2.1 The analysis of the surface subject/logical object

Bruening (2021) proposes that objects of verbs are left implicit through a head that merges with the verb and stops it from projecting its internal argument. There are two such heads. One binds the verb’s argument as an existential (indefinite implicit objects of verbs like eat), while the other binds it as a definite (implicit objects of verbs like watch and win). Here I propose that middles involve the latter. Definite implicit arguments can be bound by quantifiers (see Williams 2015). So the following sentence can mean the same thing with and without the pronounced pronoun:

I propose that in the middle, the object of the verb is an implicit definite that comes to be bound by the surface subject.

Bruening (2021) calls the definite head “ι.” It merges with the V as a head and stops it from projecting its internal argument:

I modify the denotation of ι from Bruening (2021) to treat the bound object of the verb as an index n (it can be any index; type e is individuals, s eventualities, t propositions):

This essentially treats the implicit object semantically (but not syntactically) as a pronoun. An assignment function will assign the index a referent if the discourse context is felicitous. Or, the index can be bound by a lambda operator associated with a quantifier, as in the quantificational example in (4). ι will combine with a verb like win as follows, selecting an index at random:

Which verbs ι can combine with is a matter of selection: individual verbs say in their lexical entries whether they can combine with it. Turning to the middle, I propose that ι can combine productively even with verbs that do not select for it, provided that the output can combine with Voice_m so that the index comes to be bound. The way this works is as follows. First, there is a head, Voice_m, that combines with VPs of the form in (5). Voice_m is constrained so that it must form a complex head with V and ι, which it does by virtue of the complex head V-ι undergoing head movement to Voice_m. We can state this in terms of feature selection: V_m has an ι feature, but since ι strictly selects V (see section 2.2), the only way to check off this feature is for ι to combine with V and then have the result combine with Voice_m.

Voice_m is one of several Voice heads made available by UG. There is also a Voice_tr (transitive or active Voice) and a Voice_un (unaccusative Voice), as well as a Pass head (Bruening 2013). Voice_m has some additional unique properties. First, like Voice_tr, it introduces an agent. However, unlike Voice_tr, it does not project this argument in the syntax. The argument is instead also just an index in the semantics. See section 3 for more details on this aspect of the analysis.

Second, Voice_m counts as an A-movement head for the proposal in Bruening (2019), although it does not actually trigger A-movement the way Voice_un and Pass do (so middles could never involve A-movement, since movement must be triggered). Bruening (2019) proposes that A-movement heads trigger a special abstraction rule (Bruening 2019: 28, (83–85)):

The A-Movement Rule performs lambda abstraction. The rule creates a predicate where, typically, the A-trace in the sister of the A-movement head is the open argument position. In the case of middles, there is no trace, but because of ι, there is an index in the semantic representation. The lambda operator is able to bind this index.

Voice_m is also unusual in that it projects a specifier. That is, it takes a syntactic NP argument in addition to its complement VP. However, this NP argument is not a semantic argument of Voice_m. It is only interpreted through the A-movement rule. Voice_m bears the index of its specifier (10), and this enables the specifier to bind the index introduced by ι:

The agent that Voice_m introduces is an index in the semantics, just like the argument that ι binds (see section 3). Below is a simplified semantics, ignoring modality for now (see section 4):

These two mechanisms, the ι head proposed for implicit arguments, and the abstraction rule triggered by certain Voice heads in English, combine together in the middle to permit an NP generated in Spec-Voice_mP to bind the logical internal argument. Thus, the surface subject of the middle is interpreted as the internal argument, without it ever having occupied a VP-internal position.

We do have to explain why long-distance middles are impossible in this analysis. Implicit objects can be bound long distance, as in example (4), for instance. We might expect an example like *Politicians easily believe that lobbyists bribe to be grammatical, on the meaning, ‘One can easily believe that lobbyists bribe politicians’. The answer is licensing. The ι head is typically licensed only by particular verbs, through selection (see Bruening 2021). As stated above, ι can be added productively to verbs that do not select for it just when they combine with Voice_m. Voice_m can locally license the ι head, by virtue of combining with V and ι as a complex head (and it must do this, as described above). It cannot do this across phase and clause boundaries. Long-distance middles are therefore impossible.

The rest of this section motivates the first component of the analysis by describing numerous ways in which middles pattern very differently from passives, where object-to-subject movement is very well motivated. The hypothesis that the semantic object is base-generated as a surface subject, and no object position is projected, explains these differences.

2.2 Surface subject can’t be object of P

I start with an old but underappreciated—and contentious—observation. This is that the surface subject of a middle cannot be the semantic argument of a preposition. In my judgment and the judgment of every native English speaker I have consulted informally, objects of P are completely unacceptable. Middles contrast strikingly with passives in this regard:

If middles involved a step of A-movement, exactly like passives, then we would expect that they would be well-formed with objects of prepositions. English is generally quite free in allowing movement of objects of prepositions, both in A-movement and in A-bar movement. The fact that middles do not allow objects of prepositions is a first indication that they do not involve movement.

This judgment has been contested over the years. Keyser & Roeper (1984: 400) say that prepositional middles are marginal but consider them acceptable. They give the following examples one question mark:

Fagan (1988: 194) argues that Keyser & Roeper (1984) are not correct, and states that all the native speakers she has consulted find them unacceptable. Massam (1992) gives one example and gives it a question mark:

McConnell-Ginet (1994) says that prepositional middles are unacceptable but gives no examples (and credits the observation to Massam 1992).

