1 Introduction

One of the main goals of linguistic theory is to formalize what it is we know when we know a language. As the language varies, so too does the knowledge that underlies it. This observation is obvious to the point of being trivial: speakers of Spanish know Spanish, and their knowledge differs from that of speakers of English, who know English; hence the difference in language. As linguists explore various constraints on the variation that is allowed across languages, cross-linguistic variation leads them to the principles underlying Universal Grammar. Here we address a potentially different but prima facie no less informative type of variation, that between speakers of the same language. One might call this variation “intra-linguistic.”

We take as our starting point the observation that there are many different types of speakers. There are monolingual, literate, educated native speakers who know their language perfectly—the Chomskyan ideal stated in the well-known quote: “Linguistic theory is concerned primarily with an ideal speaker-listener, in a completely homogeneous speech-community, who knows its language perfectly and is unaffected by such grammatically irrelevant conditions as memory limitations, distractions, shifts of attention and interest, and errors (random or characteristic) in applying his knowledge of the language in actual performance” (Chomsky 1965: 3). There are L1 learners, often monolingual as well: children on their way to fully acquiring the language. There are also L2 learners, (young) adults who are actively learning (or have actively learned) another language in addition to their first. From the perspective of traditional linguistic theory, these count as “typical” speaker profiles. For our purposes, the real excitement centers around “atypical” cases where the normative developmental trajectory does not attain; the speakers that result at once fit all and none of the profiles just described. We have in mind heritage speakers: (relatively) unbalanced bilinguals who shifted from their first language (their heritage language) to their dominant language early in childhood.¹

According to most definitions, heritage speakers are individuals who were raised in homes where a language other than the dominant community language was spoken, resulting in some degree of bilingualism in both the heritage language and the dominant language (Valdés 2000; Rothman 2009; Benmamoun et al. 2013a; b; Scontras et al. 2015; Montrul 2016). This relatively unconstrained definition makes it almost impossible to give a concrete model for a heritage speaker, but this is by design: heritage language proficiency falls along a continuum. On one end, heritage speakers can be so proficient that in casual conversation they pass as native speakers. On the other end, some heritage speakers are able to comprehend but unable to fluently speak their heritage language; these are the so-called “overhearers” (Au et al. 2008; Chang 2016). The variability among heritage speakers derives from the many pathways that lead to the attainment of language under bilingualism, with significant variation in the amount and quality of input.

The notion of heritage speakers typically subsumes simultaneous or early sequential bilinguals. Some heritage speakers are born to immigrant parents in a country where the heritage language is not the native tongue; these bilinguals may be exposed to both the heritage language and the dominant language of the broader community from the outset of acquisition. Other heritage speakers are born in the home country of their parents, and for the early years of their life encounter only the heritage language. They often begin acquiring their second, dominant language at the time of emigration, when they shift to a different speech community. Within sequential bilinguals, the age of emigration or first exposure to the dominant language affects the resulting language proficiency (for extensive discussion, see Montrul 2008; 2016). Still more variation arises from the amount of input and maintenance speakers receive for the heritage grammar: some speakers use the heritage language at home regularly with their families, while others have only infrequent opportunities to perform in their heritage language.²

Despite the heterogeneity of their demographic profiles and levels of proficiency, certain characteristics of heritage speakers have been shown to be remarkably consistent, setting these speakers apart both from native and L2 speakers of a given language. At the phonological level, heritage speakers diverge from their native counterparts in aspects of pronunciation and prosody (Godson 2003; Barlow 2014; Chang 2016). At the syntactic level, heritage speakers tend to impose more rigid word order where native speakers allow for flexibility (Isurin & Ivanova-Sullivan 2008; Ivanova-Sullivan 2014); relatedly, they limit their inventory of syntactic dependencies (Polinsky 2011). At the intersection of syntax, semantics, and pragmatics, evidence from Chinese and English heritage speakers suggests that heritage grammars lack quantifier scope ambiguities (Scontras et al. 2017). The current investigation follows up on findings at the morphological level, where heritage speakers are known to eliminate irregular forms and struggle with inflectional morphology (Benmamoun et al. 2013a; b).

