Besides the alternation between strong and null pronouns as complements of prepositions, introduced in Section 1 for French and illustrated in (8) for Spanish, the two languages offer further means of pronominalization in the domain of PPs (Pescarini 2021b: 125–127).

An option shared by both languages is that the preposition’s complement takes the shape of a dative clitic and appears outside of the PP (typically adjacent to the finite verb). In both (9) and (10), the dative clitic (le/lui) in (9b-10b) roughly corresponds to the strong pronoun (él/lui) in (9a-10a).

Additionally, French prepositions à and de exhibit an alternation between a strong pronoun in complement position and a clitic pronoun substituting the whole PP (cf. (11) for the preposition de). While lui in (11a) pronominalizes son frère ‘his brother’ and is the complement of de, en in (11b) can be analyzed as pronominalizing the whole PP de ce film ‘of this movie’ (Kupferman 2021: 1025). Such PP-substituting clitics (y and en) are not available in Spanish.

The options are summarized in Table 1. In the following, we will concentrate on the alternation between strong and null pronouns, and where necessary we will use the term weak as a cover term for both null pronouns as in (8) and clitic pronouns as in (9b)-(11b).

In French, and to a lesser extent also in Spanish, the alternation between strong and null pronouns has been linked to animacy. Starting with French there is a vast body of literature linking the pronoun type to the animacy of the antecedent:1 Strong pronouns are considered to (typically) refer to human antecedents (Sandfeld 1965; Pinchon 1972a; 1972b; Thun 1986; Jones 1996; Cardinaletti & Starke 1999; Dobrovie-Sorin 1999; Veldre-Gerner 2007; Adler 2012; Zribi-Hertz 2021; Heidinger 2023; Therrien 2023; Cardinaletti 2024). Null pronouns are considered to be (typically) restricted to non-human ones (Danell 1973; Zribi-Hertz 1984a, 1984b, 2000, 2021; Thun 1986; Wilmet 2003; Adler 2012; Poplack et al. 2012; Grevisse & Goosse 2016; Heidinger 2019; 2023; 2024; Tseng 2021).2 Empirical evidence for such animacy restrictions come from a variety of data types including systematic corpus studies (Danell 1973; Poplack et al. 2012; Shimanskaya 2018; Heidinger 2019) and experimental studies (Shimanskaya 2018; Heidinger 2023; 2024). Heidinger (2019: 13) reports the following frequencies in a corpus of novels: as complements of avec ‘with’ and sans ‘without’, 3^rd person strong pronouns have human antecedents in 96% of the cases while null pronouns have human antecedents in only 10% of the cases. Shimanskaya (2018) conducted a picture selection experiment to determine whether strong pronouns are interpreted as human or inanimate in ambiguous contexts. The results show a clear preference for allowing both interpretations, i.e., both a human and an inanimate antecedent, but unequivocal interpretations are more than twice as often human than inanimate (20 % vs. 8 %) (Shimanskaya 2018: 800).

As for the rigidity of these animacy restrictions, the above authors show considerable variation. The strong pronoun is categorically described as “limited to [+human] interpretation” in Cardinaletti (2024: 210), while Zribi-Hertz states that strong pronouns with an animate antecedent are “more natural” (2021: 1061) than those with an inanimate antecedent, but that strong pronouns can refer to inanimates when no weak form is available (2021: 1062).

Turning to Spanish, the existing literature has typically focused on the animacy restrictions of strong pronouns. It has been assumed that strong pronouns are limited to [+human] antecedents in subject and object positions, but not as complements in PPs. The PP context has been discussed as a syntactic context where the strong pronoun does not have a weak competitor, unlike in subject or object position. This might have to do with the fact that – unlike in French – null pronouns are not available with simple lexical prepositions such as con ‘with’ or sin ‘without’ (cf. (12), Zribi-Hertz 1986; Campos 1991; Fraga 2020; Troberg 2020) – except for cases of ellipsis/background deletion (cf. Pavón 2023; Heidinger 2024).

Minimal pairs as in (13) illustrate this difference between subject position and complement position (of P) in Spanish (Alonso-Ovalle & D’Introno 2001). While in subject position the strong pronoun cannot be used with an inanimate antecedent (13a), it does not underlie such a restriction as a complement of the preposition de ‘of/about’ (13b).

