Abstract

In English, adverbials may intervene between the verb and a selected PP. We consider three analyses of this fact: the traditional account, that the PP shifts rightward across a right-adjoined adverbial (Stowell 1981); an alternative account, that the verb moves leftward across a left-adjoined adverbial (Pesetsky 1989, Johnson 1991); and a hybrid account that assumes both extraposition and verb raising. We argue that the order of postverbal adverbials favors the extraposition analysis, provided this analysis is combined with the hypothesis that certain adverbials can directly modify other adverbials (Rohrbacher 1994, Williams 2014). We then compare two instantiations of the extraposition analysis: the traditional account and an antisymmetric account that emulates PP-extraposition through a combination of PP-intraposition and roll-up movement. While close to being notational variants, these accounts can be teased apart using the very strict locality requirement that holds of interaction with temporal only. The data then show that the symmetric account has the edge. Finally, we briefly discuss the implications of our findings for the analysis of the English VP, with a focus on the circumstances under which the verb moves.

1 Introduction

In English, PP complements may be separated from the verb by adverbials, as (1) shows.

(1)

  • Susan looked at the telegram pensively.

  • Susan looked pensively at the telegram.

The aim of this article is to determine the source (or sources) of this word order variation. Although the problem appears simple, there is a bewildering array of potential analyses to choose from, depending on one’s view of the syntax of adverbials, verbs, and PPs, and on the framework the analysis is couched in.

Our strategy for navigating this maze is as follows. The bulk of the article (sections 27) will be devoted to a comparison of three analyses rooted in a traditional symmetric theory of phrase structure (that is, a theory that allows variation in the order of sister nodes). The first analysis assumes extraposition of the PP, the second assumes movement of the verb, and the third is a dual-source analysis that assumes both extraposition of the PP and movement of the verb. We argue for the extraposition analysis on the basis of structures containing two adverbials. We show, contra Pesetsky 1989 and in line with Rohrbacher 1994, that regardless of the position of the PP the lower adverbial precedes the higher adverbial (as in (2)). The extraposition analysis provides the simplest account for this pattern.

(2) V

graphic
PP
graphic
AdvLow
graphic
PP
graphic
AdvHigh
graphic
PP
graphic

We then outline, in section 8, an antisymmetric variant of the extraposition analysis. While the traditional symmetric theory of phrase structure allows a straightforward account of the order of adverbials in (2), the antisymmetric account must rely on an additional mechanism that takes as input a structure in which adverbials occupy designated specifier positions to the left of VP and delivers as output a structure in which the base order of VP and adverbials is reversed. The movement regime responsible for this order reversal is known as roll-up movement (see Barbiers 1995, Koopman and Szabolcsi 2000, Cinque 2005, 2010). Once in place, roll-up movement permits a straightforward antisymmetric translation of the extraposition analysis, delivering an account of (2) that appears descriptively adequate.

To force a choice between the symmetric and antisymmetric analyses of PP-extraposition, we explore the syntax of temporal only (as in John promised to leave on Monday, but he onlyT left on Tuesday). We show that on the relevant interpretation of only, this particle must immediately c-command the temporal expression it interacts with (on Tuesday in the above example). In this regard, temporal only differs from exclusive only, which permits association with a focused constituent across larger distances. The hyperlocality of interaction with temporal only allows us to construct an argument against the antisymmetric analysis of PP-extraposition, leaving its symmetric counterpart as the only viable account of PP placement in English.

In section 9, we explore the implications of the extraposition analysis, given the tension between our conclusions and the evidence for verb raising in Larson 1988 and subsequent work.

2 An Overview of Symmetric Accounts of Adverbial Intervention

The traditional account of the alternation in (1), adopted explicitly in Stowell 1981, assumes that the position of the PP varies, either as a result of rightward movement (as in (3b)) or through base-generation of the PP above the adverbial (as in (3c)). We will call this account the extraposition analysis.

(3)

graphic

An alternative account is to keep the position of the PP constant and to attribute the alternation in (1) to two factors: verb raising and variation in the linearization of the adverb (see (4)). We will call this account the verb-raising analysis.

(4)

graphic

A third option is to assume that there are two sources for the order in (1b): extraposition of the PP, as in (3b–c), and raising of the verb, as in (4b). We will call analyses of this type mixed analyses.

The general idea that verb movement may be responsible for the intervention of adverbials between a verb and its complement goes back to proposals by Emonds (1976) and Pollock (1989). Its extension to examples like (1b) is due to Johnson (1991) and to unpublished but influential work by Pesetsky (1989).

The main criterion we use to decide between the various analyses is adverbial order in three conditions, as schematized in (5). In the PP-initial condition, both adverbials follow the PP. In the PP-medial condition, one adverbial precedes the PP and the other follows it. In the PP-final condition, both adverbials precede the PP. For each of these conditions, one may ask whether the higher adverbial (AdvH) precedes or follows the lower adverbial (AdvL).

(5)

graphic

As already pointed out in Pesetsky 1989, the extraposition analysis predicts that if more than one adverbial intervenes between the verb and the PP, the lower adverbial precedes the higher one (see (6a)). By contrast, the verb-raising analysis predicts that in the PP-final condition the lower adverbial follows the higher one (see (6b)).

(6)

graphic

The predictions of the extraposition and verb-raising analyses also diverge in the PP-medial condition, where the extraposition analysis predicts that the lower adverbial precedes the higher adverbial, while the verb-raising analysis predicts variable order. (The verb-raising analysis must allow variation in the linearization of relevant adverbials, as any adverbial that can precede the PP can also follow it.) Finally, the extraposition and verb-raising analyses both predict that in the PP-initial condition the lower adverbial will precede the higher adverbial.

There are various mixed analyses that differ in the height of the landing sites for verb raising and extraposition. A first option is that both movements can cross AdvH, as in (7). We will call an account along these lines the equal-height analysis.

(7)

graphic

The equal-height analysis predicts that in the PP-initial condition the only order that can be generated is AdvL-AdvH. However, in the PP-medial and PP-final conditions both AdvH-AdvL and AdvL-AdvH are permissible orders, depending on whether the PP moves and whether the adverbs are left- or right-adjoined.

It is, of course, not a logical necessity that the verb and the PP move equally high. If the verb raises above AdvH, while the PP’s highest landing site is located below AdvH, the expected word order patterns shift (see (8a)). This low-PP analysis predicts that in the PP-initial condition the lower adverbial precedes the higher one, that in the PP-medial condition the adverbials come in variable order, and that in the PP-final condition the higher adverbial precedes the lower one. Conversely, it is possible that verb raising lands in a position below AdvH, while the PP has access to a position above AdvH (see (8b)). We refer to this account as the low-V analysis. Like the extraposition analysis, it predicts AdvL-AdvH as the only permissible order in all three conditions.

(8)

graphic

We summarize the predictions that the various analyses make in table 1.

Table 1

Adverbial orders predicted by five analyses of PP-extraposition (H = high adverbial, L = low adverbial)

Mixed
ExtrapositionV-raisingEqual heightLow PPLow V
V-PP-Adv-Adv L-H L-H L-H L-H L-H 
V-Adv-PP-Adv L-H L-H/H-L L-H/H-L L-H/H-L L-H 
V-Adv-Adv-PP L-H H-L L-H/H-L H-L L-H 
Mixed
ExtrapositionV-raisingEqual heightLow PPLow V
V-PP-Adv-Adv L-H L-H L-H L-H L-H 
V-Adv-PP-Adv L-H L-H/H-L L-H/H-L L-H/H-L L-H 
V-Adv-Adv-PP L-H H-L L-H/H-L H-L L-H 

Two remarks are in order. First, all analyses predict the same order for the PP-initial condition. Including this condition in the tests we carry out will nonetheless be useful, as it gives us a general baseline for the strength of the ordering effect.

Second, the extraposition analysis and the low-V analysis seem empirically indistinguishable. This is true when we restrict discussion to AdvH and AdvL. However, if we consider two adverbials that the verb can move across on the low-V analysis, the two accounts can be teased apart. Let us call these adverbials AdvL1 and AdvL2 , and assume that AdvL2 c-commands AdvL1 . The predictions of the extraposition analysis parallel those in table 1: AdvL1 should precede AdvL2 in all three conditions (see (9a)). However, for pairs of low adverbials the low-V analysis generates predictions on a par with the equal-height analysis. This is because both verb raising and PP-extraposition have access to landing sites higher than AdvL2 (see (9b)).

(9)

graphic
graphic

For concreteness’ sake we list predictions for all five accounts in table 2; those made by the extraposition analysis, the verb-raising analysis, and the equal-height analysis remain constant, but in addition to those made by the low-V analysis, those made by the low-PP analysis shift.

Table 2

Adverbial orders predicted by five analyses of PP-extraposition (L1 = lower of two adverbials, L2 = higher)

Mixed
ExtrapositionV-raisingEqual heightLow PPLow V
V-PP-Adv-Adv L1-L2 L1-L2 L1-L2 L1-L2 L1-L2 
V-Adv-PP-Adv L1-L2 L1-L2/L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 
V-Adv-Adv-PP L1-L2 L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 
Mixed
ExtrapositionV-raisingEqual heightLow PPLow V
V-PP-Adv-Adv L1-L2 L1-L2 L1-L2 L1-L2 L1-L2 
V-Adv-PP-Adv L1-L2 L1-L2/L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 
V-Adv-Adv-PP L1-L2 L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 L1-L2/L2-L1 

Below, we report on two sets of data that can be used to test the predictions in table 1 and table 2. In section 3, we consider structures containing a time and a manner adverbial. As we will show, these behave as predicted by the extraposition and low-V analyses (the latter captures the data on the assumption that verb raising can cross manner but not time adverbials). In section 4, we consider structures with two low adverbials. It turns out that the data fit neither the extraposition nor the low-V analysis. However, further exploration in sections 5 and 6 reveals a more subtle argument in favor of the extraposition analysis. (a) The extraposition analysis, but not the low-V analysis, turns out to be compatible with an auxiliary hypothesis (adapted from Rohrbacher 1994, Ackema and Neeleman 2002, and Williams 2014) that captures the recalcitrant data. (b) Where this auxiliary hypothesis does not apply, the data follow the predictions of the extraposition analysis in table 2, rather than those of the low-V analysis.

The main data points we discuss involve the order of adverbials, as well as scopal relations between them. Where we explore scope, we rely on judgments from a panel of ten native-speaker linguists.1 Where we explore adverbial order, we base our claims on experiments conducted on Amazon Mechanical Turk (AMT). Such experiments have been shown to be as rigorous as experiments carried out in a laboratory setting (Sprouse 2011). Aggregated grammaticality judgments from AMT should therefore allow us to compare sentences in a reliable way, revealing information that can help us decide between competing theories.

3 Time and Manner Adverbs

3.1 Preliminaries

The premise of our first set of word order experiments was that time adverbials are attached higher than manner adverbials, at the very least as a matter of preference (see Jackendoff 1972, Cinque 1999, Ernst 2002). If so, we reasoned, it would be possible to test the various analyses under consideration on the assumption that AdvL is a position hosting manner adverbials, while AdvH is a position hosting time adverbials. This yielded the predictions shown in table 3.

Table 3

Adverbial orders predicted by five analyses of PP-extraposition (M = manner adverbial, T = time adverbial)

Mixed
ExtrapositionV-raisingEqual heightLow PPLow V
V-PP-Adv-Adv M-T M-T M-T M-T M-T 
V-Adv-PP-Adv M-T M-T/T-M M-T/T-M M-T/T-M M-T 
V-Adv-Adv-PP M-T T-M M-T/T-M T-M M-T 
Mixed
ExtrapositionV-raisingEqual heightLow PPLow V
V-PP-Adv-Adv M-T M-T M-T M-T M-T 
V-Adv-PP-Adv M-T M-T/T-M M-T/T-M M-T/T-M M-T 
V-Adv-Adv-PP M-T T-M M-T/T-M T-M M-T 

In designing the test, we had to take into account the circumstances under which time and manner adverbials can surface between a verb and a selected PP. While such PPs can uncontroversially follow manner adverbials, many speakers find extraposition of light PPs across time adverbials only marginally better than extraposition of light DPs. Both PP-extraposition across a time adverbial and DP-extraposition improve when the extraposed complement is heavy.

