Extraction from gerunds and the internal syntax of verbs

2.1 Restriction I: Aktionsart restrictions

The first requisite in Spanish, and the one which is most relevant to our purposes, has to do with the Aktionsart of the main predicate. The extraction is possible with achievement main verbs (see all the previous acceptable examples). Accomplishments (7) produce ungrammatical results, as do activities (8).

With respect to states, extraction is not possible simply because a gerund construction is ungrammatical with them. The examples in (10) show that these verbs are statives: they are ungrammatical as complements of perception verbs and in combination with progressive periphrases, two tests that classically identify states (avoid a relative clause interpretation of the gerund, which is anyway marked in Spanish).

The conclusion is that extraction is only possible for speakers when the finite verb is an achievement. A simple statistical analysis performed over the speakers’ judgements shows that this difference is quite sharp. A question-answer pair like (6) above received high grades (mean=7.2732, median=8, standard deviation (sd)=2.99, and a variation coefficient (vc)=0.41). In contrast, a pair like (11), with an activity as main predicate, was considered ungrammatical (mean=2.16, median=1, sd=2.06, vc=0.95).

Since Maienborn (2003, 2005), alongside the ‘classical’ activities represented by (11), a class of non-dynamic activities (also known as Davidsonian states) has been distinguished (eg., wait, shine). As illustrated in (12), the ungrammaticality is also sharp with Davidsonian states (mean=2.32, median=1, sd=2.06, vc=0.99), showing that with respect to this phenomenon there is no need to differentiate the two subgroups.

We also ran a Pearson’s chi-square test with Yates’ continuity correction to compare the number of positive answers (6 or more) vs. the number of negative answers (5 or less) in the group formed by the sentences whose finite verb was an achievement, vs. the group with accomplishments or activities. This was made, of course, to assess the likelihood that the difference in answers was not due to chance. As chi-square does not measure the effect size, we ran over those results a Cramér’s V test.

The comparison shows that there is a significant tendency, and that the effect is big (chi-squared=1059, 594; degrees of freedom (df)=1, p-value < 2.2e-16, V value=0.78). The following boxplot representation summarizes the contrast between the two groups, which is visually sharp; notice, however, that individual speakers are outliers (Figure 1).

Figure 1:

Grammaticality of extractions with achievements vs. other classes.

These results confirm the intuitions of the authors of this paper and show that they can be extended to a wider community of non-biased speakers.

Note that the results show a language-particular difference with English: in English, unlike Spanish, the transparent gerund can combine with a verb like lie (What was John lying in bed reading?, taken from Truswell [2007: 1360]). This verb belongs to the class that Maienborn (2003) calls Davidsonian states: they share properties with states and activities, but they are definitely not achievements. We address this difference in Section 6.3, where we argue that the contrast is due to a difference in the internal structure of the gerund in the two languages.

2.2 Restriction II: Adjacency

The possibility of extraction in Spanish tends to be dependent on the immediate adjacency of the main verb to the gerund. Contrast the previous sentences with those in (13). Without adjacency, these forms are degraded.

The results of the test confirm this. (6) above contrasts with (13a), minimally, in the position of the subject. If the results for (6) were high, (13a) received much lower grades, with a significant dispersion, suggesting that the sentence is also degraded for speakers (mean=5.31, median=5, sd=3.47, vc=0.65). Again, a Persons’ chi-square shows a distribution that is unlikely to be due to chance (chi-squared=16.53, df=1, p-value < 4.787e-05), although with a moderate effect size (V=0.22). The following graphical exploration shows the differences (Figure 2).

Figure 2:

Grammaticality of extractions with a gerund adjacent to main verb vs. an intervening subject.

Here we find, again, a contrast with English, where adjacency is not required (cf. What did John arrive home whistling?). Moreover, in English, a transparent gerund can receive another reading: a causative interpretation, noted in Truswell (2007), as in What_i did John enrage his neighbors [whistling t_i]? This construction is different from the one that we are studying in this section in three respects: (a) first, it is not accepted by most speakers that we have interviewed in English, and its Spanish equivalent is radically ungrammatical; (b) second, the gerund is not predicated of an internal argument: the person that whistles is the person that enrages the neighbors. We will ultimately derive these language-particular differences from the internal structure of gerunds in English and Spanish, cf. Sections 4.3 and 6.3.

2.3 Restriction III: Absence of resultatives

In the grammatical sentences in Spanish, the prepositional complement expressing the result location of movement does not appear in the sentence with the extraction out of the gerund, as in (14a). This cannot be due only to adjacency. Even when the goal phrase can be separated from the main predicate, the result is worse than the version where the goal is implicit, as illustrated respectively in (14b) and (14c).

