2.1 Items

Minimal triplets were constructed like the following:

All variants included an additional modifier, because modifiers were judged to make the sentences more natural, as mentioned above. Gendered quantified noun phrases were chosen so that there would be no question of the suitability of the pronoun covarying with the quantifier.

The task was a simple judgment of acceptability. Gordon & Hendrick (Reference Gordon and Hendrick1997) showed that subjects rate sentences lower in acceptability if they cannot resolve the referent of a pronoun. The logic of Experiment 1 is that, if binding of the pronoun her by the quantifier is unacceptable in any of (5a–c), then that pronoun will lack a referent, and subjects will rate it lower in acceptability. Note that this does not mean that subjects find such sentences unacceptable. All of these sentences have a perfectly acceptable reading where the pronoun refers to some individual not named in the sentence. What Gordon & Hendrick (Reference Gordon and Hendrick1997) showed was that, if a sentence presented out of the blue has a pronoun with no referent available in that sentence, then they rate the sentence lower in acceptability. In addition, Kush, Lidz & Phillips (Reference Kush, Lidz and Phillips2017) used acceptability judgments in experiments involving strong and weak crossover with wh-questions, and showed that subjects do indeed rate sentences that violate weak crossover lower than sentences that do not. This task should then be a valid comparison of the three sentences in (5).

The Base sentence without fronting of the PP was included as a comparison. In the Base sentence in (5a), the pronoun is neither c-commanded nor preceded by the quantifier. Subjects are expected not to permit the bound variable interpretation and to only allow a referential reading. On this reading the pronoun is not provided with a referent in the presented sentence, and so we expect, given the findings of Gordon & Hendrick (Reference Gordon and Hendrick1997) and Kush et al. (Reference Kush, Lidz and Phillips2017), that they will rate it lower. We can then compare Top and LocInv examples like those in (5b–c) to the Base condition. If Culicover & Levine (Reference Culicover and Levine2001), Doggett (Reference Doggett2004), and Diercks (Reference Diercks, Everaert and van Riemsdijk2017) are correct and topicalization and locative inversion differ in their ability to support a bound variable reading, then Top and LocInv items should be rated differently by the participants in this experiment. Specifically, Top should pattern with Base and be rated lower than LocInv.

A further check of the logic of the task was provided by an additional pair, as follows:

In the Base sentence in (5a), it might be possible for the PP to take scope over the subject, leading to higher than expected judgments of acceptability. In (6b), in contrast, the theoretical literature is in general agreement that the quantifier should be unable to take scope over the matrix pronoun across a finite clause boundary (or should be able to only with great difficulty). The Var item in (6a), in contrast, should easily allow a bound variable reading. The Var and NoVar items then provide a check on the experiment: If the subjects are performing as expected, judgments of acceptability on the Var items should be significantly higher than those on the NoVar items, given that, in Var (6a), subjects can resolve the referent of the pronoun internal to the sentence, but in NoVar (6b), they cannot (it has to refer to an entity not named in the sentence). If this is indeed the finding, then the logic of the task is confirmed and we can legitimately compare the Base, Top, and LocInv items in (5).

Note that absolute values of the acceptability judgments are meaningless and nothing will be made of them; all we can do is compare judgments on minimally different items to see if they are being treated differently. In the present experiment, the crucial comparison is between Top and LocInv ((5b) vs. (5c)), and their comparison to Base (5a). Given that Var/NoVar are not minimally different from the set of Base/Top/LocInv, no comparison across these will be attempted. All of these items are also expected to be judged to be acceptable, so it is the comparisons that are relevant, not the absolute values (low versus high, or positive versus negative z-scores, for instance).

Nine sets like those in (5) were constructed and divided into three lists so that no subject saw more than one item from each set of the triplet in (5). The Var and NoVar items in (6) were divided among the three lists so that no subject saw both members of any pair. The complete list of items appears in the appendix. Each list included three exemplars of each condition, for a total of 15 experimental items. Each list also included 33 filler items, where 16 of these were items for an unrelated experiment involving adverb placement. None of these items involved quantifiers or pronouns, and they had very different syntactic structures, so they effectively served as distractor items. The other 17 fillers were simply fillers, six of which were acceptable and six of which were unacceptable, with the other five somewhere in between. The task was to rate how acceptable each sentence was on a scale of 1 to 7, 1 being ‘completely unacceptable’ and 7 being ‘completely acceptable’.