(10)

  • Emma looked

    graphic
    ??yesterday
    graphic
    at the memorandum
    graphic
    yesterday
    graphic
    .

  • Emma read

    graphic
    *yesterday
    graphic
    the new memorandum
    graphic
    yesterday
    graphic
    .

  • Emma looked

    graphic
    yesterday
    graphic
    at the memorandum from the finance director
    graphic
    yesterday
    graphic
    .

  • Emma read

    graphic
    yesterday
    graphic
    the new memorandum from the finance director
    graphic
    yesterday
    graphic
    .

These informal judgments were confirmed through a baseline experiment conducted on AMT (Experiment 1). We recruited eighty participants, all native speakers of English with IP addresses in the United States. They judged five sets of eight test items that followed the scheme in (11).2

(11)

  • V-PPLight-AdvT vs. V-AdvT-PPLight

  • V-PPHeavy-AdvT vs. V-AdvT-PPHeavy

  • V-DPLight-AdvT vs. V-AdvT-DPLight

  • V-DPHeavy-AdvT vs. V-AdvT-DPHeavy

Each item was assigned a score on a seven-point Likert scale. The experiment had a Latin square design, so participants saw only one item per condition. The order of test sentences was randomized per participant and the test included both grammatical and ungrammatical fillers, as well as questions to check that participants were paying attention to the task.

The results are summarized in table 4. (Significance was calculated using post hoc two-tailed t-tests, with p < .05 as the threshold; throughout, standard deviations are given between parentheses.)

Table 4

Acceptability of PPs/DPs preceding/following time adverbials (N = 80)

Light PPHeavy PPLight DPHeavy DP
PP-AdvT AdvT-PP PP-AdvT AdvT-PP DP-AdvT AdvT-DP DP-AdvT AdvT-DP 
6.56 (0.87) 3.78 (1.34) 5.80 (0.76) 5.09 (1.31) 6.73 (0.58) 2.76 (1.44) 5.73 (0.87) 5.01 (1.31) 
p < .001 p < .01 p < .001 p < .01 
Light PPHeavy PPLight DPHeavy DP
PP-AdvT AdvT-PP PP-AdvT AdvT-PP DP-AdvT AdvT-DP DP-AdvT AdvT-DP 
6.56 (0.87) 3.78 (1.34) 5.80 (0.76) 5.09 (1.31) 6.73 (0.58) 2.76 (1.44) 5.73 (0.87) 5.01 (1.31) 
p < .001 p < .01 p < .001 p < .01 

A two-factor ANOVA (Complement Weight × Adverb Placement) applied to the PP data shows a significant main effect of complement weight (F1(1, 79) = 4.83, p < .05), a significant main effect of adverb placement (F1(1, 79) = 197.55, p < .0001), and crucially a significant interaction between complement weight and adverb placement (F1(1, 79) = 69.4, p < .0001). These findings parallel trends in the DP data, where a two-factor ANOVA (Complement Weight × Adverb Placement) also shows a significant interaction between complement weight and adverb placement (F1(3, 237) = 84.18, p < .0001), in addition to a significant main effect of complement weight (F1(1, 79) = 25.32, p < .0001) and a significant main effect of adverb placement (F1(1, 79) = 355.52, p < .0001).

The data in table 4 suggest that PP-extraposition across time adverbials may be an instance of heavy-XP shift. However, that cannot be the whole story. Extraposition of light PPs across temporal adverbials improves considerably in certain contexts—for example, when the PP is followed by a coordinate clause that contains a coreferential pronoun, as in (12a). As (12b) shows, this is not the case for extraposition of light DPs.

(12)

  • Emma looked yesterday at the memorandum, and it made her blood boil.

  • *Emma read yesterday the new memorandum, and it made her blood boil.

Our take on these data is that there are two distinct interpretations that license intervention of time adverbials. In one of these, the PP is in focus and has possibly undergone heavy-XP shift. The nature of the other interpretation is revealed by the intonation of examples like (12a): the PP must be destressed, suggesting that it represents given information and is most likely a continuing topic in Lambrecht’s (1994:132) sense.

Again, the informal judgments in (12) can be replicated experimentally. We conducted a second baseline experiment (Experiment 2) with the same setup as above; now, however, each test item was followed by a coordinate clause containing a coreferential pronoun, as in (12). This had a clear impact on acceptability scores, as table 5 shows. There are two noticeable effects. The presence of the coordinate clause leads to an increase of the average score for extraposition of light PPs (which rises from 3.78 to 4.92). At the same time, it inhibits heavy-DP shift, bringing the average score down from 5.01 to 4.04. Both effects are highly significant ( p < .001).

Table 5

Acceptability of PPs/DPs preceding/following time adverbials; coordinate condition (N = 80)

Light PPHeavy PPLight DPHeavy DP
PP-AdvT AdvT-PP PP-AdvT AdvT-PP DP-AdvT AdvT-DP DP-AdvT AdvT-DP 
5.99 (1.12) 4.92 (1.73) 5.79 (0.74) 5.41 (1.25) 6.21 (0.92) 3.15 (1.83) 5.92 (1.32) 4.04 (1.70) 
p < .001 p < .05 p < .001 p < .001 
Light PPHeavy PPLight DPHeavy DP
PP-AdvT AdvT-PP PP-AdvT AdvT-PP DP-AdvT AdvT-DP DP-AdvT AdvT-DP 
5.99 (1.12) 4.92 (1.73) 5.79 (0.74) 5.41 (1.25) 6.21 (0.92) 3.15 (1.83) 5.92 (1.32) 4.04 (1.70) 
p < .001 p < .05 p < .001 p < .001 

In sum, it appears that there are two basic processes by which adverbials can end up between a verb and a selected PP. The first results in intervening manner adverbials; the second can additionally result in intervening time adverbials. Closer inspection of the data suggests that the second process in turn has two variants: one with the PP in focus, the other with the PP a continuing topic.

Of the analyses under consideration, two can explain why intervention of manner adverbials and intervention of time adverbials should differ. On the low-V analysis, the V-AdvT-PP order must be derived by PP-extraposition (as the verb cannot raise across time adverbials), while the V-AdvM-PP order can be derived either by PP-extraposition or by verb raising. On the extraposition analysis, one could assume that intervention of manner adverbials may result from base-generation (as in (3b)), while intervention of time adverbials must result from rightward movement of the PP (as in (3c)) (a point we return to in section 9). The other accounts have no obvious handle on the facts uncovered above, as far as we can tell.

3.2 The Order of Manner and Time Adverbials

As mentioned, the design of our word order test had to take account of the circumstances in which PP-extraposition across time adverbials is licit. We therefore constructed two word order experiments. In the first, the clause containing the PP was followed by a coordinate clause containing a pronoun coreferential with the preposition’s complement (Experiment 3). There were ten sets of test items, each consisting of a basic sentence and five alternations, as in (13) (given with our informal grammaticality judgments).

(13)

  • Bill talked to his neighbor

    graphic
    softly
    graphic
    last night
    graphic
    *softly
    graphic
    , and she told him some news.

  • Bill talked softly to his neighbor last night, and she told him some news.

  • *Bill talked last night to his neighbor softly, and she told him some news.

  • Bill talked

    graphic
    softly
    graphic
    last night
    graphic
    *softly
    graphic
    to his neighbor, and she told him some news.

The method was otherwise as in Experiments 1 and 2: the experiment was conducted on AMT with eighty participants, who judged test items and fillers in a Latin square design.

The results are given in table 6. They show that in all three conditions there is a clear preference for orders in which manner adverbials precede time adverbials. This preference is apparently unaffected by the position of the PP.

Table 6

Acceptability of adverbial order; manner and time adverbials; coordinate condition (N = 80)

PP-initialPP-medialPP-final
AdvM-AdvT 4.95 (1.63) 5.65 (0.98) 5.43 (1.36) 
AdvT-AdvM 3.94 (2.11) 4.81 (1.95) 4.26 (1.67) 
 p < .01 p < .01 p < .01 
PP-initialPP-medialPP-final
AdvM-AdvT 4.95 (1.63) 5.65 (0.98) 5.43 (1.36) 
AdvT-AdvM 3.94 (2.11) 4.81 (1.95) 4.26 (1.67) 
 p < .01 p < .01 p < .01 

A two-factor ANOVA (Adverbial Order × Sentence Type) confirms this conclusion. It shows a significant main effect of sentence type (F1(1, 79) = 9.58, p < .001) and a significant main effect of adverbial order (F1(1, 79) = 46.71, p < .0001). However, the interaction between sentence type and adverbial order was not significant (F1(2, 158) = 0.41, p = .664).

We conducted a second experiment (Experiment 4), parallel to the previous one except that PPs were heavy and subsequent coordinate clauses were omitted, as in (14). As before, the experiment had a Latin square design and was carried out on AMT with ten sets of test items and eighty participants.

(14)

  • Bill talked to a very shy neighbor of his

    graphic
    softly
    graphic
    last night
    graphic
    *softly
    graphic
    .

  • Bill talked softly to a very shy neighbor of his last night.

  • *Bill talked last night to a very shy neighbor of his softly.

  • Bill talked

    graphic
    softly
    graphic
    last night
    graphic
    *softly
    graphic
    to a very shy neighbor of his.

The results are given in table 7. They show that in all three conditions there is a significant preference for orders in which manner adverbials precede time adverbials. Nonetheless, it is somewhat unexpected that the score for AdvT-AdvM in the PP-final condition is as high as it is (in comparison to the PP-initial and PP-medial conditions). The resulting overall pattern does not fit any of the accounts under consideration, as none predicts variable order in the PP-final condition and AdvM-AdvT elsewhere (which is what the average scores suggest).

Table 7

Acceptability of adverbial order; manner and time adverbials; heavy PP (N = 80)

PP-initialPP-medialPP-final
AdvM-AdvT 4.39 (1.51) 5.36 (1.03) 5.70 (0.64) 
AdvT-AdvM 3.16 (1.81) 3.01 (1.91) 5.10 (0.54) 
 p < .001 p < .001 p < .001 
PP-initialPP-medialPP-final
AdvM-AdvT 4.39 (1.51) 5.36 (1.03) 5.70 (0.64) 
AdvT-AdvM 3.16 (1.81) 3.01 (1.91) 5.10 (0.54) 
 p < .001 p < .001 p < .001 

That there is something to be explained here is confirmed by a two-factor ANOVA (Adverbial Order × Sentence Type), which shows a significant interaction of sentence type and adverbial order (F1(2, 158) = 15.37, p < .0001) (alongside a significant main effect of sentence type (F1(1, 79) = 61.27, p < .0001) and a significant main effect of adverbial order (F1(1, 79) = 121.43, p < .0001)).

Our suggestion is that the prosodic break preceding constituents that have undergone heavy-XP shift facilitates a parenthetical reading of the manner adverbial immediately preceding the moved PP. This effect is absent in the PP-initial and PP-medial conditions, which do not have such a facilitating prosody. To find out whether this is indeed the explanation of the anomaly, we carried out Experiment 4 a second time, but now with each test item introduced by a wh-question that was answered by the manner adverbial, as in (15) (Experiment 5).

(15)

  • How did Bill talk to a very shy neighbor of his last night?

    Bill talked to a very shy neighbor of his

    graphic
    softly
    graphic
    last night
    graphic
    *softly
    graphic
    .

  • How did Bill talk to a very shy neighbor of his last night?

    Bill talked softly to a very shy neighbor of his last night.

  • How did Bill talk last night to a very shy neighbor of his?

    *Bill talked last night to a very shy neighbor of his softly.

  • How did Bill talk last night to a very shy neighbor of his?

    Bill talked

    graphic
    softly
    graphic
    last night
    graphic
    *softly
    graphic
    to a very shy neighbor of his.

As parentheticals cannot answer wh-questions, this should have the effect that the scores in the PP-final condition fall in line.

