Control Constructions

How control constructions work

English infinitival complements often look incomplete on the surface:

John tried [ ___ to leave].
John persuaded Mary [ ___ to leave].

The bracketed clause has no overt subject — but we always know who's doing the leaving. In the first sentence John leaves; in the second Mary leaves. The matrix verb tells us which: try is a subject-control verb (matrix subject = embedded subject), persuade is an object-control verb (matrix object = embedded subject).

The standard generative analysis (since Chomsky's Lectures on Government and Binding, 1981) is that the empty slot contains a silent pronoun called PRO (capitalized to distinguish it from overt pronouns). PRO is phonologically null but syntactically real, and its reference is fixed — controlled — by an argument of the matrix verb:

John₁ tried [PRO₁ to leave].
John persuaded Mary₁ [PRO₁ to leave].

The subscripts mark coreference. PRO inherits its reference from its controller. Control theory in syntax is the body of principles that determines, for each verb, who the controller is.

Subject control vs object control

The split depends on the matrix verb. Some verbs make their subject the controller; others make their object the controller. Compare two minimal pairs:

(a) John promised Mary to leave.    → John leaves   (SUBJECT control)
(b) John persuaded Mary to leave.   → Mary leaves   (OBJECT control)

Same word order, same arguments, but radically different meanings — because promise and persuade are lexically specified for different control patterns. Promise is one of a small class of subject-control verbs that take an object: it has been a long-standing puzzle in acquisition research that children acquire object control before subject-control-with-object (Carol Chomsky 1969).

Common subject-control verbs include intentional/decisional verbs (try, want, hope, plan, decide, intend, manage, refuse, agree) and emotive verbs (love, hate, prefer). Common object-control verbs include verbs of communication and causation directed at someone (persuade, force, tell, ask, order, urge, advise, instruct, allow, permit).

Control vs raising — the central distinction

Control verbs and raising verbs both take infinitival complements, and on the surface their sentences look identical:

(a) John tried to leave.    [CONTROL]
(b) John seems to leave.    [RAISING]

But the structures are very different. In control, John is an argument of try (John is the trier) and PRO is a separate element coreferential with John (PRO is the leaver). Two distinct thematic roles. In raising, John is not an argument of seems at all; John is only the leaver, and has been "raised" from the embedded clause to the matrix subject position because the matrix verb has no role to assign. One thematic role.

	Control (John tried to leave)	Raising (John seems to leave)
θ-roles for John	Two — trier and leaver	One — leaver only
Empty embedded subject	PRO	Trace (t) of moved John
Idioms preserved?	No — "the cat tried to be out of the bag" loses the idiom	Yes — "the cat seems to be out of the bag" keeps it
Expletive subject?	No — *"It tried to rain"	Yes — "It seems to be raining"
Selectional restrictions	Matrix verb constrains subject	Only embedded verb constrains
Passive equivalence	Different meaning under passive	Same meaning under passive
Examples	try, want, hope, promise, persuade	seem, appear, happen, tend, turn out

The diagnostics are powerful because they exploit the difference in θ-role assignment. Try requires a subject capable of trying — agentive, sentient. Seem requires nothing of its subject; it just reports an appearance. So the cat-out-of-the-bag idiom survives raising but not control: the idiom needs to stay intact, and control imposes a real semantic role on its subject that's incompatible with idiomatic readings.

Worked example: applying the diagnostics

Suppose you encounter the verb tend. Is "John tends to leave early" a control or a raising structure? Apply the diagnostics:

Test	Result	Verdict
Idiom preservation: "The cat tends to be out of the bag"	Idiom survives	Suggests raising
Expletive: "It tends to rain in April"	Fine	Suggests raising
Passive equivalence: "Mary tends to hate John" ≈ "John tends to be hated by Mary"	Truth-conditionally equivalent	Suggests raising

All three tests align: tend is a raising verb. Now contrast with try: "*The cat tried to be out of the bag" loses the idiom; "*It tried to rain" is gibberish; "Mary tried to hate John" ≠ "John tried to be hated by Mary." All three tests give the opposite verdict — try is a control verb.