2.2 Subjects

In total, 82 subjects were recruited using Amazon Mechanical Turk, limited to people with IP addresses in the United States. All reported being native speakers of English. Two of the subjects were thrown out because they gave the same rating of acceptability to all of the sentences and were therefore clearly not engaged in the task. This left 80 subjects whose data entered the analysis. All 80 gave judgments on the filler items in the expected pattern. Because the age of the speakers was important to the unrelated study involving adverbs, all of the subjects were over 40 years old. There is no reason to think that judgments on variable binding should differ by age, so this should not be an issue. The average age was 46. Of the 82 subjects, 40% identified as female, 60% as male. Subjects were paid 50 cents for their participation. The entire experiment typically took subjects about six minutes to complete.

3.1 Items

Items were constructed in triplets like the following, with a forced-choice question that would go with each item in the triplet:

The two possible answers to the question were the matrix subject and a quantificational phrase in a PP in the embedded clause. The clauses of interest (those with the fronted PP) were always embedded under a verb like say or think that permits locative inversion and PP topicalization in its complement. The purpose of this was to provide a second possible referent for the pronoun, in the form of the matrix subject.

The logic of this task follows that of the experiments in Bruening & Al Khalaf (Reference Bruening and Al Khalaf2019): If the There and LocInv conditions differ in the possibility of variable binding, we should see this difference reflected in fewer choices of the quantificational antecedent in the There condition (i.e. fewer choices of every knight). If A-bar movement like topicalization never enables variable binding, we ought to expect to see the There condition patterning with the Base condition. However, there could be extra-grammatical reasons why those two might differ (preferences or processing heuristics based on linear order, for example). The crucial comparison is the There-LocInv one. If there is a grammatical difference in variable binding, as Doggett (Reference Doggett2004) and Diercks (Reference Diercks, Everaert and van Riemsdijk2017) claim, responses to these two conditions should differ. Conversely, if there is no difference in variable binding between the There and LocInv conditions, as proponents of the null expletive analysis of locative inversion assert (Postal Reference Postal2004; Bruening Reference Bruening2010), and as we might expect given the results of Experiment 1, then responses to these two conditions should not differ.

Note that the definiteness restriction is violated in the There condition, but this was judged to only make the sentences slightly awkward, not unacceptable. Subjects should be able to answer the forced-choice question even if they find the There condition degraded compared to the LocInv condition. Speakers are assumed to be able to use their grammar to compute antecedent-pronoun relations, and variable binding relations, even if a sentence is not fully acceptable (so long as the structural conditions on antecedence are met).

Six sets like those in (8) were constructed and divided into three lists so that no subject saw more than one item from each set. The complete list of items appears in the appendices. Each list included two exemplars of each condition for a total of six experimental items per subject. Each list also included twelve filler items. Six of these were items for an unrelated experiment, also involving adverb placement. None of these items had PP fronting or involved quantifiers, so they effectively served as distractor items. The other six were simply fillers. This time subjects were instructed to answer different questions for each item. The items for this experiment asked for the referent of the pronoun, as did two of the fillers (where one choice should have been excluded by Binding Condition C, and subjects did indeed never select that choice). The items for the unrelated adverb experiment asked subjects to rate how acceptable each sentence was on a scale of 1 to 7, 1 being ‘completely unacceptable’ and 7 being ‘completely acceptable’. Four of the fillers also asked subjects to rate acceptability. Subjects were expected to find one of these acceptable and three unacceptable.

3.2 Subjects

Once again, 80 subjects were recruited using Amazon Mechanical Turk, limited to people with IP addresses in the United States. All reported being native speakers of English. None of the subjects were thrown out, as they all answered as expected on the filler items. Because the age of the speakers was important to the unrelated study involving adverbs again, all of the subjects were over 40 years old. There is no reason to think that judgments on variable binding should differ by age, so this should not be an issue. The average age was 45.65. Of the 80 subjects, 33.75% identified as female, 62.5% as male (the others did not answer). Subjects were paid 60 cents for their participation.

4.1 The analysis of English locative inversion

This finding removes one of the arguments in favor of the PP-as-subject analysis of locative inversion, and argues in favor of the expletive subject analysis, given that variable binding is yet another way in which locative inversion patterns with sentences with an overt there. For a long list of ways in which they pattern the same, see Postal (Reference Postal2004), Bruening (Reference Bruening2010), and also Bruening (Reference Bruening2016), where locative inversion and there sentences pattern together but behave differently from quotative inversion sentences for a number of grammatical phenomena. To give one example, in tag questions with locative inversion, the tag is there:

There in the tag could not be a pronominal for the fronted PP, because there is not an appropriate pronoun for this PP:

There in the tag must instead correspond to a null expletive there in the locative inversion clause.