The results are as expected, as table 8 illustrates. A two-factor ANOVA (Adverbial Order × Sentence Type) shows a significant main effect of sentence type (F1(1, 79) = 4.34, p < .05) and a significant main effect of adverbial order (F1(1, 79) = 235.97, p < .0001). However, the interaction between sentence type and adverbial order is no longer significant (F1(2, 158) = 0.23, p = .79).

Table 8

Acceptability of adverbial order; manner and time adverbials; heavy PP; wh-context (N = 80)

PP-initialPP-medialPP-final
AdvM-AdvT 5.51 (1.65) 5.71 (1.07) 5.98 (0.73) 
AdvT-AdvM 3.54 (1.57) 3.65 (1.94) 3.91 (1.24) 
 p < .01 p < .001 p < .001 
PP-initialPP-medialPP-final
AdvM-AdvT 5.51 (1.65) 5.71 (1.07) 5.98 (0.73) 
AdvT-AdvM 3.54 (1.57) 3.65 (1.94) 3.91 (1.24) 
 p < .01 p < .001 p < .001 

All in all, our findings show a consistent preference for AdvM-AdvT order across all three conditions. This is as predicted by the extraposition analysis and the low-V analysis. Our results falsify the verb-raising analysis and the remaining mixed analyses, which incorrectly predict that in the PP-medial condition and/or the PP-final condition there should not be a preference for manner adverbials to precede time adverbials.

3.3 An Aside on PP Modifiers

We close this section with a brief discussion of another aspect of our findings. While the pattern in tables 6 and 8 is robust, the difference in average scores between the two adverbial orders is relatively modest (1.5 on average across all conditions; 1.49 on average in the PP-initial condition, where the predictions of all accounts converge). This is partly a fact about violations of the adverbial hierarchy: many produce only a limited penalty. But there is a second factor at play. Test sentences with multiple adverbs often appear to be reduced in acceptability, which in effect compresses the Likert scale when participants judge adverbial orders (see Payne 2018 for discussion). We can demonstrate the effect by considering the influence of category on the acceptability of sentences with a time and a manner adverbial, as in (16).

(16)

  • Bill spoke [AdvP eloquently] today.

  • ??Bill spoke today [AdvP eloquently].

  • Bill spoke [PP with eloquence] today.

  • Bill spoke today [PP with eloquence].

We conducted a test with ten sets of example sentences modeled on (16) (Experiment 6). Each set contained two items in which manner was expressed by an adverb and two in which it was expressed by a PP. Time was always expressed by an adverb or a DP. As it turned out, scores were consistently higher if the manner adverbial was a PP (see table 9), which suggests that sequences of adverbs indeed come at a cost.

Table 9

Acceptability of adverbial order; time expressed by AdvP/DP (N = 80)

Manner-TimeTime-Manner
M = AdvP 5.20 (0.85) 4.60 (1.37) p < .01 
M = PP 6.09 (0.75) 6.30 (0.93) n.s. 
 p < .001 p < .001  
Manner-TimeTime-Manner
M = AdvP 5.20 (0.85) 4.60 (1.37) p < .01 
M = PP 6.09 (0.75) 6.30 (0.93) n.s. 
 p < .001 p < .001  

Given these effects, one may think that it would be better to conduct word order tests with manner PPs. However, the data also show that a violation of the adverbial hierarchy reduces the score in the multiple-adverb condition, but not in the adverb-PP condition. This is confirmed by a two-factor ANOVA (Modifier Order × Manner Category), which shows a significant interaction between modifier order and the category of the manner modifier (F1(1, 79) = 13.16, p < .001), alongside a significant main effect of category (F1(1, 79) = 133.4, p < .0001). (The main effect of modifier order was not significant (F1(1, 79) = 2.99, p = .08), presumably because order is free in the adverb-PP condition.)

We would suggest that this pattern is found because PP adverbials can be extraposed from an underlying position in which they satisfy the adverbial hierarchy. Whatever the value of that suggestion, it is clear that the best tests for effects of the adverbial hierarchy avoid PP adverbials, as was the case in Experiments 3, 4, and 5.

4 Intentionally Twice and Again Continuously

From here onward, we restrict discussion to the extraposition and low-V analyses, which are the only accounts compatible with the results in section 3. To force a decision between these two analyses, we must consider structures with two adverbials that are low enough for the verb to move across (if it does move). As explained in section 2, the predictions generated by the extraposition analysis remain constant: regardless of the position of the PP, the lower of the two adverbs must precede the higher one (see (9a), repeated as (17a) for convenience). However, the predictions of the low-V analysis shift toward the equal-height analysis. In the PP-initial condition, the higher adverbial must still follow, but in the PP-medial and PP-final conditions, either the lower or the higher adverbial may come first (see (17b)).

(17)

graphic
graphic

Reversible adverb pairs provide one way to test these predictions. (Indeed, the behavior of such adverb pairs is among the strongest evidence for verb raising in Pesetsky 1989.) As c-command relations between reversible adverbs are not fixed (see (18)), we cannot test the extraposition and low-V analyses by looking at word order: both theories predict free word order in all three conditions. However, we can consider scope. The extraposition analysis predicts right-to-left scope across the board (see (17a), where c-command between AdvL2 and AdvL1 is right-to-left). The low-V analysis predicts ambiguity in the PP-medial and PP-final conditions, because c-command between AdvL2 and AdvL1 can be left-to-right or right-to-left (see (17b)). In the PP-initial condition, however, c-command and therefore scope between AdvL2 and AdvL1 are exclusively right-to-left.

(18) John

graphic
intentionally
graphic
twice
graphic
intentionally
graphic
knocked on the door.

As mentioned in section 2, we relied for scope judgments on a panel of ten native-speaker linguists; all were speakers of American English and all were trained at PhD level. We asked them for their judgments on adverbial scope in three sets of three pairs of sentences (Scope Test 1). Each pair corresponded to one of the conditions under discussion, with variation in the order of the adverbs, as in (19). Each set had a different combination of reversible adverbs.

(19)

  • John knocked on the door

    graphic
    intentionally
    graphic
    twice
    graphic
    intentionally
    graphic
    .

  • John knocked intentionally on the door twice.

  • John knocked twice on the door intentionally.

  • John knocked

    graphic
    intentionally
    graphic
    twice
    graphic
    intentionally
    graphic
    on the door.

A clear consensus emerged. When the adverbs are adjacent, scope is variable, but when they are separated by the PP, scope is right-to-left (see table 10). Neither analysis predicts this pattern.

Table 10

Scope between pairs of reversible adverbs (N = 10)

PP-initialPP-medialPP-final
L > RL <> RL < RL > RL <> RL < RL > RL <> RL < R
10 
PP-initialPP-medialPP-final
L > RL <> RL < RL > RL <> RL < RL > RL <> RL < R
10 

The extraposition analysis makes the wrong predictions for the PP-initial and PP-final conditions, while the PP-initial and PP-medial conditions are problematic for the low-V analysis.3

These findings are corroborated by further data involving again. While this adverb can be merged very low indeed, it cannot appear in the scope of manner adverbs like continuously. This means that the extraposition and low-V analyses make diverging predictions for sentences containing again and a suitable manner adverb. The extraposition analysis predicts that the manner adverb will systematically precede again (see (17a)). The low-V analysis predicts that in the PP-initial condition again should follow the manner adverb, as c-command among the adverbials is right-to-left (see (17b)). However, it predicts free order in the PP-medial and PP-final conditions, where c-command between the adverbials is variable.

We tested these word order predictions through an AMT experiment. We recruited eighty participants, who judged five sets of sentences of the type in (20) in the same setup as before (Experiment 7).

(20)

  • John knocked on the door

    graphic
    continuously
    graphic
    again
    graphic
    continuously
    graphic
    .

  • John knocked continuously on the door again.

  • John knocked again on the door continuously.

  • John knocked

    graphic
    continuously
    graphic
    again
    graphic
    continuously
    graphic
    on the door.

Again, the results were not as expected. When the adverbs are separated by the PP, there is a preference for AdvM-again (as predicted by the extraposition analysis), but when they are adjacent, the data are problematic for both the extraposition and the low-V analyses: in the PP-initial condition, both orders are on a par, while in the PP-final condition there is a prefence for again-AdvM (see table 11).

Table 11

Acceptability of adverbial order; again and manner adverbials (N = 80)

PP-initialPP-medialPP-final
AdvM-again 4.53 (1.15) 5.81 (0.75) 4.30 (1.04) 
again-AdvM 4.89 (1.16) 5.21 (0.75) 4.90 (1.45) 
 n.s. p < .01 p < .01 
PP-initialPP-medialPP-final
AdvM-again 4.53 (1.15) 5.81 (0.75) 4.30 (1.04) 
again-AdvM 4.89 (1.16) 5.21 (0.75) 4.90 (1.45) 
 n.s. p < .01 p < .01 

That adjacency may be the factor behind this unexpected data pattern is confirmed by a two-factor ANOVA (Adverbial Order × Adverbial Adjacency), which shows a significant main effect of adverbial adjacency (F1(1, 79) = 60.72, p < .0001) and a significant interaction between adverbial adjacency and adverbial order (F1(1, 79) = 24.03, p < .0001). The main effect of adverbial order was not significant (F1(1, 79) = 1.35, p = .25).

In sum, neither the extraposition analysis nor the low-V analysis predicts the patterns found with pairs of low adverbs. Both accounts will therefore need to invoke some auxiliary hypothesis to capture the findings in table 10 and table 11. In sections 5 and 6, we will consider what amendments are available under the extraposition and low-V analyses. We argue that on closer inspection the data support PP-extraposition.

5 Amending the Extraposition Analysis: Adverbial Clustering

One auxiliary hypothesis compatible with the extraposition analysis is that some adverbials may left-adjoin to other adverbials (as argued previously in Rohrbacher 1994 and Ackema and Neeleman 2002; see also Williams 2014). Such direct modification is not in itself controversial. It is generally assumed, for example, that only may be associated with another category at a distance, as in (21a), or be directly attached to it, as in (21b) (see Rooth 1985).

(21)

  • John only invited MARY.

  • Only MARY did John invite.

Thus, our claim is simply that direct modification is available for a larger class of adverbials than usually assumed. The problematic data then follow if we assume that the adjoined adverbial takes scope over and precedes its host.4 When the adverbs are adjacent, they may have merged independently, yielding right-to-left scope (see (22a,c)), or the first may have merged with the second, yielding left-to-right scope (see (22a′,c′)). When the adverbs are separated by the PP, however, they must have been attached independently, so that only right-to-left scope is available (see (22b)).

(22)

graphic

This is enough to capture the variation in scope in the PP-initial and PP-final conditions in table 10, as well as the scope rigidity in the PP-medial condition.

The data in table 11 show not just optionality, but in fact a slight preference for left-to-right scope in the PP-final condition. This may be the result of a general aversion, already mentioned in section 3.3, against sentences containing multiple adverbials. If the first adverb is adjoined to the second, the result is a single complex modifier, rather than a sequence of simplex modifiers.

An evaluation of the extraposition analysis in conjunction with the auxiliary hypothesis of direct adverbial modification must address three core issues: whether there is any empirical evidence for adverbial clustering (see section 5.1), how adverbial clusters are interpreted (see sections 5.2 and 5.3), and how adverbial clustering can be constrained so as to preserve our account of the data discussed in section 3 (see section 5.4).

5.1 Basic Evidence

An observation that bears on the first of these questions comes from clefting. While a combination of a time adverbial and a manner adverbial resists clefting (see (23a)), intentionally twice can be clefted (see (23b)). This suggests that intentionally twice, but not last night desperately, can constitute a syntactic unit. Note that, in line with expectations, intentionally must take scope over twice when clefted: (23b) implies that John had the intention to knock twice on the door. (Where judgments were checked with our panel of native-speaker linguists, we will from now on indicate this as a fraction that expresses how many panelists accepted a given example or a given reading.)

