Grammaticalization is a process in which operations of linguistic activity are subjected to rules of grammar. Instead of having free play at the level of discourse, where meaningful units are selected and combined into larger units in conformity with cognitive and communicative intentions, operations become dependent on factors of linguistic structure and finally on contextual conditions. A good example of this is the role of agreement in phoric relations (Corbett 2006). The syntagmatic scope of this relation is bracketed in .

In a, the personal pronoun has a regular anaphoric relation to its antecedent across a sentence boundary. In this context, semantic agreement (in feminine gender) is normal, although mechanic agreement (in neutral gender) would be possible. In #b, the relative pronoun is in the same NP as its antecedent, thus, in a relation of syntactic phora. Here, mechanic agreement is standard, although semantic agreement would occasionally be found. In #c, the adjective is an attribute of the noun which determines the gender; here, only mechanic agreement is found. Thus, at the highest level of linguistic structure, the speaker chooses the gender which corresponds to his message, while at the lowest level of this series, the grammar dictates the gender to be used; in other words, agreement inside the nominal group is fully grammaticalized.¹

Grammaticalization has often been described as a process of expansion of some linguistic unit and, thus, of increased frequency. This is, however, an automatic side-effect of increasing obligatoriness (s. the section on expansion). As semantic restrictions on the appropriateness of a certain formative drop, the factors conditioning its occurrence are strengthened. However, only if these conditioning factors belong to the language system may they lead to grammaticalization. In other words, conditioning extralinguistic context may lead to the generalization of a fashionable expression; but this does not thereby become grammatical. To give just two examples: For the past 30 years, teenage slang has vocalized positive evaluation by cool; and many young and adult German speakers have been expressing surprise at their own slip by the English word oops. In the confines of certain styles, these words have seen an expansion of use from total absence to omnipresence within a few decades. However, the factors conditioning or triggering their use are not part of the linguistic system; and therefore this is a purely lexical change and has nothing to do with grammaticalization.

also reminds us that grammaticalization is a process that proceeds along a scale of degrees. A given linguistic operation, construction or formative is not either grammatical or non-grammatical (outside the reach of grammar); instead it is grammaticalized to a certain degree. Grammaticalization is variation along a scale. On the synchronic axis, it manifests itself in the coexistence of variants of a unit which are in a relation of polysemy or polyfunctionality and one of which is more subject to grammatical constraints than the other. On the diachronic axis, the later variant is the one which is more subject to grammatical constraints than the other.

While criteria to determine the direction of some variation on the synchronic plane remain a desideratum of linguistic methodology, grammaticalization on the diachronic axis has been found to be uniform in the sense just formulated across many languages and areas of grammar. The inverse diachronic process which converts an item or a construction into a less grammatical one can be defined in theory and be dubbed ‘degrammaticalization’. However, very few cogent examples of such a process have been found. It is an empirical generalization that gram­ma­ticalization appears to be irreversible. In other words, while there is an oriented variation that transforms linguistic operations and units into more grammatical ones, a kind of variation which transforms them into less grammatical ones has no systematic place in linguistic activity. If this is so, then it would be of scientific interest to have an explanation for it.

¹ If agreement trigger and target occur in this order, linear distance between them also correlates with an increase of semantic over mechanic agreement (s. Köpcke & Zubin 2009, §5 for a few statistical data). This is, at the same time, a piece of evidence for the memory imprint difference between semantic and grammatical information shown in the section on automation.