We designed a novel lexical-interference task, where participants first must read a sentence and then are presented with two words from that sentence. Their task is to identify which of the two words appeared last. When applied to filler–gap dependencies, the idea is that, if there is reactivation of lexical material in a gap site (as shown in Wagers & Phillips 2014, among other works), we will see interference from a lexical item that is part of a filler, if the gap occurs after the other word choice. For example, in the sentence in (8a), the word patient comes last, but nurse might be reactivated at the gap site after with. If it is, then we could see interference in the form of longer reaction times and/or lower accuracy, compared to a baseline with no filler–gap dependency, such as in (8b).

We used this idea to test reconstruction of complements and adjuncts to nouns. Unlike previous work on reconstruction, which has only compared complements of nouns to adjuncts, we used a three-way comparison: the head noun itself, a complement to the noun, and an adjunct to the noun, as in (9). This allows us to compare the behavior of heads with that of complements and that of complements with that of adjuncts in a single experimental paradigm.

The lexical item at issue here is question. In the head condition, it is part of a compound head noun. In the complement condition, it is the complement of the head N. In the adjunct condition, it is an adjunct to the head N.

Applying this to the idea illustrated in (8) yields the paradigm in (10). Here, the word question appears before the word relations—thus, the correct answer to the question in (12) is relations—but if there is reactivation of question after relations (i.e., at the hypothesized reconstruction site, indicated with “—”), then we would expect participants to be slower and/or less accurate in their responses to the wh conditions in (10) than to the baseline conditions in (11).

The baseline items (base) have a structure similar to the corresponding wh items and involve mostly the same lexical items, but the first of the two words (here, question) is no longer part of a filler–gap dependency.

We expect reactivation to manifest as a longer reaction time to the question, “Which word appeared last in the sentence?” in the wh condition than in the base condition. We could (instead or in addition) find decreased accuracy in the responses to this condition. All researchers agree that the head N reconstructs, so we minimally expect reactivation in the head condition. In contrast, everyone agrees that adjuncts do not reconstruct; we therefore expect no reactivation effect for adjunct. How complement will pattern is the open question: with head (indicating that complements do reconstruct) or with adjunct (indicating that they do not)?

We ran a forced-choice comprehension experiment, with a 2×3 within-subjects, within-items design. Two factors were manipulated: Movement (base vs. wh) and Position (head vs. complement vs. adjunct), corresponding to the tokens in (10) and (11).

Data from two participants was excluded because they did not self-identify as L1 English speakers, and data from a further eight participants was excluded because they answered fewer than half of the yes–no comprehension questions correctly.1 Of the data from the remaining 198 participants, 209 trials (0.09% of the dataset) were removed prior to analysis based upon the time that was taken to read the test sentence: because these sentences were on average 20 words long, we removed trials with reading times of under 4,000 ms (an estimated mean by-word reading time below 200 ms; n = 195), and we also removed trials with long reading times of over one minute in total (n = 14).

We analyzed the remaining 2,167 trials by fitting three 2×3 mixed-effects regression models. Each model had crossed random effects for participants and items (Baayen et al. 2008) and was fit using the Lme4 package (R 4.0.2: Bates et al. 2015). Our first manipulation contrasted the wh conditions with the base conditions (Movement, two levels) and was sum coded as +1/2 for wh and –1/2 for base. The second manipulation, Position (three levels), was contrast coded using centered Helmert contrasts. The first coefficient, head versus complement + adjunct, compared the head conditions (coefficient: +2/3) with the complement and adjunct conditions, pooled (coefficient: –1/3 for each). The second coefficient, complement versus adjunct, contrasted complements (coefficient: –1/2) with adjuncts (coefficient: +1/2): this contrast compares the complement and adjunct conditions to each other, directly (the head conditions do not participate in this comparison; their coefficient is 0). We used the maximal random-effect structure justified by the experiment design that would allow for model convergence (Barr et al. 2013). p values were calculated via a Satterthwaite approximation, computed using the LmerTest function (Kuznetsova et al. 2017).