The findings of this paper are then most compatible with the null expletive analysis, which posits a null there in Spec-TP of locative inversion clauses:

The PP never occupies Spec-TP in this analysis, rather, it is in topicalized position (assumed to be Spec-CP). For a detailed exposition and defense of this analysis, see Bruening (Reference Bruening2010).Footnote ³

4.2 Disputes over data

This study also joins a growing list of cases where judgments from some professional linguists have diverged from those of the population at large. For one case involving reconstruction for Binding Principle C, see Adger et al. (Reference Adger, Drummond, Hall and van Urk2016) and Bruening & Al Khalaf (Reference Bruening and Al Khalaf2019). Other cases include multiple questions (Clifton, Fanselow & Frazier Reference Clifton, Fanselow and Frazier2006), adjectival passives of raising to object verbs (Bruening Reference Bruening2014), and nominalizations of raising verbs (raising to subject and raising to object; see Bruening Reference Bruening2018). It is becoming increasingly important for linguists to support any judgments that they report in their work using multiple sources of evidence, such as corpus evidence and survey results. It should no longer be acceptable practice for one linguist to simply assert that the judgments of another are faulty; it is incumbent upon anyone who cares about the validity of data to ascertain what the facts actually are using every available source of data.

Of course, this raises the question of why some professional linguists disagree about the data. One possibility is that there are actually two grammars, with some speakers permitting variable binding from a topicalized PP and others not. The two camps of professional linguists would differ in their grammars. To examine this possibility, I further analyzed the data from Experiment 2. Seven participants out of 80 did not choose the quantificational antecedent for the pronoun on either of the two There sentences they saw. It is possible that these seven participants rule out variable binding from a fronted PP in sentences with an overt there. Of these seven, four did answer with the quantificational antecedent on at least one of the LocInv sentences. This means that four out of 80 subjects show the pattern published by Doggett (Reference Doggett2004) and Diercks (Reference Diercks, Everaert and van Riemsdijk2017). If there is such a grammar, then, it belongs to a tiny minority: only five percent of English speakers have this grammar (assuming the participants of Experiment 2 are representative).

Another possibility is that the acceptability of variable binding is related not to the A- versus A-bar status of the fronted PP, but to information structure. It is possible that the speakers who do not like variable binding with an overt there have a harder time constructing the relevant information structure for the clause. I discuss this possibility next.

4.3 The nature of weak crossover

The standard analysis of weak crossover treats it as arising from A-bar movement (for an overview, see Safir Reference Safir, Everaert and van Riemsdijk2017). In the standard treatment, a quantificational element can only bind a pronoun as a variable from its highest A-position. If it crosses a pronoun by A-bar movement, it is unable to bind that pronoun. This is exactly what happens in the topicalization examples used in Experiment 1:

This is presumably derived from the Base sentence in (13) by A-bar movement:

In the standard analysis, the quantifier in the topicalized PP in (12) should be able to bind the pronoun as a variable only to the extent that it can in (13), where it presumably occupies the equivalent of an A-position for a PP.

However, this is not what was found in the two experiments reported here. In Experiment 1, participants rated the Top sentence in (12) significantly higher than the Base sentence in (13). In Experiment 2, participants chose the quantificational antecedent much more frequently in the topicalization items (the There items) than they did in the Base items. This seems to indicate that topicalizing the PP actually enables variable binding, contrary to the predictions of the standard account.

An alternative to the standard account is proposed by Eilam (Reference Eilam2011). Eilam (Reference Eilam2011) argues that weak crossover is not about the A- versus A-bar status of the position occupied by the quantifier; rather, the condition for a quantifier to bind a pronoun as a variable is that the quantifier has to be construable as a topic, while the pronoun has to be contained in an XP that can be construed as a focus. For instance, the reason that variable binding is not very acceptable in the Base sentence in (13) is that the default topic of a clause is the subject. The pronoun is then contained in a constituent that is a topic rather than a focus. At the same time, postverbal material in English is typically focused, making the quantifier a focus rather than a topic. In contrast, binding from a subject is uniformly acceptable, because the subject is the default topic and postverbal material is usually focused:

Here, every princess is easily (and most naturally) construable as a topic, while her lady in waiting can be construed as a focus (and is, by default).