The meaning-preservation diagnostic

The cleanest diagnostic for control is meaning preservation under passivization of the embedded clause. Compare:

(1) John seems to like Mary.        ≈  Mary seems to be liked by John.    [SAME MEANING — raising]
(2) John tried to like Mary.        ≠  Mary tried to be liked by John.    [DIFFERENT — control]

In (1), seem assigns no role to John, so swapping the embedded clause's voice doesn't change who's doing what. In (2), try assigns John the role of trier — when we passivize, the trier shifts to Mary, and now Mary (not John) is the one trying. The shift in trier identity changes the truth conditions. Meaning preservation under embedded passivization is the gold standard for raising; meaning shift signals control.

Variants and related cases

• Arbitrary PRO. "It's important to be honest" — PRO has no specific controller, gets a generic-everyone reading.
• Partial control. "John wanted to gather at noon" — PRO refers to a group including John (Landau 2000). One controller can't gather alone.
• Long-distance control. Some embedded structures allow PRO to be controlled by an argument of an even higher clause: "John knew that it would be hard to get a job" — PRO is John, even though know intervenes.
• Implicit control. "It was decided to postpone the meeting" — PRO has no overt controller; an implicit deciders-group is inferred.
• Adjunct control. "John left without saying goodbye" — PRO in the gerund is controlled by the matrix subject. Classic adjunct-control puzzle: usually subject-controlled, but exceptions exist.
• Inflected infinitive (Portuguese). European Portuguese has infinitives that show subject agreement: "É importante saberMOS isso" (It's important [for us] to know this). Less reliance on PRO.
• Equi-NP deletion. The pre-PRO transformational analysis: an overt embedded subject identical to the matrix subject was deleted. Replaced by PRO theory in the 1970s but the term still appears in older textbooks.
• Movement theory of control (Hornstein 1999). Modern proposal that PRO is just a trace of A-movement — control is reduced to raising. Controversial; defenders point to economy, critics point to data PRO theory handles cleanly.

Why control matters

• θ-theory. Control is the testbed for theta-role assignment — every theory of how predicates assign semantic roles must explain control patterns.
• Universal grammar. Children learn subject vs object control without explicit instruction, suggesting innate mechanisms; the promise-acquisition delay (around age 9 in English) is a famous data point.
• Cross-linguistic typology. Languages without infinitives (e.g., many Balkan languages) use subjunctives; the underlying control patterns transfer.
• Anaphora and binding. PRO is part of a broader empty-category inventory (PRO, pro, traces, NP-trace) that anchors the theory of binding.
• Computational parsing. NLP systems must resolve control to build correct logical forms — "Mary persuaded Bob to call her" needs Mary as caller-recipient and Bob as caller.
• Argument structure. Control verbs are diagnostic for studying argument-structure alternations and lexical-semantic class boundaries.

Common pitfalls

• Confusing control with raising. The surface looks identical. Always run the idiom and expletive diagnostics before assuming. Tend, seem, appear, happen are raising; try, want, hope are control.
• Treating PRO as a free pronoun. PRO is obligatorily controlled in most contexts — "John tried to leave" cannot mean Mary left. Free reference appears only in arbitrary-PRO contexts.
• Calling all infinitival gaps "ellipsis." Ellipsis is recoverable from prior linguistic context; control is structurally determined by the matrix verb. They're different mechanisms.
• Forgetting promise. Most object-position controllers are object-control, but promise is a subject-control verb that happens to take an object. "John promised Mary to leave" → John leaves, not Mary.
• Assuming infinitives always involve control. ECM verbs (Exceptional Case Marking) like believe, expect, want can take an overt embedded subject: "John believes Mary to be honest." Mary is here a real syntactic argument, not PRO.
• Mixing up partial and exhaustive control. "John managed to leave" is exhaustive (John alone leaves); "John wanted to gather" is partial (John plus others gather). Different verbs, different control flavors.