Again, the statistical analysis confirms this. A sentence like (14b), vs. (14c) – identical to (6) – got lower grades (mean=2.79, median=1, sd=2.69, vc=096). A chi-square comparison of the same sentences with and without an overt goal gives x-squared=87.48, df=1, p-value < 2.2e-16, with V=0.52, that is, a highly significant difference with a considerable effect. Note that there are, however, some outliers (Figure 3).

Figure 3:

Grammaticality of extractions without an overt result vs. with an overt result.

This contrasts again with English, where an overt result can appear (What did John come home whistling?, taken from Truswell [2007: 1366]). In Section 6.2 we suggest that this difference can be related to the availability of strong resultative phrases in English, but not in Spanish. ^[6]

2.4 Restriction IV: The secondary predicate must refer to an internal argument

A final condition that has been identified has to do with the argument structure of the main predicate. This property is uncontroversial in Spanish, as reported by Demonte (1987/1988), and in English according to Truswell (2007): only if the main predicate has an internal argument is the extraction possible. Contrast in this respect (15a), containing an unergative verb, with (15b), involving an unaccusative.

However, one could blame the contrast on the different aspectual nature of each predicate. We have to concentrate, therefore, on transitive achievement verbs to be sure that this property holds. In (16) the extraction is accepted and the gerund clearly takes the internal argument as its subject.

Contrast this with (17), where the gerund has to take the external argument as its subject and the extraction is impossible. All speakers that were consulted for this in the preliminary phase of this investigation has sharp judgements about this, and the question is uncontroversial in the literature, so we did not include this contrast in the questionnaire. ^[7]

2.5 Interim summary

Let us summarize the conditions under which extraction from inside a gerund is possible in Spanish:

1. The gerund secondary predicate tends to be immediately adjacent to the main predicate when extraction is possible.

2. For extraction to be possible, the main predicate must have an internal argument, which is the subject of the gerund.

3. The main verb must be an achievement.

4. Constituents expressing the result location must be absent.

Equally relevant are the differences between the results that we have obtained here and the facts reported by Truswell (2007) for English (as opposed to Spanish):

1. English allows transparent gerunds which denote the cause of the main verb; these gerunds take the main verb’s external argument as subject.

2. In English, the main verb can be a Davidsonian state like lie.

3. In English, the constituent expressing the result location can be present.

4.1 The structure of Aktionsart in a first phase syntax

The first piece of the puzzle is to spell out our assumptions about the decomposition of Aktionsart inside a constructionist approach to lexical aspect. Here we will basically follow Ramchand (2008: 38–56) in her proposal that the three main subevents independently identified in the tradition (Dowty 1977, 1979; Pustejovsky 1991, among others) correspond to distinct syntactic heads. Ramchand’s proposal is that there are three event-related heads. They are:

1. Initiation: the head that codifies the cause component inside an eventuality. A predicate that has an argument expressing an initiator –cover term used by Ramchand to group agents and causers – contains this projection; that is, John broke the glass has it, but The glass broke does not have it.

2. Process: the head that provides the eventuality with the event part, denoting a dynamic part. Verbs that denote events have it; that is, John learnt English has it, but John knows English does not have it.

3. Result: the head that expresses the subsequent state after a telic event has arrived to its endpoint. This head is only present when the verb must compulsorily be telic and is able to license result interpretations of prepositions and other items that do not express change in themselves. That is, in The glass broke into one thousand pieces there is a ResP that licenses the reading of into one thousand pieces as the subsequent state attained after the breaking event; in contrast, there is no ResP in John ate the ham.

It is important to note that a fourth head can be involved in the structure. Process can take a Rheme argument as its complement (25), which in turn could take a ResP as its complement (for instance, when Rheme is interpreted as Path, as in John ran into the water).

This rheme introduces an entity which defines part of the properties of the event denoted by Proc. Rhemes do not introduce individuated, referential arguments that are predicated over, but entities whose internal topological properties are identified with the event described by the main verb; rhemes are descriptors or modifiers of the main event that must unify their properties with the internal topological properties of the event (Ramchand 2008: 46).

Paths are the most studied instantiations of rheme arguments, but not the only ones. A path is an entity with internal mereological structure, bounded or unbounded. When combined with an event, the internal mereology of the path identifies one-to-one with the internal parts of the event, codescribing it. Although gerunds are not paths proper, they also have an internal mereology; our contention in this section and the following is that in the cases where the gerund allows for subextraction, the gerund occupies a rheme position: it codescribes the event, and identifies its internal mereology – the unbounded process it denotes – with the event of the main verb, giving rise to an interpretation of simultaneity (Section 4.3).

