The basic operations are those of segmentation and classification […]. Both of these operations are performed upon an utterance or upon its parts, but always on the basis of some comparison between these and some other utterances: e.g. morpheme segmentation is carried out before and after /s/ in 'What books came?' but not in 'What box came?' because of comparison with 'What book came?' and so on. If we were analyzing a corpus without any interest in its relevance for the whole language, we could list all the environment of each tentative segment in all utterances of the corpus, and on this basis decide the segmentation in each utterance. Usually, however, we are interested in analyzing such a corpus as will serve as a sample of the language. For this purpose we bring into our corpus controled material for comparison. Given 'What books came?' we do not compare it with arbitrary other utterances, but search for utterances which are partially similar, like 'What book came?' 'What maps came?' 'What books are you reading?'. Ideally, we seek a group of minimally different utterances for comparison. In eliciting such comparative utterances from an informant, or from oneself, or from some arranged or indexed body of material, we have an experimental situation in which the linguist tests variations in the utterance stock in respect to a selected utterance; the only danger that the utterance stock may be artificially modified due to the experimentally asked question (as when an informant accepts an utterance proposed by the linguist even though it is a bit different from anything he would say on his own as a speaker of the language) […]. As a result of these operations, we not only obtain initial elements, but are also able to define new sets of elements as classes or combinations (sequences, etc.) of old ones.
- Harris (1951), a pag.367