(23)

  • *It was last night DESPERATELY that Mary looked for her puppy. (0/10)

  • It was intentionally TWICE that John knocked on the door. (10/10)

It is also predicted, correctly as it turns out, that again continuously can undergo clefting. However, what we can conclude from this observation is not immediately clear, as again in (24) could be an independent modifier in the top part of the cleft, something that is unlikely to be true of intentionally in (23b). In section 5.3 we will show, though, that examples like (24) are grammatical on exactly the reading in which again modifies continuously.

(24) It was again CONTINUOUSLY that John knocked on the door. (10/10)

A second way to test our auxiliary hypothesis is to replace the initial adverb in a pair of adverbs that permit postverbal left-to-right scope with a near-synonymous PP. While we have hypothesized that in the structures at hand the adjoined adverbial must precede the category it attaches to, PP modifiers tend to follow in almost all circumstances. Therefore, judgments are predicted to change when a PP replaces the first modifier in an adverb-adverb sequence. When adverbial clustering is ruled out, a pattern of judgments should emerge that is reminiscent of judgments for pairs of time and manner adverbials.

Indeed, when intentionally in (23) is replaced by with intention, the result is degraded.

(25) *It was with intention TWICE that John knocked on the door. (0/10)

The effect extends to adverbial scope in the PP-initial and PP-final conditions. We asked our panel of native-speaker linguists to judge scope between a PP modifier and an adverb in three sets of three examples, one of which is given in (26) (Scope Test 2). The expected change in judgments was evident, as all ten reported that they could only get right-to-left scope, regardless of condition (see table 12). This is of course exactly the pattern predicted by the amended extraposition analysis. (NB: The number of test sentences was relatively low, as there are few PPs whose meaning sufficiently approximates that of the relevant adverbs.)

(26)

  • John knocked with intention twice on the door.

  • John knocked with intention on the door twice.

  • John knocked on the door with intention twice.

Table 12

Scope judgments for PPmod-adverb pairs (N = 10)

PP-initialPP-medialPP-final
L > RL <> RL < RL > RL <> RL < RL > RL <> RL < R
10 10 10 
PP-initialPP-medialPP-final
L > RL <> RL < RL > RL <> RL < RL > RL <> RL < R
10 10 10 

A second time is an expression whose interpretation approximates again, but which cannot directly modify other adverbials, as shown by the ungrammaticality of (27).

(27) *It was a second time CONTINUOUSLY that John knocked on the door. (0/10)

The amended extraposition analysis therefore predicts that when again in the examples in (20) is replaced by a second time, only the orders that do not rely on adverbial clustering will survive. Informal judgments suggest that this is correct. Regardless of condition, continuously a second time is the only acceptable order for the native speakers we have consulted.

(28)

  • John knocked on the door

    graphic
    continuously
    graphic
    a second time
    graphic
    *continuously
    graphic
    .

  • John knocked continuously on the door a second time.

  • *John knocked a second time on the door continuously.

  • John knocked

    graphic
    continuously
    graphic
    a second time
    graphic
    *continuously
    graphic
    on the door.

To validate these judgments, we conducted a version of Experiment 7 (Experiment 8) in which again was replaced by a second time. Otherwise, the setup was unchanged. The results presented in table 13 show that there is a significant preference in all conditions for the order in which a second time follows the adverb, as predicted by the extraposition analysis (see (17a)).

Table 13

Acceptability of adverbial order; manner adverbial and a second time (N = 80)

PP-initialPP-medialPP-final
AdvMa second time 4.8 (1.26) 4.74 (1.73) 4.28 (1.05) 
a second time–AdvM 3.8 (1.94) 3.45 (1.71) 3.38 (1.54) 
 p < .01 p < .01 p < .01 
PP-initialPP-medialPP-final
AdvMa second time 4.8 (1.26) 4.74 (1.73) 4.28 (1.05) 
a second time–AdvM 3.8 (1.94) 3.45 (1.71) 3.38 (1.54) 
 p < .01 p < .01 p < .01 

As predicted, the effects of adverbial adjacency disappeared in Experiment 8. A two-factor ANOVA (Adverbial Order × Adverbial Adjacency) shows no significant main effect of adverbial adjacency (F1(1, 79) = 0.04, p = .84) and no significant interaction between adverbial order and adverbial adjacency (F1(1, 79) = 1.23, p = .27). However, the main effect of adverbial order is significant, as expected (F1(1, 79) = 54.65, p < .0001).

We conclude (a) that there is sufficient empirical support for adverbial clustering, and (b) that where adverbial clustering is ruled out, the data are as predicted by the extraposition analysis, rather than the low-V analysis.5

5.2 Interpretive Effects: Intentionally Twice

We now turn to the semantic effects of attaching one modifier directly to another, starting with adverbial clusters introduced by an adverb like intentionally (as in intentionally twice). We propose that intentionally and its kin allow association with focus. This is not a novel claim. Williams (2014) argues the point in some detail. The effect is easy to see with accidentally, the antonym of intentionally and the adverb we concentrate on below. Take an example like John accidentally murdered BILL. Murder is an intentional act, and so one would expect this sentence to be a contradiction. Its coherence is due to accidentally associating with BILL. The ordinary value of the sentence is that John murdered Bill; its focus value consists of the presupposition that there is an alternative x to Bill, such that John intended to kill x.

The role of focus can be further illustrated with the examples in (29).

(29)

  • Susan accidentally gave Bill ASCI-FINOVEL.

  • Susan accidentally gave BILL a sci-fi novel.

  • (i) Susan gave Bill a sci-fi novel; (ii) Ǝx, x an alternative to a sci-fi novel, Susan intended to give Bill x.

  • (i) Susan gave Bill a sci-fi novel; (ii) Ǝx, x an alternative to Bill, Susan intended to give x a sci-fi novel.

The sentence in (29a) permits the interpretation in (29c), but not that in (29d). Conversely, the sentence in (29b) permits the interpretation in (29d), but not that in (29c). (There are other interpretations of these examples that are not relevant here; they could, for instance, be used when speaker and addressee know that Bill hates sci-fi novels, but Susan was not aware of this.)

The same pattern can be observed in examples more directly relevant to the question under discussion. On a parse of the examples in (30) in which accidentally takes scope over twice, (30a) comes with the presupposition that Susan intended to knock twice on something other than the door, while (30b) presupposes that Susan intended to knock on the door, but either fewer or more times than two.

(30)

  • Susan accidentally [knocked on the DOOR twice].

  • Susan accidentally [knocked on the door TWICE].

Adverbials that associate with focus may often directly attach to the focused constituent (as already illustrated for only in (21)). We suggest that this is what lies behind adverbial clustering with accidentally: this adverb may merge with a second adverbial if the latter constitutes its associated focus. Thus, when accidentally is merged with twice in (31a), the interpretation that obtains is parallel to that in (30b) (see (31b); for related discussion, see Bobaljik 2017).

(31)

  • Susan knocked

    graphic
    on the door
    graphic
    [accidentally twice]
    graphic
    on the door
    graphic
    .

  • (i) Susan gave two knocks on the door; (ii) Ǝn, n an alternative to 2, Susan intended to give n knocks on the door.

Again, we follow Williams (2014) here. Williams argues explicitly that focus-sensitive adverbs may either be merged in their scopal position or attach to the associated focus.6

The hypothesis that accidentally is a focus-sensitive adverb makes an important prediction: if it attaches to a second adverb, this has a disambiguating effect, as it precludes any other element in the clause from acting as the associated focus. To show that this prediction is correct, we must first establish what interpretation (31) has when accidentally and twice are merged independently, as in (32). In this structure, accidentally does not take scope over twice, and therefore it cannot generate a presupposition about the intended number of knocks on the door. Rather, it triggers the presupposition that Susan twice intended to perform an action other than knocking on the door (the nature of which is partly dependent on where stress is placed within knock on the door).

(32)

  • Susan [[knocked

    graphic
    on the door
    graphic
    accidentally] twice]
    graphic
    on the door
    graphic
    .

  • (i) On two occasions, Susan knocked on the door; (ii) on each occasion Ǝa, a an alternative to knock on the door, Susan intended to perform a.

Thus, we predict that Susan knocked accidentally twice on the door and Susan knocked on the door accidentally twice are ambiguous between the readings in (31b) and (32b), but do not permit an interpretation on a par with (30a), where accidentally triggers the presupposition that Susan intended to give two knocks on some object other than the door. This is because accidentally precedes twice and so can only take scope over twice if it is attached to it, as in (31). But if it is attached to it, twice must be the associated focus, which precludes association with the door. The relevant reading is indeed unavailable, as confirmed by the unanimous judgment of our panel of ten native-speaker linguists.

5.3 Interpretive Effects: Again Continuously

We next consider adverbial clusters introduced by again (such as again continuously). Like accidentally, again triggers a presupposition. As argued extensively in the literature, one crucial factor that governs the nature of this presupposition is the c-command domain of the adverb. Of particular interest here is the contrast in the interpretation of (33a) and (33b).7

(33)

  • Oliver [again [showed the second book to Louise]].

  • Oliver showed the second book [again [to Louise]].

Suppose that Oliver is a rare-book seller who has two antique tomes on offer. He shows these to a select group of customers, one of whom is Louise. In that context, (33a) can have the interpretation in (34a), but not that in (34b). By contrast, (33b) has the interpretation in (34b), as well as that in (34a), albeit that the latter is harder to access. (Our panelists were unanimous in these judgments.)

(34)

  • (i) Oliver showed the second book to Louise; (ii) Oliver previously showed the second book to Louise.

  • (i) Oliver showed the second book to Louise; (ii) Oliver previously showed the first book to Louise.

The example in (33a) is unremarkable. Again is attached to the bracketed constituent and therefore triggers the presupposition that someone ( possibly Oliver) previously carried out the action described by this constituent. We assume that in (33b) again is attached to to Louise. This means that the VP is not part of again’s c-command domain, so that the presupposition triggered is that Oliver previously performed some unspecified action directed toward Louise. In the context at hand, this action is most easily construed as show the first book. It may also be construed as show the second book, but this is pragmatically odd, as that construal is explicitly encoded by (33a).

The underspecified nature of the presupposition triggered by again in examples like (33b) is brought to the fore by the contrast in (35).

(35) Oliver seems to be showering Louise with attention and ignoring everyone else. He introduced himself to Louise. Then he read a poem to Louise.

  • #And then he [again [showed pictures of a romantic sunset to Louise]]. (0/10)

  • And then he showed pictures of a romantic sunset [again [to Louise]]. (10/10)

The sentence in (35a) requires accommodation of some sort, as the context does not provide an earlier instance of show pictures of a romantic sunset to Louise. There is no such effect in (35b), where again merely signals that Oliver previously performed actions directed toward Louise, a presupposition supported by the context given.

Clefts can be used to show that clustering is necessary to get the interpretive effect observed in (35). In the context at hand, the cleft in (36a) is awkward and requires accommodation, but the one in (36b), where again and to Louise presumably form a cluster, is fully acceptable.

(36)

  • #And it was to Louise that he showed pictures of a romantic sunset again. (0/10)

  • And it was again to Louise that he showed pictures of a romantic sunset. (10/10)

It is a small step to assume that adverbial clustering with again is motivated by the same interpretive effect. If so, we expect that structures whose well-formedness relies on clustering will not imply that there is a previous instance of the action described by the VP. This crucial prediction is correct. Suppose that Field Commander Cohen was our most important spy, and that in the course of one of his adventures he agreed to knock on two doors in a particular manner to signal whether the coast was clear. In this context, (37a) can have the interpretation in (37b).

(37)

  • Cohen knocked

    graphic
    on the second door
    graphic
    [again continuously]
    graphic
    on the second door
    graphic
    .

  • (i) Cohen knocked on the second door continuously; (ii) Cohen previously knocked on the first door continuously. (10/10)

In fact, because Cohen’s earlier actions are left unspecified, the following is also felicitous:

(38) Cohen talked continuously for an hour. Then he played the piano continuously for 45 minutes. And then he knocked

graphic
on the door
graphic
[again continuously]
graphic
on the door
graphic
. (10/10)

These examples can be contrasted with examples in which continuously and again are merged separately. In such examples, the VP is part of the c-command domain of again and must therefore be mapped to the presupposition. Thus, (39a) requires that Cohen previously knocked on the second door continuously, as in (39b). The sentence cannot presuppose that Cohen continuously knocked on the first door, as in (39b′).