Hence, while the literature on French has focused on the diverging animacy restrictions of strong and null pronouns in one and the same syntactic slot, namely as complements of prepositions, the literature on Spanish has focused on the diverging animacy restrictions of strong pronouns in different syntactic contexts: complement position in PPs vs. subject and object position.

The Spanish data in (12) and (13) already point to competition effects in the domain of animacy restrictions.3 In the present case, the animacy restrictions of the strong pronoun seem to depend on the availability of a weak alternative. With Spanish simple prepositions, a null pronoun is not available and the strong pronoun can be freely used for inanimates (cf. Fernández Soriano 1999: 1220; Alonso-Ovalle & D’Introno 2001: 404; Pomino 2017: 702 for introspection-based confirmations, and Jensen 1973 and Heidinger 2024 for quantitative data from questionnaires and experiments respectively). Animacy restrictions in positions where weak competitors are available have been reported by Jensen (1973), Fernández Soriano (1999), Alonso-Ovalle & D’Introno (2001) and Pomino (2017).4

These empirical observations align with Zribi-Hertz’s (2000) competition-based account of animacy restrictions in French: Only in the absence of a weak competitor can French strong pronouns be used irrespective of the antecedent’s animacy. Recall (5) in Section 1 where the strong pronoun can be used with inanimates as a complement of par ‘by’ and vers ‘towards’ which do not allow for null pronouns. A case that seems particularly interesting for the role of competition in Spanish, and which has so far not been analyzed from this perspective, are spatial prepositions like encima (de) ‘above’, debajo (de) ‘below’ etc. (Pavón 2003; Fábregas 2007; Troberg 2020). The minimal pair in (14) is interesting in two respects: it shows (i) spatial prepositions which allow for both strong or null pronouns, and (ii) that the availability of the null pronoun does not degrade the acceptability of the strong pronoun with inanimates (cf. also Zribi-Hertz 1986: 625, and (8) above).

We thus face the paradoxical situation that in Spanish PPs the availability of a null pronoun does not seem to trigger a competition effect, unlike in French (where it does). The empirical landscape in Spanish is however still unclear both with respect to the availability of the null pronoun and its consequences for the animacy restrictions of strong pronouns. Since both languages exhibit prepositions which allow (e.g., Sp. debajo, Fr. avec), and prepositions which disallow null pronouns as complements (e.g., Sp. con, Fr. par), we can test animacy-related competition effects in French and Spanish in a parallel fashion.

Although we treat prepositions like French avec and Spanish debajo on a par as concerns the availability of null pronouns, we acknowledge that there exist important differences between them (e.g., morphological complexity or necessity of de to introduce overt complements). We address the categorial status of expressions like encima (de) ‘above’, debajo (de) ‘below’ etc. in more detail in Section 2 of the Appendix (specifically the question whether they can be categorized as prepositions), available at https://osf.io/yufbv/files/gw6fp.

Pronouns are flexible expressions due to their minimal lexical content. A pronoun like it is restricted by person (3^rd), number (singular), or animacy (non-human) but not by proper lexical content. Thus, it is already more flexible than lexical noun phrases such as the house, the cat etc. carrying the same grammatical features. This flexibility is further enhanced whenever grammatical or semantic distinctions are not constant. If one assumes that French and Spanish strong pronouns are [+human] to account for the preferences of these pronouns for human antecedents, then this feature cannot be constant given the many instances where strong pronouns have inanimate antecedents (cf. (3), (5), (12), (14) above). Similarly, if one assumes that null pronouns are [–human] or [–animate] this feature cannot be constant given that the null pronoun sometimes has a human antecedent (cf. (4), repeated as (15), and (16)).