Most recently, Newman (2020) claims that middles are well-formed with objects of prepositions. While she notes that many speakers do not like them, she says that many do, and marks the following examples with “(?)”:

Because judgments on prepositional middles have been contested since 1984, I decided to do a large-scale survey using Amazon Mechanical Turk. For this purpose I made use of the free tools described in Gibson et al. (2011) and available at http://tedlab.mit.edu/software/, modified for the purposes of this experiment. The experiment had a two by two design, crossing factors Passive-Middle and Verb-Preposition, as follows:

There were eight sets of items (all given above), and each subject saw only one member of each set. Subjects answered a comprehension question about every sentence and also rated every sentence on a scale of one to five (1: Extremely unnatural, 2: Somewhat unnatural; 3: Possible, 4: Somewhat natural, 5: Extremely natural). Subjects also rated 16 fillers. Eight of these were items for an unrelated study on adverbs and adjectives. The other eight were true fillers, created by modifying sentences from online newspaper articles. All sixteen fillers came in pairs of grammatical and ungrammatical, and each subject saw only one member of each pair. Subjects rated a total of 24 items.

83 workers were recruited from within the USA, limited to those classified as “masters” by Amazon Mechanical Turk. Several completed the survey more than once; in those cases, all beyond the first were thrown out. Subjects’ data were also discarded from the analysis if the subject was not from the USA, did not identify their native language as English, or got fewer than 75% of the comprehension questions correct. Data from 6 subjects were thrown out for these reasons (most for repetitions). This left 77 subjects whose data entered the analysis. 42 identified as male, 33 as female, one as other (one did not answer). 5 were age 20–30, 48 were 30–45, 10 were 45–55, and 13 were over 55 (one did not answer).

Mean ratings and standard deviations are shown below, both raw and z-scores:

Participants behaved as expected on the fillers: They rated the grammatical ones very high, above 4, and the ungrammatical ones very low, close to 2. The two Passive conditions were also rated above 4, while the two Middle conditions were rated lower. In fact, the PMiddle condition had the lowest ratings I have ever seen from Amazon Mechanical Turk participants. Ratings below 2 are virtually unheard of.

Statistical analysis was run using R (R Core Team 2012). Z-scores were analyzed by means of linear mixed-effect modeling using the R-package lme4. In the model, the two fixed effects were Factor 1 (verb versus preposition) and Factor 2 (middle versus passive), and subjects and items were included as random slopes and intercepts. The full model failed to converge, so it was simplified until it did.¹ Reported p-values were extracted from the fitted model objects using the Satterthwaite approximation implemented by the lmerTest package. Both factors were significant, as was their interaction. Middles were consistently rated worse than passives, and items with prepositions were consistently rated worse than items with just verbs. At the same time, prepositional middles were rated even worse than would be expected from the combination of these two main effects.

I conclude from this survey that prepositional middles are very unacceptable. This matches my judgments and the judgments of every native speaker I have asked.

Newman (2020) also claims that there is an effect of repeated exposure, such that speakers start to accept prepositional middles after hearing multiple examples of them. I was unable to test this hypothesis in the current experiment because each participant saw only two tokens of prepositional middles (a statistical model with presentation order as a factor found no effect, not surprisingly). If this were true, however, it would be unclear what its significance would be. One possibility is that speakers’ initial unhappiness with prepositional middles is only the result of never having encountered them before, and repeated exposure relieves this unhappiness. On this view, prepositional middles are produced by the grammar, they just happen not to have been produced by a single speaker of English until two linguists (Keyser & Roeper 1984) decided to try them out. I find this extremely unlikely, since English speakers are generally quite productive with prepositional passives and preposition stranding in general. They even do it in word-formation, producing such adjectival passives as unheard of and unasked for. If the grammar did allow prepositional middles, surely some creative English speaker would have used them before now. The counter hypothesis, which seems to me to be more likely, is that the grammar does not allow prepositional middles. Only linguists ever thought to try them out, since no native speaker merely using their grammar (rather than experimenting with it) ever would. I suspect that only linguists, who are used to experimenting with their grammar, would find improvement under repeated exposure to prepositional middles. However, it will take additional experimentation to be able to evaluate this, and this additional experimentation I will have to leave for future work. For the moment, I conclude that prepositional middles are not grammatical, and our model of the grammar of middles will have to rule them out.

Having concluded that prepositional middles are ungrammatical, let us consider why. Massam (1992) suggests that the restriction on middles here is just like the restriction on object drop in recipes and instructions. They often cannot be the object of a preposition, either:

Massam points out that null objects in recipe contexts are also not allowed to be the theme of a double object construction:

As we will see in more detail in section 2.7, middles are not allowed with double object constructions. However, middles contrast with null objects as the goal of a double object construction in this respect. Middles are not well-formed with goals, but null objects are fine as goals:

It therefore does not appear fruitful to try to explain all the restrictions on middles by positing a null pronoun in object position.

Additionally, while the examples of null objects of prepositions in (26b) and (27b) are indeed unacceptable, other cases do seem to be allowed:

I therefore conclude that we cannot explain the preposition ban in middles by positing a null object of the type that appears in instructions. However, there is something right about the idea. Objects of prepositions may not be implicit outside of instructions the way objects of verbs can be. The entire PP can be implicit, but not just the object of P:

So, while objects of P can be the type of null object that appears in instructions, they cannot be the type of implicit argument that figures in the current analysis of middles (see Bruening 2021 and section 3.5 for discussion of the differences between these two types of null objects). This means that the head ι (and its indefinite counterpart) can only combine with Vs, and never with Ps. We can model this as selection: ι strictly c-selects category V. It follows that a middle may never be formed with the object of a preposition.