We first analyzed participants’ accuracy in responding to the probe question (i.e., identifying which of the two words appeared last in the sentence). On average, participants answered this correctly 81.7% of the time, but the 2×3 logistic-regression model revealed no significant main effects or interactions (all ps > .13), meaning that participants’ accuracy levels did not change as a function of our experimental manipulations.

Next, we analyzed time taken to respond to the probe question (“Which word appeared last?”).2 The mean reaction times are shown in figure 2. Descriptively, we note that only the head condition is affected by the manipulation of Movement; the complement and adjunct conditions are not. For the purposes of statistical analysis, these values were log transformed and analyzed via a 2×3 mixed-effects linear regression following the procedure described above.

We found a main effect of Movement, where reaction times for wh conditions were slower overall than base conditions (4,136 ms vs. 4,012 ms; β = –0.014, SE = 0.007, t = –2, p = .048). There were no other significant main effects (both ps > .6). Most importantly, however, we find an interaction of Movement with head versus complement and adjunct, pooled (β = –0.035, SE = 0.015, t = –2.4, p = .019), but critically no interaction of Movement with complement versus adjunct (p = .98). The full model is shown in table 1. We then unpacked the interaction by running a second model with Movement nested within Position (this single model takes into account the variability in the full dataset). This model revealed a significant effect of Movement for head nouns, where movement triggers slower reaction times (4,164 ms vs. 3,846 ms; β = –0.04, SE = 0.01, t = –3.1, p = .002), but not for complements or adjuncts (ps > .84).3

We find evidence that reaction time in identification of a target word is sensitive to interference due to lexical reactivation at gap sites. Critically, however, only the head noun causes this interference: complements and adjuncts to the noun do not. This indicates that only the head noun is reactivated at the gap site; neither complements nor adjuncts to the moved N are reactivated, since neither gave rise to an interference effect. There is, however, a potential alternative explanation for this result. Given that the NPs in the head conditions (e.g., discussion question) are less structurally complex than those in the complement and adjunct conditions (e.g., discussion of a question, discussion after a question), it is possible that more time is required to read the sentences in the complement and adjunct conditions (i.e., before participants exited the sentence display to respond to the probe question). This extra processing cost might well be enhanced in the sentences with wh movement, because one must hold the relevant wh filler (e.g., which discussion of a question) in memory before forming the relevant wh dependency. Critically for our present purposes, such extra reading time would lead to a better encoding of word order in the sentence, and this could have shortened participants’ response times to the probe question independently of any reconstruction effects (or absence of them).

We examined this possibility by analyzing the time taken by participants to read the sentence displays (recall that these were presented in full). The mean reading times, by condition, are shown in figure 2. Like with response times to the probe question, these response times were log transformed prior to statistical analysis. There were two main effects. First, participants took less time to read sentences in the head conditions than in the complement and adjunct conditions, pooled (10,510 ms vs. 10,798 ms/11,318 ms; β = –0.026, SE = 0.007, t = –3.7, p = .0002). Participants also took longer to read sentences in the complement conditions than in the adjunct conditions, overall (β = 0.025, SE = 0.008, t = 2.6, p = .008). These differences in reading times are likely to reflect differences in the length of the sentences in the head, complement, and adjunct conditions: the head conditions were generally the shortest, and the adjunct conditions were the longest.4 Returning to the overall sentence reading times, we note further that these reading-time effects were across the board; that is, they occurred for both wh and base conditions and therefore are not tied to wh movement specifically. Importantly, neither of the two interactions was significant (ps > .89), and in fact, the numerical trend for the interaction of head versus complement + adjunct with Movement patterns in the opposite direction from what we would expect if the patterns in figure 1 were driven by longer reading times for the complement and adjunct conditions than for the head condition, when wh movement occurs. (That is, numerically, participants took more time to read the sentences in the wh–head condition than in the wh–complement and wh–adjunct conditions, by comparison with the base–head, base–complement, and base–adjunct sentences.)