In the nominal domain, heritage speakers are known to perform differently from native speakers in the production and comprehension of agreement in gender (see Håkansson 1995 for Swedish; Montrul et al. 2008 for Spanish; Polinsky 2008a for Russian), marking of definiteness (Håkansson 1995 for Swedish; Bolonyai 2016 for Hungarian), case-marking (Polinsky 1995; 2006; 2008a; b for Russian; Song et al. 1997 for Korean), and topic-marking (Laleko & Polinsky 2013; 2016 for Japanese and Korean). Despite the vulnerability of these categories—many of which also pose a challenge to L1 and L2 learners—there are also linguistic domains in which heritage speakers demonstrate remarkably resilient knowledge, for example, tense (even in recessive bilinguals, as shown by work on Inuttitut; see Sherkina-Lieber et al. 2011) or numerical expressions (Polinsky 2016a).

The question thus arises as to precisely what heritage speakers know about their heritage morphology. The problem lies in using observed differences in linguistic behavior to diagnose and distinguish differences in linguistic knowledge. The null hypothesis holds that heritage speakers possess the same knowledge as native speakers: a heritage speaker of Spanish has the same grammatical knowledge of Spanish as a native speaker of Spanish. If the knowledge is the same, then any differences we observe in behavior arise from differences in the usage of that knowledge. Faced with the costly task of performing in their less dominant language, heritage speakers, even the most proficient ones, would diverge from native speakers because of the overwhelming processing load. Under the null hypothesis, observed differences are symptomatic not of a diverging heritage grammar, but of the relative scarcity of online resources.

The competing hypothesis, which we investigate experimentally below, holds that differences in performance are symptomatic of deeper, structural differences in the heritage grammar. Finding evidence that points to a true divergence between native and heritage grammars, one must then confront the question of what pressures led to the difference. To answer this question, we must arrive at a clear characterization of the structural difference: does the heritage grammar reduce or augment structure relative to the native baseline? The pressures driving each outcome are interestingly different. Finding reduced structure relative to the native baseline, heritage speakers will have likely prioritized representational economy, restructuring their grammar in favor of lighter-weight linguistic representations. Less articulated, more parsimonious structures (e.g., structures with fewer explicit agreement features or syntactic projections) could ease the load on working memory and might therefore be preferred to their fully-articulated brethren. Pressure from representational economy would therefore lead to a general slimming-down of the heritage morphology.

Conversely, heritage speakers may prioritize analyticity, restructuring native grammars with more-fully-articulated syntactic structures. Increasing analyticity with syntactic structures that allow one-to-one correspondences between underlying features and surface forms would lead to a richer, more rigid morphology (Sorace 2005; Keating et al. 2016; Montrul 2016). The research to date suggests that increased analyticity is commonly the trigger for restructuring in heritage grammar. It bears noting that we use the terms “representational economy” and “increased analyticity” without advancing any specific agenda on theories of sentence processing (for example, theories of “shallow” vs. “full” parsing; Ferreira et al. 2002; Sanford & Sturt 2002; Clahson & Felser 2006). Our intent is merely to operationalize these notions in such a way as to deliver diverging predictions with respect to levels of feature articulation: a less-well-articulated feature space (i.e., fewer feature values and less structure) would evidence pressures from representational economy, while increased analyticity predicts fully-specified feature values and fully-articulated structure.

To investigate the content of heritage speakers’ knowledge about their morphology, together with the pressures that could shape differences in that knowledge relative to the native baseline, here we extend the experimental methodology developed in Fuchs et al. (2015) to examine number and gender agreement in heritage Spanish. We focus on number and gender features because they are expressed on independent exponents in the native Spanish grammar, and because theoretical approaches to the structure of these categories provide two analytical possibilities for their representation in syntax. By comparing the structuring of these agreement categories in native and heritage Spanish, we aim to demonstrate that the loss of agreement in heritage languages is not superficial or accidental, but rather follows from principled and predictable structural changes.

