Two routes to determine pronoun form

In the conceptual-lexical route (referred to by Meyer & Bock 1999 as the “conceptual hypothesis”) the speaker determines pronoun form in a similar way to noun form: by activating a concept like ‘multiple apples’, which is then grammatically encoded as a lemma and phonologically encoded as a lexeme. The lemma that serves as the input for the phonological encoding stage would just be a pronoun rather than a noun. Within the conceptual-lexical route, two paths are potentially available to determine pronoun form (Figure 1). On the direct path, the pronoun’s lemma is accessed directly from the concept. This may be possible when all the features that determine pronoun form have conceptual correlates. For example, a speaker may refer to some apples with the plural pronoun they because the concept has a multiplicity feature. Alternatively, on the mediated path, the pronoun form is accessed from the concept via an associated noun in the mental lexicon (Jescheniak, Schriefers & Hantsch 2001; Meyer & Bock 1999; Schmitt, Meyer & Levelt 1999). In this case, the concept ‘multiple apples’ activates the lemma apple with a plural number feature, which in turn activates the lemma for the pronoun. This form of mediation may be required when pronoun form is influenced by grammatical features without conceptual correlates, such as grammatical gender. For example, in Spanish, the pronoun referring to a set of apples is not only plural, but also feminine, as in ¿Te las vas a comer? (‘Are you going to eat them._FEM?’). The pronoun las reflects the feminine gender of the noun manzana (‘apple’), which cannot be attributed to apples being conceptually female.

The conceptual-lexical route may be the only one available when pronouns are used deictically, without a linguistic antecedent, as when a speaker points to some apples and says, They are ripe. However, it may also apply when there is a linguistic antecedent, as in the sentence Those apples are ripe, aren’t they?, where the pronoun they co-refers with the antecedent those apples. In this case, the pronoun and antecedent would be encoded fully separately: the concept of ‘multiple apples’ would first be grammaticality encoded as apples — and phonologically encoded as /æplz/ — and later in the sentence, the concept would be grammatically encoded as they — and phonologically encoded as /ðeɪ/. In this case, the antecedent and the pronoun end up sharing grammatical and/or morpho-phonological features as a collateral effect of encoding the same concept.

In contrast, in the syntactic route (similar to Meyer & Bock’s 1999 “lexical hypothesis”), a pronoun receives its features from the antecedent rather than the concept, similar to how verbs receive their features from a sentence’s subject phrase (Figure 1). Under this route, speakers access what has been previously said in the sentence or discourse and perform a matching operation between the antecedent and the pronoun’s features. The syntactic route may be important in cases when a concept has more than one possible grammatical encoding. For instance, the concept ‘scissors’ may be referred to using the plural noun scissors or the singular noun phrase pair of scissors. In this case, how a concept has previously been referred to in a discourse is important for selecting the correct pronoun form — they or it. This route to pronoun form selection can also capture grammatical gender agreement, since the linguistic antecedent has grammatical gender. The syntactic route may be readily applied in cases of intra-sentential pronominalization, when speakers could already be monitoring the content of the sentence in order to abide by constraints on sentence formulation, such as binding constraints and avoidance of noun phrase repetition.

To summarize, a key difference between the two routes is whether pronoun form is influenced by other parts of the sentence or previous discourse (syntactic route) or not (conceptual-lexical route). Our study assesses whether the syntactic route is used during the production of pronouns with intra-sentential antecedents in Spanish, a language in which pronouns mark the number and grammatical gender of their referent. We address these questions by investigating whether pronouns are subject to agreement attraction.

Agreement attraction as a tool to arbitrate between the two routes

Observing what speakers say is usually not enough to arbitrate between the conceptual-lexical route and the syntactic route: If everything goes well during production, a pronoun’s form will be the same regardless of the route taken. However, speakers sometimes make errors, and these errors can help reveal which route was taken. Here we focus on agreement attraction errors, as in (1), where the pronoun they does not agree with the agreement controller actor but rather with the local attractor noun soap operas. In production studies, attraction effects are indexed by a higher number of agreement errors when the agreement controller and attractor have different features (e.g., the actor in the soap operas) than when they have the same features (e.g., the actor in the soap opera) (Bock, Nicol & Cutting 1999; Bock, Eberhard & Cutting 2004; Bock et al. 2006; Kandel & Phillips 2022; Kandel, Wyatt & Phillips 2024).

(1) *The actor in the soap operas rehearsed, didn’t they?

The finding that pronouns are subject to agreement attraction demonstrates that pronoun form can be influenced by elements that are neither the concept referred to by the pronoun nor its associated lemma. This finding is unexpected under the conceptual-lexical route but could arise naturally from the syntactic route, as retrieving an antecedent for feature-matching opens up an opportunity for other nouns in the sentence to interfere.