In Eilam’s account, weak crossover arises in a standard A-bar movement scenario like wh-movement because in wh-movement, the quantifier (the wh-phrase) is the focus, not the topic:

At the same time, the pronoun is contained in the subject, which is the default topic, not the focus.

In this account, if a topicalized quantifier can be construed as a topic, while the constituent that contains the pronoun can be construed as a focus, then topicalization (A-bar movement) will enable variable binding. I suggest that this is what is going on in (12): participants can construe every princess as a topic, while her lady in waiting can be construed as the focus. Her lady in waiting is the subject, which is the default topic, but if the fronted PP is construed as the topic instead, then the subject can be in focus. I suggest that this is exactly what is going on, and it is why participants readily permit variable binding. They can easily access the information structure that is required in Eilam’s account of weak crossover.

In Experiment 2, the same information structure can be constructed for sentences with an overt there:

In the embedded clause, every knight can be construed as a topic, while the postverbal NP his squire can be construed as a focus. Hence, subjects answer with every knight as the antecedent for the pronoun a high percentage of the time. In contrast, in the corresponding Base sentence, the pronoun is contained in a preverbal subject, which is the default topic, while the quantifier is postverbal and hence a default focus:

Participants therefore respond much more rarely with the quantifier as the antecedent for the pronoun.

I take the findings of Experiments 1 and 2 to indicate that the standard account of weak crossover is not correct. A-bar movement does not uniformly lead to weak crossover. In some cases, it enables variable binding instead. The analysis of weak crossover proposed by Eilam (Reference Eilam2011) is the best extant account of this phenomenon. It also makes sense of the pattern of results obtained in the two experiments reported here.

Turning to locative inversion, it is certainly compatible with the requisite information structure. Consider the following line from The Three Billy Goats Gruff:

Here, the bridge has just been mentioned and so it is a topic, while the great ugly troll has not yet been mentioned and is being introduced as a presentational focus. Having the subject postverbal in fact makes it a focus by default. Given that the PP is or can be topical, while the postverbal subject is a focus, locative inversion then easily allows variable binding in Eilam’s account:

Even out of the blue, the fronted PP can be construed as a topic, while the postverbal subject is the focus.

It is also possible to construe a preverbal subject as the focus, in the presence of a topicalized PP. We saw this in (12) above. In such cases, it may be that adding modifiers to the sentence helps to make this information structure more salient. This may be why (20b) sounds more natural than (20a) (examples repeated from the introduction):

In my judgment, simply adding focal stress to collar also helps in (20a). I take this to support the account proposed in Eilam (Reference Eilam2011).

Let us now return to the topic of possible speaker variation, and the question of whether there might be two different grammars, one which permits variable binding from a fronted PP and one which does not. As noted in the previous subsection, if there is a grammar that does not permit such binding, it belongs to a very small minority of speakers. Another possibility, however, is that such speakers have a harder time constructing the relevant information structure. Suppose that such speakers have a very strong preference for a preverbal subject to be the topic, even in the presence of a fronted PP. Then such speakers will not readily allow variable binding of a pronoun in the preverbal subject.

As for the distinction that such speakers are purported to make between clauses with and without an overt there, there may be multiple factors at play. Recall that Postal (Reference Postal2004) and Bruening (Reference Bruening2010) argued that any degradation in variable binding with an overt there is due to the definiteness effect with overt there rather than to the failure of variable binding. This is probably one factor. Another might be information structure again. We therefore predict that, if the pronoun can be contained in an indefinite focused phrase, while the fronted PP is saliently made a topic, then even speakers in this minority group should allow variable binding. Consider the following attempt:

In this pair, the context sentence sets up the dogs as the topic. The postverbal NP is an indefinite, meaning that there should be no definiteness effect with overt there. The prediction is that every English speaker will allow variable binding in both members of this pair, even the English speakers who do not like variable binding with a fronted PP and an overt there. As of yet, I have not been able to test this prediction.

4.4 Overall conclusion

This paper has reported on two experiments involving weak crossover and locative inversion. It found that the majority of English speakers permit a quantifier inside a fronted PP to bind a variable in both a preverbal subject and a postverbal subject with an overt there, contra some claims in the theoretical literature. These findings are most compatible with the information-structural analysis of weak crossover proposed by Eilam (Reference Eilam2011), rather than the standard account where it is triggered by A-bar movement; and they are most compatible with the null expletive analysis of locative inversion, rather than the analysis where the PP is the subject.