1.1 Antecedent accessibility and grammatical sensitivity

It is often assumed that candidate antecedents are identified by their morphological features (e.g. Carreiras et al. 1996; Arnold et al. 2000; Rigalleau, Caplan & Baudiffier 2004; Chow, Lewis & Phillips 2014). For example, when the parser first comes across the pronoun it in (1) and (2), many models hold that it should probe recent memory for an eligible antecedent with 3^rd-person, singular, inanimate features.¹ As the NPs the medal and every medal match those features, their representations should be retrieved.

Morphologically-driven retrieval is a rapid way of finding potential antecedents, but a naïve version of the mechanism is susceptible to error if it ignores the structural position of matching phrases vis-à-vis the pronoun. In addition to identifying acceptable potential antecedents, such a mechanism is liable to identify matching NPs that are disqualified from antecedence by structural constraints. For example, the QP every medal in (3a) would be identified as a potential antecedent for the pronoun it, even though it is not allowed to grammatically bind the pronoun due to lack of c-command.

Recent studies have investigated whether antecedent retrieval erroneously accesses structurally illicit phrases as potential antecedents based on morphological feature-match (Chow, Lewis & Phillips 2014; Cunnings, Patterson & Felser 2015; Kush, Lidz & Phillips 2015; 2017; but see Badecker & Straub 2002). The results of these studies suggest that retrieval does not access matching NPs in grammatically unacceptable positions.

Kush et al. (2015) tested whether non-c-commanding QPs are accessed as antecedents for pronouns that they are grammatically prohibited from binding. In one experiment the authors had participants read sentences like those in (5), which contained a potential antecedent phrase (scout) inside a relative clause and a critical pronoun (him) in the matrix clause. The design manipulated the gender of the potential antecedent (boy/girl scout) such that it either matched or mismatched a critical pronoun. The design also manipulated whether the potential antecedent was referential (5a) or quantification (5b).

The authors found a gender-mismatch effect at the critical pronoun in (5a): participants had more difficulty reading the pronoun when it mismatched the referential NP than when it matched. The effect was interpreted as evidence that participants could access the matching referential antecedent and had trouble when there was no matching antecedent. Gender-match did not, however, affect the processing of the pronoun when the potential antecedent was a QP (5b). Kush and colleagues concluded that the asymmetry in gender-mismatch effects indicated that antecedent retrieval does not access non-c-commanding QPs that are not acceptable antecedents (see also Cunnings et al. 2015).

Moulton and Han (2018) pointed out that the inaccessibility of the non-c-commanding QPs in the above studies is compatible with two interpretations. The first interpretation is that retrieval uniformly ignores non-referential phrases that do not c-command a pronoun. The alternate possibility is that retrieval simply ignores non-referential phrases that are not acceptable antecedents for a pronoun, without direct reference to c-command. Under this interpretation accessibility would not be narrowly determined by syntactic scope (c-command), but rather semantic scope (see Barker 2015; Ruys & Winter 2011 for discussion of semantic scope).

Moulton & Han (2018) tested the processing of exceptional covariation dependencies to tease these two possibilities apart. The authors tested the accessibility of QPs in constructions like (6a), where a pronoun (he) co-varies with a QP (each boy), even though the QP does not c-command the pronoun from within the fronted temporal adjunct. Two self-paced reading experiments compared the processing of such exceptionally co-varying pronouns to standard bound and co-referential pronouns. Both experiments varied gender-match between a critical pronoun he/she and an antecedent phrase (each/the boy). Experiment 1 compared exceptionally-co-varying pronouns in (6a) to (6b), where each boy c-commanded the pronoun and syntactic binding was possible. Experiment 2 compared (6a) to (6c), in which co-reference between the boy and he was possible.