However, the interpretation that the syntactic route is supported by attraction phenomena crucially depends on the assumption that errors are driven by the presence of an attractor noun within the sentence. This is not obviously the case in most previous studies on pronoun production, which focus on number attraction (Bock, Nicol & Cutting 1999; Bock, Eberhard & Cutting 2004; Bock et al., 2006; Kandel & Phillips 2022; Kandel, Wyatt & Phillips 2024). Since grammatical number is usually associated with conceptual number (also referred to as notional number), in a sentence like (1), it is not necessary for attraction to come from the grammatical features of the attractor noun soap operas itself; rather, attraction may derive from the message-level representation of the plural concept encoded by this noun. For instance, soap operas should activate a conceptual representation of ‘multiplicity’ that could be erroneously combined with the concept ‘actor’, explaining why the pronoun appears as they. Indeed, pronouns are sensitive to conceptual number, leading speakers to produce plural pronouns to reference antecedents that are conceptually plural but grammatically singular (e.g., collectives such as fleet; Bock et al. 1999). Importantly, the conceptual properties of attractors, such as natural gender, have also been found to influence pronoun form (Slevc, Lane & Ferreira 2007). Thus, while number attraction indicates that pronoun form can be influenced by something other than the underlying concept or lemma, ruling out the conceptual-lexical route, it does not unequivocally support that the influence comes from a linguistic element, as required by the syntactic route.

One way to assess whether pronoun attraction can indeed derive from a linguistic element is to test whether pronouns show attraction from the grammatical gender of inanimate nouns, e.g., manzana._FEMININE in Spanish. Since grammatical gender doesn’t have systematic conceptual correlates (though see Boroditsky, Schmidt & Phillips 2003 and references therein for a different view), grammatical gender attraction cannot be attributed to the influence of the natural gender of other concepts in the message. Rather, grammatical gender attraction would demonstrate that pronoun form can be influenced by other parts of the sentence or previous discourse, as predicted by the syntactic route.

In support of this hypothesis, Meyer and Bock (1999) observed evidence of grammatical gender attraction with Dutch pronouns using a preamble completion paradigm. In this paradigm, participants heard a sentence containing two nouns (e.g., aardappel._COMMON ‘potato’; badpak._NEUTER ‘swimsuit’) and were shown a predicate that only matched one of them (e.g., gaar ‘cooked’). Participants were instructed to repeat the preamble sentence and add a continuation sentence with the structure [pronoun] is [predicate], which required using a gender-marked pronoun (e.g., Die._COMMONis gaar). The results showed more pronoun form errors when the two nouns in the preamble mismatched in gender than when they matched, consistent with gender attraction.

While the presence of gender attraction in Meyer and Bock’s (1999) study provides some evidence in favor of a syntactic route, the observed effects may have in part arisen as a consequence of the elicitation paradigm used. This paradigm differs from natural production in ways that could inflate errors and/or favor the use of the syntactic route over the conceptual-lexical route. First, participants must interpret the grammatically-encoded representations of both nouns in the preamble sentence (the antecedent and the attractor) in order to determine which option should be pronominalized. Requiring speakers to consult the attractor as part of the pronominalization process may have artificially raised error rates relative to natural speech when the attractor would be irrelevant to pronoun form. In addition, the paradigm involves processes related to both parsing and working memory, as participants must remember, interpret, and correctly recall the provided preamble in order to repeat it and produce a continuation. This process might lead participants to plan constituents at a different time than they normally would if they are initially focused on recalling and repeating the preamble (Kandel & Phillips 2022). Also, elicited errors could reflect misinterpretations of the preamble structure rather than an influence of the attractor noun (Ryskin et al. 2021). Crucially, in Meyer and Bock’s (1999) task, speakers did not generate a prelinguistic message that they later transformed into linguistic output. Rather, the content of the message was predetermined by the provided preamble sentence and predicate, and the majority of the linguistic structure was already provided to participants — the only part of the utterance that speakers needed to plan for themselves was the pronoun. This may have led speakers to rely less on the message-level representations of the utterances to be produced, making them more likely to use a syntactic route than a conceptual-lexical route to determine pronoun form.