(39)

  • Cohen [

    graphic
    again
    graphic
    [knocked on the second door continuously]
    graphic
    again
    graphic
    ].

  • (i) Cohen knocked on the second door continuously; (ii) Cohen previously knocked on the second door continuously. (10/10)

  • (i) Cohen knocked on the second door continuously; (ii) Cohen previously knocked on the first door continuously. (0/10)

In line with this, the final sentence in (40) is awkward and requires accommodation.

(40) Cohen talked continuously for an hour. Then he played the piano continuously for 45 minutes. #And then he [

graphic
again
graphic
[knocked on the door continuously]
graphic
again
graphic
]. (0/10)

Clefts confirm that the interpretive effect required by the context in (38)/(40) relies on clustering. The cleft in (41a) is infelicitous in the relevant context, but the one in (41b) is perfectly natural.

(41)

  • #And it was continuously that he knocked on the door again. (0/10)

  • And it was again continuously that he knocked on the door. (10/10)

The examples of independent attachment of again and continuously all involve structures without PP-extraposition. However, it is important to also look at structures with PP-extraposition, as this may place the PP outside the c-command domain of again, thereby removing the obligation to map it to the presupposition. Thus, in the scenario sketched above, (42a) permits either of the interpretations in (42b,b′).

(42)

  • Cohen [

    graphic
    again
    graphic
    [knocked continuously]
    graphic
    again
    graphic
    ] on the second door.

  • (i) Cohen knocked continuously on the second door; (ii) Cohen previously knocked continuously on the second door. (10/10)

  • (i) Cohen knocked continuously on the second door; (ii) Cohen previously knocked continuously on the first door. (10/10)

This implies that the data in (37) provide evidence for the interpretive effects of adverbial clustering in the PP-initial condition, but not in the PP-final condition. For that, we must consider whether the verb is obligatorily mapped onto the presupposition. This should be the case under independent attachment of again and continuously, but not under adverbial clustering. The data in (43) are in line with this: the final sentence in (43a) requires accommodation, but the final sentence in (43b) does not.

(43)

  • Cohen knocked continuously on the window. #Then he [

    graphic
    again
    graphic
    [banged continuously]
    graphic
    again
    graphic
    ] on the door. (0/10)

  • Cohen knocked continuously on the window. Then he banged [again continuously] on the door. (10/10)

In sum, there is a signature interpretive effect that is found in all examples whose grammaticality relies on again clustering with a second adverbial.

5.4 Time Adverbials

The conclusion from sections 5.2 and 5.3 is that there are clear interpretive effects of adverbial clustering with accidentally and again that have to do with the presuppositions triggered by these elements. We propose that it is these effects that license adverbial clustering in the first place.

This means that we are now in a position to consider whether the extraposition analysis can still account for the data of section 3 if combined with the auxiliary hypothesis that adverbials may cluster. This hypothesis explained the existence of left-to-right scope in the PP-initial and PP-final conditions with adverbs like accidentally and again.

The data in section 3 involved pairs of time and manner adverbials, and the core observation was that time adverbials follow manner adverbials regardless of condition (that is, whether the adverbials are sandwiched between V and PP, are separated by the PP, or appear sentence-finally). To account for this, we must assume that time adverbials cannot adjoin to other adverbials to form an adverbial cluster. As (23a) has already shown, this assumption is correct.

The findings of sections 5.2 and 5.3 give a clear sense of why time adverbials should resist adverbial clustering. Such adverbials do not trigger the kind of presuppositions associated with accidentally and again; they simply specify the time at which a proposition holds. Therefore, they cannot have the kind of privileged relationship with a second adverbial that accidentally and again may enter into. And in the absence of an interpretive license for adverbial clustering, temporal adverbials must be merged with an appropriate category in the extended verbal projection, yielding the predictions tested in section 3.

6 Amending the Low-V Analysis

The unamended extraposition analysis made incorrect predictions in the PP-initial and PP-final conditions for pairs of two manner adverbs or a manner adverb and again. A single auxiliary hypothesis could fix these problems, as the relevant conditions are similar in one important respect: the adverbials are adjacent. The low-V analysis makes incorrect predictions for the same pairs of adverbs in the PP-initial and PP-medial conditions. As there is no obvious factor shared by these conditions (to the exclusion of the PP-final condition), it is hard to make do with a single auxiliary hypothesis.

Recall the basic shape of the account. (a) There are three (relevant) adverbial attachment sites (AdvL1 , AdvL2 , AdvH in (44)), each of which may in principle be linearized to the left or right of its sister. (Time adverbials occupy the AdvH position; manner adverbials and the like occupy the AdvL positions.) (b) The PP may extrapose across any of these adverbial positions. (c) The verb moves across AdvL2 , but not across AdvH.

(44)

graphic

This correctly predicts the distribution of time adverbials, but does not capture the existence of descending pairs of low adverbs in the PP-initial condition, nor the absence of such pairs in the PP-medial condition—if the left instance of AdvL2 and the right instance of AdvL1 are chosen and if the PP remains in situ, the resulting order is V-AdvL2-PP-AdvL1.

There is a way to reconcile the low-V analysis with the facts of section 3. To begin with, one could allow verb raising to pied-pipe the PP complement, as in (45). This has the consequence that descending pairs of nontime adverbials may now appear sentence-finally, which fixes the problem with the PP-initial condition (as V-PP-AdvL2-AdvL1 can now be generated).

(45)

graphic

A second auxiliary hypothesis is required for the PP-medial condition. A simple way of ruling out descending pairs of adverbs in this condition is to remove one of the adverbial positions in (44) and (45): namely, the right instance of AdvL1 . This yields the schemas in (46a) and (46b) for bare verb raising and pied-piping, respectively.

(46)

graphic
graphic

The resulting analysis captures the problematic data.

However, this is not enough. Our discussion of adverbial clustering has uncovered several new facts. Since these fall out from the amended PP-extraposition analysis, it is reasonable to ask whether they also have a place in the amended low-V analysis.

If one accepts our conclusion that there is adverbial clustering, one should reject the low-V analysis. This is because adverbial clustering removes the evidence for verb raising by providing an alternative explanation of descending adverbial pairs in the PP-final condition. Therefore, the low-V analysis requires a substitute account of the data in sections 5.15.3.

We will not discuss this matter in detail, but simply note that finding such an alternative account may not be straightforward. As an example of the difficulties that present themselves, consider how the following sentences are analyzed on the the amended low-V analysis ((47b) is a variant of (37a)):

(47)

  • Cohen knocked [[continuously [VPtv on the second door]] again].

  • Cohen [VP knocked on the second door] [again [tVP continuously]].

The sentence in (47a) is a verb-raising structure, with continuously in AdvL1 and again in the right instantiation of AdvL2 . The sentence in (47b) is derived by raising of VP; continuously still appears in AdvL1 , while again now appears in the left instantiation of AdvL2 . Thus, the sentences are identical in terms of their underlying syntax; they differ only in the linearization of continuously and in whether or not the PP is pied-piped. As linearization and pied-piping are typically irrelevant for interpretation, one would expect the two sentences to have the same meaning. That is not the case, however: (47a) triggers the presupposition that Cohen previously knocked continuously on the second door, but as shown in section 5.3, (47b) does not. Contrasts of this type remain unexplained and therefore require additional assumptions.

7 Interim Conclusion

We have shown that, if suitably amended, both the extraposition analysis and the low-V analysis can capture the order of adverbials in the PP-initial, PP-medial, and PP-final conditions as described in sections 3 and 4. The extraposition analysis must be combined with an auxiliary hypothesis of adverbial clustering.

(48)

  • PP complements can extrapose across right-adjoined adverbials.

  • Adverbs may left-adjoin to other adverbials (if there is an interpretive license).

The low-V analysis is a mixed analysis that assumes verb raising, as well as PP-extraposition. In addition, in order to capture the data in sections 3 and 4, it must rely on two auxiliary hypotheses, given in (49c) and (49d).

(49)

  • The verb moves leftward across manner but not time adverbials.

  • PP complements can extrapose across right-adjoined adverbials.

  • Verb raising may pied-pipe PP complements.

  • The lowest VP-external adverbial position precedes VP.

As things stand, the data in section 5 receive an explanation under the (amended) extraposition analysis, but not under the low-V analysis.

Two conclusions may be drawn regarding (symmetric) analyses of adverbial intervention. First, the fact that certain adverbial pairs can come in descending order in the PP-final condition does not provide evidence for verb raising, as a plausible alternative analysis is available. Second, of the two analyses that can capture the data, the PP-extraposition analysis demonstrably covers more ground using fewer assumptions and must therefore be assumed to be correct unless evidence to the contrary is uncovered. This is the article’s take-home message; we now turn to issues of implementation and integration.

8 Extraposition through Roll-Up Movement

8.1 An Antisymmetric Translation of the Extraposition Analysis

The pattern uncovered in the previous sections is summarized in the schema in (50): the order of postverbal adverbials corresponds with increasing scope or increasing height on the adverbial hierarchy, regardless of the position of the selected PP.

(50) V

graphic
PP
graphic
AdvL1
graphic
PP
graphic
AdvL2
graphic
PP
graphic
AdvH
graphic
PP
graphic

Syntactic structure is traditionally taken to be symmetric: the order between sister nodes is subject to crosslinguistic and language-internal variation. This allows a straightforward account of the fact that in (50) higher adverbials follow lower adverbials. Following the verb’s base position, linear order corresponds with increased height of attachment as a matter of course. The fact that adverbial order is not affected by the position of the PP then follows from the extraposition analysis, as the PP can be shifted rightward and upward.

In this section, we consider an alternative version of the extraposition analysis that rejects the assumption of symmetry. The background to this is the rise in the 1990s of theories postulating that syntactic structure is fundamentally asymmetric, with constituents farther to the right systematically located lower in the tree. The best-known proposal of this type is the antisymmetry framework developed in Kayne 1994.8

Antisymmetric analyses of word order cannot account for the order of postverbal adverbials through base-generation in an ascending structure. They must postulate an alternative mechanism that explains why higher adverbials follow lower adverbials. The hypothesis most commonly pursued makes use of so-called roll-up movement (see Barbiers 1995, Koopman and Szabolcsi 2000, Cinque 2005, 2010).

Cinque (1999) argues that adverbials are specifiers in a uniformly descending base structure. As specifiers precede the node they combine with, the position of postverbal adverbials is the result of movement. The relevant movement regime reverses the order of verb and adverbials. For example, a sequence of two ascending postverbal adverbials can be generated by VP moving across the lower adverbial, followed by movement across the higher adverbial of a constituent containing VP and the lower adverbial. Thus, the base structure is “rolled up,” as shown in (51).

(51)

graphic

Roll-up movement makes available a straightforward translation of the extraposition analysis. For concreteness’ sake, let us assume that selected PPs surface in the specifier of a functional projection IIP, whose position with respect to functional projections hosting adverbials is variable. (It is immaterial for our current purposes whether PPs move to Spec,IIP. It is also immaterial whether IIP is decomposed into a range of functional projections, each with a fixed position in the verbal spine.) The antisymmetric version of the extraposition analysis combines these assumptions (which are summarized in (52a)) with the movement rules in (52b) and (52c).

(52)

  • Roll-up movement around adverbials is optional.

  • Roll-up movement around PPs is obligatory.

  • Selected PPs can shift (leftward) across manner and time adverbials.

Given an input structure [AdvH [AdvL [PP VP]]], roll-up movement around the PP and the adverbials generates the surface representation in (53a). Given an input structure [AdvH [PP [AdvL VP]]], roll-up movement generates (53b). Finally, given an input structure [PP [AdvH [AdvL VP]]], roll-up movement creates (53c). (In these structures, functional projections other than IIP are numbered consecutively.)