In sum, it appears that our critical interaction for response time (see figure 1) is unlikely to be due to participants spending more time reading the wh–complement and wh–adjunct sentences compared with the wh–head ones. We conclude that the most likely explanation for our findings is that only the head N is reactivated at the gap site.

As discussed in the introduction, there is currently a debate regarding whether complements inside moved NPs reconstruct for Binding Condition C. We have found that the complement of an N is not reactivated at the gap site. This finding is consistent with the majority of the experimental work on reconstruction for Binding Condition C in English, which has found that neither complements nor adjuncts to N trigger a Condition C effect. It is not consistent with Stockwell et al. 2021 and 2022, which claimed to find reconstruction of complements for Condition C. However, it should be noted that both of these works only compared complements and adjuncts to each other and did not compare either of those to a baseline without wh movement. Finding a difference between complements and adjuncts does not show that complements reconstruct for Binding Condition C. Stockwell et al. 2022’s results are also consistent with Tollan & Heller 2022’s findings regarding the complement–adjunct asymmetry in pronoun resolution discussed in the introduction: recall that, for pronouns in subject position (the same configuration used in Stockwell et al.’s experimental items), speakers of Niuean showed higher rates of co-reference with a prior adjunct antecedent than with a complement antecedent, in simple declarative sentences with no movement/reconstruction involved.

Furthermore, in Stockwell et al. 2022, acceptance rates for co-valuation with a complement to a moved N were three times as high as a surface Condition C violation (Stockwell et al.’s “Bad” item; 2.19 vs. 0.72). As the authors note, their participants seem to have responded with preferences, with their ratings for each pair of interpretations adding up to approximately 8. Taking the scores out of 8, co-valuation with a complement was accepted at rates comparable to what was found in other experiments: 2.19/8 = 27%. In Stockwell et al. 2021, the rate of acceptance was similarly high: 1.95/8 = 24% (in the Short condition; much higher in the Long condition). These rates are almost identical to what was found in Leddon & Lidz 2006 and Bruening & Al Khalaf 2019. As such, the results of Stockwell et al. 2021 and 2022 are actually consistent with the other experimental findings and our own. All of these experiments indicate that complements to N, just like adjuncts, do not reconstruct. Only the head noun itself reconstructs.

We conclude that the body of empirical evidence thus far points towards a view in which neither complements nor adjuncts of Ns reconstruct. As we have shown, neither is reactivated in sentence comprehension, but head nouns are. This of course leaves open the possibility that complements reconstruct in a manner that is not observable through either pronoun interpretation or reactivation, but, unless there is some other way to detect it, we do not see any way of knowing whether such reconstruction could exist.

Within theories of syntax that utilize movement, there are basically three approaches to representing movement. In the Government and Binding approach (e.g., Chomsky 1981), what was left behind in a gap site was a special element, a trace, which had properties distinct from those of the moved item. Chomsky 1993 proposed the second approach, the copy theory, which says that what occupies the gap site is a copy of the moved element. Finally, there is also a multidominance approach to movement (e.g., Citko 2005). In the multidominance approach, a phrase is dominated by two nodes. In wh-movement examples like ours, an object of a verb would simultaneously be dominated by VP (it would be sister to V) and CP (sister to C-bar or its equivalent).

Our findings rule out the multidominance approach. We can see no way in such an approach for the wh filler to differ from what is present at the gap site. They are literally the same phrase, in this approach. If our interpretation of our findings is correct and only the head noun is represented at the gap site, then multidominance approaches cannot be correct.

Our findings are compatible with the trace theory, provided that traces trigger reactivation of the head of the filler. Cross-modal lexical priming has always found reactivation of the head of a filler at the gap site (e.g., Swinney 1991, Hickok et al. 1992, Love & Swinney 1996), and (as far as we know) this has always been taken to be compatible with the trace theory. One could view a trace as essentially like a pronoun, and pronouns have also been found to reactivate the lexical content of their antecedents (Leiman 1982, Shillcock 1982, Nicol 1988).