Gender-mismatch effects were found at the syntactically bound (6b) and co-referential (6c) pronoun, but not on the exceptionally co-varying pronouns (6a). Based on the absence of a gender-mismatch effect in (6a), Moulton and Han concluded that the non-c-commanding QP was inaccessible to antecedent retrieval and that the general accessibility of non-referential antecedents was strictly governed by syntactic scope/c-command rather than semantic scope. In this regard, the co-varying QP was not treated as the true antecedent, as the pronoun did not establish a direct anaphoric dependency with the phrase. Co-variation was assumed to arise via a different route (e.g. via establishing a relationship with an alternative mediating situation variable, as proposed by Moulton 2017).

Moulton and Han (2018) speculated that a similar processing profile should be observed with other exceptionally co-variant dependencies, such as donkey pronouns. If this prediction is correct, then the indefinite phrase with which the donkey pronoun co-varies should not be accessible to antecedent retrieval. We tested this possibility.

2.1 Participants

32 native-speakers of Norwegian (23 Female, mean age = 30.5) from the greater Trondheim community participated in the study. Participants were compensated with a gift certificate good for one movie ticket.

2.2 Materials

Test sentences contained a critical object pronoun and an indefinite NP that was a potential antecedent for the pronoun. Critical pronouns were either the masculine human-denoting object pronoun ham ‘him’ or the feminine object pronoun henne ‘her’ (Faarlund et al. 1997). Critical pronouns immediately followed the main verb. Antecedent NPs were human-denoting NPs (e.g. ei datter ‘a daughter’ in Table 1). Antecedent NPs were embedded within a relative clause (RC) attached to the main clause subject.

24 sets of experimental sentences were created in accordance with a 2 × 2 factorial design that crossed the factors GENDERMATCH and DEPENDENCYTYPE. An example item is given in Table 1 and a full set of materials can be found at the project’s OSF page: https://osf.io/g5u48/. GENDERMATCH controlled whether the critical pronoun matched the indefinite antecedent NP in grammatical gender. The gender of the critical pronoun was counter-balanced across items. DEPENDENCYTYPE determined whether the indefinite antecedent NP could license a co-referential or a donkey reading of the pronoun. DEPENDENCYTYPE was controlled by varying whether the main subject NP was quantificational (Hver far ‘every father’ in Table 1) or referential (Faren ‘the father’).² The main subject was two words in the quantificational conditions, but only one word in the referential conditions. This difference reflects the fact that the definite article in simple NPs is a bound morpheme that suffixes to the head noun. The suffix takes different forms depending on the number and gender of the noun (Julien 2005).

When the subject NP was quantificational, the antecedent NP could license a donkey reading of the pronoun. When it was referential, the antecedent NP could only establish co-reference with the pronoun.

We wished to avoid the possibility that the post-verbal position of the pronoun would allow participants to anticipate features of a likely object based on the semantics or selectional properties of the main verb. Under such a scenario, it would be difficult to attribute gender-mismatch effects to retrieval, rather than violations of a prediction. In order to minimize this possibility we chose main verbs that were not close semantic associates of the antecedent NP and that did not strongly predict reference to the potential antecedent NP. We used main verbs that could take a diverse range of direct objects (e.g., få ‘get’) or verbs whose canonical object was not the antecedent NP (e.g., kjøre ‘drive’ in Table 1).

Materials were written in Bokmål, one of the two official standardized forms of written Norwegian (alongside Nynorsk; Venås 1993; Vikør 1995). Bokmål is the preferred written standard for an estimated 85–90% of the total population, and all students in Norwegian schools receive instruction in both standards (Staalesen 2014).

Test sentences were distributed in a Latin Square design across 4 separate lists containing 78 additional sentences for a total of 102 sentences. 54 of these sentences were fillers of varying length and complexity. 24 were sentences from an unrelated experiment on the processing of relative clauses. All sentences were followed by a yes-no comprehension question to encourage participants to read carefully. One item was removed from analysis due to a typo in the pre-critical region.