There are in principle two analytical options to account for both the default mappings between pronoun types and animacy and for the mismatches, i.e., strong pronouns referring to inanimates and null pronouns referring to humans. The first option would be to assume that animacy features are part of the feature content of these pronouns, and that in the case of mismatches some type of feature manipulation occurs (deletion of animacy features, non-interpretation of features etc.). Cardinaletti & Starke (1999) claim that strong, weak, and clitic pronouns differ with respect to their internal syntactic structure, with decreasing structural richness from strong to clitic pronouns. The [+human]-feature results from a C head, which is part of the structure of strong pronouns but is missing in weak and clitic pronouns. Dobrovie-Sorin (1999) attributes a special role to the grammatical gender features of strong pronouns. The gender features on strong pronouns identify an empty N-head as meaning male/female individual respectively, which causes the animacy restriction. The gender features of strong pronouns are thus “contentful”, unlike the grammatical gender features of weak and clitic pronouns. Hence in both of these proposals animacy is encoded in the structure of strong pronouns. What is missing, however, is an account of deviations from the default, i.e., the use of strong pronouns for non-humans and inanimates.

The second analytical option would be to assume no animacy features on pronouns, hence there cannot be a mismatch between the feature content of the pronoun and the animacy of the antecedent. Building on insights from Zribi-Hertz (2000) we will argue that strong pronouns do not carry animacy features. This accounts in a straightforward manner for the possibility of using strong pronouns for inanimates. In such a proposal, however, one needs to find a way to account for the preference of strong pronouns for human antecedents. We will argue that this preference results from a faithfulness constraint which is violated whenever the semantic gender features of a human antecedent are not matched by grammatical gender features of the pronoun. Both the animacy- and the gender-based analysis are introduced in more detail in Section 4.5

We conduct two acceptability judgment experiments on French and Spanish and follow the same procedure to design the experimental material in both languages. We use a 2×2 design manipulating two variables: PRONOUN TYPE and ANIMACY. The PRONOUN TYPE variable refers to the type of pronoun used as complement of a preposition. It has two levels: strong pronoun (lui, elle in French, él, ella in Spanish) or null pronoun. The ANIMACY variable refers to the animacy of the antecedent of the manipulated pronoun. This variable also has two levels: human or inanimate.6 The design results in the 4 conditions in (17).

We create 16 items following this 2×2 design in each language. To test the relevance of competition we compared two contexts: one with competition between strong and null pronouns, the other without competition, allowing only strong pronouns. These two contexts are obtained by the manipulation of preposition type. We use introspective judgements drawn from the literature to establish these two sets of prepositions (Zribi-Hertz 1984a; 1984b; 1986; Troberg 2020; Tseng 2021 for French; Pavón 2003; 2023 for Spanish): with-competition prepositions avec ‘with’, sans ‘without’ in French, delante ‘in front’, detrás ‘behind’ in Spanish vs. without-competition prepositions malgré ‘despite’, vers ‘towards’ in French, con ‘with’, sin ‘without’ in Spanish. Hence, in addition to the two within-item variables PRONOUN TYPE and ANIMACY our design includes the between-item variable COMPETITION, which again has two levels: with-competition and without-competition items, with 8 items for each level.

All items follow the same structure, which we illustrate using (18). First, an initial frame-setting adjunct (Près des falaises ‘near the cliffs’) provides some general orientation. Next, a context part (Rémi rejoint sa bouée ‘Rémi joins his buoy’) introduces two singular lexical NPs (NP1 and NP2) in a transitive structure, NP2 being manipulated for animacy (human/inanimate). Finally, the critical part (puis il nage avec ‘then he swims with it’), contains the preposition (Prep) manipulated for competition (with/without-competition) and the pronoun (Pro) manipulated for form (strong/null). Additionally, NP1 and 2 always differ in grammatical gender: if NP1 is feminine, NP2 is masculine, and conversely, if NP1 is masculine, NP2 is feminine (these two possibilities are equally distributed across the items).

The item in (18) thus has the following properties in terms the experimental variables: inanimate×null×with-competition. The Spanish item in (19) displays the following properties: human×strong×with-competition.

To summarize, we created 16 items for each language, using a 4-conditions design (ANIMACY × PRONOUN TYPE). Among those, 8 items contain with-competition prepositions (avec ‘with’, sans ‘without’ in French, delante ‘in front’, detrás ‘behind’ in Spanish), and 8 items contain without-competition prepositions (malgré ‘despite’, vers ‘towards’ in French, con ‘with’, sin ‘without’ in Spanish).

