, to be common to participants in the discourse» (p.98). Given that the context is "constantly changing as things are said" (p.101), a dialogue is a constant adjustments of the presuppositions prompted by the sentences which are uttered by the speakers.
The main problem with Davidson's proposal is that it reduces language to a set of idiolects of speakers; further criticism given by Dummett 1992 is mainly devoted to the lack of analysis of language as social institution and social practice; even if communication is important, to restrict the function of language to communication is to give a fairly incomplete view of language, which is also a vehicle of thought in a community.
If I have well understood Stalnaker's proposal, every proposition, relevant or not, that is taken for granted by the speakers contributes to the definition of the context. The difficulty with this proposal is that the context so conceived is "too big". A context embeds all the presuppositions of the speakers in the dialogue; while they change subject, the presuppositions are still there and we have to take them into account to ascertain the validity of the context set. Therefore the overall project seems bound to a computational complexity which is beyond any reasonable formal treatment.
There is another form of holism developed in the tradition of semantic networks. For Quillian 1967 each node of a semantic network activates the nearest nodes following some path in the network and is indirectly connected with every other node. Meaning can be computed at different degrees, depending on the number of nodes activated, and similarity of meaning is computed as similarity in activation of nodes starting from the node under consideration. However the idea that, given any system, the elements of the system are in principle connected is an aspect of holism which does not necessarily imply meaning holism; in fact the claim behind semantic networks does not entail that that to compute the meaning of an expression we have to run through the entire semantic network, even if there is no clear assumption about restrictions in the numbers of nodes activated for individuating meaning.