2.3 Procedure

Participants were run on a desktop PC using the Linger software package (Doug Rohde, MIT) in a self-paced phrase-by-phrase moving window paradigm (Just, Carpenter & Woolley 1982). Phrases for presentation consisted of one or two words, as demarcated with slashes in Table 1. Each trial began with a sentence masked by dashes appearing on the screen. Letters and punctuation marks were masked. Spaces were not masked, leaving word-boundaries visible. As the participant pressed the spacebar, a new word appeared on the screen and the previous word was re-masked. All sentences were followed by a yes/no comprehension question.

2.4 Analysis

Reaction times (RTs) below 100 ms were removed, affecting less than 0.1% of all data. RTs were subsequently log-transformed and the resulting logRTs were used as the dependent variable in analysis.

Analysis was restricted to two regions of interest: the critical pronoun and the region immediately following the pronoun (henceforth the post-pronoun region). The composition of the post-pronoun region varied across items, but in most items it contained either a lone preposition (e.g. til ‘to’) or a preposition plus an infinitival marker (e.g., om å) presented simultaneously. Statistical analysis was conducted using linear mixed effects models estimated using the lme4 package (Bates, Maechler & Bolker 2011) in R (R Core Development Team 2014). Models included simple-difference sum-coded fixed effects of DEPENDENCYTYPE (Donkey: 0.5, Referential: –0.5) and GENDERMATCH (Mismatch: 0.5, Match: –0.5). Instead of using models with “maximal” random effects structures (Barr et al. 2013), we determined the “parsimonious” random effects structure for each individual model according to the model selection procedure described in Matuschek et al. (2017; see also Bates et al. 2015; Baayen et al. 2017) using the RePsychLing package. All models contained random intercepts for participants and items. The parsimonious model used for analysis of reading times in the pronoun region also contained a random slope for DEPENDENCYTYPE by item. The model used for analysis of the post-pronoun region contained by-subject and by-item random slopes for DEPENDENCYTYPE and GENDERMATCH, and a by-item random slope for their interaction. Estimated p-values were calculated using the Satterthwaite approximation implemented in the lmerTest package (Kuznetsova, Brockhoff & Christensen 2017).

2.5 Results and discussion

2.5.2 Reading times

Log-transformed reading times are plotted in Figure 1. A summary of statistical effects for the critical and post-pronoun region are in Table 2.

Figure 1

Region-by-region log-transformed reading times beginning from the antecedent NP. Error bars indicate standard errors of the mean.

Table 2

Summary of statistical effects for the Pronoun and Post-pronoun regions.

	Pronoun Region			Post-Pronoun Region
	Beta	SE	t	Beta	SE	t
DEPENDENCYTYPE	–0.01	0.03	–0.39	–0.01	0.03	–0.23
GENDERMATCH	0.08	0.02	3.14**	0.20	0.03	6.34***
TYPE:MATCH	0.01	0.05	0.25	0.03	0.06	0.48

• **, *** indicate significance at p < .01, .001, respectively.

Clear and sustained gender-mismatch effects began at the pronoun and continued throughout the remainder of the sentence for both donkey and co-referential sentences. Statistical analysis confirms that the gender-mismatch effect was significant at the critical pronoun (t = 3.14, p < .01). Matching pronouns were read more quickly than mismatching pronouns (average raw RT difference = 32 ms). There was no sign of a DEPENDENCYTYPE × GENDERMATCH interaction, indicating that the gender-mismatch effects were comparable across the donkey and the co-referential sentences. The gender-mismatch effect was even larger in the post-pronoun region (t = 6.34, average raw RT difference = 85 ms). Again, the absence of an interaction suggested that there were no differences in the magnitude of the effect across different sentence types.