A natural consequence of the idea that rhemes act as modifiers of the event expressed by Proc is to propose that there is a division of labor between Proc and Rheme. Proc introduces the event argument <e>, but Proc, by itself, does not fully define the event’s internal mereological properties. Rhemes are syntactic positions that, given a head-complement configuration, transfer further descriptive properties to the event through identification with their internal mereological structure. Consider the case of paths: (26a) denotes an event because it contains ProcP, but we interpret that this event occupies an extended time period (specifically, movement across some space that takes some time to take place) because Proc takes a PathP as a complement (26b), which gives extension to the event. This is possible, to begin with, because the path (through the forest) has an internal composition that is itself extended; mapping the event to that extended structure forces a reading of the event where it denotes an extended period of time itself. This is independent of whether PathP is spelled out overtly or not; that is, we assume that in any case where an event has internal temporal extension there is structurally a PathP (an instance of Rheme), independently of whether it is spelled out or not. ^[9]

The direct consequence of this idea is that some Aktionsart classes will necessarily project Rheme in order to define the internal development of the event. The presence of Rheme is going to be compulsory if one must have activities – which emerge when the Rheme defines a path which does not have any natural boundary, as in (26b) – or accomplishments – if the Rheme is a path that has a natural boundary or is itself followed by ResP. This is what happens in (27), where the transition arrives to some specific point once it is concluded; see also Krifka (1989), Verkuyl (1993), Rothstein (1999) and Mateu (2002), among many others, for how (un)boundedness in other domains is transferred to verbal aspect.

In contrast, other Aktionsart classes do not occupy through the main predicate the Rheme position, which will then be available for a gerund structure to fill it.

Rhemes, of course, do not need to be defined in the spatial domain. It is accepted at least since Tenny (1986) and Krifka (1987) that individuals denoted by noun phrases with different quantifiers and determiners can also be used to define the internal mereological properties of an event (to eat an apple; to eat apples), and the scale structure of adjectives can also be used to measure a change through the different degrees below and above the standard value (to whiten, to become red, etc.). This notion of generalized rheme explains that verbs other than movement verbs can denote temporally extended transitions, provided that their Proc takes as its complement a noun (28a) or an adjective (28b) with properties able to identify with an event.

A structure like (29a) is the maximal syntactic instantiation of an event (for instance, to derive the accomplishment in [29b]).

Note that (29a) contains a causation component, a process of running which happens along the path that starts in the original position of the subject and ends in the position of the water, and finishes with John in the water. Rheme and Res do not need to co-occur: an achievement verb like enter (John entered the room) allows a construction [Init [Proc [Res]]], that is, where there is no extended path leading to the inside of the room, as the relevant point is only the moment in which John crosses the door of the room. However, Rheme and Res co-occur when the process leading to the result location happens along an extended path.

Structures like (30), where the spine of the tree contains two identical heads in strict adjacency are impossible:

These configurations are impossible due to vacuous projection: at some point in the structure we would have a head that selects itself; even assuming that its selectional requisites would be somehow satisfied in that situation, the result would be that two contiguous levels of the structure would have identical properties. Any property satisfied by the second head would be redundant with the satisfaction of the same properties by the first head. The second reason is that such repetition would not allow for selection. In (30b), for instance, the lower instance of Proc could not be selected by Init, making it unclear which syntactic property of Init it is satisfying.

What could be possible, because it avoids vacuous projection and allows for univocal selection, is a recursive structure like (31).

Init would combine with the first (leftmost) Proc as its causation component, and Proc with Res as its result, but Res would not be able to identify with the second Proc because it has already been integrated in a structure with the first Proc, and the same result cannot be the consequence of two distinct processes. Thus, the second Proc would integrate with the second Res, and the result would be two distinct complex event structures.

4.2 Achievements do not spell out Rheme

The next piece in the puzzle is to show that achievements – which are the only verbs that allow extraction from a gerund clause – do not fill the Rheme position available inside the verbal structure, and as a result they do not denote changes that occupy a temporal extension co-described through the mereological properties of another entity. Two pieces of evidence, originally due to Piñón (1997), support the idea that achievements lack internal duration. Both accomplishments and achievements allow in-phrases setting a limit.

It seems, however, that these two modifiers behave differently. In (32a), with an accomplishment, the modifier measures the duration of the event, which is possible because the event has some internal duration – because it contains path, we claim, here manifested as the cumulative DP la carta ‘the letter’. In contrast, in (32b), as the event lacks any internal duration, the same modifier measures the time that elapsed between some arbitrarily set point in time, previous to the beginning of the event, and the single point in time where the summit is reached. Consequently, in (32b) the meaning is identical to an after-modifier, which can substitute it without change in meaning, but in (32a) the substitution means something else.