2 Background: Feature representation in the native Spanish baseline

Grammatical agreement happens when features of one constituent (e.g., a sentential predicate) purposefully align with the features of another (e.g., a sentential subject). The specific form that a constituent expresses is determined by the features that the constituent holds (e.g., I vs. they; runs vs. run). Languages vary in the quality and quantity of agreement they express; the recurrent feature categories that participate in agreement are person (e.g., first vs. second), number (e.g., singular vs. plural), and gender (e.g., feminine vs. masculine). Once we assume that agreement features are represented structurally and merged into a syntactic tree, the question shifts to precisely how those structures are built.

Explorations of feature geometry investigate the structural organization and hierarchical relationships in agreement systems (Ritter 1993; Harley 1994; Harley & Ritter 2002). Feature geometry identifies the feature categories of person, number, and gender; here we limit our investigation to number and gender only. This narrowing of focus is motivated in two ways. First, person stands apart from the other features in its distribution (for example, unlike the other features, person agreement never appears on adjectives; see Baker 2008). Second, the hierarchical position of person relative to the other feature categories has been well established: it dominates the other features (Harley & Ritter 2002; among others). Meanwhile, the relationship between gender and number is less clear.

Under the assumption that both number and gender are represented in the syntax, there are two competing models of their structural relationship.³ One model analyzes the two features as bundled together (cf. Ritter 1993; Carstens 2000), hosted on the head of the same projection:

Under this bundling model, gender and number features are projected together, which means they are dependent on each other (Antón-Méndez et al. 2002). The effect is that valuation of one feature presupposes valuation of the other; number and gender agreement is part of a single process whereby one head probes for both features at the same time.

Empirical support for the bundling model begins with the observation that languages regularly combine number and gender, fusing the two features into a single morpheme that expresses changes in one or both features. Few languages have systems in which both number and gender features participate in agreement but remain completely independent of each other, and it is hard to find a language that reliably this description; at the writing of this paper, candidate languages include Abkhaz (Hewitt 1979), Amharic (Kramer 2015: 44), Yimas (Foley 1991), and Romanian, at least according to some researchers (Giurgea 2008; Croitor & Giurgea 2009).

The alternative to the bundling model splits number and gender features; they are projected and valued independently (Picallo 1991; Antón-Méndez et al. 2002; Carminati 2005). Under the split model, gender morphology hosted on the nominal spine heads its own gender projection (GenP), which is dominated by the number projection (NumP), as in (2).⁴ Thus, number and gender agreement are valued separately, independently of each other.

Evidence for the split model comes from the observation that number and gender morphology is consistently linearly ordered. In those languages, including Spanish, where the two features can be pre-theoretically separated into independent morphemes, the morpheme order is Stem-Gender-Number:

Under the split model—with the structure shown in (2) above—the morpheme order of Stem-Gender-Number is a direct consequence of the fact that number hierarchically dominates gender. Under a bundling model, number and gender are bundled together and therefore leveled—there is no clear hierarchical relationship between the features that could deliver a consistent linear order.

With arguments in favor of both the bundled model and the split model, the task is to determine which model is appropriate for which language—it need not be the case that every language approaches number and gender features in the same way (for a discussion of yet more arguments, see Fuchs et al. 2015 and the references therein). In Fuchs et al. (2015), we used experimental evidence from agreement phenomena to establish the baseline feature representation and specification in the native Spanish grammar. Our assumption, then and now, is that grammar and its parser are in an isomorphic relation, so that observing the parser allows for observation of the grammar (Phillips 2013). The work in this approach centers around developing testable hypotheses about behavior on the basis of articulated theories of grammar. Our strategy was to elicit responses to agreement failures in a way that would diagnose the potentially independent contributions of number and gender features. Before presenting the details of our experiment, we will present a quick overview of the relevant facts concerning Spanish agreement.