(53)

graphic
graphic

Thus, the analysis in (52), like its symmetric counterpart, predicts ascending order for any pair of adverbs in the PP-initial, PP-medial, and PP-final conditions.

8.2 Pas

As things stand, the extraposition analysis and its antisymmetric counterpart appear to be identical in empirical content. This conclusion is strengthened by the observation that the gross constituency of roll-up structures is identical to traditional left-branching VPs (see Abels and Neeleman 2009 for detailed discussion). The structures in (53) group overt material together in the same way as those in (54).

(54)

  • [[[V PP] AdvL] AdvH]

  • [[[V AdvL] PP] AdvH]

  • [[[V AdvL] AdvH] PP]

However, an argument can be constructed in favor of the view that postverbal adverbials are part of a traditional left-branching structure. The argument replicates an argument to the same effect developed in Neeleman 2017 for Dutch. While Dutch is a head-final language, it allows PPs to surface to the right of the verb, in an order suggestive of an ascending structure (see Koster 1974, Barbiers 1995, Neeleman 2017). Thus, the same question that we asked for English is pertinent for Dutch: are postverbal PPs generated in a traditional left-branching structure, or is surface order the result of roll-up movement?

An answer to this question can be based on the distribution of the particle pas. This particle can interact with a temporal expression to indicate that the event described took place at a time later than expected (see Barbiers 1995 for discussion). Thus, the example in (55) implies that the speaker would have expected Olivia to see the import of the relevant facts earlier than on Sunday, in line with the context given. (Here and below, we indicate interaction with pas through underlining.)

(55)

  • [Olivia has been working on this problem for years.]

  • Ze heeft het belang van deze feiten pasop zondag begrepen.

  • she has the import of these facts PAS on Sunday understood

  • ‘She understood the import of these facts only on Sunday.’

The interest of pas for our current purposes lies in the extreme locality required for interaction with a temporal modifier. Barbiers (1995) and Neeleman (2017) argue that pas must immediately c-command the temporal modifier, as stated in (56).9 (The notion of asymmetric c-command in is defined as follows: α asymmetrically c-commands β iff α c-commands β, and β does not c-command α.)

(56) Pas must c-command the XP it interacts with, and there can be no YP such that pas asymmetrically c-commands YP, and YP asymmetrically c-commands XP.

The main evidence for the immediate c-command condition is that pas must be adjacent to the temporal modifier it interacts with in examples like (55). Thus, there is a sharp contrast between and (57a) on the one hand and (57b) on the other.

(57)

  • Ze heeft pasop zondag het belang van deze feiten begrepen.

    she has PAS on Sunday the import of these facts understood

  • *Ze heeft pas het belang van deze feiten op zondag begrepen.

    she has PAS the import of these facts on Sunday understood

The adjacency requirement between pas and a preverbal temporal modifier holds quite generally. For instance, a so-called R-pronoun extracted from a PP can land above pas or below the temporal PP it interacts with, but not between the two (see (58)). Similarly, a contrastively focused PP cannot move to a position between pas and the temporal PP (see (59)).

(58)

  • Eva heeft er pas [na drie jaar] wat tR over gelezen.

    Eva has there PAS after three year something about read

    ‘Eva has only read something about that after three years.’

  • *Eva heeft pas er [na drie jaar] wat tR over gelezen.

    Eva has PAS there after three year something about read

  • Eva heeft pas [na drie jaar] er wat tR over gelezen.

    Eva has PAS after three year there something about read

(59)

  • Eva heeft daar-over pas [na drie jaar] wat tPP gelezen.

    Eva has there-about PAS after three year something read

    ‘Eva has only read something about that after three years.’

  • *Eva heeft pas daar-over [na drie jaar] wat tPP gelezen.

    Eva has PAS there-about after three year something read

  • Eva heeft pas [na drie jaar] daar-over wat tPP gelezen.

    Eva has PAS after three year there-about something read

Strikingly, pas can interact with a postverbal PP, even if other material intervenes.

(60)

  • dat Jan pas [zonder blozen] praatte [na tien jaar therapie]

    that Jan PAS without blushing talked after ten year therapy

    ‘that Jan talked about without blushing only after ten years of therapy’

  • dat Jan pas praatte [zonder blozen] [na tien jaar therapie]

    that Jan PAS talked without blushing after ten year therapy

It can be demonstrated that in such configurations pas must still be extremely local to the postverbal PP. For example, the position of pas fully determines the scope of an extraposed temporal PP. Consider the structures in (61). If postverbal PPs were exempt from the immediate c-command condition, (61a) might permit a reading in which the PP takes scope over AdvP, while (61b) might permit a reading in which AdvP takes scope over the PP.

(61)

  • AdvP pas V PP

  • pas AdvP V PP

However, the fact of the matter is that such readings do not exist. The order in (61a) requires that AdvP take scope over the PP (suggesting the structure [AdvP [ pas V PP]]), while the order in (61b) requires that the PP take scope over AdvP (suggesting the structure [ pas [AdvP V] PP]). This is demonstrated by the data in (62). The example in (62a) unambiguously expresses that Gordon often wants to see two successful trials of a dish before he prepares it for guests. The example in (62b) unambiguously expresses that only after two successful trials will Gordon regularly prepare a dish for guests. Thus, even when the temporal PP is postverbal, pas must immediately c-command it, as required by (56).

(62)

graphic

We are now in a position to put together the argument against an antisymmetric analysis of PP-extraposition in Dutch. We first consider the ungrammaticality of structures in which pas is separated from a preverbal modifier (as in (57b), (58b), and (59b)). Both the traditional and the antisymmetric analyses of Dutch word order can capture this data point. In the antisymmetric representation in (63a), the intervening element induces a violation of the immediate c-command condition: pas asymmetrically c-commands XP, and XP asymmetrically c-commands PP. The same holds of the more traditional representation in (63b).

(63)

graphic

We next consider the cases in which pas interacts with a postverbal modifier (as in (60) and (62)). The traditional account faces no problem here. The relevant examples have a structure in which pas immediately c-commands the temporal PP (see (64b)), which explains why interaction is possible. As it turns out, however, the antisymmetric account assigns (60) and (62) a structure that violates (56): in (64a), pas asymmetrically c-commands the fronted VP, and that VP asymmetrically c-commands PP. Thus, antisymmetry fails to distinguish the grammatical examples in (60) and (62) from the ungrammatical ones in (57)–(59).

(64)

graphic

It is hard to see what could remedy this difficulty within standard Antisymmetry.10 As the structures in (63a) and (64a) are isomorphic, any attempt to accommodate the grammaticality of (64a) is likely to also incorrectly permit (63a).

8.3 Temporal Only

English does not have an unambiguous counterpart of pas. However, only can interact with a temporal modifier to yield a reading in which the event described takes place later than expected. For example, the most natural interpretation of (65) is that Sunday was late for Susan to understand the central question of the paper she was studying, in contradistinction to the more common exclusive interpretation of only, which in the case at hand would imply that Sunday was the only day on which Susan understood the paper’s central question (she understood it neither before nor after; compare She onlyEXCL ate fish on SUNDAY, which implies that she didn’t eat fish on any other day). Here and below, we will refer to this pas-like use of the focus-sensitive particle as “temporal only.”

(65)

  • [Susan has been studying this paper for months.]

  • She onlyT understood its central question last Sunday.

There is reason to think that the immediate c-command condition in (56) extends to temporal only. First, temporal only and the temporal expression that it interacts with must be clausemates, as predicted if interaction requires immediate c-command. Thus, on a temporal construal of only, last Sunday in (66a) must be interpreted as modifying the embedded clause rather than the matrix clause; by contrast, in (66b) it must be interpreted as modifying the matrix clause rather than the embedded clause. The same pattern can be observed with control complements. This is relevant because control complements are smaller than finite complements, and therefore the data in (66c–d) demonstrate a stricter locality effect.

(66)

graphic

Second, temporal modifiers like then can appear preverbally, which makes it possible to test whether the adjacency effect observed in Dutch (57)–(59) is also found in English. The data in (67) suggest that it is: while there is some flexibility in the placement of temporal only and preverbal then, the two cannot be separated.

(67)

  • She may onlyTthen have understood the paper’s central question. (10/10)

  • She may have onlyTthen understood the paper’s central question. (10/10)

  • *She may onlyT have then understood the paper’s central question. (1/10)

Finally, just as with Dutch, pas, the position of temporal only determines the scope of the expression it interacts with vis-à-vis preverbal adverbials. Thus, (68a), like (62a), means that Gordon often requires two successful trials of a dish before he prepares it for guests, while (68b), like (62b), unambiguously expresses that only after two successful trials will Gordon regularly prepare a dish for guests.

(68)

graphic

If the immediate c-command condition indeed extends to temporal only, then the argument against roll-up movement based on Dutch can be replicated for English. On a standard symmetric view of phrase structure, onlyT immediately c-commands the temporal expression in examples like (65), but not in examples like (67c) (where have is a closer c-commanding category). By contrast, an analysis in which the generation of postverbal modifiers involves roll-up movement cannot make the right cut. If anything, temporal onlyT is closer to the expression it interacts with in (69a) (which represents (65)) than it is in (69b) (which represents (67c)). Therefore, both configurations violate the immediate c-command condition.

(69)

graphic

Given this difficulty, it is doubtful that antisymmetry can provide an adequate account of the order of postverbal adverbials. By implication, this means it cannot shed much light on the core observations established in sections 35 either, leaving the symmetric version of the extraction analysis as the more successful contender.

9 The Bigger Picture: Verb Movement in English

The main conclusion that can be drawn from the preceding sections is that if an adverbial intervenes between a verb and a selected PP, this is a consequence of PP-extraposition, rather than the result of the verb moving away from the PP. This conclusion creates a curious problem: how can the verb-raising analysis be wrong, given the widely acknowledged evidence for verb movement elsewhere?11 In double object constructions, for instance, it is very likely that the verb moves across the indirect object (see Larson 1988 and much subsequent work).12

(70) I gave [VP Isabella [V′tV [a copy of War and Peace]]].

One way to resolve the tension between our findings and the evidence from double object constructions and the like is to assume that there are no adverbial attachment sites below the position in which the verb lands. This would make it impossible for verb raising to generate V-Adv-PP orders. We will show, however, that an account along these lines is unlikely to be correct (see Experiments 9 and 10 below).

An alternative solution would be to argue that verb raising does not take place across the board, but is restricted to specific environments. It takes place in double object constructions, for example, but not in structures of PP complementation.

A proposal that makes this cut can be found in Janke and Neeleman 2012. It is based on two assumptions, neither of which is particularly controversial. The first is that objects must meet a condition of Case Adjacency, as already argued in Stowell 1981. The second is that objects do not have to be sisters to the verb, but can in fact be structurally separated from it by elements that may be merged sufficiently low.

This rejection of verb-object sisterhood is in line with our suggestion in section 3.1 that PP-extraposition across manner adverbials can be the result of base-generation. We may simply assume that there is variation in the base position of PP complements with respect to manner adverbials, which may occupy the position marked XP in (71).

(71)

graphic

Evidence for this analysis comes from do so–ellipsis. Crucially, the constituent replaced by so cannot contain a trace bound from outside the ellipsis site (see Haddican 2007 and references mentioned there). We illustrate this restriction for wh-movement in (72a) and for heavy-XP shift in (72b). But if so cannot contain a trace, the PP in (72c) cannot have escaped the ellipsis site through movement—it must have been base-generated outside of it (data from Janke and Neeleman 2012:187).

(72)

    • I [read a novel]1 every week without fail.

    • Really? *So, [which novel] did you do [so]1 last week?

  • *John [read t2 carefully]1 [most of Ecclesiastes]2 and Bill did [so]1 [the entire Song of Solomon].

  • Jordan [met secretly]1 [with his lawyer], and William did [so]1 [with his accountant].

If so, we would expect it to be possible to also generate a structure in which a DP object is structurally separated from the verb, as in (73b).

(73)

graphic

Note, however, that while (73a) satisfies Case Adjacency (Stowell 1981), the alternative structure in (73b) does not. Janke and Neeleman (2012:153) formulate this constraint as in (74); it is violated in (73b) because DP is preceded by XP in its case domain,

graphic
XP, DP
graphic
.