The acceptability judgment task relies on a 5-point Likert scale ranging from Absolutely unacceptable to Perfectly acceptable with no additional labels in between. Participants read sentences one by one on a screen and rated the acceptability using the Likert scale. The item presentation follows the Latin-square method: each participant sees all the conditions and items, but is presented with each item in only one condition. The conditions are counterbalanced across participants, and the presentation order is randomized.

We recruited 32 participants in each language using the platform Prolific. We pre-screened participants based on demographic data to study a more homogeneous sample. We used the following criteria respectively: participants must be born and currently live in France/Spain, have French/Spanish nationality, have spent most of the time before turning 18 in France/Spain, have French/Spanish as their primary language and must be between 22 and 60 years old. As we did not use a control task in our experiments, we analyzed data for all 64 participants.

The group of 32 French participants includes 13 women and 19 men, with a mean age of 32 (SD = 8.1). The group of 32 Spanish participants includes 10 women, 1 non-binary and 21 men, with a mean age of 34 (SD = 9.7). Anonymized data and R scripts for the statistical analysis and supplementary results can be found on the OSF repository of this study: https://osf.io/yufbv/.

We found effects of competition in both languages (Figure 1). In French, competition affects the interaction between animacy and pronoun type. In contexts with competition, where the two pronoun types are available, strong pronouns are more acceptable with human antecedents and null pronouns with inanimate antecedents. In contexts without competition, null pronouns are unacceptable, and strong pronouns are as acceptable with human as with inanimate antecedents.

In Spanish, all factors (animacy, pronoun type and competition) interplay in two-way interactions. First, competition affects the acceptability of strong/null pronouns and inanimate/human antecedents. In contexts without competition, animacy does not play a role, but pronoun type does: strong pronouns are acceptable, and null pronouns are not. In contexts with competition, both types of pronouns are acceptable, with strong pronouns being still more acceptable, and animacy also has an effect: human antecedents are more acceptable. Finally, there is also an interaction between animacy and pronoun type: strong pronouns are more acceptable with human antecedents.

Taken together, the results for French and Spanish suggest that competition is a crucial factor to understand animacy restrictions on strong pronouns in PPs: animacy restrictions are present only in with-competition contexts, where null pronouns are available. In without-competition contexts, strong pronouns do not exhibit animacy restrictions in either language.

The main difference between French and Spanish, however, lies in the acceptability pattern of null pronouns in competition contexts. In French, null pronouns are more acceptable with inanimate antecedents than with human antecedents. As such, it seems that the availability of an acceptable null pronoun for inanimate antecedents makes the strong pronoun less acceptable in the same context. However, in Spanish, null pronouns are mildly acceptable, with no clear animacy preference. Despite this, even if the alternative null form is not better, strong pronouns in Spanish still exhibit a dispreference for inanimate antecedents. This shows that competition effects not only arise when a competitor is better (as in French), but also when a competitor is merely possible and even less acceptable (as in Spanish).

To further evaluate whether the acceptability of strong pronouns with inanimate antecedents correlates with the acceptability of the null competitor we define two measures: First, a measure of animacy restrictions on strong pronouns, Δ_Acc, corresponding to the acceptability difference between human and inanimate strong pronouns (20a). Second, Acc_Null, a measure corresponding to the mean acceptability of null pronouns (20b). The relationship between these two variables indicates how animacy restrictions on strong pronouns change with the acceptability of null pronouns.

Figure 2 represents the relationship between Δ_Acc and Acc_Null in French and Spanish across items (16 observations per language). We observe positive correlations: when the null pronoun gets more acceptable, the acceptability difference between human and inanimate gets bigger, i.e., the animacy restriction gets stronger. We report the Pearson coefficients of linear correlation: 0.42 in French and 0.30 in Spanish. These can be considered as weak to moderate correlations (Cohen 1992; Gignac & Szodorai 2016). In addition to the stronger correlation found in French, we also note that the slope of the regression line is bigger in this language (β = 0.37) than in Spanish (β = 0.17). Overall, without-competition items cluster at the left of the Acc_Null scale (Acc_Null < 2), with a null Δ_Acc (Δ_Acc ≈ 0), corresponding to the absence of animacy restrictions. For with-competition items, both Acc_Null and Δ_Acc are higher, and they show more variation.

