Semantic Message Detection for Machine Translation, Using an Interlingua*

The purpose of this paper is to draw attention to a semantic feature of language which must be well understood before it is even imaginable that machines should be programmed to do genuine M.T.M.T., that is, of which the input is both unrestricted, and also composed of heterogeneous and randomly-chosen texts. This feature is the argument, or message, of any piece of discourse. It has been shown that failure to detect this can produce dangerously misleading "garbage"1), even when the machine, though it does not detect message, is programmed to detect many features of the input language's grammar and syntax, and when it is provided, as well, with a considerable bilingual dictionary. In my view, the present "critical situation" in M.T.,2) is not due to the fact that genuine Mechanical Translation is inherently impossible, as Bar-Hillel thinks, 3) but to the fact that the mechanizable techniques at present being used to analyse language are not powerful enough to detect the message, or argument, of any particular text. Linguistics deals only with syntactic groupings of words picked up unilingually from text, and with the interrelations of these. Information Theory deals with codes, not with languages; it is assumed throughout that the message itself, before it is coded and after it is decoded, is intuitively comprehensible and that it need not be further analysed. Metamathematics deals only with the interrelations of sentences, assuming to start with that these are