Concerns about the preamble paradigm have led some researchers to adopt description tasks to elicit agreement attraction effects (see Kandel & Phillips 2022; Kandel, Wyatt & Phillips 2024 for pronoun attraction; see Kandel & Phillips 2022; Kandel, Wyatt & Phillips 2022; Nozari & Omaki 2022; Veenstra, Acheson & Meyer 2014 for verb attraction). Scene description tasks alleviate some of the concerns of the preamble paradigm that might affect pronoun production: These tasks do not provide participants with any pre-packaged linguistic material to interpret or recall, speakers generate a message themselves based on the events of the scene, and they do not need to reference the attractor noun when deciding to pronominalize. Because of this, the salience of the message-level and linguistic-level representations of the antecedent and attractor noun are likely more comparable to natural speech production than in a preamble completion task, and speakers may be less biased to use the syntactic route. In line with this, two previous studies that used a description task to test number attraction in pronouns (Kandel & Phillips 2022; Kandel, Wyatt & Phillips 2024) obtained a smaller attraction effect than observed in previous preamble studies (Bock et al. 2006; Bock, Nicol & Cutting 1999; Bock, Eberhard & Cutting 2004), in which errors more closely resembled the attraction rates observed for verbs. This is consistent with the possibility that description tasks are less likely to lead to elevated errors and/or artificially favor using the syntactic route, particularly to the extent that this involves a matching process similar to subject–verb agreement.

However, prior scene description studies only elicited pronouns in one language (English) and only tested number attraction. Therefore, more research is needed to understand how pronoun form is determined cross-linguistically and from what source pronoun attraction errors arise, which provides insight into the pronoun planning route used. First, it is unclear whether the number attraction effect would replicate in other languages, particularly in languages like Spanish or Italian, whose speakers are known to be more sensitive to conceptual number during subject–verb agreement compared to speakers of languages with more limited inflectional morphology like English (Vigliocco 1996; Vigliocco, Butterworth & Garrett 1996; Vigliocco, Butterworth & Semenza 1995). Second, as discussed above, the presence of number attraction on its own cannot provide conclusive insight into the route used to determine pronoun form, as number has conceptual correlates. Grammatical gender attraction would be a better test, as it would show that pronoun attraction errors arise due to the influence of a linguistic element, which can be parsimoniously explained by the syntactic route but not the conceptual-lexical route. It is unclear whether the grammatical gender attraction effects that have so far only been elicited with a preamble task (Meyer & Bock 1999) would persist in a scene description task that is less likely to artificially favor the syntactic route to pronoun form planning.

The present study

The present study investigates how speakers determine pronoun form, using agreement attraction effects to arbitrate between the conceptual-lexical and the syntactic routes. We elicited pronouns using a scene description task inspired by that of Kandel, Wyatt & Phillips (2024), allowing us to investigate how pronoun form is determined when the challenges for the speaker are more similar to those in natural speech than in prior preamble completion experiments (e.g., Bock, Nicol & Cutting 1999; Meyer & Bock 1999). The study comprises two experiments that tested pronoun agreement attraction in Spanish, a language with rich inflectional morphology whose pronouns bear both number and grammatical gender features.

Experiment 1 tested whether the number attraction findings observed in prior studies replicate in Spanish, allowing us to assess the reliability of the effect cross-linguistically and whether there are cross-linguistic differences in pronoun form planning. As discussed above, findings from subject–verb agreement studies suggest that speakers of languages with richer inflectional morphology than English rely more on conceptual number. If this extends to antecedent–pronoun agreement, Spanish speakers may use the conceptual-lexical route more and potentially avoid attraction. Alternatively, the findings with subject–verb agreement might not extend to pronouns — these findings have been attributed to the availability of tacit subjects in Spanish/Italian, but this may not be relevant for antecedent–pronoun dependencies. If so, other outcomes are conceivable. For instance, Spanish speakers may be more likely than English speakers to use a syntactic route to pronoun form, potentially showing more attraction: This is because Spanish pronouns not only contain features with conceptual correlates (number, natural gender) but must also agree with the grammatical gender of their antecedent (a linguistic feature).

Experiment 2 tested whether Spanish pronouns show grammatical gender attraction, providing additional insight into the source of pronoun attraction effects. If we observe both number and gender attraction effects, that would provide evidence of an influence from linguistic representations, suggesting that pronouns are planned using the syntactic route. By contrast, if we observe number attraction but no grammatical gender attraction, it is possible that pronoun attraction errors only arise due to interference at the conceptual level, since number (but not grammatical gender) has conceptual correlates. By eliciting both forms of attraction using the same paradigm within the same language, we can be confident that any differences we observe between Experiments 1 and 2 are due to the pronoun feature tested (number vs. gender) as opposed to differences between studies or languages.