The empirical results that need to be accounted for in the subsequent formal analysis are thus the following:

In our constraint-based analysis, we adopt an optimality-theoretic approach to competition (Bresnan 2001a, 2001b). We follow standard assumptions about the components of a grammar in Optimality Theory (Prince & Smolensky 1993) according to which a grammar includes:

In the context of optimality-theoretic syntax, inputs and outputs have received different representations depending on the specific formal framework used together with OT. In transformational OT frameworks, inputs and outputs correspond to derivational steps in a syntactic structure before the PF/LF interfaces (Grimshaw 1997; Legendre et al. 1998; Müller & Sternefeld 2001; Speas 2001). In OT-LFG (Bresnan 2000), inputs are grammatical feature matrices (f-structures), which are paired in the outputs with constituency tree realizations (c-structures). Using Distributed Morphology, Trommer (2001) considers inputs to be sets of morphological features, and outputs to be inserted Vocabulary Items. We remain here agnostic regarding these different options, hoping that our analysis can be easily transposed into different frameworks.

To first introduce the two analyses and their respective constraints, we will simplify the picture offered by our experimental results, considering only the preference for either the strong or null pronoun (cf. Table 2), rather than relying on the gradience of acceptability data. We postpone the integration of the actual gradient acceptability judgments to the analysis within Stochastic Optimality Theory (Section 4.2).

Given that this article deals with the alternation between strong pronouns and null pronouns as complements in PPs, we assume here that inputs consist of a preposition and an abstract pronoun (annotated as Pro, following Speas 2001), the latter being potentially associated with some features (such as animacy). We consider two candidates as possible realizations for Pro: a strong pronoun (represented as LUI in French and ÉL in Spanish) or a null pronoun (Ø).

We start by introducing two structural constraints which are directly linked to the two types of pronouns in the candidate set.7 In the case of strong and null pronouns, two opposing constraints favor either one type or the other: *STRONG and *NULL. Crosslinguistic distribution suggests that strong pronouns constitute the unmarked type of pronoun compared to weaker pronouns (weak, clitic, null etc.): while no language lacks “free” pronominal forms, some languages lack weak, clitic, or null pronominal forms (Bresnan 2001b). This can be explained if *NULL, which penalizes null pronouns, is ranked above *STRONG in those languages where only strong pronouns are found. At the same time, we observe that in languages which do have weak pronominal forms there is a tendency to use weaker pronouns (weak, clitic, null etc.) instead of strong ones. This preference can be considered as a particular instance of a more general economy principle (e.g., Avoid Pronoun (Chomsky 1981), Structural Economy Principle (Safir 1993) or Minimize Structure (Cardinaletti & Starke 1999); see also Zribi-Hertz (2000) for its relevance in the prepositional domain). In those languages, *STRONG would be ranked above *NULL, hence disfavoring the use of strong pronouns.

Obviously, these two constraints are not sufficient to account for the data since either ranking would result in only strong pronouns or only null pronouns to be used (or be acceptable), whereas our data show that both types are possible in French and Spanish. The remaining necessary constraints will be introduced in the following paragraphs.

A key observation about the distribution of null pronouns in French and Spanish PPs is that it is linked to the preposition itself (this was also the premise of distinguishing with-competition from without-competition contexts in the experiments reported in Section 3). Still, it is not entirely clear which factors determine whether a preposition (dis)allows null pronouns (cf. Heidinger & Troberg (accepted) for a recent comparative overview).

In Spanish, the morphological structure of the preposition seems to be a good predictor in that only morphologically complex prepositions (encima, delante, detrás etc.) allow for null pronouns (Zribi-Hertz 1984b; 1986; Pavón 2003; Fraga 2020; cf. Campos (1991) on para ‘for’ being a potential exception). The nominal component of complex prepositions seems to be relevant in that the null pronoun in such cases is the complement of the noun (within the morphologically complex preposition)8 (cf. Fernández Soriano & Ordóñez (2024) for analysis of detrás, delante etc. containing an internal preposition plus a nominalizing null suffix, and Troberg (2020) on the nominal component of complex prepositions in French). Another proposal for Spanish null pronouns with complex prepositions (which also relies on morphological complexity) comes from Fraga (2020: 28) who suggests that “big Ps” allow for null pronouns because the omission of the complement of a head must be “supported by a clitic”, which according to her is de- in the case of complex prepositions.