As for the copy theory of movement, the original arguments given for it in Chomsky 1993 have not been borne out. As mentioned in the introduction, there are no convincing cases of reconstruction of anaphors inside moved NPs, since anaphors inside NPs are exempt from Binding Condition A. All experimental evidence indicates that there is no reconstruction for Binding Condition C, either.5 Our findings now show that there is no reactivation of dependents of a moved N at the gap site. One might be tempted to take this as evidence against the copy theory. We do not think that it is, however, for several reasons.

First, there are proposals within the copy theory that are compatible with the facts. These proposals allow copies in a chain to differ from each other. For instance, Lebeaux 1988 and Chomsky 1993, among many others, have proposed that adjuncts can late-merge counter-cyclically, merging with the head noun after it has moved. This would only need to be extended to complements of N (or else Ns do not take complements, only adjuncts, as in Reuland 2011, Adger 2013, Grimm & McNally 2013). Another proposal comes from Bruening & Al Khalaf 2019: that the derivation proceeds left to right, so all of the material in the filler is merged first in the moved position. At the gap site, copying is minimal, such that it copies only what is needed. This is only the head noun itself. A third idea is that there can be vehicle change (Fiengo & May 1994) in lower copies, with R-expressions being replaced by pronouns (e.g., Safir 1999, Sauerland 2003). However, Hunter & Yoshida 2016 argues that there cannot be vehicle change in movement, only in ellipsis, and Adger et al. 2017 argues that vehicle change cannot be the right explanation for the lack of Condition C reconstruction the authors found in their experiments. It is also not clear whether vehicle change is compatible with our results. As noted above, cross-modal-lexical-priming studies have found that pronouns activate the lexical content of their antecedents (Leiman 1982, Shillcock 1982, Nicol 1988); if so, then we would not expect R-expressions and any pronouns co-valued with them to differ on measures of lexical reactivation. We therefore reject vehicle change as a possible explanation for our findings. However, since there are at least two analyses within the copy theory of movement that are compatible with our results, we do not think that our study should be taken as evidence against the copy theory of movement.

The second reason not to rule out the copy theory comes from predicate fronting. Complements of moved predicates (verbs, adjectives, prepositions), in contrast with complements of moved Ns, have been found to reconstruct for Binding Condition C (Leddon & Lidz 2006, Adger et al. 2017, Bruening & Al Khalaf 2019). If the gap position were just a trace, this would not be expected. One could suggest, as Huang 1993 did, that the moved predicate contains a representation of the subject of the predicate. On this analysis, Binding Condition C is violated internally to the moved predicate, not in the reconstructed position. However, this cannot account for the fronted PPs that Bruening & Al Khalaf 2019 studied. It was found that an adjunct inside the fronted PP did not reconstruct, while the complement of the P did. If binding took place entirely in the fronted constituent, the adjunct should have given rise to a Binding Condition C effect. These findings require the copy theory of movement, along with less than a full copy in the gap position (in particular, adjuncts are not represented in the gap position).

Third, anaphors and pronouns that undergo A-bar movement themselves can reconstruct, for instance in clefting:

In the trace theory, the trace of A-bar movement is supposed to be an R-expression subject to Binding Condition C, regardless of what the moved NP is. The grammaticality of the anaphor in (15) and (16a) is then unexpected, as is the grammaticality of the pronoun in (16b). One could perhaps claim that the anaphor is being used as a logophor rather than as an anaphor, but this would not help to explain why the trace of the pronoun does not act like an R-expression in (16b). In contrast, on the copy theory, the anaphor and the pronoun are the heads of the moved NP, and so there is a copy of them in the gap position. We then expect Binding Conditions A and B to apply so that the anaphor and the pronoun act as though they have not moved, which is exactly the case.

Given all of these considerations, we conclude that the copy theory is currently the best approach to representing movement chains. However, our data and all the other experimental data regarding Condition C indicate that copies in gap positions are not complete copies. In particular, arguments and adjuncts of moved Ns are not represented in the gap site; only the head N is.