In the passive, in contrast, a head Pass combines with Voice_tr and blocks the projection of the external argument (e.g., Bruening 2013):

Some NP then needs to move to the surface subject position (in English, which does not allow impersonal passives). An object of a P can serve for this purpose as well as an object of a V can. A trace as the object of a P can also be abstracted over by the A-movement rule.

2.3 Idiom chunks

Middles also contrast quite strikingly with passives when it comes to verb-object idioms. As is well-known, many verb-object idioms can be passivized quite easily. However, none of them form middles:²

This is even true of verb-object idioms that appear canonically in the passive:

Newman (2020) claims that some idioms are acceptable as middles. She gives the following examples (her judgments):

Once again these judgments do not match mine, or those of anyone I have asked. Note also that keep up appearances is not particularly idiomatic, as all items in it are interpreted literally. Additionally, break has an unaccusative use and could be being used in that way for speakers who accept this example (in my judgment, the idiom break the ice is necessarily transitive and cannot be an unaccusative). I conclude that idioms do not form good middles, even when they do form good passives.

We can understand this fact if idioms are stored as partial syntactic structures, as in O’Donnell (2015); Bruening (2020b). For instance, spill the beans is stored as a VP:³

These partial syntactic structures can be altered once merged into the syntax, but they must be merged with the structure they are stored with. In the passive, the VP can be merged into the structure and then the NP object can move away (near obligatorily, with the canonically passive ones). This is allowed. However, if in the middle the head ι stops the projection of the argument of V, the idiom will not be usable:

Thus, the current analysis explains why idioms chunks can passivize but cannot form middles.

In contrast with the idioms illustrated above, if the surface subject of the middle corresponds to an open slot in an idiom, then that idiom can form a middle:

Instrumentals adjoin to a projection of Voice (see Bruening 2013). I therefore assume that this idiom is stored as the following partial structure (where Voice is unspecified, it can be Voice_tr or Voice_m):

I assume that X just indicates an argument that needs to be saturated. It does not necessarily need to be projected. In the passive, it is projected, and it then moves to the surface subject position. In the middle, it is instead suppressed by ι, but the argument role is later saturated by the base-generated surface subject. If the only constraint on open slots is that they need to be saturated, then the middle will be well-formed. All the rest of the structure in (48) besides X will be present:

Thus, the current analysis correctly predicts that open slots in idioms can be the surface subject of a middle, but NPs that are a fixed part of the idiom cannot be.

2.4 Expletives

Expletives are particularly revealing, as they often constitute the strongest argument for A-movement. Expletives readily undergo object-to-subject movement in the passive, with believe-type verbs (or “ECM verbs”):

I have been able to find three ECM verbs that seem to form middles well, when they are used just with an NP object. These are calculate, compute, and explain (since calculate and compute are close synonyms, I only illustrate with calculate). As shown in the (a) examples below, they do form middles. As shown in the (b) examples, they also form passives as ECM verbs taking a non-finite complement. An expletive can raise from object to subject in the passive in the (b) examples. However, the (c) examples show that this is completely unacceptable with a middle. To the best of my knowledge, this is a new observation.

It expletives are just as unacceptable as there:

In the current analysis, the head ι blocks projection of the argument of V, but this argument comes to be bound from a higher position. An expletive fills no argument role, so there is nothing that the head ι could suppress but allow to be bound. The current analysis therefore explains why middles differ from passives in this respect.

2.5 ECM in general

Turning to ECM in general, Keyser & Roeper (1984: 407), Ackema & Schoorlemmer (1995: 183), Ackema & Schoorlemmer (2006: 172) state that middles are not well-formed with ECM verbs. However, they only give examples with believe and expect. These do not form middles even with NP objects. As noted above, I have only found three ECM verbs that seem to be compatible with the middle, namely, calculate, compute, and explain. Speakers that I have consulted find that ECM is acceptable with calculate/compute, but not with explain:⁴

I suggest that some verbs are ambiguous between raising and control. The verbs that work here, calculate and compute, have a control derivation available to them. They can take an internal NP argument in addition to an infinitive. The NP controls the null subject PRO of the infinitive:

ι can combine with such a verb, blocking the projection of the NP argument. The argument filling that role is then projected higher, as explained above. From the higher position, it can still control PRO in the embedded clause. This control derivation does not work with expletives, as shown in the previous subsection, because expletives are not capable of control.

2.6 Non-selected arguments in resultatives

Resultatives are well-formed with middles (Hoekstra & Roberts 1993):

This might at first be taken to indicate that the surface subject must be an underlying object, since it is generally thought that only underlying objects may be the subjects of resultatives (Levin & Rappaport Hovav 1995). I will show momentarily that this does not follow. First, it is important that resultatives do not form middles when the NP involved is not a selected argument of the verb:

Middles again contrast with passives in this respect. To the best of my knowledge, this is a new observation.