This follows if (32a) does not have any internal duration, and therefore, if the event does not contain internal topological properties that need to be identified.

A second piece of evidence is the meaning of the progressive form. With verbs like (32a), here (34a), since there is internal duration, the progressive denotes any point during the progression of the event; in (32b), here (34b), the verb does not have an internal progression that can be denoted and the interpretation that obtains is a preparatory stage reading (cf. also Marín and McNally 2011), where we denote any point in a period defined between an arbitrary point previous to the event and the event itself. As a consequence of it, (34b) denotes the same as (35b), with an inceptive periphrasis, but (34a) is distinct in meaning from (35a).

In Ramchand’s theory, this means that with achievements Proc does not spell out RhemeP; consequently, the internal argument of an achievement will not co-describe the event. Achievements are punctual transitions, and as such they do not contain internal topological properties that require identification with another entity. As can be seen in (36), an achievement combines directly with a Result state as the argument of Proc; the result of this is that, instead of identifying the internal topological properties of the event, the process is interpreted as lacking internal structure and leading immediately into a result state: hence the punctuality.

Now, if achievements do not fill the Rheme position (advancing a bit of Section 5.1), what this means is that an achievement defines a non-maximal event space where there is still room for a RhemeP projection. If transparent gerunds can be projected as RhemePs, a gerund structure would be able to integrate in the same structural space as an achievement finite verb, yielding (37) as a result. With achievements, then, there is a possibility of merging gerund and finite verb in the same structure, which makes extraction possible.

In contrast, with activities or accomplishments, the main verb already fills the RhemeP position, because a path is required to define the internal development of the event. Consequently, trying to integrate a gerund in the same predicate structure in Spanish would produce a structure like (38), which is ungrammatical because two rhemes are projected: a path and a gerund.

Since (38) is impossible, when the main verb already requires a Rheme, the gerund has to be introduced as an adjunct, which makes – as expected – extraction impossible.

The next step is, thus, to show that gerunds are plausible materializations of RhemePs.

4.3 Gerunds as RhemePs

In this section we argue that there are reasons to support the idea that in the Spanish transparent gerund constructions studied here, gerunds are introduced as RhemePs. Take a structure like (39). where the gerund fills the Rheme position (Ger is for the time being a descriptive label; we will be more specific about its internal structure in Section 4.4).

If our proposal is right, we would expect the situation denoted by the gerund to co-describe the event. However, given that the achievement verb does not select for the gerund (in contrast with, say, an aspectual auxiliary), we would not expect there to be event identification between the gerund and the main verb. As one anonymous reviewer points out to us, it is possible to say John carefully cleaned the table sloppily singing the Marseillaise without contradiction, which means that the two actions are not necessarily identified in the interpretation component. ^[10]

But first we need to address the question of what kind of internal structure a gerund has. It is relatively uncontroversial that gerunds are the designated form in the verbal paradigm to express the ongoing event denoted by the verbal predicate. That is, in informal terms, gerunds express a series of points that correspond to the process denoted by the predicate they select, excluding in principle the possible initial and final boundaries of that process. Several pieces of evidence have been provided in favor of this idea. Consider (40), which is a contrast first noticed – to the best of our knowledge – in Vendler (1967).

With a perception verb that has a temporal duration, like see, both the infinitive and the gerund are possible as complements expressing the perceived event, but with a subtle difference in meaning. When the infinitive is used, it is strongly implied that John witnessed the whole series of actions that Mary performs in order to steal the jewel, culminating with the moment in which Mary puts it in her pocket. On the other hand, when the gerund is used, the sentence just means that John saw Mary at some point during the ongoing action of stealing, but it does not follow that he actually saw her putting the jewel in her pocket, or begin to open the safe. The idea is that the infinitive codifies the whole sequence of subevents, while the gerund focuses on the process subevent (which might eventually conclude with Mary getting to possess the jewel, or not).

Given this reasoning, the contrast in (40b) receives a natural explanation. As spot is an instantaneous perception verb, without temporal extension, it is not possible to combine it with the infinitive, because the latter denotes a sequence of event phases and it is impossible to perceive that sequence in an instant. In contrast, the gerund is possible because it denotes an ongoing process and any instant in the series of temporal points defined by the gerund denotes a part of the process.

What does this mean for the internal structure of gerunds? We propose that gerunds are projections of an aspectual head with imperfective value. This head selects the event denoted by the base verb and focuses on a time interval that excludes the initial and (when available) the culmination point of the event.