Following Kandel, Wyatt and Phillips (2024), we diagnosed attraction using two dependent measures. The first was the presence of agreement errors (the typical measure used to assess attraction). The second was the duration of the post-attractor region in error-free sentences. Previous work with the scene description paradigm has shown that the time taken to produce an agreement target can provide insight into how often attraction pressures are active, even in cases when no error is made (Kandel & Phillips 2022; Kandel, Wyatt & Phillips 2022). Consequently, by assessing attraction with two different measures, we may be able to derive greater insight into how often and in what contexts the processes and pressures that underlie pronoun number and gender attraction effects are active.

Methods

Materials

Participants described videos of inanimate objects touching other objects and turning them black. This action was referred to with the made-up verb pipear (‘pipping’). This word was based on the nonce word mimmed used by Kandel et al. (2024), except that the initial phoneme was replaced by a plosive in order to facilitate onset detection in the latency analysis. We used a non-word to discourage participants from preceding the direct objects of their sentences with the differential object marking preposition a (see Von Heusinger & Kaiser 2007 for the relationship between differential object marking and the lexical semantics of verbs in Spanish). The use of differential object marking may contribute to participants’ perception of objects as animate and, relatedly, female or male (see von Heusinger & Kaiser 2003 for the connection between differential object marking and animacy, and Fábregas 2013 for an overview of differential object marking in Spanish). This was important for Experiment 2, which aimed to test whether grammatical (rather than natural) gender elicits agreement attraction.¹

The elicited sentences had the target structure: NP1 (antecedent) + verb pipear + NP2 (attractor) + above/below + pronoun. We manipulated the number of the antecedent noun (singular/plural) and whether the antecedent and attractor nouns matched in number (match/mismatch), resulting in four experimental conditions (2 match conditions and 2 mismatch conditions; Table 1).

Table 1

Example target sentences in Experiment 1.

CONDITION	TARGET SENTENCE
SS – match	El chaleco ha pipeado el candado (de) debajo de él The vest has pipped the lock below it
SP – mismatch	El chaleco ha pipeado los candados (de) debajo de él The vest has pipped the locks below it
PP – match	Los chalecos han pipeado los candados (de) debajo de ellos The vests have pipped the locks below them
PS – mismatch	Los chalecos han pipeado el candado (de) debajo de ellos The vests have pipped the lock below them

Note. The antecedent and coreferential pronoun are bolded, while the attractor is underlined. The preposition “de” before the adverb “debajo” is shown between parentheses to reflect its optional status: In pilot testing, some Spanish speakers prefered to produce it, while others didn’t (Supplemental file 1). Therefore, both utterances with and without the preposition were accepted as target responses in Experiment 1. Abbreviations: SS = singular antecedent, singular attractor, SP = singular antecedent, plural attractor, PP = plural antecedent, plural attractor, PS = plural antecedent, singular attractor.

Sixteen nouns were used as antecedents and attractors (8 masculine and 8 feminine), leading to 112 target sentences (28 per participant per condition; half with masculine nouns and half with feminine nouns; within a sentence, all nouns were the same gender). The nouns were all three syllables with stereotypical gender suffixes (-o for masculine and -a for feminine). Because the same nouns were used in Experiment 2 (which manipulated gender), care was taken to ensure that masculine and feminine nouns had similar frequency values. Based on the Subtlex-ESP database (Cuetos et al. 2011), masculine nouns had an average raw frequency per million words of 17.47 (range: 3.03–41.54) and feminine nouns of 17.01 (range: 4.52–43.07) (t(14) = –0.07, p = 0.95).²

To create videos corresponding to each target sentence, we used images corresponding to each noun from the MultiPic database (Duñabeitia et al. 2018). The images had intermediate visual complexity on a 1–5 scale (average: 1.92, range: 1.4–2.29), and there was no evidence that complexity differed across masculine and feminine nouns (masculine nouns average: 1.89, range: 1.5–2.29; feminine nouns average: 1.98, range: 1.4–2.29; t(14) = 0.62, p = 0.54). The images all had high naming consistency (average: 99%, range: 94–100%), though naming consistency was slightly higher for feminine nouns (masculine nouns average: 98%; range: 94–100%; feminine nouns average: 100%, range: 100–100%; t(14) = 2.17, p = 0.048). This difference is unlikely to have affected the study results, as participants were introduced to the target names for all images at the start of the experiment (see Procedure below).

Each video was divided into a target and an alternative display. This was done to prevent participants from planning their responses before the pipping action occurred. In each display, a central object set (consisting of one or two objects of the same type) was positioned with two identical sets of objects above and below it (Figure 2). In the target display, this central object set (corresponding to N1; the antecedent) pipped the object set above or below it (corresponding to N2; the attractor). The position of the target display (left/right) was counterbalanced across items. If the target display corresponded to a match condition, the alternative corresponded to a mismatch condition, and vice versa. This was done to avoid some trials being perceived as more difficult than others. We made sure that the position of the attractor in the target display (above or below the antecedent) was evenly distributed across trials. The nouns representing the on-screen objects always had the same gender (across both the target and alternative displays), to avoid mixing number and gender interference.

