3.1 Experiment 1: CPs within NPs

Experiment 1 examined wh-movement of NPs that include CPs within them. We designed experimental items in a $2\times 2$ design with factors argument versus adjunct (whether the CP was an argument of the N, or a relative clause) and wh-movement versus no wh-movement. The following is a sample set from the paradigm.

We used embedded wh-questions for two reasons: first, to provide another possible referent for the pronoun, in the form of the matrix subject; second, so that subjects would not be confused about what question they should be answering. Rather than trying to ask them directly about coreference possibilities, we gave them a forced choice question about who the referent of the pronoun was. For instance, for the set of four sentences above, the question was Who authorized the announcement? A: the staffer; B: Hillary Clinton. In half of the experimental items the relevant R-expression was the second choice of the two, and in half it was the first choice, to guard against any bias for picking the first or second of two choices.

This method has the disadvantage that we cannot know for sure from any results whether subjects truly disallow a given referent. However, it has the advantage that it is a very natural task and does not require that the subjects try to engage in any metalinguistic analysis, which could be faulty or trigger reasoning outside the grammar. In practical terms, we believe that this method can provide a reasonable amount of evidence for or against a grammatical constraint on coreference. If a referent truly is ruled out by the grammar, then we should see choices of that referent at a rate close to zero. If there is no grammatical constraint, then we should see subjects behaving at around chance, which in a two-choice task is 50%.Footnote ^[4] Since the experimental items are set up to directly compare argument and adjunct CPs, we will also be able to see whether naive subjects treat them differently, as some of the theoretical literature would lead us to expect.Footnote ^[5]

We distributed the experimental items into four lists so that each subject saw only one of the four sentences in the set in (15). We created eight such sets, so that each subject judged two of each type. The complete set of items appears in the Appendix. We also included 16 filler items with comprehension questions, to check that subjects were engaged in the task. Subjects therefore read 24 sentences and answered one question about each. Within each list, the order of sentences was randomized.

Using Amazon Mechanical Turk, we recruited 20 subjects for each list, for a total of 80 subjects. Subjects were paid 60 cents for participating. We limited recruitment to subjects with IP addresses in the United States, and also asked them to say what their native language was and what country they were from. Subjects who reported a language other than English as their native language were excluded from the analysis. Subjects were also excluded if they got more than two questions wrong on the filler items. Five subjects were excluded for one of these reasons, leaving a total of 75 subjects in the analysis: 49 male, 25 female, 1 other. The age range was 20–62 (median age $=$ 30).

Results are shown in Table 1, in the form of percentage of ‘B’ responses. This is the response that violates Binding Condition C in the NoWh conditions, and would violate it if there were complete reconstruction in the Wh conditions. As can be seen, where there is no wh-movement, subjects choose the Condition C violating answer at a rate close to zero. In contrast, in the two wh-movement conditions, the answer that would violate Condition C if there were complete reconstruction is chosen at a rate close to chance (50%). There is a difference between arguments and adjuncts, such that we see a higher percentage with adjuncts, but this difference is small.

Table 1 Results of Experiment 1, CPs.

Statistical analysis was run using R (R Core Team 2012). Responses were analyzed by means of linear mixed-effect modeling using the R-package lme4 (using glmer with family binomial). The two fixed effects in the analysis were wh-movement and the argument/adjunct distinction. Following the recommendations in Barr et al. (Reference Barr, Levy, Scheepers and Tily2013), we first included by-subject random intercepts and slopes and by-item random intercepts. However, the maximal model failed to converge, so we ended up simplifying to a model using only random intercepts for both subjects and items. Table 2 shows the results.

Table 2 Summary of fixed effects in the mixed-effects model for Experiment 1.

The last column lists the $p$ -values. As can be seen, only a main effect of wh-movement is significant. There is no main effect of the argument/adjunct distinction, and no interaction. This means that, in these types of examples, wh-movement simply bleeds Binding Condition C. R-expressions contained within adjuncts and R-expressions contained within arguments in a fronted wh-phrase are free to corefer with a lower pronoun.