Keir Moulton (personal communication) asked whether the pairwise differences that we observe could perhaps stem from a confound unrelated to gender-mismatch: In a subset of our materials the pronoun in the mismatch sentences was the same gender as the matrix subject (e.g. Hver far/Faren ‘every/the father’ and ham ‘him’ in the example in Table 1). If the critical pronoun were to retrieve the matrix subject in the mismatch condition, then the difficulty that we labeled a gender-mismatch effect could alternatively be attributed to a violation of Principle B. Instead of slowing down because the pronoun failed to match any accessible NPs in the sentence, participants might have slowed down under this interpretation because they recognized they attempted to link the pronoun to the local subject and then recognized that such a dependency would be grammatically illicit.

We consider this alternative explanation unlikely because most recent reading experiments suggest that participants do not consider a matching local subject as an antecedent for an otherwise unheralded object pronoun in violation of Principle B (Lee & Williams 2008; Chow, Lewis & Phillips 2015; though see Kennison 2003). Nevertheless, we conducted a post-hoc analysis to determine whether the supposed confound could account for our results.

We identified experimental items in which the mismatching pronoun and the matrix subject were the same gender. Of our set of 24 items, there were 10 items that conservatively fit the description. Six of the ten items had matrix subjects with definitional gender (father, princess, man, woman) that matched the gender of the critical pronoun. The matrix subjects of the other four items were nouns that were not male or female by definition, but which may have been biased towards the matching gender (e.g., student, school kid, officer).

We removed the ten items from our data set and reran the analysis above. If the mismatch effect had been due to the proposed confound, we would expect the pairwise effect of gender-mismatch at the pronoun and the spill-over region to disappear with the subset of items. This is not what we observed. Even with the reduced power, we still observe significant gender-mismatch effects in the pronoun and spillover regions (t = 2.08, 4.46; ps < .05, .001, respectively). Thus, we conclude that difficulty in the mismatch conditions can safely be apportioned to gender-mismatch.

3.1 Relation to formal analyses of Donkey Anaphora

Insofar as gender-mismatch effects reliably diagnose antecedent retrieval – and there is some degree of parser-grammar isomorphism (Boland 2005; Marantz 2005) – our results suggest that donkey pronouns enter into a dependency with the indefinite phrase. We briefly discuss how this finding aligns with the different formal analyses of donkey pronouns. As we discuss below, the idea that the indefinite serves as an antecedent is, in principle, compatible with variants of both of the dominant semantic approaches to donkey pronouns. Thus, our goal is not to use our results to adjudicate between the two theoretical accounts.⁴ Instead, we wish to point out how the results could be accommodated under each theory and in which cases alignment is harder than in others.

Broadly speaking, there are two approaches to interpreting donkey pronouns. Covert definite (sometimes called e-type or D-type) accounts hold that donkey pronouns are interpreted as covert definite descriptions, where apparent covariation between the indefinite and the pronoun does not occur via binding by the indefinite phrase (Cooper 1979; Evans 1980; Heim 1990; Elbourne 2001; 2005; a.o.). Instead covariation requires quantification over a separate hidden variable, the identity of which differs by theory: some (e.g. Cooper 1979) posit an individual variable bound by the QP every skier as in (9a), while others (Heim 1990; Elbourne 2001) invoke covert quantification over situation variables, as in (9b).

In contrast to covert definite accounts, Dynamic accounts reject a formal distinction between donkey, traditional bound, and co-referential pronouns, treating them all as instances of semantically bound individual-denoting variables (Kamp 1981; Heim 1982; Groenendijk & Stokhof 1991; Brasoveanu & Dotlacil 2016; a.o.). Pronouns are identified with discourse referents introduced by quantificational and non-quantificational noun phrases alike. Discourse referents “live” – i.e. can serve as antecedents – as long as their corresponding phrase has semantic scope. Under this interpretation, the [ACCESSIBLE] feature would correspond directly to semantic scope.