Figure 2 

Example of a visual display in Experiment 1.

Note. The display corresponds to the target sentence “El chaleco ha pipeado los candados debajo de él” (‘The vest has pipped the locks below it’). A trial consisted of a 1 second preview, the pipping action, and then a 5 second response window in which participants had to describe the action.

Analysis

Latency analysis

The latency analysis assessed whether there were articulation slowdowns in utterances with correct pronoun forms across conditions. The latency analysis excluded all non-target utterances as well as utterances with disfluencies and number errors — i.e., only complete and correct responses were included in the analysis. These responses were aligned to their transcriptions at the word boundary level using the Montreal Forced Aligner (v.2.0.0; McAuliffe et al. 2017). We used the word onsets and offsets identified by the forced aligner to compute the duration of the post-attractor segment.³ The post-attractor segment duration was computed by subtracting the offset of the pronoun from the offset of the attractor noun (Figure 3). This segment was chosen because the attractor offset was the earliest point in time in which the pronoun form could be planned without interference from the planning of the two preceding noun phrases. The performance of the forced-aligner was evaluated by comparing its alignments to those of two humans and computing inter-rater agreements for word boundaries (Supplemental file 2). The interquartile range criterion was used to detect outliers (in ms), such that observations below the first quartile or above the third quartile by 1.5 times the interquartile range were excluded (Hawkins 1980).

Figure 3 

Example of a target response with its division into segments.

The statistical analysis examined post-attractor segment latencies with linear mixed effects regression, using the same predictors as in the error analysis, with the addition of Syllable Count (the number of syllables of the post-attractor segment). Syllable Count was included in the model to account for the fact that plural pronouns were one syllable longer than singular pronouns. The dependent variable was the log-transformed duration of the post-attractor segment.

In both the error and latency analyses, if there was a significant interaction, the model was re-fit such that the interaction term was replaced by the nested effects of the critical factor (e.g., the effect of Match for singular and plural antecedents separately).

Results

Out of 5,264 utterances, 8.03% were incomplete (range across conditions: 4.94–10.3%). Of the remaining 4,841 complete utterances, 5.41% were excluded from the error analysis due to containing non-target responses (range across conditions: 4.41–6.02%), and 12.95% were excluded from the latency analysis due to containing non-target responses and/or number errors (range across conditions: 10.2–15.3%). No observations were excluded from the latency analysis as outliers after the application of the interquartile range criterion. Thus, 4,579 trials were entered into the error analysis, and 4,214 trials were entered into the duration analysis.

Error analysis

The distribution of number error rates across conditions is shown in Figure 4. The statistical analysis showed an effect of match (Table 2): Number errors were more likely in mismatch than in match conditions, consistent with agreement attraction. Further, there was an interaction between antecedent number and match. Nested comparisons showed that the effect of match was larger with singular antecedents (estimate = 2.230, z = 5.384, p < 0.001) than with plural antecedents (estimate = 1.074, z = 3.585, p < 0.001). This suggests that plural attractors (as in the SP condition) caused more attraction than singular attractors (as in the PS condition), consistent with a number markedness asymmetry.

Figure 4 

Descriptive summary of error rates and durations of the post-attractor segment in Experiment 1.

Note. Diamonds show averages across participants in match (SS, PP) and mismatch (SP, PS) conditions. Points show by-participant averages. Abbreviations: SS = singular antecedent, singular attractor, SP = singular antecedent, plural attractor, PP = plural antecedent, plural attractor, PS = plural antecedent, singular attractor.

Table 2

Output of the Experiment 1 error analysis model.

COEFFICIENT	ESTIMATE	STANDARD ERROR	z-value	p-value
Intercept (grand mean)	–4.430	0.234	–18.939	< 0.001
Trial Order	–0.004	0.003	–1.368	0.171
Antecedent Number	0.487	0.309	1.577	0.115
Match	1.651	0.260	6.350	< 0.001
Antecedent Number × Match	–1.158	0.512	–2.259	0.024

Note. Model formula: Number Error ~ Trial Order + Antecedent Number * Match + (1 + Antecedent Number * Match || Participant) + (1 + Match || Item). A positive coefficient for Antecedent Number reflects more number errors with plural than singular antecedents. A positive coefficient for Match reflects a greater likelihood of number errors for mismatch than match conditions. The double bars in the model formula represent the removal of the correlation between random slopes and intercepts.

