The cohesion¹ of a discourse is the set of relations among its components. Taken as a qualitative notion, it is such a set which is sufficient to shape these components into a unified whole.

The relations in question are, in the first place, syntagmatic relations. These are relations between neighboring – typically, adjacent – components of the discourse which connect them into larger units. Secondarily, they are paradigmatic relations, for instance, relations of referential identity or hyponymy, between subsequent discourse components.

The relations in question obtain at different levels of linguistic structure which may be conceived as subdomains of the functional domain of discourse structure:

• At the level of its sympractic embedding, an utterance is related to the entire situation which includes the speech situation. It may, e.g., be a reaction to a non-linguistic event or trigger a non-linguistic act.
• At the level of rhetorical structure, utterances are related, typically in the form of couples, like a question-answer pair. A finer categorization of illocutionary forces allows pairs such as the claim-justification relation or the grounds-conclusion relation. The enunciative relation between a framing utterance and a reported utterance may be considered a special form of rhetorical structure.
• At the level of information structure, presuppositions are related to assertions; elements have topical or focal status contrasting with neighboring elements.
• In the universe of discourse, referents are anchored, individuated and (re-)activated, producing topic continuity and topic change.

The components related in these ways are of different nature. Entire utterances are related in rhetorical structure. In information structure, propositions and components of propositions are related. The components that constitute continuous topics are referential expressions. These include not only thing-like referents, but also propositions which may be taken up.

This means that two neighboring discourse components, e.g. two propositions, may be the relata of an interpropositional relator. Alternatively, they may feature some property, e.g. a modality, or internal component which is taken up identically or somehow matched between the two propositions, like a topic being talked about.

• The most important linguistic means coding syntagmatic relations between two components of a text are relational signs. These are linguistic signs which provide two (or more) argument places to be occupied by such components. At lower levels of syntactic structure, relational signs comprise verbs, relational nouns, adpositions and others. Above the sentence level, they are interpropositional relators which specify, e.g., a causal or temporal relation between adjacent sentences. With the exception of the identity operator, a relational sign generally takes two arguments which are different in some respect, thus allowing the speaker to procede in the discourse.
• At the same time, a coherent discourse comprises more than one bit of content associated with a given item. Since these bits of content are assigned one after the other, the target of these assignments would have to be repeated each time. For different reasons, this is not always feasible. There are therefore linguistic means dedicated to the task of indicating that an item already established in the universe of discourse is being meant again. These are operations of endophora and, specifically, of reference tracking. The linguistic means used comprise, for instance, personal pronouns. Endophora itself is, again, a syntagmatic relation between components – in this case, referential expressions. At the same time, such expressions bear a paradigmatic relation to each other – in the simplest case, the relation of referential identity.

These are the two fundamental linguistic operations that create cohesion in a text. When a semantically specific relation between two adjacent sentences is represented by a linguistic sign, it typically combines these two operations. For instance, in a sequence of two sentences ‘p; thereby q’, the word thereby consists of the relational item by and the anaphoric item there. The latter represents p, which by connects with q.

Both the construction consisting of two sentences connected by such an interpropositional relator and the construction consisting of two sentences sharing a referent are subject to grammaticalization. Grammaticalization converts discourse structure into grammar. Thus, the interpropositional relator may grammaticalize into a subordinative conjunction, connecting a subordinate clause with its superordinate clause. Likewise, as the syntagmatic relation between contiguous clauses becomes a syntactic dependency, a endophoric relation between the referential expressions contained in them becomes a relation of phoric control governed by syntax. In the present framework, the targets of these grammaticalization processes are comprised by different functional domains. In this perspective, nexion is the counterpart of discourse structure at the sentence level, and likewise reference is treated at the sentence level.

¹ Part of the literature makes a distinction between coherence and cohesion of the text; then the present topic might be called ‘coherence’, instead.