Thus, the temporal trace of the gerund is an interval (represented by <... +++... >) describing the progression part of an event, excluding its initial and final boundary (Dowty 1977; see also Bolinger 1971; Borer 2012). This is consistent with a number of existing proposals about the semantics of the progressive form: Vlach (1981) analyses the truth conditions of the progressive as true in world w at time i if in w at i there is an ongoing process which would eventually attain a culmination. However, the culmination is not included in the denotation, and the truth value is independent of whether the event is interrupted before the culmination or not: Landman (1992) proposes that the progressive is true in w if in some world on the continuation branch of w (cf. Dowty’s 1979 concept of inertia worlds) some event realizes the event type of the predicate. What is crucial for our purposes is that the denotation of the gerund does not include an initial or a final point in the event’s internal structure.

Given this situation, consider a sentence like (42):

The interpretation of this sentence is that, at a particular point in time, two events have co-occurred in the same situation: Juan’s entrance and Juan’s singing the Marseillaise. This is precisely what is expected if the gerund acts as a rheme inside the event’s internal structure. The situation expressed by the gerund consists of a series of extended points, while the main predicate – lacking a path – denotes a punctual transition. This punctual transition defines a point in time, call it t^e. Becoming one single verbal structure, we propose, involves identifying this t^e with one of the points inside the interval denoted by the gerund (tg):

The result is that the truth conditions only require that at the precise moment of entrance, Juan was singing; nothing is entailed about whether Juan was singing before or after that event. Notice that this procedure, in fact, is more successful than Truswell’s at accounting for the semantic interpretation of the gerund. In Truswell’s analysis, as we noted earlier, the relation between the gerund and the main verb was one of immediate temporal precedence: the gerund described a preparatory phase of the main event. However, it is not clear what ‘preparatory phase’ means in this context: whistling is difficult to categorize as a preparation for the arriving event. Our analysis, in contrast, precisely predicts that the entailment will be only that the subject was whistling at the point of arriving.

We propose the specific structure in (44) for a Spanish gerund introduced as a Rheme: the gerund contains verbal structure, but is introduced as Rheme via an additional head. This head, following among others Fábregas (2008) and Gallego (2010b), is a relational head, and more specifically a preposition. Thus, as other Rhemes, like Path Phrases, the gerund projects as a PP. That preposition introduces the subject of the gerund in its specifier (cf. Bowers 2000), and defines the aspectual value of the gerund as imperfective (see Hale 1986 and Hale and Keyser 2002 for the view of prepositions as aspectual-denoting heads). ^[11]

Unlike Gallego (2010b), we propose here that the gerund’s preposition in Spanish is a Path preposition, not a Central Coincidence Preposition. The difference between our proposal and Gallego’s might be, to some extent, purely terminological: Gallego’s proposal is couched in a theory where imperfectivity has to be associated with Central Coincidence relations (cf. Hale 1986); the opposite kind of P, Terminal Coincidence, defines a change of state and triggers telicity, which is clearly the wrong value for a gerund. In this system there is no, per se, Path preposition as denoting an extended sequence of points (but see Koopman [1997]; den Dikken [2003]; Ramchand [2008] and Svenonius [2010] for a different view of this kind of element).

While all transparent gerunds introduced as Rhemes are projections of this PathP in our proposal, not all gerunds are the same, and we in fact would like to propose that most of the empirical differences between Spanish and English derive from differences in the internal structure of gerunds in each language. This is the topic of the next section.

4.4 Gerunds in Spanish and English

It is well known that English gerunds can be used as nominal arguments; in contrast, Spanish gerunds cannot (Borgonovo 1994). Borgonovo (1996) has argued that gerunds with perception verbs are never arguments. Consider the two pairs of sentences in (45) and (46).

Borgonovo’s explanation is the following: the infinitive in (45a) and (46a) heads a constituent that is taken as an argument by the perception verb. In contrast, the gerund in (45b) is an adjunct; in (45b), the argument of the perception verb is the DP María. The reason for the ungrammaticality of (46b) is that the perception verb does not have an internal argument: the gerund is an adjunct, and its subject, being non-referential and expletive, cannot be used by ver ‘see’ as its internal argument (see also Di Tullio 1998).

This difference is also confirmed by a contrast like (47):

As one anonymous reviewer points to us, this suggests that English gerunds are themselves nominal constituents (although the label is orthogonal to our purposes, NPs), whereas Spanish gerunds are always PPs, as Fábregas (2008) and Gallego (2010b) have argued, and specifically, in our approach, PathPs. The proposal that follows was suggested by the anonymous reviewer, to whom we are indebted.