(74)

  • Case Adjacency

    A case-marked DP cannot be preceded by any XP in its case domain.

  • The case domain of a DP consists of the DP itself and any constituents linearly intervening between the DP and the head licensing its case.

The problem with Case Adjacency can be solved by merger of the accusative DP to the left of V′, followed by verb movement across it. In the VP-shell structure thus derived, the DP is right-adjacent to the verb, so that it comes first in its (trivial) case domain

graphic
DP
graphic
.13

(75)

graphic

On this analysis, VP-shell formation is case-driven: the verb moves if and only if a case-marked internal argument is not the first constituent to merge with it. In all other circumstances, verb raising is blocked by economy considerations. This is of course relevant to structures of PP complementation, as PPs do not require case licensing. The structure in (76) is blocked by the more economical structure in (71b) (which is derived without movement and is characterized by the same hierarchy, PP > XP > V).

(76)

graphic

As verbs whose sole argument is a prepositional complement are predicted to remain in situ, the above is sufficient to reconcile our case against the verb-raising analysis of V-Adv-PP orders with the case for short verb movement in VP-shell structures.

The distribution of particles provides independent evidence for the conclusion that verbs may move in the context of selected DPs, but not in the context of selected PPs. It is widely assumed that separation of verb and particle is a function of verb raising (see Koster 1975 and much subsequent work; see Larsen 2014 for an overview of recent literature on particle constructions). If so, the contrast in (77) seems highly relevant.

(77)

  • Sophia looked

    graphic
    up
    graphic
    the information
    graphic
    up
    graphic
    .

  • John walked

    graphic
    out
    graphic
    on Mary
    graphic
    *out
    graphic
    .

Janke and Neeleman (2012) develop the argument as follows. They assume (a) that particles form a complex predicate with the verb (see Johnson 1991, Roeper and Keyser 1992) and (b) that particles project optionally. This allows for two structures, [V Prt] and [V PrtP], each of which can merge with a selected DP or PP. If a particle verb selects a DP and the particle does not project, as in (78a), Case Adjacency is satisfied (notice that the DP’s case domain in (78a) is

graphic
Prt, DP
graphic
; (74) specifically mentions XPs, which implies that only intervening maximal projections violate Case Adjacency). If in the same structure the particle were to project, Case Adjacency would be violated (the case domain in (78b) is
graphic
PrtP, DP
graphic
). As before, English responds to this threat by moving the verb in a VP-shell structure (see (78c)).

(78)

  • Sophia [VP [V looked upPrt] [DP the information]].

  • *Sophia [VP [V looked upPrtP] [DP the information]].

  • Sophia [V′ looked [VP [DP the information [VtV upPrtP]]]].

As only particles that project can host modifiers and complements, it follows that any such extra material is excluded in the V-Prt-DP order, but permitted in the V-DP-PrtP order. This well-known fact is illustrated in (79) using the prepositional modifier right (see Den Dikken 1995).

(79)

  • *Sophia [VP [V looked [PrtP right up]] the information].

  • Sophia [V′ looked [VP the information [VtV [PrtP right up]]]].

If a particle verb selects a PP, Case Adjacency does not come into play. This has two implications. First, there is no longer a trigger for verb movement, not even when the particle projects. In the absence of a trigger, verb movement is blocked, so that the particle must surface adjacent to the verb. Second, modification of the particle is unproblematic, even though it appears between the verb and a selected category.

(80)

  • John [VP [V walked (right) out] on Mary].

  • *John [V′ walked [VP on Mary [VtV (right) out]]].

In section 3.1, we suggested that time adverbials differ from manner adverbials in that they resist merger below selected PPs, so that the V-AdvT-PP order results not from base-generation, as in (81b), but from movement, as in (81c). (We adopted this account because PP-extraposition across time adverbials requires an interpretive license not necessary for PP-extraposition across manner adverbials.)

(81)

graphic

If time adverbials differ from manner adverbials in resisting merger below a selected category, they should also not appear in the lowest position of a VP-shell.

(82)

graphic

This prediction can only be tested if there is a way of forcing a VP-shell structure. Janke and Neeleman (2012) argue that this can be achieved through insertion of an object-oriented floating quantifier. These can appear in VP-shell structures, but not in traditional ascending VPs, because floating quantifiers must be c-commanded by their antecedent DP and (as a language-specific property of English) precede the node they are merged with.

(83)

graphic

The claim that VP-shell formation is necessary to license object-oriented floating quantifiers captures an observation by Maling (1976). Maling notes that object-oriented floating quantifiers are grammatical only if followed by some other element (say, a manner adverb).

(84) Mary read the booksboth *(carefully).

This follows, as VP-shell formation takes place in order to circumvent a potential violation of Case Adjacency. If the only element following the verb is the object, verb-object adjacency is guaranteed, and VP-shell formation is blocked. (Recall that floating quantifiers precede the node they combine with and can therefore not motivate VP-shell formation.)

VP-topicalization can be used to show that the element that “saves” the floating quantifier must be merged low. The structures in (73a) and (75) differ in that (73a) allows XP to be stranded under VP-topicalization, but (75) does not, as verb and object do not form a constituent. If object-oriented floating quantifiers indeed require VP-shell formation, it is predicted that the presence of such an element blocks stranding of XP. This turns out to be correct. Consider the case where XP is a manner adverbial. The examples in (85) show that VP-topicalization can strand or pied-pipe a manner adverbial.

(85)

  • Mary promised she would read the two books I sent,

    and read both books she did carefully. (10/10)

  • Mary promised she would carefully read the two books I sent,

    and read both books carefully she did. (10/10)

Insertion of an object-oriented floating quantifier leads to a sharp drop in the acceptability of the stranding derivation (see (86a–b)). By contrast, pied-piping remains possible (see (86c)).

(86)

  • *Mary promised she would read the two books I sent,

    and read the books both she did carefully. (0/10)

  • *Mary promised she would read the two books I sent,

    and read the books she did both carefully. (0/10)

  • Mary promised she would carefully read the two books I sent,

    and read the books both carefully she did. (9/10)

If object-oriented floating quantifiers indeed diagnose VP-shell structures, then the hypothesis that temporal adverbials resist merger below selected categories predicts that there should be an acceptability contrast between (87a) and (87b), on a par with the contrast involving PP-extraposition across manner and time adverbials.

(87)

  • Charlotte studied the letters both carefully.

  • ??Charlotte studied the letters both yesterday.

To test this prediction, we conducted an AMT experiment (Experiment 9) in which eighty participants were asked to judge ten pairs of test sentences modeled on (87) (as before, the test had a Latin square design and included fillers and questions to check that participants were paying attention to the task). The results show that, as expected, object-oriented floating quantifiers are less acceptable when followed by a time adverbial (see table 14). A one-factor ANOVA confirms that the main effect of adverb type is significant (F1(1, 79) = 30.68, p < .0001).

Table 14

Acceptability of floating quantifiers followed by a manner/time adverbial (N = 80)

V-DP-FQ-AdvMV-DP-FQ-AdvT
4.29 (1.05)  3.11 (1.58) 
 p < .0001  
V-DP-FQ-AdvMV-DP-FQ-AdvT
4.29 (1.05)  3.11 (1.58) 
 p < .0001  

There is a second set of data that confirm that manner adverbials, but not time adverbials, may appear within VP. In our informal judgment, the former may appear in pre-particle position—for instance, in dative constructions—but the latter may not. By contrast, both manner and time adverbials may appear between the particle and the dative PP.

(88)

  • Matt passed the sharp tools foolishly down to Raven.

  • Matt passed the sharp tools down foolishly to Raven.

  • *Matt passed the sharp tools yesterday down to Raven.

  • Matt passed the sharp tools down yesterday to Raven.

The pre-particle attachment site is located below an internal argument and should therefore only be able to host low adverbials. By contrast, strings in which an adverbial surfaces between the particle and the PP complement are generated by PP extraposition. As extraposition can cross time adverbials (with the provisos discussed in section 3.1), we expect post-particle adverbial attachment to be more liberal than pre-particle attachment.

To test these predictions, we conducted an AMT experiment (Experiment 10). The experiment had a Latin square design, with eighty participants who judged ten sets of four test sentences that differed in the order of manner and time adverbials with respect to particles.

As table 15 shows, the results confirm our predictions. A two-factor ANOVA (Adverbial Type × Particle-Adverbial Order) shows that there is no significant main effect of adverbial type (F1(1, 79) = 1.27, p = .26). However, there is a significant main effect of particle-adverbial order (F1(1, 79) = 11.39, p < .001) and, as predicted, a significant interaction between adverbial type and particle-adverbial order (F1(1, 79) = 7.38, p < .001). Post hoc t-tests confirm the nature of the interaction: the two orders of manner adverbials with respect to particles do not attract significantly different scores (p = .65), but there is a clear preference for placing time adverbials in post-particle position (p < .0001).

Table 15

Word order preferences for time and manner adverbials with respect to particles in dative constructions (N = 80)

Prt-AdvMAdvM-PrtPrt-AdvTAdvT-Prt
4.06 (2.06)  3.91 (2.16) 4.43 (2.15)  3.04 (1.76) 
 p = .65   p < .0001  
Prt-AdvMAdvM-PrtPrt-AdvTAdvT-Prt
4.06 (2.06)  3.91 (2.16) 4.43 (2.15)  3.04 (1.76) 
 p = .65   p < .0001  

That manner adverbials can be merged VP-internally is relevant not just for the analysis of PP-extraposition, but crucially also for the issue of verb raising. As already mentioned, the tension between our findings regarding PP-extraposition and the standard view that verbs in English uniformly raise to some low functional head could be resolved if the position in which the verb lands were located below the lowest adverbial attachment site. However, this cannot be true. As Experiments 9 and 10 show, there are adverbial attachment sites low enough for the verb to move across.

We conclude, then, that English has verb raising in some structures but not others. One way to make the right cut is to treat VP-shell formation as triggered by case. The verb remains in situ unless there is a case-marked internal argument merged subsequently to a postverbal XP.14

Notes

1 Judgments from our panel were collected through a series of written questionnaires. Panelists were given the opportunity to ask questions of clarification. We worked with linguists because there is evidence that training on related tasks improves judgments (see Culbertson and Gross 2009), something that might help in the case of scope judgments. The data explored in the article are not part of the standard linguistic curriculum, so panelists would not have received training specific to the questions we are interested in.

2 Experiments 1–8 replicate the findings of earlier experiments that had a different setup (all subjects judged all test sentences; see Neeleman and Payne 2019b). This, we would suggest, underlines the robustness of our findings.

AMT participants were different for each experiment that we report on. In none of the experiments did we have to exclude any participants, as no one rated the fillers in unexpected ways or gave other unusual responses.

The test items we used are available upon request.

3 The data regarding the PP-initial condition go against the long-standing claim that scope among sentence-final adverbs is right-to-left; see Andrews 1983 and much subsequent work. However, this generalization has been called into doubt before, most recently by Bobaljik (2017). Our findings corroborate Bobaljik’s assessment of the data.

4 We do not know why adverbs should precede adverbs they are adjoined to. However, notice that the same restriction holds of adverbs that are adjoined to adjectives.

  • He saw his face in the mirror—sad and [

    graphic
    suddenly
    graphic
    old
    graphic
    *suddenly
    graphic
    ].

5Pesetsky (1989) suggests that reducing the weight of the extraposed PP favors left-to-right scope among adverbials sandwiched between it and the verb. If so, we would suggest an explanation based on prosody. Consider the examples in (i) (where breaks and primary and secondary stress are indicated).

    • {John knocked continuously again} {on the DOOR}.

    • {John knocked} {again continuously} {on the DOOR}.

In (ia), continuously and again are merged independently, yielding right-to-left scope. In (ib), the adverbs cluster, yielding left-to-right scope. The thing to note is that (ib) has a more balanced prosody than (ia), where the PP follows a large prosodic unit. This would favor adverbial clustering if the PP is light.