Agreement in number and gender in Spanish is pervasive: determiners, adjectives, and participles must all agree with a head noun in both number and gender. The Spanish plural is formed by suffixing -s onto a noun, while the singular is understood as the absence of the plural morpheme (in other words, it appears morphologically bare).⁵

Our focus here is on morphologically-transparent inanimate nouns with uninterpretable gender (as opposed to animate and human nouns whose gender is often interpretable and may be encoded separately, represented higher in the structure of a nominal; see footnote 3). Spanish has two surface genders: masculine and feminine; they are relatively equally distributed in the nominal lexicon (Bull 1965). The two genders also appear to be equally specified morphologically: the theme vowel -o is most often associated with the masculine gender, and the theme vowel -a is most often associated with the feminine gender.⁶

To summarize: the presence (or absence) of a word-final -s provides cues to that word’s number, and the vowel (-a- vs. -o-) that would precede -s provides cues to the word’s gender. As the number- and gender-agreement morphemes are in principle morphologically independent, we could manipulate their combination to produce sentences with different kinds of agreement errors. Because the bundling and split models of feature representation make different commitments regarding the valuation of agreement features, the predictions of the two models pull apart in cases where one or both features are incorrectly matched. These are cases of agreement attraction (Bock & Miller 1991; Bock & Eberhard 1993).

In agreement attraction, errors in agreement are masked by the features of a distractor constituent. For subject-verb agreement, the predicate fails to agree with its subject, but rather agrees with the features of some other noun present in the same clause (typically, but not necessarily, closer to the agreeing verb in the surface order). The paradigm example appears in (5), where the predicate (shown in SMALL CAPS) should be singular to match the singular feature of the head noun (goal, shown in bold), but instead it is plural, matching the plural feature of the so-called “local” noun (distractor, italicized). The agreement error yields an ungrammatical sentence, yet such cases are regularly produced and perceived as grammatical.

Agreement attraction has been widely studied in behavioral experiments, and a substantial body of literature has identified factors that affect the probability that a case of attraction will be perceived as grammatical. Fuchs et al. (2015) put agreement attraction to use in exploring the difference between bundling and split approaches to feature representation. If number and gender features are bundled together, they ought to be valued simultaneously. Thus, the number and gender features of a noun (more accurately, the single feature bundle of that noun) should determine agreement together, at the same time. With simultaneous feature valuation, when an incorrect noun enters into agreement with an adjective, both its number and gender features should effect agreement attraction. To illustrate this point, consider the following ungrammatical sentences of Spanish:

In each sentence, the local noun (i.e., revistas or periódicos) has spread its features to the predicate (i.e., aburridos/-as), which may lead to the illusion of grammaticality.⁷ If number and gender are projected and valued together, as per bundling approaches, then when the predicate (erroneously) gets a feature (e.g., number) from the local noun, it should get the other feature (e.g., gender) as well. In other words, agreement attraction in one feature ought to precipitate agreement attraction in the other feature, with the result that both of the above sentences should be rated equally high (or equally low).

If number and gender are split, then the features are projected and valued independently, and agreement attraction in number can proceed independently of attraction in gender. The violations are evaluated (or ignored) on their own merits. If attraction in number does not carry the local noun’s gender features along for the ride, then the sentence in (6a)—where the predicate agrees with the local noun in number but matches the head noun in gender—should be rated higher than (6b), where neither feature of the head noun correctly appears on the predicate. It was precisely these sorts of sentences that we set out to test.

Fuchs et al. (2015) designed an auditory sentence-rating task in which participants rated the acceptability of sentences they heard. Each stimulus was based on the syntactic frame in (7).

(7)	(Subject) Verb NP1 PP NP2 Adverb Adjective/Participle

Sentences contained an animate subject and a verb that introduced a small clause (Contreras 1987; Jiménez-Fernández & Spyropoulos 2013). Within each small clause was a noun (NP1) modified by a prepositional phrase that contained a distractor noun (NP2); the head noun (NP1) served as the subject to a predicative adjective (or participle; we label this element ADJ).⁸ For such a sentence to be grammatical, ADJ had to inflect for number and gender, agreeing with NP1.⁹ In the design of our stimuli, we manipulated the number and gender of NP1, NP2, and ADJ, creating grammatical and ungrammatical sentences.