What this means for our purposes is that, when projected as Rheme, the English gerund is the NP complement of a phonologically empty Path preposition (48a), while in Spanish the gerund morphology itself spells out that PathP (48b):

In other words, this means that a Spanish gerund will contain PathP in any context, while the English gerund is expected to have a more flexible syntax: as NP it will be able to combine as an argument with a verb, and also to combine with different sets of (empty) prepositions. We will see how this difference explains the empirical contrasts between English and Spanish with respect to transparent gerunds in Section 6.3, but for the time being let us show one additional piece of evidence that PathP is hardwired in the Spanish gerund, but not in the English one. Consider (49).

The interpretation of (49a) is the expected one if a Spanish gerund is always a projection of PathP. It is interpreted as a Creswellian-location (Creswell 1978): it introduces a place which is reached after traveling through a path, which is the standard interpretation of Path prepositions embedded under locative verbs, as in (50).

In contrast, if the English gerund does not encode a PathP, the ungrammaticality of (49b) is expected. In Section 6.3 we will see how these differences account for the distinct properties of Spanish and English transparent gerunds.

5.1 Extraction is only possible when the main verb does not spell out RhemeP

Recall that extraction from the gerund construction is not possible in Spanish when the main predicate is an accomplishment or an activity. What these two classes have in common, in terms of the syntactic decomposition of the verbal constituent, is that they fill RhemeP with a path, which gives their events internal development and therefore identifies their temporal extension. InitP is present only if the event is causative, and ResP only if it culminates in a result state. In such cases, the RhemeP position the gerund would occupy is already filled.

Conversely, extraction is possible if the event is an achievement, which, as we have seen, lack internal duration. In such cases RhemeP is filled in Spanish by the gerund construction.

Gerund constructions are projected as RhemeP. Given this, a gerund can combine with an achievement, as in (53), in one single verbal structure, because the main verb does not project Rheme.

In the case of accomplishments and activities, the integration is not possible, because two RhemeP would have to be projected. Either we get a tree like the one in (54), which is illicit because the same label is duplicated and the event would only unify with the closest Rheme, or two distinct verbal constituents have to be defined. In either case, the gerund cannot integrate with the finite verb in one structure, and extraction of an argument of the secondary predicate is expected to be ungrammatical.

Beyond the general ban on having a head selecting itself, one distinct problem in this configuration would be that the rheme’s nature as a co-describer of the event could not be performed by the lower projection of RhemeP, given that it is not the complement of the head Proc.

This accommodates, in our syntactic analysis, Truswell’s (2007) generalizations about why Aktionsart delimits the acceptable and unacceptable extractions: as his analysis made clear, the two verbs must be part of the same event structure inside the same syntactic expansion of the verbal phrase. Here we capture the restriction structurally. There is, therefore, no need to postulate an operation of semantic reanalysis that acts independently of syntax. The two predicates, as far as syntax is concerned, are only one verb, because the maximal expansion of the verbal domain is filled by both forms at the same time. We can, thus, propose the informal principle in (55), which explains what reanalysis tried to explain in previous accounts:

5.2 Stative verbs do not license gerunds

Remember that pure stative verbs do not license gerund clauses, making it (vacuously) impossible to extract arguments from them. Of course, the fact that pure stative verbs reject depictive predicates – as well as other modifiers and adjuncts – has been repeatedly noted in the literature (e. g., Demonte and Masullo [1999: 2475–2477]), but our analysis can suggest an explanation to this restriction. ^[12]

Pure stative verbs, like hate, know or exist, lack an event argument; this means that they lack ProcP. If the gerund construction – perhaps, more in general, all depictive secondary predicates – is introduced as RhemeP, and the role of Rheme is to co-describe the situation expressed by the main verb, then we expect gerunds not to be available as Rhemes of stative verbs. In other words, (56) is expected to be ungrammatical.

Why would this be so? We have seen that gerunds (Section 4.3) denote extended time periods. Being in a rheme position, this would force the main predicate to be co-described by the gerund, identifying one-to-one the internal parts of the event’s development with at least a subset of the gerund’s internal properties. However, the stative verb does not contain an event, and therefore there is no dynamic part that can be mapped into the gerund’s extended time period. The homomorphism that needs to take place between the state and its rheme would not be attained. This could explain, in general, why stative verbs reject co-descriptors, gerunds or other kinds of secondary predicates, as in (57): if these elements are projected as rhemes, and given their internal properties they have an extended temporal structure, they are incompatible with states because they do not share a mereological structure that can be unified. However, we will not elaborate on this any further and leave it for future investigation.