In French, the distribution of null pronouns in PPs is more complicated (Zribi-Hertz 1984a; 1984b; 1986; Troberg 2020; Tseng 2021; Heidinger & Troberg accepted). Morphologically complex prepositions (e.g., en face) seem to generally allow for null pronouns and simplex but semantically weak prepositions (e.g., à, de, en) disallow null pronouns. There is however a third and heterogeneous class: simplex but semantically strong prepositions, which show a mixed behavior as some of them (e.g., avec, sans) allow for null pronouns, while others (e.g., parmi, chez) do not.

Given this situation and the fact that our goal is not to account for which prepositions (dis)allow null pronouns, we will use a constraint LEX which delegates the availability of null pronouns on the idiosyncratic lexical properties of individual prepositions (Bresnan 2001b).

The three constraints introduced so far are relevant to both analyses that will be introduced in the following. The type of constraint that distinguishes these two analyses from each other are faithfulness constraints. Such constraints demand that features specified in the input are also present in the output (the output needs to be ‘faithful’ to the input). Setting up these constraints requires to assume features on the input and to assign specific features to strong and null pronouns (as the two pronoun types of the output candidates). Based on the features that are decisive for the analyses we dub the analyses animacy faithfulness vs. gender faithfulness.

In the animacy faithfulness analysis, pronouns are specified with animacy features: strong pronouns are [+human] and null pronouns are [–animate]. The faithfulness constraint evaluating the correspondence between input and output animacy is FAITH_ANIMACY (23).

With this additional constraint, we can now account for the use of strong and null pronouns in French PPs with human and inanimate antecedents. Beginning with the expression of human and inanimate complements with with-competition prepositions (here: avec), the empirical observation to account for is that strong pronouns are preferred over null pronouns with human antecedents and that null pronouns are preferred over strong pronouns with inanimate ones (cf. Section 3.2 and Table 2). Tableaux 1 and 2 show the constraint violations and constraint rankings necessary to account for these data.9 Just considering the constraint violations of the candidates, Tableau 1 shows that either FAITH_ANIMACY or *NULL is ranked above *STRONG. The violations in Tableau 2 show that either FAITH_ANIMACY or *STRONG is ranked above *NULL. Taken together the violations in Tableaux 1 and 2 show that FAITH_ANIMACY is ranked higher than both *NULL and *STRONG. This is represented in Tableaux 1 and 2 by the relative positions of the constraints (the highest ( = leftmost) position of LEX will be motivated below). The ranking between *NULL and *STRONG cannot be determined based on the data from Table 2.

The second type of preposition, i.e., the one which does not allow for null pronouns, is exemplified with vers ‘towards’ in Tableaux 3 and 4. Recall that with such prepositions the use of the null pronoun violates LEX, i.e. the prepositions’ ban on null pronouns.10 Given that FAITH_ANIMACY is ranked higher than both *NULL and *STRONG the constraint violations in Tableau 3 show that LEX is ranked higher than FAITH_ANIMACY. The candidate LUI_[+human] is optimal despite it being unfaithful to the Pro_[–animate] input. Tableau 4, with a human input, does not provide further insights on constraint rankings, but it confirms the ones established so far. Hence, the relative ranking of *NULL and *STRONG remains open, and it will be the analysis of the gradient acceptability data (Section 4.2) that informs about their ranking.

Taken together Tableaux 3 and 4 also capture the competition effect on the animacy restrictions of the two types of pronouns. Crucially, the acceptability of one form (strong pronouns) depends on the acceptability of another form (null pronouns). The absence of a null competitor enhances the functional range of the strong pronoun with prepositions like vers or malgré. As for the specific type of competition effect, this scenario corresponds to what we have called boosting in Section 1: a form specified for a restricted context (here: [+human]) appears in a context which is normally not meant for it (here: [–animate]).