Discussion

The results of Experiment 1 show that the number agreement attraction effect observed in prior experiments in English replicates in Spanish. We observed evidence for agreement attraction in both error and timing measures: Number agreement errors were more likely in the mismatch conditions, and speakers were slower to articulate the post-attractor region containing the pronoun in these conditions even when the correct pronoun was produced, suggesting that the process leading to errors is active on more than just the trials in which errors occur (see Kandel, Wyatt & Phillips 2022 for discussion of how to interpret attraction timing effects for verbs). The error effect showed a similar markedness asymmetry to that observed in verb attraction studies, with stronger attraction from plural attractors than singular ones (e.g., Bock et al. 2006; Bock & Miller 1991; Bock, Nicol & Cutting 1999; Eberhard 1997; Thornton & MacDonald 2003; Vigliocco & Nicol 1998).

The results illustrate that representations other than the pronoun antecedent can interfere when determining pronoun number. As discussed in the Introduction, this influence is suggestive of the use of a syntactic route to determine pronoun features. However, since number has conceptual correlates, we cannot conclusively infer that it was the representation of the attractor at the sentence-level rather than the message-level that caused attraction. Grammatical gender, which doesn’t have systematic conceptual correlates, thus serves as a stronger test of whether attraction can arise from linguistic representations other than that of the antecedent. Experiment 2 assessed the presence of grammatical gender attraction.

Methods

Analysis

Responses were transcribed and coded in the same way as in Experiment 1, except that we coded gender errors instead of number errors (note that participants did not produce any pronoun number errors). The statistical analysis also followed Experiment 1, except that the predictor Antecedent Number was replaced with Antecedent Gender (sum-coded, –0.5 masculine/0.5 feminine). Since the gender of the antecedent noun did not vary within items, no by-item random slope for Antecedent Gender was entered in the random structure of the statistical models.

Results

Out of 8,064 utterances, 4.3% were incomplete (range across conditions: 3.72–4.81%). Of the remaining 7,720 complete utterances, 1.68% were excluded from the error analysis due to containing non-target responses (range across conditions: 1.08–2.08%), and 12.23% were excluded from the latency analysis due to containing non-target responses and/or gender errors (range across conditions: 10.6–15.6%). The removal of outliers for the latency analysis using the interquartile range criterion excluded 2.64% of the remaining data (range across conditions: 2.48–2.99%). Thus, 7,590 trials were entered into the statistical analysis of gender errors, and 6,597 trials were entered into the analysis of post-attractor segment durations.

Error analysis

The distribution of gender error rates across conditions is shown in Figure 5. The statistical analysis showed positive main effects of Match and Antecedent Gender (Table 5). Errors were more likely in mismatch than in match conditions, consistent with agreement attraction, and with feminine than masculine antecedents.

Figure 5 

Descriptive summary of error rates and durations of the post-attractor segment in Experiment 2.

Note. Diamonds show averages across participants in match (MM, FF) and mismatch (MF, FM) conditions. Points show by-participant averages. Abbreviations: MM = masculine antecedent, masculine attractor, MF = masculine antecedent, feminine attractor, FF = feminine antecedent, feminine attractor, FM = feminine antecedent, masculine attractor.

Table 5

Output of the Experiment 2 error analysis model.

COEFFICIENT	ESTIMATE	STANDARD ERROR	z-value	p-value
Intercept (grand mean)	–3.437	0.149	–23.003	< 0.001
Trial Order	0.001	0.002	0.768	0.443
Antecedent Gender	0.567	0.236	2.406	0.016
Match	0.730	0.144	5.068	< 0.001
Antecedent Gender × Match	0.051	0.229	0.222	0.824

Note. Model formula: Gender Error ~ Trial Order + Antecedent Gender * Match + (1 + Antecedent Gender * Match || Participant) + (1 + Match || Item). A positive coefficient for Antecedent Gender reflects a greater likelihood of gender errors with feminine than masculine antecedents. A positive coefficient for Match reflects a greater likelihood of gender errors for mismatch than match conditions. The double bars in the model formula represent the removal of the correlation between random slopes and intercepts.

Discussion

Experiment 2 showed evidence of grammatical gender attraction in the error analysis: Gender agreement errors were more likely in the mismatch than match conditions. Since grammatical gender doesn’t have conceptual correlates, grammatical gender attraction is indicative of influence from the linguistic representation of the attractor, suggesting that speakers use a syntactic route to determine pronoun form (at least at times).