To explain these facts, we need a minimal analysis of resultatives. I propose that resultatives involve a head Res(ult) that mediates between the adjective and the VP (on different versions of a Res head, see Ramchand & Svenonius 2002; Jung & Choi 2023 and references there):

The argument of the adjective can either be saturated within AP, or not. Res can take either an argument of type <e,st> or type <s,t>. Either way, it introduces resultative semantics (the adjective state comes about as a result of the VP event). If Res takes an AP argument of type <e,st>, then its V argument must also be of type <e,st>. In this case, it unifies the two individual arguments to create a predicate of type <e,st>. The output of this then combines with an NP, the metal in the tree above. This NP is interpreted as both the object of the verb and the subject of the adjective. (The V moves to Voice, not shown, putting the V before the object.)

If the argument of A is instead saturated within AP, then Res takes an AP argument of type <s,t> and also a V argument of type <s,t>. That is, whichever type Res’s first argument is, its second argument has to be that, too. This gives us non-selected arguments in resultatives:

Object drop with verbs like drink is analyzed in the theory adopted here as suppression of the object of the verb by a head that merges with the verb. In this case, the implicit object of drink is indefinite, so the head ∃ merges with the verb rather than the head ι (Bruening 2021). Merger of ∃ completely suppresses the argument of the V, so that there is no NP object of drink in the syntax. This makes drink exactly like crow: they do not have objects in the syntax. Res takes an AP argument of type <s,t>, and a V argument of type <s,t>:

There would be nothing semantically wrong with Res combining with a verb after it has combined with its object. The output of that would be type <s,t>, and so the AP could also be type <s,t>:

The problem here is not semantic, but syntactic. In an English transitive, only one object can be case-licensed. Spelling this out formally, Voice_tr can case-license exactly one VP-internal NP, through Agree (Chomsky 2000). *They drank beer, liquor, and wine the pub dry is ungrammatical because Voice_tr can only case-license one of the two NPs (beer, liquor, and wine and the pub) that need case-licensing. In contrast, they drank the pub dry is well-formed because Voice_tr can case-license the pub. The object of drink was existentially closed by the head ∃ and was not projected into the syntax at all.

Turning back to middles, they are well-formed where the arguments of the A and the V are unified. In the analysis proposed here, the head ι must merge with the combination of V and ResP:

ι fills in an index for the unified arguments of hammer and flat, which then comes to be bound by an NP merged in Spec-Voice_mP, exactly as before. This is well-formed.⁵ V moves to ι and on to Voice_m, satisfying Voice_m’s requirements.

The reason non-selected arguments are not well-formed (*This pub drinks dry easily) is that the NP is an argument of the adjective, not the V. ι only combines with verbs, not with As. It therefore cannot suppress the argument of the A. The NP cannot be projected in Spec-AP and then move to Spec-VoiceP as in the passive, because Voice_m does not trigger A-movement the way Pass does. Thus, the current analysis explains the facts of resultatives, and why passives and middles differ.

Note that the generalization that emerges from PPs, resultatives, ECM, and expletives is that the surface subject in a middle must be a semantic argument of the verb. The current analysis captures this generalization: the head ι can only combine with a V and in doing so it binds the V’s internal (individual) argument. If the V does not have an internal argument, ι cannot combine with it.

2.7 Double object verbs

As mentioned briefly above, double object verbs do not form middles (Kayne 1982). Middles are well-formed with the prepositional dative, but they are unacceptable with the double object construction:

As can be seen, double object constructions do not form middles with either the goal or the theme as the surface subject.

These data follow naturally from the current use of ι from Bruening (2021) plus the analysis of the double object construction from that same paper. Bruening (2021) adopts the proposal that the double object construction involves an Appl (licative) head between Voice and V (Bruening 2001; 2010):

Bruening (2021) proposes that both the verb and the Appl head take two arguments (the verb optionally). The verb moves to Appl and Appl unifies both their arguments, such that the goal is both a goal in a verbal event and a possessor in a possession eventuality, while the theme is both the theme in the verbal event and the possessee in the possession eventuality. Importantly, V cannot combine with one of the heads that creates implicit arguments, ι or ∃, because these heads would close one of the arguments of V, which would then block argument unification with Appl. See Bruening (2021) for details.

This same problem is what blocks middle formation in a double object construction. The ι head needs to combine with a V and bind its internal argument. If it were to do so, however, then Appl could not unify its argument with the verb’s internal argument. ι cannot combine with the output of V and Appl combining, because it strictly selects Vs, as we have already seen. (Note that implicit arguments are possible in double object constructions, but, according to Bruening 2021, this is only possible if an individual verb specifies that it can combine first with Appl and then with ι or ∃. The middle is not selected, it is a productive syntactic combination, and so this will never be possible for a middle.)

In contrast, the prepositional dative has the NP and the PP as arguments of the V. ι can combine with V and block projection of the NP, while allowing the PP to combine prior to the merger of Voice_m:

ι only blocks projection of the first argument of the verb, so the second must still merge with the V to satisfy its requirements.

The current analysis therefore explains why middles are well-formed with the prepositional dative but not with the double object construction.