The finding that donkey pronouns directly retrieve indefinites as antecedents can be reconciled with covert definite accounts if retrieval functions to provide the content of the covert description (rather than supplying an individual or binder). The finding is perhaps most difficult to reconcile with models such as Cooper (1979), where the content of the definite description is supplied by a “contextually-salient” relation that is available to the semantics and not via connection to the indefinite itself. By contrast, Elbourne’s (2002; 2005) D-Type analysis of donkey anaphora is relatively easy to align with the assumption. Elbourne proposes that donkey pronouns are definite descriptions whose nominal content has been elided. The content of the elided noun must be recovered under identity with the indefinite antecedent as shown in (8).

If donkey pronouns use the indefinite to provide content for a definite description (and co-referential and syntactically bound pronouns are not interpreted as covert definite descriptions), then the purpose of retrieval varies as a function of dependency. However, the fact that different dependencies are computed does not threaten the assumption that the same antecedent retrieval mechanism underlies the processing of all pronouns.

Squaring the uniform retrieval hypothesis with dynamic accounts is relatively trivial, as the default assumption of dynamic accounts is that the same mechanisms that underlie co-reference and syntactic binding support the interpretation of donkey anaphora. Insofar as the processing of other pronoun types involves antecedent retrieval – which corresponds, in this case to the retrieval of a discourse referent – then so too should the processing of donkey pronouns.⁵

3.2 Reconciling present and past results: Other cases of exceptional covariation

We close by considering the differences between our findings and those of Moulton and Han (2018) in light of the uniform retrieval hypothesis. All else equal, uniform retrieval holds that if a pronoun co-varies with an overt QP, then processing that pronoun involves retrieving the QP (or some chunk in memory corresponding to the QP). Insofar as reliable gender-mismatch effects index successful retrieval, Moulton and Han’s results suggest that the non-c-commanding QPs in their items were inaccessible to antecedent retrieval even though they seem to license exceptional covariation.

As the c-commanding QPs were not accessible in Moulton & Han’s non-c-commanding QP sentences, the question arises of how participants interpreted pronouns in these sentences. Under the uniform retrieval hypothesis, the absence of a reliable gender-mismatch effect entails that the pronouns were initially interpreted as lacking an explicit antecedent. The alternative, endorsed by Moulton and Han, is that the pronouns immediately retrieve an antecedent, but that that antecedent does not correspond to the QP: it is instead a “situation binder”. The absence of the gender-mismatch effects supposedly follows from the fact that situation binders lack gender features, unlike nominal antecedents.

Moulton and Han concluded matching pronouns in the non-c-commanding condition immediately find an antecedent based on two pieces of evidence: First, when participants were asked to choose a possible reading of matching test pronouns in an untimed interpretation task, they chose the co-varying reading in the non-c-commanding condition roughly as often as they chose binding under c-command or co-reference in other test sentences. The authors inferred therefore that participants in the self-paced reading task immediately resolve the pronoun as covariant with the non-c-commanding QP. Second, the average RT for the matching pronoun in the non-c-commanding QP condition was comparable to RTs for matching bound or co-referential pronouns. If antecedent-less pronouns are always read more slowly than pronouns with antecedents, then the absence of a pairwise difference between the matching pronouns is consistent with the assumption that matching pronouns in the non-c-commanding QP condition immediately find an antecedent.

There are reasons, however, to question this conclusion: First, participants’ offline preferences do not offer a direct window onto immediate incremental processing. The offline results are equally compatible with the interpretation that participants do not access the non-c-commanding QP during initial antecedent retrieval, but eventually recognize that covariation may be possible when given more time to reflect. Second, the similarity in average RTs between the matching pronoun in the non-c-commanding QP condition and matching pronoun in the syntactic binding (or co-reference) condition is only informative if mismatching pronouns in the non-c-commanding QP condition were more difficult than matching pronouns. However, there were not reliable differences between the matching and mismatching pronouns in the non-c-commanding QP condition. As such, we cannot conclude that matching pronouns were easier to process than pronouns that did not have a sentence-internal antecedent.