Consistent with previous work on agreement attraction that has identified an asymmetry between number features in determining agreement attraction (Bock & Miller 1991; Bock & Eberhard 1993; Vigliocco et al. 1995; 1996; Vigliocco & Nicol 1998; Bock et al. 2001; Hartsuiker et al. 2003; Alcocer & Phillips 2009; Bock et al. 2012; Acuña-Fariña et al. 2014; Jegerski 2016), we found that native speakers of Spanish exhibit clear effects of attraction for number: ungrammatical sentences with a singular NP1 but plural NP2 and ADJ were rated higher than other cases of ungrammaticality. Also consistent with the previous literature, we failed to find evidence of attraction for gender. Crucially, we found that within the potential attraction conditions (i.e., sentences with singular NP1, but plural NP2 and plural ADJ, as in (6a; b) above), conditions with feminine head nouns in which agreement attraction occurred in two features were rated significantly lower than conditions with agreement attraction in only one feature. The analogous difference for masculine head nouns was not significant. Figure 1 plots average ratings for these potential attraction conditions, thus representing the baseline.

Figure 1

Average ratings for potential attraction conditions from the 50 native speakers of Spanish tested in Fuchs et al. (2015).

We interpreted these results as evidence for a split model for number and gender in the native Spanish grammar. Native speakers treat attraction in two features separately from attraction in one. Because the two features participate in agreement separately, attraction in number does not precipitate attraction in gender.¹⁰

In addition to testing bundling vs. split models of number and gender, we set out to determine the content of each feature type. We used evidence from grammaticality effects—the ability to distinguish ungrammatical sentences from grammatical ones—in non-attraction sentences to determine feature content. Finding sensitivity to both singular and plural agreement errors, we concluded that number in the native baseline is a multi-valued feature (both singular and plural are specified feature values; cf. Sauerland 2003; Scontras 2013a; b). Finding sensitivity only to agreement errors involving feminine adjectives, we concluded that gender is single-valued (feminine is the specified feature value, while masculine serves as the absence of gender specification, consistent with the classic analysis proposed by Harris 1991; see also Kramer 2015).

Having mapped, as it were, the terrain of number and gender in the native Spanish baseline, we now shift the question to heritage speakers: does their knowledge of Spanish morphology match that of the native baseline? If it does not, what sorts of pressures likely led to the difference? Concretely, do heritage speakers show evidence of bundling or splitting number and gender? And, relatedly, how is each category specified: does number require that both singular and plural be specified? Does gender require that both masculine and feminine be specified? Or are single-valued oppositions sufficient? If heritage speakers are driven by pressures to increase analyticity, we would expect them to converge on a maximally-articulated feature representation where number and gender are split, with each feature value specified. Such a convergence would entail maintaining the structure observed in the native grammar and enhancing it with additional feature values.

If, however, representational economy influences the grammar of agreement in heritage speakers, we would expect them to converge on a lighter-weight, less-well articulated feature representation where number and gender are bundled together, and with the minimal number of feature values specified. Of course, there also remains the possibility that there is no difference between the native and heritage grammars of agreement, which we assume as the null hypothesis (although this would run counter to the well-documented divergence of heritage language morphology surveyed above).

3 Testing heritage speakers

To investigate the content and structure of heritage agreement categories, we ran a direct replication of the auditory sentence-rating task from Fuchs et al. (2015), this time testing English-dominant heritage speakers of Spanish. Previous studies have looked at number or gender agreement in heritage speakers of Spanish, most commonly in agreement between two expressions that are adjacent (e.g., Montrul et al. 2008; 2014; Foote 2011; Montrul 2013); ours is the first to apply the agreement attraction paradigm in an attempt to diagnose the structure and content of these speakers’ knowledge of their agreement morphology. Building on the results of previous studies, we find it imperative to increase the difficulty in comprehension by introducing linear (as well as structural) distance between the noun bearing agreement features and the agreeing expression. We also wanted to ensure that the target of agreement was the same for both gender and number (i.e., the predicative adjective), thus allowing for direct comparison across feature types (cf. Acuña-Fariña et al. 2014, who also compared two feature types).