The final set of constraints assumed in the Animacy Faithfulness analysis is thus the following:

And their ranking in French is the following: LEX ≫ FAITH_ANIMACY ≫ {*STRONG, *NULL}

The alternative analysis, inspired by Zribi-Hertz (2000), considers that neither strong nor null pronouns are specified for animacy, i.e., animacy is not part of the pronouns’ feature content. In this analysis, the observed animacy effect of strong/null pronouns relies on gender, and the asymmetric specification of the two pronoun types in this respect: strong pronouns, not null ones, carry gender features.11 This establishes a privileged relation between human referents and strong pronouns since only human antecedents carry semantic gender features (female/male) and only strong pronouns carry grammatical gender features (masculine/feminine). We thus hypothesize that the gender feature of strong pronouns is faithful to the semantic gender of their antecedent.12 This second analysis thus comes with a different faithfulness constraint: FAITH_GENDER (24a.).

The empirical facts that need to be accounted for are the same as with animacy faithfulness. First, in the case of avec, null pronouns are preferred for inanimate antecedents. This follows from the *STRONG ≫ *NULL ranking (Tableau 5). In the case of human antecedents, the use of the null pronoun is circumvented by the higher ranking of FAITH_GENDER (Tableau 6), which forces the use of a strong pronoun. It is thus better to be faithful to the gender specification (represented by the [+gender] feature), than to avoid strong pronouns. The position of LEX relative to the other constraints will be described in Tableaux 7 and 8.

Since in this analysis strong pronouns do not carry animacy features, animacy cannot be the reason why strong pronouns are not preferred (or commonly used) for inanimate antecedents. The use of the strong pronoun is instead blocked by a candidate that does not violate *STRONG, i.e., the economy principle striving for lighter pronominal forms. There is thus a striking difference between the two analyses with respect to the manifestation of competition. In the animacy faithfulness analysis, strong pronouns which are specified as [+human] are boosted to being used for inanimate antecedents if the null pronoun is not available (e.g., with prepositions like vers or malgré). In the gender faithfulness analysis, strong pronouns are blocked from being used for inanimate antecedents with prepositions like avec or sans because these prepositions allow for a better candidate in terms of economy. Hence, we are dealing with boosting in the animacy faithfulness analysis, and with blocking as the type of competition effect in the gender faithfulness analysis. Another noteworthy difference between the two analyses is where a competition effect manifests itself. In the animacy faithfulness analysis, competition manifests itself with prepositions which do not allow for null pronouns, while in the gender faithfulness analysis it is with prepositions which do allow for null pronouns.

To complete the picture, let us briefly consider within the gender faithfulness analysis those prepositions which do not allow for null pronouns (Tableaux 7 and 8). For inanimate antecedents, LEX favors strong pronouns. For human antecedents, both LEX and FAITH_GENDER favor strong pronouns (although these two constraints remain unranked with respect to each other at this point).

The ranking so far under the gender faithfulness analysis is the following: {LEX, FAITH_GENDER} ≫ {*NULL, *STRONG}.

After considering the two analyses in light of the French data, we can attempt a first evaluation. Since both analyses capture the preferences observed in the experimental data, the evaluation must rely on conceptual considerations. In this respect the animacy faithfulness analysis might at first sight seem more appealing, since the empirical observation we need to account for are animacy and not gender restrictions in the use of strong and null pronouns. There are, however, several advantages of the gender faithfulness analysis which in our view weigh heavier. First, the gender faithfulness analysis derives the same pattern with fewer specifications (as animacy features on pronouns are no longer necessary). Second, it builds on gender features that are needed independently (e.g., for agreement), and whose presence/absence on the pronouns is undisputable, because they are morphologically realized (unlike the animacy features on pronouns). Third, it is compatible with more general assumptions about the feature specification of pronouns. After all, one questionable assumption of the animacy faithfulness analysis is that the null pronoun presents as many feature specifications as the strong one, whereas it is often assumed that null pronouns are less specified forms (Cardinaletti & Starke 1999; Alonso-Ovalle & D’Introno 2001; Bresnan 2001a; Siewierska & Bakker 2006).