2.8 CP subjects

Middles also behave very unlike cases of A-movement when it comes to CP complements. CP complements of verbs can become the subject in the passive:

They can even undergo further A-movement, with raising:

In contrast, CPs cannot be the surface subject of a middle at all. To the best of my knowledge, this is a new observation:

This is even true where the verb does form a middle with an NP complement, as well as a passive with a CP:

The restriction here cannot be about syntactic category. As shown by Alrenga (2005), CPs behave as though they are NPs in the passive. They are only allowed as the subject of a passive if the verb permits an NP object in the active. According to Alrenga (2005), this is an instance of the generalization that in leftward displacement, like topicalization and clefting, CPs can only be related to NP positions (Higgins 1973; Kuno 1973; Kaplan & Bresnan 1982; Postal 1994; Bresnan 1995; Takahashi 2010). This means that in some sense, the CPs in the above examples of the passive are actually NPs. If CPs can be NPs in the passive, they should be able to be NPs in the middle, too. So we cannot rule them out by saying that the middle only allows NPs. The restriction against CP subjects of middles must be about semantic categories, instead. It appears that ι requires an individual argument of type e to suppress when it combines with the V. It cannot suppress a propositional argument (type t or something similar).

Importantly for the discussion in this section, the restriction against CPs indicates that middles are behaving very differently from instances of A-movement like passive and raising. Passive and raising do not care about semantics at all, either in terms of semantic types or thematic roles. Middles do. If middles involved A-movement, we would not expect this restriction.

2.9 Reconstruction

The first component of the proposed analysis is the base-generation of the surface subject in a high position, with no representation for it in the VP. The analysis therefore predicts that reconstruction of the surface subject into a VP-internal position should be impossible. This does not mean that we should not see reconstruction at all, however. For instance, the high base-generation position is still lower than the surface position of negation and modals, so we expect that the surface subject could reconstruct below negation and modals. To test whether reconstruction into VP is possible, then, we need to look for reconstruction specifically below the lexical verb.

One such case involves intensional verbs. Intensional verbs do not entail the existence of their object, but they do entail the existence of their subject:

The speaker can therefore both deny the existence of the object, and still assert that a subject is V-ing the object, as in (73a). The object is in the scope of the intensional verb. In contrast, the speaker cannot consistently deny the existence of the subject, and assert that the subject is V-ing the object in (73b). The subject is outside the scope of the intensional verb.

Many intensional verbs do not form middles, but ones like blame and scapegoat do. With these verbs, we see the same pattern as with look for: the subject is outside the scope of the intensional verb, but the object is inside its scope:

The passive subject patterns with the active object. It can be interpreted within the scope of the verb, through reconstruction:

In contrast, I get very uncertain judgments from speakers on middles. Speakers do not find examples like the following patterning clearly with either active subjects or active objects:

On the analysis proposed here, the middle has a paraphrase like the following:

I believe that this paraphrase also gives rise to uncertainty about judgments of contradiction:

Thus, the middle is patterning as predicted by the proposed paraphrase, where the surface subject did not start as the underlying object.

The judgment becomes clearer if the middle is made episodic. It contrasts clearly with the passive in such cases:

It therefore appears that the surface subject of a middle is not able to reconstruct into a VP-internal position, exactly as the current analysis predicts.

2.10 Summary

This section has presented data—some old, much of it new—which indicates that there is no movement from object to subject in the middle. Much earlier literature also argued for this conclusion, in particular Ackema & Schoorlemmer (1994; 1995). Taken all together, the evidence indicates that there is no object-to-subject movement in the English middle, contrary to A-movement analyses like those of Keyser & Roeper (1984); Stroik (1992); Hoekstra & Roberts (1993); Newman (2020). The surface subject has to be base-generated in a position external to the VP. See the appendix for additional support for the proposed analysis from VP ellipsis.

It should also be noted that all of the data presented here also argue against another logical possibility, where the object of the V is projected and moves as a null operator. Bruening (2014) proposes such an analysis for adjectival passives. In the middle, one could propose that a null object moves and abstracts, forming a predicate that gets predicated of the (base-generated) surface subject. All of the data presented above indicates that this is incorrect. Middles act both unlike passives and unlike constructions where null operators are well motivated (e.g., relative clauses, tough-movement). All of the data indicate that there is no object internal to the VP at all. See the appendix on putative parasitic gap licensing in middles.

3.1 The external argument is entailed

First, middles are like passives in that they entail the action of an external argument, unlike unaccusatives:

Mauner & Koenig (2000) showed this experimentally: subjects judged sentences like The donated antique vase had sold immediately, yet no one had sold it to be contradictory.

More than just entailing an external argument, Ackema & Schoorlemmer (1994) argue that middles require agentive external arguments.⁶ They show that only verbs that take agent subjects form middles. Verbs that have non-agent subjects do not (Ackema & Schoorlemmer 1994 gave the Dutch equivalents):

Earlier work on middles often posited an “affectedness constraint,” according to which the logical object needs to be affected. Ackema & Schoorlemmer (1994) show that the actual requirement is an agent requirement on the logical external argument. Verbs that do not have affected objects can easily form middles, so long as the logical external argument is agentive (all examples from Ackema & Schoorlemmer 1994; the first set were the Dutch equivalents):

None of the above verbs have affected objects, but they all have agentive subjects.

Middles then entail an agentive external argument. The question is whether this external argument is present in the syntax, or is only present semantically. I begin with phenomena that clearly indicate that the logical external argument is incapable of participating in syntactic dependencies, before turning to binding.

3.2 Depictive secondary predicates

The unexpressed external argument of a middle cannot be modified by a depictive secondary predicate:⁷

Middles contrast with passives in this regard. Passives permit a depictive secondary predicate to modify the unexpressed logical external argument, as the following corpus examples show (see Pitteroff & Schäfer 2019: 160–161 for recent discussion and references):

The logical external argument of a middle is even weaker than that of a passive, then, since the two contrast in this respect. I will argue that neither is actually present in the syntax, but they are missing in different ways, which accounts for this contrast.