We conclude from this that there is no reconstruction for Condition C at all with CPs that front as part of wh-NPs. While the literature has claimed that argument CPs do reconstruct for Binding Condition C, we find no evidence for such reconstruction in the broader population. Our subjects chose the Condition C violating antecedent for the pronoun at a rate close to chance, indicating that there is no grammatical constraint against that coreference.

3.2 Experiment 2: PPs within NPs

Experiment 2 examined wh-movement of NPs that include PPs within them. As noted above, Safir (Reference Safir1999) claimed that the argument–adjunct distinction is most robust with PP dependents of fronted NPs. We followed the same experimental design, and constructed experimental items in a $2\times 2$ design with factors argument versus adjunct (whether the PP was an argument of the N, or an adjunct) and wh-movement versus no wh-movement. The following is a sample set from the paradigm.Footnote ^[6]

As before, subjects were given a forced-choice question regarding the referent of the pronoun. For the set of four sentences above, the question was Who considers the portrait valuable? A: the chambermaid; B: the countess. Once again, in half of the experimental items the relevant R-expression was the second choice of the two, and in half it was the first choice, to guard against any bias for picking the first or second of two choices.

As in Experiment 1, the experimental items were distributed into four lists so that each subject saw only one of the four sentences in the set in (16). Eight such sets were created, so that each subject judged two of each type. See the Appendix for the full set of items. Sixteen fillers were used, so subjects read 24 sentences and answered one question about each. The order of the items was randomized within each list.

We recruited 20 subjects for each list, for a total of 80 subjects, again using Amazon Mechanical Turk. The recruitment was limited to subjects with IP addresses within the United States. Subjects were excluded from the study if they reported that their native tongue was not English or if they answered more than two questions wrong on the filler items. Five subjects were excluded for one of these reasons, leaving 75 subjects in the analysis: 34 male, 41 female. The age range was 21–69 (median age $=$ 31).

Results are shown in Table 3, in the form of percentage of ‘B’ responses. This is the response that violates Binding Condition C in the NoWh conditions, and would violate it if there were complete reconstruction in the Wh conditions. As can be seen in Table 3, where there is no wh-movement, subjects choose the Condition C violating answer at a rate close to zero. In contrast, in the two wh-movement conditions, the answer that would violate Condition C if there were reconstruction is chosen at a higher rate, though now at a lower rate than in Experiment 1 (lower than chance, but still much higher than zero). Note that the rates here roughly match what Leddon & Lidz (Reference Leddon and Lidz2006) found for adults on similar items.

Table 3 Results of Experiment 2, PPs.

We analyzed the results using R (R Core Team 2012). As before, the responses were analyzed via linear mixed-effect modeling, using the R-package lme4. Wh-movement and the argument/adjunct distinction were the fixed effects in the analysis. As in Experiment 1, we included by-subject random intercepts and slopes and by-item random intercepts and slopes, following the recommendations in Barr et al. (Reference Barr, Levy, Scheepers and Tily2013). Because the maximal model failed to converge, we simplified the analysis to include only random intercepts for subjects and items. The results are shown in Table 4.

Table 4 Summary of fixed effects in the mixed-effects model for Experiment 2.

As can be seen from the last column in Table 4, there is a main effect of wh-movement, but no main effect of the argument–adjunct contrast. Contrary to the expectations of the standard view in the literature, there is also no interaction between the argument–adjunct contrast and wh-movement. As with CPs, we find no statistically significant difference between argument and adjunct PPs regarding reconstruction for Condition C. We see only a main effect of wh-movement: Wh-movement bleeds Condition C.