In order to better understand the nature of the restriction, consider what kinds of rhemes a stative verb can combine with. Ramchand (2008: 55–56) observes that in a structure like (58), the rheme is a complement to Init, a stative projection, not Proc, an event projection. In a situation like this, any element that co-describes the eventuality (a state) will have to denote a location of sorts, and its internal mereological structure, if any, would need to be ignored, because there is no internal development or punctual change that can be identified with that extended structure.

The crucial problem with a gerund – perhaps also with an adjectival predicate – in this context is that their internal extended structure cannot be ignored: they denote precisely situations that hold of entities for non-trivial time intervals. And, in any case, there is no event to identify that extended temporal structure with, so the result is that they just cannot be introduced as rhemes of states.

6.1 Extraction is only possible if the gerund modifies the internal argument

A property noted in Demonte’s (1987/1988) analysis of adjectival secondary predicates that allow extraction and noted but not analyzed in Truswell (2007) is that the extraction is possible in Spanish if the secondary predicate modifies the internal argument of the main verb. ^[13]

We show now that this restriction naturally follows from the structure proposed. Consider, first, what it means to be an internal argument. In a system like Ramchand’s (2008), one and the same argument can occupy more than one theta-position through movement (see also Boeckx et al. 2010 for evidence in favor of movement from one theta-assigning position to another). In her analysis, the only argument of a verb like arrive occupies the three specifier positions in (59). The argument is generated in the lowest available position of the structure and moves up to fill the other positions, where it gets new entailments about its involvement in the event: an entity that triggers some displacement, is displaced itself and ends up in a result location.

In this system, being an internal argument roughly means this: being in spec,ProcP (independently of whether the same element eventually moves to spec,InitP – as in arrive – or a different element is introduced in that position); being an external argument, in contrast, is being base-generated in InitP, without previous merge in spec,ProcP.

Given this, the reason why gerund structures must take the internal argument as subject follows from an intervention effect. As RhemeP is lower than ProcP, the specifier of RhemeP can become the specifier of ProcP by movement. From there it can further move to spec,InitP or not, as in (60).

What is impossible, however, is that spec,RhemeP moves directly to spec,InitP without landing first in spec,ProcP. If ProcP and RhemeP have different specifiers, then the one in spec,ProcP will land in InitP, since that movement is more local. Otherwise, intervention yields the ungrammatical structure in (61):

Consequently, we derive the restriction that gerunds only allow extraction if they are predicated of internal arguments. ^[14] If the extraction happens, the gerund construction must be integrated in the same syntactic structure as the main verb. As the gerund construction projects Rheme, ProcP will always be closer to Init than the gerund. Thus, if the specifiers of Rheme and Proc are different, the one belonging to Proc will always move to spec,InitP.

For the same configurational reason, the adjacency between the main verb and the gerund that is necessary when the gerund allows for subextraction in Spanish also follows trivially: they are the two members inside the same structure, so, by virtue of the restricted syntactic space they share, they are expected to be adjacent to each other. We just need to assume that no movement operations take place in such a way that RhemeP moves above InitP, for instance – and as far as we know, nobody has made such proposal.

6.2 Absence of overt result phrases

As the reader will have certainly noticed by now, there is no linear adjacency between the three heads required by a predicate like arrive. RhemeP, lexicalized by the gerund, is projected between ProcP and ResP (cf. for instance 59). In a Ramchandian system, when a lexical item lexicalizes a series of heads, the procedure used in order to lexicalize several heads with a single exponent is spanning, which is defined as in (62).

Obviously, this procedure cannot be applied in our case to lexicalize achievements, because Proc does not take ResP as complement if RhemeP is in between. If we try using a different procedure to capture the synthetic expression of a series of heads – such as standard head movement, morphological fusion (Halle and Marantz 1993) or phrasal spell out (Caha 2009) – the same problem arises. The situation is illustrated in (63).

In other words: the exponent arrive, which is generally associated also with a ResP, will not be able to lexically spell out that head in the syntactic configuration where a gerund is in the rheme position.

This situation, in fact, is consistent with our data. If arrive, in the configuration, cannot identify the ResP layer, then we have an explanation for why, in structures where the gerund is integrated with the main verb, to overtly express as a syntactic constituent the result location is ungrammatical (cf. the contrast between (64a) and (64b)).

Although the analysis of this particular fact is still unclear to us, we would like to suggest an explanation that connects it precisely with the spell-out procedure. The idea is that, in fact, ResP is not projected in the structure in cases where Rheme is projected as a gerund with an achievement verb. The impossibility of overtly expressing a result is explained because the projection that licenses it, ResP, is not projected in the structure.