3.3 Control

Whether the logical external argument in a middle can control PRO has been a matter of debate. There is one obligatory subject control verb that also takes an object, namely promise. However, a middle seems ill-formed with this verb, regardless of control (see the finite clause in 86c):

It-passives of object control verbs permit control by the implicit argument of a passive, but the corresponding middle is unacceptable:

However, it is not clear whether this is due to the failure of control, or the inability of an expletive to be the subject of the middle (see section 2.4).⁸

The literature has concentrated on control into adjuncts. Stroik (1992) argues that the unexpressed external argument can control into an adjunct clause on the basis of examples like the following:

However, Ackema & Schoorlemmer (1995) show that PRO in these adjunct clauses does not require syntactic control:

Control into adjunct clauses is therefore not telling.

Middles do not license purpose clauses, except of the kind that are good with anything (Roberts 1987; Stroik 1992):

Middles contrast with passives in this respect, as is well known (Roeper 1987):

The contrast between middles and passives with respect to purpose clauses was verified experimentally by Mauner & Koenig (2000) (but note that some of their middle items were actually unaccusatives, which also do not license purpose clauses). (For more discussion of control by the implicit external argument of a passive, see Pitteroff & Schäfer 2019.)

So far, then, it appears that the logical external argument of a middle is not present syntactically. It cannot be modified by a depictive secondary predicate, and it cannot participate in control.

3.4 Binding

Binding at first appears to contradict the evidence from depictive secondary predicates and control. Middles do appear to permit an anaphor bound by the logical external argument, as Stroik (1992) claimed:⁹

The passive equivalents of these also seem to be acceptable:

One might conclude from this that the logical external argument of both the passive and the middle is syntactically represented. This conclusion would be too hasty. Anaphors can sometimes be used logophorically, meaning that they can take as an antecedent an entity that does not locally c-command them and may not even be present in the syntax (Pollard & Sag 1992; Reinhart & Reuland 1993). Charnavel & Sportiche (2016) offer a way to rule out this possibility: using inanimate anaphors. Anaphors used logophorically can only take sentient entities as their antecedents. Inanimate anaphors require a local c-commanding antecedent and do not permit a logophoric use. When we use inanimate anaphors in passives, we find that they are unacceptable:

I take this to indicate that putative examples of binding by the missing external argument of a passive are not actual examples of binding (contra, e.g., Roberts 1987; Baker et al. 1989; Collins 2005). The missing external argument of a passive is not capable of binding an anaphor. An anaphor can be acceptable in the passive, but when it is, it is being used logophorically.

Unfortunately, we cannot use the same test in middles, since the logical external argument of a middle is necessarily animate (human or animal). However, it seems likely that middles and passives are behaving the same with respect to binding. The missing external argument of a passive is capable of more syntactic effects than that of a middle: it can be modified by a secondary depictive predicate and it can control PRO (see above). It would be very strange if the logical external argument of a middle could bind an anaphor but the logical external argument of a passive could not. If anything, it should be the other way around. Since apparently bound anaphors in the passive actually seem to be logophors, I conclude that they are in middles, as well. All of the data are consistent with this conclusion, and taking syntactic binding to be possible in the middle is inconsistent with the data from depictive secondary predicates and control shown above.

Further strengthening this conclusion, unaccusative examples that are similar to the examples of the middle in (92) also permit an anaphor. Unaccusatives do not have an external argument, so they must involve anaphors being used logophorically:

It is clear that the examples of the middle in (92) could involve anaphors being used logophorically, since very similar examples of unaccusatives must.

It appears that, when there is no logical external argument in the syntax, it is possible for any anaphor in the verb phrase to be interpreted logophorically rather than as a syntactic anaphor. Why this would be the case is not entirely clear.

Regardless, if binding were possible, it would perhaps be the strongest possible evidence for the syntactic presence of the logical external argument in the middle. Conversely, if the logical external argument is not capable of syntactic binding, then the logical conclusion is that it is not present in the syntax. Since middles and passives only seem to allow logophoric “binding” and not syntactic binding, I conclude that in neither case is the logical external argument represented syntactically as an NP.

3.5 Accounting for the facts

Implicit objects of verbs in English divide into two types (see Bruening 2021 for recent discussion). Those that are licensed by particular lexical items behave as though they are completely absent from the syntax, since they cannot bind an anaphor, they cannot be modified by a depictive secondary predicate, and they cannot control PRO (Rizzi 1986):

In contrast, null objects that are licensed in instruction contexts (e.g., Massam & Roberge 1989) can bind anaphors, can be modified by depictive secondary predicates, and can control PRO:

The logical external argument of the middle seems to behave exactly like the first type of implicit object: it cannot bind an anaphor, it cannot be modified by a depictive secondary predicate, and it cannot control PRO. I will accordingly treat it in a similar manner, and not project it in the syntax at all. The logical external argument of the passive, in contrast, behaves in a non-uniform manner: it cannot bind an anaphor, but it does seem to be able to control PRO and it can be modified by a depictive secondary predicate.