We conclude that the literature that has claimed that there is a difference between argument and adjunct PPs is incorrect. We find no effect of the argument–adjunct distinction. Instead, we see only a main effect of wh-movement, such that fronting an NP that contains a PP bleeds Condition C for any R-expressions contained within the PP.Footnote ^[7]

There is still the question of why the rates are below chance in this experiment, as opposed to Experiment 1 with CPs, where the rates were within a few percentage points of chance (50%). We suspect that the difference is the linear distance between the R-expression and the pronoun. In the wh-movement conditions in Experiment 1, there are multiple words in between the R-expression in the fronted wh-phrase and the pronoun in the lower clause (typically 4–5 words). In contrast, in Experiment 2, the R-expression and the pronoun are either adjacent or they are separated by only 1–2 words. Adger et al. (Reference Adger, Drummond, Hall and van Urk2016) find an effect of linear distance on coreference between a pronoun and an R-expression. The further the distance between the R-expression and the pronoun is, the more likely subjects are to choose coreference. Our guess is that this difference in linear distance is what is behind the lower rate in Experiment 2. In any case, although the rate is below chance, it is still significantly higher than zero, which we take to indicate that there is no grammatical constraint on coreference. Most importantly, there is no difference between argument PPs and adjunct PPs.

3.3 Interim summary

Experiments 1 and 2 have found no support for the argument–adjunct asymmetry in reconstruction for Binding Condition C that previous literature has reported. Rather, speakers freely permit coreference between an embedded pronoun and an R-expression contained within a fronted wh-phrase.

We might conclude at this point that A-bar movement simply bleeds Condition C. Given this possibility, we decided to check a data point from Reinhart (Reference Reinhart1976) that was very important to the development of syntactic theory. It will turn out that not all A-bar movement bleeds Condition C.

3.4 Experiment 3: Fronted PPs

The important data point involves fronted PPs. Reinhart (Reference Reinhart1976, Reference Reinhart1983) used these to argue for the importance of c-command in Binding Condition C. These present a contrast like the following example.

According to Reinhart (Reference Reinhart1976, Reference Reinhart1983), (17b) is a Condition C violation, with the pronoun subject c-commanding the R-expression to its left. Bruening (Reference Bruening2014) showed that this explanation is untenable, and instead analyzed (17b) as involving reconstruction of the fronted PP. Since the PP semantically modifies the VP, it reconstructs to the edge of the vP phase (in the phase theory of Chomsky Reference Chomsky, Martin, Michaels and Uriagereka2000; see (Bruening Reference Bruening2014: 361–363) for details). In this reconstructed position, the pronoun subject binds the R-expression. One of the arguments for this analysis is the fact that further embedding the R-expression obviates Condition C.

Further embedding the R-expression would not change backward c-command, so Reinhart’s analysis would still expect a Condition C violation. In contrast, if a relative clause inside a fronted constituent does not need to reconstruct (as we saw above), then the reconstruction analysis predicts this contrast.

Given our findings that most subjects do not show reconstruction for Condition C at all, we began to wonder whether Reinhart’s original contrast is actually real. We decided to test both it and Bruening’s embedding effect using the paradigm and method from Experiments 1 and 2. The two factors were now R-expression versus pronoun and embedded versus non-embedded. All sentences in the paradigm had a fronted PP. As before, the clause with fronting was embedded, to provide another possible antecedent for the pronoun. In this case, the forced-choice question had to vary depending on the item. We show a sample set of items, with the question corresponding to each item, below.

Given the literature, we now expect the following pattern of responses. First, in the Pro NoEmb and Pro Emb conditions, there should be no grammatical condition ruling out either R-expression as a referent for the pronoun, so we expect around chance performance, or 50% ‘B’ responses (but note that in the Pro Emb condition, it is most natural to take the pronoun to refer to the higher NP, and this is what we find in the results). Second, if the Rexpr NoEmb condition really is a Condition C violation, we should expect close to zero ‘B’ responses. In contrast, in the Rexpr Emb condition, if the relative clause does not need to reconstruct, we expect around 50% ‘B’ responses again. If the contrast is not real and there is no reconstruction for Condition C at all, we expect roughly 50% ‘B’ responses in all four conditions.