One argument in favor of this account is that English does allow overt result locations in our contexts (What did John arrive home whistling?). Note, crucially, that English is also able to license strong result adjectival phrases, something that is impossible in Spanish.

Ramchand (2008) argues that this contrast is due to English having a phonologically-zero Res head, while in Spanish Res has to be spelled out as part of the verb exponent. Thus, in both languages the gerund would interrupt the relation between Proc and Res, making it impossible for the main verb to spell out Res; but in English, Res would be projected as an independently available zero head, and an overt result phrase can be introduced. In Spanish, since such head is not available, overt result phrases would be unavailable because the gerund prevents the main verb from spelling the relevant head out.

An independent question, raised by two anonymous reviewers, is why if ResP is not projected in Spanish we still associate with the main verb the entailments that the process must have culminated and has a result component, whereby the subject ends in the intended result location. Here we can only speculate. One option is to treat this entailment as an effect of the main verb’s lexical entry, that is, as part of the lexical information that the verb exponent brings with it when it is introduced in the structure. This could in principle resolve the tension between the presence of a semantic entailment and the impossibility of projecting in the syntax the constituent that would express overtly that entailment. However, it would bring up questions about the relation between the truth conditions of a lexical item and its syntactic licensing conditions: it would seem that an item can be introduced in a configuration where not all its semantic ingredients have been structurally expressed. We are not ready to present a theory of the mapping between syntax and semantics where this would be a natural result; we would be forced to admit a relatively arbitrary mismatch between the information of a lexical entry and the syntactic structure where it is introduced.

Another alternative would be to propose a slightly different structure for our gerund constructions, where the Res head is introduced as an adjunct to the Proc head, and RhemeP is the only complement of the Proc head ([InitP Init [ProcP [Proc+Res] [RhemeP]]]); sharing one single projection, Res would be unable to project its own complement, but the result component would be expressed in the structure.

For the time being, though, we consider that this aspect of the analysis is not settled and we hope that further research on the licensing conditions of result secondary predicates in Spanish will be able to throw light on this problem.

6.3 The differences between English and Spanish

It is now time to compare Spanish with English. There are noticeable differences between English and Spanish when it comes to these structures:

1. With gerunds subordinate to perception verbs, English accepts, but Spanish rejects, the extraction:

1. In causative constructions where there is no adjacency between the gerund and the main verb and the subject of the transparent gerund is the main verb’s external argument, English accepts, but Spanish rejects, the extraction (cf. 68, previously 18)

Following a suggestion by one anonymous reviewer, we have proposed that English gerunds are NPs that, when projecting as Rhemes, combine with an empty Path preposition; in contrast, Spanish gerunds are always headed by a Path prepositions. This has two automatic consequences: the first is that English gerunds, as NPs, will be able to be selected as arguments by verbs, as in (67), where the extraction is expected. Spanish gerunds, not being nominal, will not be able to combine with the verb as arguments, so the equivalent of (67) is correctly expected to be ungrammatical.

Second, as the anonymous reviewer points out, this difference can be behind the rejection in Spanish of transparent gerunds denoting the causing subevent, as in (68). Assume that in the same way that the gerund has to be headed by a Path preposition to project as Rheme, in order to project as the causing subevent and integrate with the main verb, it has to be headed by a Source preposition – treating causation as the initial boundary of the event, or in other words the starting point where it is initiated. As the English gerund is NP, it can combine with an (empty) Source preposition and project as the causing subevent. Due to locality, its closest argument will be the verb’s external argument, and being at a higher position, the material in the specifier of the main verb will be able to break the adjacency between the transparent gerund and the main verb.

In contrast, in Spanish, the gerund is a PathP; given that it is already headed by a preposition, we expect correctly that another preposition will not be able to select it, as prepositions take nominal complements. Hence, the Spanish gerund will not be able to project as a Source preposition inside the spine of the tree, and it will not be transparent.

But as another anonymous reviewer points out to us, treating English gerunds as nominal arguments does not straightforwardly account for all contrasts. Consider (69), from Truswell (2007: 1360), where the main verb is what Maienborn (2003) describes as a Davidsonian state (a verb with some stative properties that, however, accepts place and manner modification): the gerund combined with such verbs is not transparent in Spanish, but it is in English:

The problem is that the internal structure of D-states is unclear; if it is more complex than the one of pure stative verbs, then we could argue that the gerund is not projecting as Rheme here, but introduced as an argument of one of the heads forming the internal structure of D-states. This would explain why Spanish rejects the equivalent sentence. However, lacking a proposal for the structure of D-states, we leave the matter open for further research.