6 Notice that there are syntactic restrictions on attachment to the focus. As argued in the main text, accidentally can form an adverbial cluster with twice. However, it cannot attach to a DP argument (cf. *John murdered [accidentally BILL]). We do not know whether this should be captured through a selectional requirement, or can be derived from more general principles.

7 Although there is an extensive literature on again, the readings of interest here are rarely discussed (and deserve further exploration). There is general agreement, however, that the attachment site of again (co)determines the presupposition it triggers. For discussion and references, see Beck and Johnson 2004 and Pedersen 2015.

The string in (33b) can also be derived by extraposition of to Louise.

  • Oliver [[[showed the second book] again] to Louise].

This, however, would not yield the interpretation discussed in the text; rather, it yields one in which Oliver showed the second book to Louise and previously showed the second book to someone other than Louise (on a par with Oliver showed the second book again, this time to Louise).

8 Other important work in the same vein can be found in Haider 2010, 2013 and Larson 2014. For reasons of space, we do not discuss this here.

9Barbiers (1995) in fact argues that pas must immediately c-command its semantic argument, which in turn must contain a focus associate. This distinction is not relevant for the data discussed in the main text.

10Barbiers (1995) develops an ingenious alternative roll-up analysis of Dutch PP-extraposition that captures the distribution of pas. He argues that the landing sites of roll-up movement are PP-internal: the lower segment of VP moves and adjoins to a preverbal PP. Thus, the base structure in (ia) is transformed into the derived structure in (ib). We cannot evaluate this analysis here, and so the criticism in the main text applies to the standard roll-up regime, which employs functional specifiers as landing sites, rather than to roll-up movement in general.

    • [VPc PP2 [VPb PP1 VPa]]

    • [VPc [PP2 [VPb [PP1 VPa PP1] tVPa] PP2] tVPb]

11 The discussion in this section abstracts away from the notion that a verb consists of an acategorial root and a categorizing functional head (see Marantz 2001 and much subsequent work). Our conclusions can be reconciled with this hypothesis as long as the categorizer combines with the root before any arguments are merged.

12Larson (1988) uses binding asymmetries and related data to argue for VP-shell structures (basing his conclusions on Reinhart’s (1976) claim that binding is conditioned by c-command). There is reason to think, however, that binding is not always a reliable structural diagnostic (for relevant discussion, see Williams 1997, Hoeksema 2000, Barker 2012, Janke and Neeleman 2012, Bruening 2014). A criterion that has more appeal to us is based on the observation that in OV languages indirect objects precede direct objects in unmarked word order. If this is a function of the preferred order of merger (such that direct objects are merged before indirect objects), then English double object constructions must have a VP-shell structure.

13 In the double object construction, XP is a DP case-marked by the verb’s trace, in line with (74).

14 This account would in principle allow for raising to object and verb movement in accusative-and-infinitive constructions (see Postal 1974, Johnson 1991, Runner 1998, and Lasnik 1999, among many others). That is, one could hypothesize that in examples like John believed Mary sincerely to be the winner, the embedded subject raises into the main clause, and the matrix verb subsequently moves across it to license its case (see (i)). However, we demonstrate elsewhere (Neeleman and Payne 2019a) that such an analysis is defeated by an alternative in which the embedded subject and the matrix verb remain in situ and intervention of matrix material results from extraposition of the embedded predicate, as in (ii).

  • John [believed [Mary [sincerely [tV [StDP to be the winner]]]]].

  • John [[[believed [S Mary tH]] sincerely] [H to be the winner]].

Acknowledgments

Earlier versions of the arguments in this article were developed during a BA course, Current Issues in Syntax, at University College London in 2016 and 2018; we would like to thank the participants for useful suggestions and criticism. Various people have given us food for thought, or have helped us in other ways. In addition to two anonymous reviewers, these include Klaus Abels, Peter Ackema, Stefan Bartell, Abigail Carlson, Jane Chandlee, Mike Donovan, Gordon Hemsley, Alyssa Kampa, Marie Lynagh, Taylor Miller, Larson Stromdahl, Kristina Strother-Garcia, Misako Tanaka, and Hans van de Koot. Special thanks are in order for Benjamin Bruening, who suggested the collaboration that led to this article and provided valuable feedback as it progressed.

References

References
Abels,
Klaus
, and
Ad
Neeleman
.
2009
. Universal 20 without the LCA. In
Merging features: Computation, interpretation, and acquisition
, ed. by
José M.
Brucart
,
Anna
Gavarró
, and
Jaume
Solà
,
60
79
.
Oxford
:
Oxford University Press
.
Ackema,
Peter
, and
Ad
Neeleman
.
2002
. Effects of short term storage in processing rightward movement. In
Storage and computation in the language faculty
, ed. by
Sieb
Nooteboom
,
Fred
Weerman
, and
Frank
Wijnen
,
1
51
.
Dordrecht
:
Kluwer
.
Andrews,
Avery
.
1983
.
A note on the constituent structure of modifiers
.
Linguistic Inquiry
14
:
695
697
.
Barbiers,
Sjef
.
1995
.
The syntax of interpretation
.
Doctoral dissertation, Leiden University
.
Barker,
Chris
.
2012
.
Quantificational binding does not require c-command
.
Linguistic Inquiry
43
:
614
633
.
Beck,
Sigrid
, and
Kyle
Johnson
.
2004
.
Double objects again
.
Linguistic Inquiry
35
:
97
123
.
Bobaljik,
Jonathan David
.
2017
. Knocking twice intentionally. In
A pesky set
, ed. by
Claire
Halpert
,
Hadas
Kotek
, and
Coppe
van Urk
,
25
30
.
MIT Working Papers in Linguistics 80
.
Cambridge, MA
:
MIT, MIT Working Papers in Linguistics
.
Bruening,
Benjamin
.
2014
.
Precede-and-command revisited
.
Language
90
:
342
388
.
Cinque,
Guglielmo
.
1999
.
Adverbs and functional heads: A cross-linguistic perspective
.
Oxford
:
Oxford University Press
.
Cinque,
Guglielmo
.
2005
.
Deriving Greenberg’s Universal 20 and its exceptions
.
Linguistic Inquiry
36
:
315
332
.
Cinque,
Guglielmo
.
2010
.
The syntax of adjectives: A comparative study
.
Cambridge, MA
:
MIT Press
.
Culbertson,
Jennifer
, and
Steven
Gross
.
2009
.
Are linguists better subjects?
British Journal for the Philosophy of Science
60
:
721
736
. 10.1093/bjps/axp032.
Dikken,
Marcel den
.
1995
.
Particles: On the syntax of verb-particle, triadic, and causative constructions
.
Oxford
:
Oxford University Press
.
Emonds,
Joseph
.
1976
.
A transformational approach to English syntax: Root, structure-preserving and local transformations
.
New York
:
Academic Press
.
Ernst,
Thomas
.
2002
.
The syntax of adjuncts
.
Cambridge
:
Cambridge University Press
.
Haddican,
Bill
.
2007
.
The structural deficiency of verbal pro-forms
.
Linguistic Inquiry
38
:
539
547
.
Haider,
Hubert
.
2010
.
The syntax of German
.
Cambridge
:
Cambridge University Press
.
Haider,
Hubert
.
2013
.
Symmetry breaking in syntax
.
Oxford
:
Oxford University Press
.
Hoeksema,
Jack
.
2000
. Negative polarity items: Triggering, scope, and c-command. In
Negation and polarity: Syntactic and semantic perspectives
, ed. by
Laurence
Horn
and
Yasuhiko
Kato
,
115
146
.
Oxford
:
Oxford University Press
.
Jackendoff,
Ray
.
1972
.
Semantic interpretation in generative grammar
.
Cambridge, MA
:
MIT Press
.
Janke,
Vikki
, and
Ad
Neeleman
.
2012
.
Ascending and descending VPs in English
.
Linguistic Inquiry
43
:
151
190
.
Johnson,
Kyle
.
1991
.
Object positions
.
Natural Language and Linguistic Theory
9
:
577
636
.
Kayne,
Richard
.
1994
.
The antisymmetry of syntax
.
Cambridge, MA
:
MIT Press
.
Koopman,
Hilda
, and
Anna
Szabolcsi
.
2000
.
Verbal complexes
.
Cambridge, MA
:
MIT Press
.
Koster,
Jan
.
1974
.
Het werkwoord als spiegelcentrum
.
Spektator
3
:
601
618
.
Koster,
Jan
.
1975
.
Dutch as an SOV language
.
Linguistic Analysis
1
:
111
136
.
Lambrecht,
Knud
.
1994
.
Information structure and sentence form: Topic, focus, and the mental representation of discourse referents
.
Cambridge
:
Cambridge University Press
.
Larsen,
Darrell
.
2014
.
Particles and particle-verb constructions in English and other Germanic languages
.
Doctoral dissertation, University of Delaware, Newark
.
Larson,
Richard
.
1988
.
On the double object construction
.
Linguistic Inquiry
19
:
335
391
.
Larson,
Richard
.
2014
.
On shell structure
.
London
:
Routledge
.
Lasnik,
Howard
.
1999
. Chains of arguments. In
Working Minimalism
, ed. by
Samuel David
Epstein
and
Norbert
Hornstein
,
189
215
.
Cambridge, MA
:
MIT Press
.
Maling,
Joan
.
1976
.
Notes on quantifier-postposing
.
Linguistic Inquiry
7
:
708
718
.
Marantz,
Alec
.
2001
.
Word and things
.
Ms., MIT, Cambridge, MA
.
Neeleman,
Ad
.
2017
.
PP-over-V meets Universal 20
.
Journal of Comparative Germanic Linguistics
20
:
3
47
.
Neeleman,
Ad
, and
Amanda
Payne
.
2019a
.
On matrix-clause intervention in accusative-and-infinitive constructions
.
Syntax
. 10.1111/synt.12174.
Neeleman,
Ad
, and
Amanda
Payne
.
2019b
.
PP extraposition and the order of adverbials in English
.
Ms., University College London/Haverford College. https://ling.auf.net/lingbuzz/003663
.
Payne,
Amanda
.
2018
.
Adverb typology: A computational characterization
.
Doctoral dissertation, University of Delaware, Newark
.
Pedersen,
Walter
.
2015
.
A scalar analysis of again-ambiguities
.
Journal of Semantics
32
:
373
424
.
Pesetsky,
David
.
1989
.
Language-particular processes and the Earliness Principle
.
Ms., MIT, Cambridge, MA
.
Pollock,
Jean-Yves
.
1989
.
Verb movement, Universal Grammar, and the structure of IP
.
Linguistic Inquiry
20
:
365
424
.
Postal,
Paul
.
1974
.
On raising: One rule of English grammar and its theoretical implications
.
Cambridge, MA
:
MIT Press
.
Reinhart,
Tanya
.
1976
.
The syntactic domain of anaphora
.
Doctoral dissertation, MIT, Cambridge, MA
.
Roeper,
Thomas
, and
Samuel Jay
Keyser
.
1992
.
Re: The abstract clitic hypothesis
.
Linguistic Inquiry
23
:
89
125
.
Rohrbacher,
Bernhard
.
1994
.
English main verbs move never
.
Ms., University of Pennsylvania, Philadelphia
.
Rooth,
Mats
.
1985
.
Association with focus
.
Doctoral dissertation, University of Massachusetts, Amherst
.
Runner,
Jeffrey
.
1998
.
Noun phrase licensing
.
New York
:
Garland
.
Sprouse,
Jon
.
2011
.
A validation of Amazon Mechanical Turk for the collection of acceptability judgments in linguistic theory
.
Behavior Research Methods
43
:
155
167
.
Stowell,
Tim
.
1981
.
Origins of phrase structure
.
Doctoral dissertation, MIT, Cambridge, MA
.
Williams,
Edwin
.
1997
.
Blocking and anaphora
.
Linguistic Inquiry
28
:
577
617
.
Williams,
Edwin
.
2014
.
Compositionality and derivation
.
Ms., Princeton University, Princeton, NJ. https://semanticsarchive.net/Archive/GYzZmI5O/paper.pdf
.