As before, we distributed the experimental items into four lists so that each subject saw only one of the four sentences in the set in (19). We created eight such sets, so that each subject judged two of each type. The complete set of items appears in the Appendix. We also included filler items with comprehension questions, to check that subjects were engaged in the task. Within each list, the order of sentences was randomized.

Eighty subjects were recruited via Amazon Mechanical Turk. Ten subjects were excluded for getting more than two comprehension questions wrong on the filler items, leaving 70 subjects whose data entered the analysis. Of these, 33 were female and 37 male; the age range was 23–70 with a median age of 36.5.

Results are shown in Table 5, in the form of percentage of ‘B’ responses. This is the response that would violate Binding Condition C if there were complete reconstruction of the fronted PP in the Rexpr conditions. We see a high percentage of ‘B’ responses in the Pro NoEmb condition, and a lower but still high percentage of ‘B’ responses in the two Emb conditions. An embedded R-expression does not seem to differ from an embedded pronoun, as Bruening (Reference Bruening2014) claimed. In the original Reinhart sentence, the Rexpr NoEmb condition, the rate of ‘B’ responses is much lower, closer to zero.

Table 5 Results of Experiment 3, fronted PPs.

We take these results to support the reconstruction account. We find a high percentage of ‘B’ responses in all conditions except the Rexpr NoEmb condition, exactly the one that Reinhart (Reference Reinhart1976) claimed was a Condition C violation and Bruening (Reference Bruening2014) analyzed as Condition C under reconstruction.

As in Experiments 1 and 2, statistical analysis was run using R, following the same procedures. The two fixed effects in the analysis were R-expression versus pronoun and embedding versus no embedding. Once again, the maximal model failed to converge, so we simplified to a model with only random intercepts for both subjects and items. Table 6 shows the results.

Table 6 Summary of fixed effects in the mixed-effects model for Experiment 3.

We see significant main effects of both factors and a significant interaction in the last column of Table 6. The effects go in the opposite direction. For pronouns, embedding reduces the number of ‘B’ responses because, as noted above, it is most natural in the Pro Emb condition to take the pronoun to refer to the higher NP. In contrast, embedding increases the number of ‘B’ responses with R-expressions. We interpret this to mean that reconstruction takes place in the NoEmb condition, but embedding the R-expression results in it no longer being reconstructed in the Emb condition.

We conclude that Reinhart (Reference Reinhart1976, Reference Reinhart1983) was correct to analyze Near Dan, he saw a snake as a Condition C violation on the coreferential reading, and Bruening (Reference Bruening2014) was correct to analyze this as Condition C under reconstruction. The fronted PP reconstructs to a position lower in the sentence, where it is in the command domain of the pronoun subject. A relative clause within the PP does not need to reconstruct, so if the R-expression is contained within such a relative clause, there is no Condition C violation.

At this point, we can also say something about the response percentages. In a two-choice task, chance performance is 50%. However, we only expect to see 50% if there is no grammatical constraint at work and all other factors are equal, in particular if there is no pragmatic reason to prefer one referent or another. As we saw in Experiment 2, decreasing the linear distance between the R-expression and the pronoun decreased the coreferent responses. In Experiment 3, we now see the same thing: in two of the conditions, it is most natural to take the pronoun to refer to the other NP, and so we see choice of the ‘B’ response at levels between 20 and 30%. It therefore appears that if there is no grammatical constraint but there is some pragmatic bias toward one of the readings (based on, e.g., linear order and distance), then we seem to get between 20 and 30% selection of the choice that is less preferred. In contrast, where there is a grammatical constraint, choice of the ungrammatical referent is low and close to zero. In Experiment 3, it is 8.6%. This might seem to be high, but it is significantly lower than 20–30% and is probably just noise.

To summarize to this point, we previously saw no reconstruction of dependents of fronted Ns. Now, we can see that this is not general: it is not the case that A-bar movement simply bleeds Condition C. We do see reconstruction for Condition C with fronted PPs. The complement of a fronted P does reconstruct and give rise to a Condition C violation. Condition C is only bled completely with dependents of fronted Ns.