If we find that the heritage speakers lose sensitivity to specific feature values, or that they bundle rather than split their number and gender features, then we have an instance of grammar divergence. More importantly, we can use any differences between the heritage and native grammars to help explain and thereby predict the gradual loss of agreement in heritage languages.

3.3 Results

To simplify presentation and ease interpretation, we follow Fuchs et al. (2015) and present the results in two parts. First, we investigate grammaticality and attraction effects for each of the four potential feature values (i.e., singular, plural, masculine, and feminine); analyses are grouped by the features of NP1 (i.e., the head noun), which ought to determine agreement. Second, we focus in on potential attraction conditions to test the predictions of bundled vs. split models of number and gender.

3.3.1 Head-noun analysis

Figure 2 plots average acceptability ratings grouped by the relevant feature of NP1.¹² Perhaps the most striking aspect of Figure 2 is the relatively high ratings across the board: on average, participants rated critical items 3.7 out of a possible 5 points on the acceptability scale. This may initially seem surprising, especially given that half of the conditions tested were ungrammatical. Moreover, previous research has demonstrated clear sensitivity to agreement violations in heritage speakers of Spanish (Foote 2011). However, recall that target ungrammaticality is the result of a mismatch in agreement features on NP1 and ADJ, and that these two constituents are contained within a much larger frame. The agreement error in ungrammatical conditions is just one small part of an otherwise lengthy and grammatically correct sentence. The fact that these agreement errors did not yield extremely low average ratings is therefore a product of the experimental design, which prioritized taxing participant memory by increasing distance between key sentential elements. This design yielded the same overall high ratings when the experiment was conducted with native speakers in Fuchs et al. (2015). The crucial metric in the previous and in the present study is therefore not the absolute value of an average rating, but rather how ratings compare to each other across different grammatical and ungrammatical conditions.

Figure 2

Average acceptability ratings grouped by NP1 features. Error bars represent bootstrapped 95% confidence intervals.

Considering first the most basic of these differences, heritage speakers gave grammatical conditions in the study an average rating of 3.95, and ungrammatical conditions an average rating of 3.66; this difference was found to be significant (Wilcoxon test, p < 0.001). This difference suggests that participants were attending to the experimental items and sensitive to differences in grammaticality. We turn next to further differences across various grammatical and ungrammatical conditions, grouping analyses by the relevant features of NP1.

Singular NP1 To avoid possible effects of gender mismatches, gender was held constant across NP1, NP2, and ADJ—all were either feminine or masculine. Keeping NP1 singular, there are four possible combinations for the number features of NP2 and ADJ, as in Figure 2. To investigate the effect of these feature combinations on the perceived acceptability of the sentences that contained them, we fit a linear mixed-effects model predicting acceptability ratings by NP2 (SG vs. PL) and ADJ (SG vs. PL), as well as their interaction. This model and the models that follow included random intercepts for participants and items. The model found a main effect of ADJ (β = –0.74, t = –3.57, p < 0.05): ungrammatical sentences in which ADJ appeared in the plural were rated lower than their grammatical counterparts. The model also found a marginally significant interaction between NP2 and ADJ (β = 0.56, t = 1.93, p < 0.06): when both NP2 and ADJ were plural, ungrammatical sentences were rated higher than the ungrammatical baseline. These results replicate the native baseline observed by Fuchs et al. (2015) and indicate that attraction from the plural is present in heritage speakers as well.

Plural NP1 As with the singular-NP1 analysis, we held gender constant. Keeping NP1 plural, there are four possible combinations for the number features of NP2 and ADJ, as in Figure 2. We fit a linear mixed-effects model predicting acceptability ratings by NP2 (SG vs. PL) and ADJ (SG vs. PL), as well as their interaction. The model did not find a significant effect of ADJ (β = –0.27, t = –1.24, p < 0.22): ungrammatical sentences in which ADJ appeared in the singular were not rated significantly lower than their grammatical counterparts. The lack of a significant difference between grammatical and ungrammatical conditions was a reflection of variation in participants’ responses; in particular, participants were divided in their ratings of ungrammatical plural-NP1 conditions. These results deviate from the native baseline observed in Fuchs et al. (2015), where native speakers demonstrated a grammaticality effect for plural NP1. The model also did not find a significant interaction between NP2 and ADJ (β = –0.22, t = –0.70, p < 0.49): as with the native baseline, there is no evidence of attraction from the singular.