We therefore take the lack of a reliable gender-mismatch effect in the non-c-commanding QP conditions to indicate that participants tended to interpret the matching pronoun as though it lacked a sentence-internal antecedent. What remains to be explained, under this interpretation, is why the QP would be treated as inaccessible to antecedent retrieval during incremental processing, given its apparent ability to license exceptional covariation in Moulton & Han’s offline task.

Following our conjecture from above: the accessibility of QPs in fronted adjuncts depends on participants’ unconscious assessment of whether exceptional covariation is likely in, or supported by, the context. If cues in the sentence or context lead participants to judge a QP unlikely to license exceptional covariation, then that QP will not be kept accessible beyond its c-command domain. We suspect that this is what happened in Moulton & Han’s (2018) items: the test sentences contained enough cues to discourage participants from keeping the QP accessible up to the pronoun.

Providing a deterministic account, as we did for indefinites, of how to assess whether adjunct-internal QPs are likely to license exceptional covariation is not trivial. It is known that the availability of exceptional covariation is governed by extra-syntactic (e.g. semantic, pragmatic) conditions (e.g., Poesio & Zucchi 2002; Anderssen 2011; Barker 2012), but the conditions are not exhaustively determined by a single feature of the local syntactic context, such as whether the QP is in the scope of a downward-entailing quantifier. Instead, the availability of the readings is tied to subtler, less deterministic cues of the local and non-local context.

Although we do not have a formal analysis of why participants treated the QPs as inaccessible in Moulton and Han’s sentences, we can point to some of the cues that might have influenced their assessments. Intuitions suggest that extending the non-c-commanding QP’s accessibility beyond a fronted adjunct depends on how easy it is to adopt a quantificational, distributive, or multiple-event reading of the fronted adjunct. Such readings feel rather unnatural in examples such as those below (taken from the appendix in Moulton & Han 2018).

Features such as tense-marking, predicate choice, and even the presence of the intervening raising predicate it seems also appear to influence what readings participants consider likely. The infelicity of a mismatching pronoun is intuitively more pronounced in sentences like (10), which have present tense, an additional indefinite, which lack an intervening raising predicate, and which are easily interpreted as following a script (Poesio & Zucchi 1992):

We suspect that when participants read sentences like (9), they judged exceptional scope to be unlikely and marked the QP inaccessible either at the edge of the adjunct or shortly thereafter. Thus, matching and mismatching pronouns are treated alike in those sentences. In sentences like (10), however, we expect that participants would keep the QP accessible beyond the adjunct and a gender-mismatch effect would arise at the pronoun. We encourage future researchers to test this prediction.

Finally, since the availability of exceptional scope is governed by multiple, less-than-deterministic cues, we speculate that the decision to extend a QP’s accessibility past its c-command domain may best be modelled as probabilistic in certain configurations. Thus, the decision to treat a QP as licensing exceptional co-variation may vary item-by-item (and perhaps speaker-by-speaker). We think that this explanation provides a promising avenue for interpreting some of Moulton & Han’s secondary findings. If participants inconsistently extended the scope of the QP beyond the fronted adjunct across items, then we would predict a mixture of trials where the QP was accessible to retrieval and trials where it was not accessible to retrieval.⁶ Thus, the average response pattern observed in the average region-by-region reading times would similarly reflect a mixture across inconsistent trials, yielding an attenuated gender-mismatch effect in the non-c-commanding QP conditions relative to other conditions with consistently accessible antecedents. Such a behavioral response, in fact, is what Moulton and Han (2018) observed in their Experiment 3: a non-significant trend towards a gender-mismatch effect was observed in the non-c-commanding QP sentences at the post-pronoun region and a marginally significant gender-mismatch effect was observed in the following region (t = –1.79, p < .10). This marginal gender-mismatch effect is consistent with our interpretation that non-c-commanding QPs were actually accessible on a subset of trials.