To account for these facts, I propose that control and depictive secondary predicates do not require the syntactic presence of an NP. They do, however, require a head that would have projected such an NP as an argument. Following Bruening (2013), the passive involves the ordinary transitive Voice_tr head, which normally projects the logical external argument in its specifier. In the passive, however, another head, Pass, stops this argument from being projected and instead existentially quantifies over it:

In this structure, there is a node, namely the mother of Voice_tr and VP, that includes an open predicate λx. If depictive secondary predicates and purpose clauses can adjoin to this node, then they can unify their own open arguments with x, prior to it being existentially closed by Pass. This will have the result that the implicit logical external argument of the passive can control PRO in a purpose clause and can be modified by a depictive secondary predicate (cf. Pitteroff & Schäfer 2019: note 20).

In contrast, an anaphor requires a c-commanding NP in order to be syntactically bound. In (98), there is no NP that saturates the external argument role, instead that argument is existentially quantified over by Pass. Hence, the implicit external argument of a short passive cannot bind an anaphor.¹⁰

Turning now to middles, they have a head, Voice_m, which does not co-occur with Voice_tr. Voice_m does not take the logical external argument as an argument, instead it includes the semantics of such an argument as part of its semantic representation (see 12b). The mother of Voice_m therefore does not include an open predicate λx…, and so it is impossible for depictive secondary predicates and purpose clauses to adjoin to this node and unify their arguments with its argument. This explains why the logical external argument of a middle cannot be modified by a depictive secondary predicate or control into a purpose clause. Since this argument is also never projected as an NP, it also cannot bind an anaphor.

As for implicit objects of verbs (96), Bruening (2021) proposed that they are bound by a head that adjoins to V, as described in section 2.1. Assuming that depictive secondary predicates and purpose clauses cannot adjoin to V as a head, prior to the adjunction of the head that binds V’s argument (since they are phrases and not heads), they will also fail with implicit arguments. At any phrasal node above V, V’s argument will have been closed off already, and so argument unification will be impossible.

In contrast, null objects in instruction contexts must be full NPs, identical in all respects except pronunciation to a regular pronounced object. They can do everything an overt NP can, because they are ordinary NP arguments.

3.6 The interpretation of the external argument

The logical external argument of the passive is generally regarded as being existentially quantified (see Bhatt & Pancheva 2006; Bruening 2013). In contrast, that of the middle is typically interpreted as something like generic one. Paraphrases in the active with one and an ability modal are typically synonymous with the middle:

However, we have also just seen examples where the logical external argument seems to be able to be interpreted referentially. The following example is repeated from above:

In cases of logophoric binding like this, the logical external argument is interpreted as the logophoric center, in this case, Melissa. We can also find cases where the logical external argument appears to be bound by a higher quantifier (the first two sentences below modify examples from Stroik 1992: 133, (17a,c)):

It is also common to have the logical external argument be interpreted as the speaker:

This seems to be the most common interpretation for episodic middles.

I therefore hypothesize that the logical external argument is simply an index in the semantics. Voice_m introduces an agent in the semantics (but not the syntax), and the entity filling the agent role is an index. This index will be assigned an interpretation by an assignment function. In typical generic middles, it will be interpreted as generic one or you or anyone. However, the context can make it be interpreted in a variety of ways. The index can even be bound by a non-referential quantifier, as in example (101a).

3.7 By-phrases and for-phrases

Middles do not allow by-phrases (Keyser & Roeper 1984):

Much literature has taken this to indicate that the external argument of the middle is not syntactically present. In the current analysis, it follows from the fact that the logical external argument is not projected at all. In Bruening (2013), the by-phrase adjoins to a projection of Voice_tr and saturates the open argument of Voice_tr with its own argument. In the current analysis of the middle, the logical external argument is never an open argument. It is only a semantic argument of Voice_m. It is therefore impossible for a by-phrase to appear in the middle.

Stroik (1992) claims that the implicit external argument in a middle can be expressed by a different PP, one headed by for:¹¹

Some researchers have argued that these PPs are not saturating the external argument role, but are instead experiencers (Hoekstra & Roberts 1993; Ackema & Schoorlemmer 1995; Klingvall 2007). An idea that is similar in spirit is that they are benefactives, as for-phrases typically are. Whatever we call the role, it does seem to be true that the complement of for does not necessarily saturate the external argument role:

This contrasts with the by-phrase in the passive, which does necessarily saturate the external role of the verb:

I assume that the for-phrase, if present, is not saturating the external argument role. It introduces either an experiencer or a benefactive. In most contexts, this experiencer/benefactive will be understood to also be the one filling the external argument role. In the current analysis, an assignment function will assign the complement of for as the value of the index that is saturating the external argument role in the semantics of Voice_m.

3.8 Summary

To summarize this section, the missing logical external argument of the middle behaves as though it is not present in the syntax. It cannot be modified by a depictive secondary predicate and it cannot control PRO. Apparent examples of syntactic binding are most plausibly analyzed as logophoric. In the current analysis, there is a Voice head, Voice_m, which introduces an external argument in the semantics as an index. This argument has no syntactic representation.

See the appendix for additional facts and arguments regarding external-argument-oriented modifiers, donkey anaphora, allomorphy, and Stroik’s (2005) argument that middles lack vP/VoiceP.

4.1 Middles are not necessarily generic or stative

It is often claimed that middles are necessarily generic (e.g., Condoravdi 1989; see Ackema & Schoorlemmer 2006). However, it is clear that not all instances of middles are generic. As previous literature has noted, episodic middles exist (e.g., Keyser & Roeper 1984; McConnell-Ginet 1994):

There is no sense in which the above sentences are generic. Yet they have all the properties of middles discussed here: the logical internal argument is in subject position but there is no passive morphology; there is a missing but entailed logical external argument.