Masculine NP1 Turning now to gender, to avoid possible effects of number mismatches, number was held constant across NP1, NP2, and ADJ—all were either singular or plural. Keeping NP1 masculine, there are four possible combinations for the gender features of NP2 and ADJ, as shown in Figure 2. We fit a linear mixed-effects model predicting acceptability ratings by NP2 (M vs. F) and ADJ (M vs. F), as well as their interaction. The model found a significant effect of ADJ (β = –0.46, t = –2.07, p < 0.05): ungrammatical sentences in which ADJ appeared in the feminine were rated lower than their grammatical counterparts. The model did not find an interaction between NP2 and ADJ (β = –0.02, t = –0.07, p < 0.95): there is no evidence of attraction from the feminine. These results replicate the native baseline.

Feminine NP1 As with the masculine-NP1 analysis, we held number constant. Keeping NP1 feminine, there are four possible combinations for the gender features of NP2 and ADJ, as in Figure 2. We fit a linear mixed-effects model predicting acceptability ratings by NP2 (M vs. F) and ADJ (M vs. F), as well as their interaction. The model did not find a significant effect of ADJ (β = –0.16, t = –0.73, p < 0.47): ungrammatical sentences in which ADJ appeared in the masculine were rated no lower than their grammatical counterparts. The model also did not find an interaction between NP2 and ADJ (β = 0.13, t = 0.44, p < 0.66): there is no evidence of attraction from the masculine. These results replicate the native baseline.

3.3.2 Attraction condition analysis

Next, we focus in on potential attraction conditions where a singular NP1 occurs with a plural NP2 and ADJ. Figure 3 plots average responses to these conditions grouped by the gender of NP1.

Figure 3

Average acceptability ratings for potential attraction conditions grouped by NP1 gender. Error bars represent bootstrapped 95% confidence intervals.

We fit a linear mixed-effects model predicting ratings by NP1 (F-SG vs. M-SG), the number of agreement errors (1 vs. 2), and their interaction. The model found a main effect of NP1 (β = –0.72, t = –2.40, p < 0.05): attraction conditions with a feminine NP1 were rated significantly higher than attraction conditions with a masculine NP1. The model did not find an effect of the number of agreement errors (β = –0.23, t = –0.77, p < 0.45), or an interaction between the number of agreement errors and NP1 (β = 0.42, t = 1.03, p < 0.31).¹³ Thus, for the feminine head nouns, conditions with agreement attraction in only one feature (F-SG F-PL F-PL) were rated as high as conditions with attraction in two features (F-SG M-PL M-PL); for masculine head nouns, conditions with agreement attraction in two features (M-SG F-PL F-PL) were rated as low as conditions with agreement attraction in only one feature (M-SG M-PL M-PL).

4 General discussion

A comparison of the results of the present study with those reported in Fuchs et al. (2015) points to grammatical divergence between native and heritage Spanish. First, concerning the content of number and gender features, English-dominant heritage speakers of Spanish appear to be losing sensitivity to the singular, treating it as the unspecified member of the opposition. The result is a single-valued featural representation of number in heritage Spanish where the native baseline has one that is multi-valued. Second, whereas the native baseline demonstrated differential sensitivity to single- vs. multi-error attraction conditions, heritage speakers lack a similar sensitivity and instead treat the two types of errors on a par. Recall that for heritage speakers in cases of number attraction, gender also participates; this indicates that the two features are projected and valued together, thus supporting a bundling model of their representation, as in (1) above. All told, with respect to gender and number features of Spanish, the native baseline and the heritage grammar show genuine divergence. Returning to our predictions from Table 1 in Section 3.2, our results support the hypothesis that pressures from representational economy drive the observed restructuring in the heritage grammar.

