The solutions given above of the paradox of holism have been all criticized for some relevant aspects. We might doubt that the solutions given above exhaust the varieties of possible solutions. I suggest that Wittgenstein's ideas, which seem to be at the heart of most of these solutions, have been partly misunderstood. When Quine quotes Wittgenstein as a heir of an holistic position, he quotes from the Blue Book. However the complete quotation from Philosophical Investigations says not only that "to understand a sentence is to understand a language", but also that "to understand a language is mastering a technique". How is to be interpreted the first sentence? Is this second assertion relevant to our discussion?
To answer the first question we should take care to understand "a language" not as "the" language. In Wittgenstein's perspective language itself, as a universal representation, disappears. We are left with a multifarious interconnection of language games, situations in which language and action are interwoven. Therefore if we want to give a definition of meaning, we have to give it always relative to a language game. And if we want to define meaning holistically, we cannot define meaning relative to the entire language, but relative to single and specific language games.
As far as the second sentence is concerned, a technique is something which is given by a social practice - we master a technique if we follow some common rules. Therefore understanding a language is ability to follow rules (social practices). This passage suggests a shift from sharing contents to sharing rules. What do we share? Not mainly contents but rules; Brandom 1994 tries to develop a similar point in speaking of "sharing the structure, not the content". The picture of communication as sharing of contents can be abandoned with a picture of communication as sharing of rules.
Which kind of rules? Here there is no clear answer in Wittgenstein; however his overall picture of the working of language avoids postulating a unique set of rules; apparently different rules apply in different language games. Language itself is a highly abstract idealization; you may postulate a language instinct and a language faculty, but the varieties of linguistic activities cannot be embedded in a unique formal representation. Wittgenstein's attitude is a reaction against a unique representation of language as a single unit; language is a social enterprise and different aspects are developed in different ways; but if there cannot be a unique representation of the working of language, there can be a careful description of different language games. Their description must be completed by an analysis of the connections and transformations from one language game to another. In Wittgenstein's Philosophical Investigations the study of the relationships among language games is left to a very general level of programmatic remarks. This point seems to me the place where artificial intelligence may give some suggestion to philosophy. There is an apparent similarity between the Wittgenteinian picture of language games and the basic ideas of multi-context theories in artificial intelligence. In multi-context theories you cannot develop a single formalization for "the" language, but you are bound to describe different contexts as different theories, each one with its language, axioms and rules.
Somebody might ask whether it is sensible to make a comparison between an old-fashioned philosopher of the beginning of the last century and the work of contemporary a.i. I can only suggest that the motivations of the formal treatment of context in a.i. are similar to the motivation behind Wittgenstein's philosophy. The answers are partly different. The motivation is the problem of generality: at any time you may find people interpreting what you say in an unexpected way.
McCarthy 1987, 1993 says that, given any sentence, you may always find, with some ingenuity, some axioms where the sentence can be interpreted as false or however different from the intended interpretation; his answer insists on the importance of relativizing any assertion to a specific context and studying the rules which permit operations among contexts. Wittgenstein 1953 says that given any sentence (or any rule) you may always find an interpretation different from yours. Wittgenstein's answer insists on the difference between interpreting a rule and following a rule. Rules are social practices which are followed before being interpreted; therefore in understanding language you cannot rely only on interpretations, but you have to rely on social practice. Social practices are laid down in language games.
The detailed study of the rules which work across contexts is exactly what is missing in Wittgenstein's approach, even if his philosophy clearly go towards this clarification. This kind of study is also what is missing in the different attempts to face the problem of holism. All attempts to solve the problem of holism end up with a search of shared contents: communication is either the sharing of meanings or a convergence towards some shared meanings or contents. No question has been posed on the means to attain this aim; Davidson 1986 (p. 445) speaks of the "mysterious" aspect of the communicative success. On the contrary, the suggestion stemming from artificial intelligence is that there is no mistery at all: we share and we may explicitly study general rules to navigate across contexts. For a communication being successful, we need to share these high level rules, and the formal study of this kind of rules may help to understand exactly the strategies used in successful communication.
There are many presentations of rules for navigating across contexts; Mc Carthy 1993 speaks of rules of entering and exiting contexts, or of lifting a sentence from one context to another; Giunchiglia 1993 speaks of bridge rules which permit to draw a conclusion in a context from a premise in another context; Fauconnier 1997 (ch.6) speaks of rules for blending concepts from different contexts; Benerecetti et alia 2000 summarize most of the research in three kinds of mechanisms of contextual reasoning: localized reasoning, push and pop, shifting. These three mechanisms correspond to three basic aspects of a contextual representation as