It therefore seems that genericity is not a crucial component of middles. Middles do lend themselves well to generic statements, especially with bare plural subjects and/or present tense. However, these are factors that license generics outside of middles. There is therefore no reason to posit a generic operator as a crucial part of the middle. Rather, clauses with bare plural subjects and/or present tense license a generic operator, and so any middle clause with a bare plural subject and/or present tense can also have a generic operator. But the syntactic analysis of middles should not include any reference to a generic operator.

It is also often claimed that middles are stative (e.g., Ackema & Schoorlemmer 1995; Lekakou 2005). The episodic examples above, and many of those illustrated throughout this paper, also show that this is not correct. Middles can be eventive. When they are, they can appear in the progressive, something that is not possible for statives in English:

I therefore do not include either stativity or genericity in the syntactic analysis of middles. Middles, like actives and passives, can be either stative or eventive, or generic or episodic.

4.2 Middles are not necessarily dispositional

Many researchers have also noted that middles typically describe a disposition of the surface subject. This has been codified as a “responsibility condition” in van Oosten (1977), and in the form of disposition ascription has been argued to be one of the defining properties of middles (Lekakou 2005; Klingvall 2007; Fábregas & Putnam 2014). It does seem to be true that many middles ascribe a property to the surface subject. However, like the tendency to be generic and stative, this also seems to be only a tendency and not a requirement. Consider the following example from Newman (2020):

In this example, it is not inherent properties of Sheet #1 that led it to wash better, it was accidental properties of the washing event. Other examples can also be constructed that show that disposition ascription is not necessary in a middle:

In the above example, it is properties of the knife, not the surface subject of the middle (this soft white bread), that are relevant. Many such examples can easily be constructed:

In this example, the ease of stowing is explicitly contrasted with inherent properties of the surface subject (its “disposition”).

Lekakou (2005) also argues that disposition ascription is part of the truth-conditional semantics of the middle, which makes it truth-conditionally distinct from the corresponding passive. She gives the following example:

In this example, it is possible to assert the middle while denying the corresponding passive, without contradiction. Lekakou takes this to show that the middle has an additional component of meaning that the passive lacks, which she argues is disposition ascription. In the current analysis, however, what the middle has that the passive lacks is a modal semantics of ability. The appropriate comparison for the middle is therefore the corresponding passive with an overt ability modal. It is a contradiction to assert the middle while denying the corresponding passive with an ability modal:

This shows that what middles have that passives do not is a modal semantics of ability, not disposition ascription.

4.3 Modality: Ability

The contradictions in (113) strongly suggest that middles include an ability modal. If we look at middles that are not generic, we can see this quite clearly:

While these sentences entail that the the butter was spread and that the elliptical machine is being assembled in the actual world, they also entail some kind of ability. In this they are like the modal verbs manage and be able to (e.g., Hacquard 2006). Appropriate paraphrases are, ‘I was able to spread the butter easily after I microwaved it’, and ‘I am (currently) able to assemble the elliptical machine’.

This ability meaning is present in all generic middles, as well. Consider some examples:

These have the meaning that, in general, agents have the ability to easily bribe politicians, and, in general, agents have the ability to button the dress in the back. This is particularly clear in the ‘bribe’ example. With the ‘button’ example, ability becomes very clear with negation, and this is true of other examples as well:

These have the meaning that agents are unable to button the dress or handle the car.

Our contradiction test also works in the other direction, denying the middle while asserting the passive with an overt modal:

Since these sentences are also contradictions, we can be reasonably secure in the conclusion that middles include the semantics of an ability modal.

Newman (2020) argues that middles do not have to involve any modality at all. She gives the following example (repeated from above), and claims that there is no detectable modality:

However, examples like this are consistent with modality. The meaning here is exactly like that in I was able to wash sheet #1 better than sheet #2, which is also perfectly felicitous in the given context. The sentence with was able to clearly has an actuality entailment, which means it makes an assertion about the real world, but it also involves modality in that it quantifies over possible worlds, namely those in which a washer has certain abilities.

Another example that Newman (2020) provides to argue against modality in middles is the following:

This sentence is also compatible with modality; an appropriate paraphrase is that an agent (here, Wallace) was not able to cut the bread.

Newman (2020) argues that, rather than modality, there is instead an “implicit attempt” in a middle. However, what Newman is describing is modality. An attempt is something that may or may not succeed in the actual world; to talk about an attempt, we have to talk about possibilities. This is exactly what modality is. Newman’s objections are therefore without force.

Given these considerations, I propose that a modal semantics of ability is a crucial component of all middle sentences, as was also proposed by Fagan (1992). In the past tense or the progressive aspect, this ability modal has an actuality entailment, exactly like the modal be able to (see Bhatt 1999; Hacquard 2006; 2020).

I propose that this modal semantics is part of the semantics of Voice_m. Voice_m does two things semantically: (i) it introduces an agent, as an index; and (ii) it introduces a modal semantics of ability, and says that the agent it has introduced has the ability to initiate events of the type denoted by the internal argument of Voice_m (the VP). This seems to adequately capture the semantics of the English middle.

4.4 Summary

This section has argued that English middles involve a modal semantics of ability. They are not necessarily generic or stative, and they also do not necessarily ascribe a disposition to the surface subject. I have proposed that Voice_m, in addition to introducing an agent, also has the semantics of an ability modal. In the past tense and the progressive aspect, it has an actuality entailment, like the overt modal be able to.