In what follows, we would like to offer three additional considerations. First, we will add a note of caution to the interpretation of our dependent measure. Then we will consider in more detail the pressures that likely led to the observed divergence between the heritage grammar and the native baseline. Finally, we will discuss the implications of our findings for the loss of morphological richness more generally.

5 Conclusions

We have used differences in linguistic behavior—differential response patterns to acceptability judgments for potential attraction conditions—to diagnose differences in linguistic knowledge of gender and number in adjectival agreement at a distance. Previous studies have explored agreement in heritage Spanish (Montrul et al. 2008; 2013; 2014; Alarcón 2011; Montrul 2016), but ours is the first to add linear and structural distance between the agreeing goal and probe while extending the agreement attraction paradigm to heritage speakers. This innovation serves to provide more information about the heritage grammar of agreement while simultaneously suggesting a clear path of restructuring from the monolingual baseline to the heritage grammar.

Our main findings are twofold. First, we show that whereas native Spanish has a multi-valued opposition in the number category (singular and plural are equally specified), heritage speakers are moving toward a single-valued opposition, losing sensitivity to the singular. Second, we find that whereas native speakers of Spanish split their number and gender features, heritage speakers bundle them, projecting and valuing the two features together. The restructuring that leads to this difference in knowledge likely proceeds under systematic pressure from representational economy: heritage speakers shrink their feature inventory and prune their structure so that it includes fewer nodes. A possible competing force is the desire for increased analyticity in the heritage grammar, but pressures from increased analyticity made predictions not supported by the experimental results; this in turn suggests that increased analyticity is not the critical goal of the heritage system’s restructuring in this domain. Crucially, the restructuring of gender and number in heritage Spanish is not a random process, but rather proceeds according to the structural constraints of the grammar in which it operates. The bundling of number and gender is possible because these feature categories are projected close to each other on the nominal spine; no structural rearrangement is necessary for these features to merge.

Returning to the observations that initially motivated the current study, we have found that within Spanish, speakers with varying levels of proficiency possess diverging but related knowledge of the Spanish grammar. Performing the costly task of speaking in a less-dominant language, heritage speakers show evidence of a grammar of agreement that has succumbed to pressures from representational economy. But should we conclude from this finding that those heritage speakers are speaking a different language from the native baseline: that they know a language other than Spanish? Our claim is that we should not. Both groups are speaking Spanish. However, there are different types of Spanish speakers (indeed, speakers of any language) based on the amount of input, maintenance, and interference from other languages. The resulting structuring of grammar in these speakers may differ, as we showed in this paper, but it is a coherent grammar, one that can be accommodated and predicted under the existing models of linguistic variation. Furthermore, the emergence of principled differences of the type documented here underscores the need for comprehensive models of intra-linguistic variation.

This result speaks to the advancing desire to include heritage speakers’ knowledge under the rubric of native speakerhood; such a desire is particularly pronounced in sociolinguistic work, sometimes compelling researchers to reject the term “heritage language” altogether (e.g., Otheguy 2016: 309–311). Although we do not wish to abandon the conception of heritage language, which has served us well and has stimulated a great deal of sophisticated research, we believe that the work presented here gives novel support to the idea that heritage speakers can be included in the intra-linguistic variation within native grammars.

Finally, having documented pressures from representational economy—in addition to the well-established pressures to increase analyticity—shaping the heritage grammar, a new question looms large: is it possible to predict which language domains may be likely to restructure to deliver more parsimonious representations, and which domains may be likely to restructure to increase analyticity? We do not yet have answers to this question, but the results presented in this work compel us to pose it.

Additional Files

The additional files for this article can be found as follows:

Abbreviations

1 = first person, 2 = second person, 3 = third person, ADJ = adjective, ADV = adverb, CL = clitic, COND = conditional, DAT = dative, F = feminine, IPFV = imperfective, INF = infinitive, M = masculine, NEG = negation/negative, PL = plural, PRS = present, SG = singular

Competing Interests

The authors have no competing interests to declare.