In terms of cybernetics, natural language is considered as a black box for the researcher. A black box is a device with observable input and output but with a completely unobservable inner structure. In the framework of this type of model, language is thought to be an imaginary “speaking device”: the researcher asks the device some questions and records its answers.
The problem of the cybernetic modeling of natural language is more difficult than in other cases, since there are two such boxes, the analyzing and synthesizing ones, working in opposite directions. The analyzing block processes the utterances and the synthesizing block produces the reactions to them.
A researcher observes the input of the analyzing block and the output of the synthesizing block, and tries to reconstruct the inner structure of each block separately. Unfortunately, the output of the analyzer is not directly used as the input of the synthesizer. There is a block of reasoning in between, and its behavior is not described in linguistic terms, so that it is not so easy to recognize either.
The main method of linguistics is to construct a model of natural language, based on the observable input and output texts, and on the linguist’s intuition, or introspection. The linguists analyze their own intuition, put forward hypotheses, build models and test them on additional linguistic material. In theoretical linguistics, the novel approaches can be tested against (compared with) intuitions of other linguists, while in computational linguistics these approaches can be also tested through various applications.
In this way, linguists have proposed functional models of language. These models are intended to give the rules of conversion of the input linguistic information to the output information, without any attempt to directly reproduce the internal mechanisms of brain activity. No anthropomorphic features of the processing steps are searched for, and no direct emulation of brain’s responses is adduced. However, the ultimate results of all processing steps should be as near as possible to those of the human brain.
So far, functional models have proven to be the most successful linguistic models, probably because they are based on real data with conceivable structure, easily accessible and available in unlimited supply, namely, on texts and recorded speech.