To our knowledge, Experiment 2 represents the first time that a gender attraction effect in pronouns has been elicited using a scene description paradigm, as opposed to a preamble paradigm (e.g., Meyer & Bock 1999). Although the overall rate of pronoun gender errors was lower in Experiment 2 (3–8%; see Supplemental file 3) than in Meyer and Bock (1999, Experiment 1) (9–30%), the size of the attraction effect was not smaller: Compared to the match conditions, the mismatch conditions elicited 1.8–1.9x more errors in our Experiment 2 and 1.2–1.5x more errors in Meyer and Bock’s (1999) most analogous sentence construction (Experiment 1, with definite determiners and the antecedent as the first of the two noun phrases in the sentence).

The gender attraction errors in our Experiment 2 were also not accompanied by a latency effect, in contrast to the number attraction errors in Experiment 1. This might indicate that the pressures leading to gender errors are weaker and/or not as active as the pressures leading to number errors. Gender attraction also differed from number attraction in that it showed no markedness asymmetry, i.e., attraction rates did not differ across genders. We discuss potential sources of these differences further in the General Discussion.

Differences between number and grammatical gender attraction

There were differences between the observed number and gender attraction effects, both in size and shape. Specifically, attraction was larger for number than grammatical gender, both in error and latency measures. These results suggest that the pressures leading to number attraction are more prevalent and/or stronger than those leading to gender attraction (Eberhard, Cutting & Bock 2005). This increased pressure could in part result from the fact that number has conceptual correlates, meaning that the attractor number may be able to influence agreement at both the message level and the linguistic level (see next section for discussion). The asymmetry could also be related to the fact that, in our experiments, number is extrinsic while grammatical gender is intrinsic, i.e., non-inflectional and deriving from the word itself (Eberhard, Cutting & Bock 2005). It is possible that the process of inflecting a word boosts its representation in working memory (Bock et al., 2001), resulting in more interference from attractors with extrinsic features (e.g., Bock, Eberhard & Cutting 2004) — in this case, the larger number attraction effect would be primarily driven by an increase in errors when the attractor is plural (note that we observed greater attraction from plural attractors than singular ones, more below).

Further, number and gender attraction effects displayed a different pattern. In Experiment 1, we observed a markedness asymmetry for number, similar to that previously observed for verb attraction (Bock et al. 2006; Bock & Miller 1991; Bock, Nicol & Cutting 1999; Eberhard 1997; Thornton & MacDonald 2003; Vigliocco & Nicol 1998): Attraction effects were larger with plural than singular attractors, as indicated by the interaction between match and antecedent number in the error analysis. We did not observe an analogous markedness effect with grammatical gender: The interaction between match and antecedent gender was not reliable. Previous investigations of grammatical gender asymmetries for verb agreement are mixed, with some studies finding asymmetries (Antón-Méndez, Nicol & Garrett 2002; Badecker & Kuminiak 2007; Franck et al. 2008; Slioussar & Malko 2016; Vigliocco & Franck 1999, Experiment 1) but others not (Igoa, García-Albea & Sánchez-Casas 1999; Vigliocco & Franck 1999, Experiment 2). Eberhard, Cutting & Bock (2005) suggest that a reduced gender-based asymmetry could arise due to differences in contrastiveness for gender and number — similar to the argument used by Bock et al. (2001) to explain differences between intrinsic and extrinsic attraction.

Alternatively, a gender markedness asymmetry may have been obscured by a tendency for participants in Experiment 2 to generalize the use of the masculine pronoun, as indicated by the main effect of antecedent gender (more errors with feminine antecedents). If speakers were more likely to produce masculine than feminine pronouns, this would lead to more errors in sentences with feminine antecedents and masculine attractors and fewer errors in sentences with masculine antecedents and feminine attractors. Because this asymmetry works in the opposite direction from what would be expected of a gender markedness asymmetry (more errors in sentences with masculine antecedents and feminine attractors), it may have canceled it out. Such generalization of the masculine pronoun could arise because masculine is considered the default gender in Spanish (Prado 1982; see Beatty-Martínez & Dussias 2019 for review). If grammatical gender is encoded as a feature node associated with lemmas (Levelt, Roelofs & Meyer 1999) and masculine is the default, the masculine feature may have higher resting activation than the feminine feature in working memory, thereby leading to a higher overall likelihood of producing masculine pronouns. Note, however, that by the same logic, we might have expected participants to generalize the use of the singular pronoun, to the extent that singular is the default unmarked number in Spanish. This was not the case in the present study, though this pattern was observed by Kandel et al. (2024).

