In linguistics, control is a construction in which the understood subject of a given predicate is determined by some expression in context. Stereotypical instances of control involve verbs. A superordinate verb "controls" the arguments of a subordinate, nonfinite verb. Control was intensively studied in the government and binding framework in the 1980s, and much of the terminology from that era is still used today. In the days of Transformational Grammar, control phenomena were discussed in terms of Equi-NP deletion. Control is often analyzed in terms of a null pronoun called PRO. Control is also related to raising, although there are important differences between control and raising. Most if not all languages have control constructions and these constructions tend to occur frequently.
Standard instances of (obligatory) control are present in the following sentences:
Each of these sentences contains two verbal predicates. Each time the control verb is on the left, and the verb whose arguments are controlled is on the right. The control verb determines which expression is interpreted as the subject of the verb on the right. The first three sentences are examples of subject control, since the subject of the control verb is also the understood subject of the subordinate verb. The second three examples are instances of object control, because the object of the control verb is understood as the subject of the subordinate verb. The argument of the matrix predicate that functions as the subject of the embedded predicate is the controller. The controllers are in bold in the examples.
Control verbs have semantic content; they semantically select their arguments, that is, their appearance strongly influences the nature of the arguments they take. In this regard, they are very different from auxiliary verbs, which lack semantic content and do not semantically select arguments. Compare the following pairs of sentences:
The a-sentences contain auxiliary verbs that do not select the subject argument. What this means is that the embedded verbs go, do, and lie and cheat are responsible for semantically selecting the subject argument. The point is that while control verbs may have the same outward appearance as auxiliary verbs, the two verb types are quite different.
Control verbs (such as promise, stop, try, ask, tell, force, yearn, refuse, attempt) obligatorily induce a control construction. That is, when control verbs appear, they inherently determine which of their arguments controls the embedded predicate. Control is hence obligatorily present with these verbs. In contrast, the arguments of many verbs can be controlled even when a superordinate control verb is absent, e.g.
In one sense, control is obligatory in these sentences because the arguments of the present participles singing, understanding, and holding are clearly controlled by the matrix subjects. In another sense, however, control is non-obligatory (or optional) because there is no control predicate present that necessitates that control occur. General contextual factors are determining which expression is understood as the controller. The controller is the subject in these sentences because the subject establishes point of view.
Arbitrary control occurs when the controller is understood to be anybody in general, e.g.
The understood subject of the gerunds in these sentence is non-discriminate; any generic person will do. In such cases, control is said to be "arbitrary". Any time the understood subject of a given predicate is not present in the linguistic or situational context, a generic subject (e.g. 'one') is understood.
Theoretical linguistics posits the existence of the null pronoun PRO as the theoretical basis for the analysis of control structures. The null pronoun PRO is an element that impacts a sentence in a similar manner to how a normal pronoun impacts a sentence, but the null pronoun is inaudible. The null PRO is added to the predicate, where it occupies the position that one would typically associate with an overt subject (if one were present). The following trees illustrate PRO in both constituency-based structures of phrase structure grammars and dependency-based structures of dependency grammars:
The constituency-based trees are the a-trees on the left, and the dependency-based trees the b-trees on the right. Certainly aspects of these trees - especially of the constituency trees - can be disputed. In the current context, the trees are intended merely to suggest by way of illustration how control and PRO are conceived of. The indices are a common means of identifying PRO and with its antecedent in the control predicate, and the orange arrows indicate further the control relation. In a sense, the controller assigns its index to PRO, which identifies the argument that is understood as the subject of the subordinate predicate.
The details of this tree are, again, not so important. What is important is that by positing the existence of the null subject PRO, the theoretical analysis of control constructions gains a useful tool that can help uncover important traits of control constructions.
Control must be distinguished from raising, though the two can be outwardly similar. Control predicates semantically select their arguments, as stated above. Raising predicates, in contrast, do not semantically select (at least) one of their dependents. The contrast is evident with the so-called raising-to-object verbs (=ECM-verbs) such as believe, expect, want, and prove. Compare the following a- and b-sentences:
The control predicates ask and force semantically select their object arguments, whereas the raising-to-object verbs do not. Instead, the object of the raising verb appears to have "risen" from the subject position of the embedded predicate, in this case from the embedded predicates to read and to have said. In other words, the embedded predicate is semantically selecting the argument of the matrix predicate. What this means is that while a raising-to-object verb takes an object dependent, that dependent is not a semantic argument of that raising verb. The distinction becomes apparent when one considers that a control predicate like ask requires its object to be an animate entity, whereas a raising-to-object predicate like expects places no semantic limitations on its object dependent.
The control predicates cannot take expletive there because there does not fulfill the semantic requirements of the control predicates. Since the raising-to-object predicates do not select their objects, they can easily take expletive there.
Control and raising also differ in how they behave with idiomatic expressions. Idiomatic expressions retain their meaning in a raising construction, but they lose it when they are arguments of a control verb. See the examples below featuring the idiom "The cat is out of the bag", which has the meaning that facts that were previously hidden are now revealed.
The explanation for this fact is that raising predicates do not semantically select their arguments, and therefore their arguments are not interpreted compositionally, as the subject or object of the raising predicate. Arguments of the control predicate, on the other hand, have to fulfill their semantic requirements, and interpreted as the argument of the predicate compositionally.
This test works for object control and ECM too.