JAY L. GARFIELD
Convention -- social agreement -- determines linguistic meaning and practice at countless levels, including the fixation of lexical meaning; the adoption of particular syntactic options from among those permitted by universal grammar; the determination of politeness and felicity conditions; and most broadly, the determination of the character of the background of human activities and purposes against which most discourse takes place, and in the context of which it is intended to be interpreted. In assigning meaning to, or understanding, a particular utterance or inscription, it is frequently, if not always, the case that convention insinuates itself into the process at each of these levels, or so many philosophers of language and linguists have imagined. This view has, however, been recently challenged by theoreticians in the Artificial Intelligence (AI) community, who, under the banner of "Procedural Semantics" have argued, sometimes explicitly ( Moore and Hendrix 1982, Miller and Johnson-Laird 1976), and sometimes merely implicitly ( Winograd 1973) that the meanings of natural language expressions are to be identified with the computational procedures to which they give rise in successful artificial or biological interpreters, and that understanding an expression just constitutes executing the appropriate procedure in its presence. On this view, a language understander could be ignorant of and isolated from any social conventions ostensibly relevant to the semantics of the expressions upon which it operates.
The semantic ideology embodied by procedural semantics and the success of artificial intelligence research in understanding natural language have often been taken to be mutually supporting (see e.g., Winograd, op. cit., Fodor 1978). After all, if the meanings of words just are the programs they call, and if the compositional rules for computing the meanings of larger expressions just are rules for composing programs into larger programs, then, since computers are program runners and composers par excellence, there would be good reason to hope that computers could compute the meanings of any