Determining pronoun form

Our results provide insight into how speakers determine pronoun form. In particular, participants in our study appeared to use a syntactic route to pronoun form, identifying the required features through a feature-matching process with a linguistic antecedent. Accessing the antecedent for feature-matching opens up the opportunity for other sentence material to interfere, which can explain the observed attraction effects. Nevertheless, participants still used the correct pronoun form on most trials. These low error rates could occur if the syntactic route is very robust to interference or if speakers use a mix of routes to determine pronoun form (Kandel, Wyatt & Phillips 2024). As mentioned in the Introduction, speakers may need different routes for different contexts in which they produce a pronoun (e.g., pronominalization when there is no linguistic antecedent vs. pronominalization with an intra-sentential antecedent), so both routes may be available for speakers. If participants in our experiments used the conceptual-lexical route on a subset of trials, then they would make fewer errors than if they routinely used the syntactic route.

Although a syntactic route to pronoun form has been considered in prior literature (e.g., Meyer & Bock 1999), the real-time mechanisms used to implement the pronoun–antecedent feature matching process within this route remain unclear. Eberhard, Cutting & Bock (2005) propose that pronouns derive their features in part through a marking-and-morphing operation with the antecedent, similar to that used for subject–verb agreement. In particular, they suggest that during structural integration pronouns reference a representation of the antecedent’s number value that is continuous from singular to plural (rather than having a binary singular or plural distinction), with the plurality of this value depending on the notional and morphological number of the noun(s) within the antecedent phrase. Agreement attraction arises when this value is more ambiguous (e.g., when an antecedent phrase contains multiple nouns of different number), leading agreement errors to be probabilistically produced (Eberhard, Cutting & Bock 2005). However, marking-and-morphing does not readily account for our results because the antecedents in our experiments contained a single noun (i.e., the attractor noun was not part of the subject phrase), and thus their number should have been unambiguous to speakers.⁴

Kandel, Wyatt & Phillips (2024) propose an alternative model of pronoun formulation, in which speakers engage a retrieval process for feature matching. Speakers retrieve the linguistic representation of the antecedent from memory and then use its features to determine pronoun form (for similar retrieval accounts of subject–verb agreement in production see Badecker & Kuminiak 2007; Lorimor, Jackson & Foote 2015; inter alia). Kandel, Wyatt and Phillips (2024) adopt the pronominalization mechanism from Schmitt, Meyer & Levelt (1999), in which an ‘in focus’ feature of the message-level representation of the referent (indicating prominence/salience in the discourse) cues the speaker to produce a pronoun. They suggest that the retrieval process targets representations that have a similar prominent mental status in working memory (e.g., high activation or an ‘in focus’ feature at the lexical level). Within this model, attraction errors arise when the attractor noun’s representation matches (or partially matches) the retrieval cues used to access the antecedent (e.g., if it is similarly highly prominent due to recency of mention), leading it to be erroneously retrieved instead of the antecedent (Kandel, Wyatt & Phillips 2024). This account can explain both the number and grammatical gender attraction effects in the present study, as the attractor noun in the elicited sentences may have had high prominence due to recency of mention and/or the fact that it was being individuated by the prepositional phrase containing the pronoun.

Kandel, Wyatt & Phillips’ (2024) account can also explain the markedness asymmetry observed with number. If inflecting a noun to make it plural (as is necessary for extrinsic number) increases its activation in working memory (Bock et al. 2001), this boost would lead to increased attraction from plural nouns. By contrast, the grammatical gender of the nouns in our study was intrinsic, and thus no comparable inflectional adjustment was necessary, making a gender-based asymmetry less likely. Interestingly, an attraction boost for extrinsic plurals predicts that gender attraction should be greater from plural attractors than from singular ones, assuming that both number and gender agreement rely on the same retrieval process — this prediction can be assessed in future work (Experiment 2 only included singular nouns).

To explain the larger attraction effect for number than gender across experiments, a retrieval account could be integrated with the ideas proposed in the previous section. Specifically, extrinsic features may lead to greater competition from the attractor than intrinsic features (Bock et al. 2001; Eberhard, Cutting & Bock 2005) or conceptual number may also be able to influence agreement. An influence of conceptual number can be incorporated into a retrieval account by combining it with a lexical competition framework, similar to that proposed by Nozari and Omaki (2022) for subject–verb agreement. Within this framework, agreement errors arise due to competition between forms during lexical selection. If candidate pronoun forms are activated not only by the features of the retrieved antecedent (consistent with Kandel, Wyatt & Phillips 2024) but also by the features of nearby nouns (consistent with Nozari & Omaki 2022), then attraction errors can arise either in the case of a retrieval error or if an incorrect pronoun form receives enough activation from a nearby attractor noun to successfully out-compete the correct form. If the message-level number features of nearby nouns contribute to the activation of candidate forms, this will increase activation of competing incorrect forms for number but not gender, thereby leading to a greater number attraction effect.