Another interesting outcome of the gender faithfulness analysis is that null pronouns for humans but not strong pronouns for inanimates violate the faithfulness constraint (FAITH_GENDER). In the animacy faithfulness analysis, however, both null pronouns for humans and strong pronouns for inanimates violate the faithfulness constraint (FAITH_ANIMACY). The asymmetry predicted under the gender faithfulness analysis is reflected in our gradient acceptability data: inanimate strong pronouns are more acceptable than human null pronouns (Figure 1 in Section 3.2).13 Given that the analysis so far is sketched in a classical OT framework, which does not allow for gradient phenomena such as acceptability differences, we cannot directly account for this result. One of the motivations for applying a Stochastic Optimality Theory analysis in Section 4.2 is to overcome precisely this limitation.

Before that, we briefly turn to the Spanish data. The only preference to account for in a classical OT is that strong pronouns are preferred over null pronouns independently of the preposition type (competition or no competition) and the referent’s animacy (human or inanimate, see Table 2 preferences). This can be achieved just by the ranking of the structural constraints: *NULL ≫ *STRONG.

Although the classical OT framework helps us to capture the general preferences observed in our data it has its limits in accounting for the gradient acceptability data from our experiments and for some of the crucial findings. In the case of Spanish, this becomes most obvious with the prepositions which do allow for null pronouns (delante and detrás in our experiment). We observed that null pronouns are more acceptable with such prepositions than with con and sin, but strong pronouns are nevertheless the preferred option. This difference with respect to the status of null pronouns is not captured in the ranking *NULL ≫ *STRONG; and it could not be captured even if we were to include more constraints (LEX and faithfulness constraints) in a classical OT. Further, we observed that the availability of null pronouns has an impact on the acceptability difference between strong pronouns for humans and for inanimates. Recall that the difference in acceptability between strong pronouns for humans and for inanimates is bigger with delante and detrás than with con and sin, which suggests a competition effect also in Spanish: the increased acceptability of null pronouns has an impact on the acceptability of strong pronouns for inanimate antecedents, even if null pronouns are not the preferred option. We consider such cases of ‘blocking from below’ as special instances of blocking.

To align our analyses with the gradient data, and to further evaluate the animacy and the gender faithfulness analysis, we turn in Section 4.2 from classical OT to Stochastic OT. Applying all the constraints discussed so far to our gradient acceptability data will allow us to evaluate the rankings of also those constraints which remained unranked in the classical OT analysis and to verify how well the two analyses capture the empirical data respectively.

Stochastic Optimality Theory (SOT) is an OT-based framework meant to allow for both categorical and gradient grammaticality contrasts (Boersma 1999; Hayes 2000; Boersma & Hayes 2001). Its major differences compared to classical OT are the following:

SOT comes with a learning algorithm, called the Gradual Learning Algorithm (GLA), which takes as input a set of constraints and a frequency distribution, and outputs a constraint ranking with weights adjusted to this frequency distribution.15

We use the GLA to analyze our experimental data and produce constraint rankings for French and Spanish under the two analytical options introduced in Section 4.1: animacy faithfulness (25a) and gender faithfulness (25b).

The GLA requires a frequency distribution as input. Since our data consist of acceptability judgements only, we apply a pre-processing step to transform the acceptability differences between strong and null pronouns into estimated frequencies (see Boersma & Hayes 2001: 82 for details). We then use these frequency distributions together with the set of constraints defined in (25) to derive rankings for French and Spanish.

The resulting constraint rankings are shown in Figure 3, with the left panel corresponding to the animacy faithfulness analysis and the right to the gender faithfulness analysis.

To evaluate the respective analyses and check whether they correctly accounted for the observed data, we use the GLA to produce frequency distributions given the different constraint rankings. We then turn back these frequencies into predicted acceptability judgment differences and we compare them against the actual ones. We measure the accuracy of the predictions with the R² statistics, which estimates the linear fit between the actual and predicted acceptability judgment differences. The results are given in Table 3.

Both analyses yield correct accuracy scores for Spanish (R² = 0.99). However, for French, the gender faithfulness analysis better accounts for the data (R² = 0.97) than the animacy-based one (R² = 0.81). While the animacy faithfulness analysis yields overall less accurate acceptability differences, its biggest error concerns the ‘malgré/vers Pro_[+human]’ input, where it predicts that the strong pronoun has 3.34 acceptability points more than the null pronoun, contra 2.47 predicted by the gender faithfulness analysis, and 2.16 in the observed experimental data.

To understand this error, we need to have a look at the rankings in Figure 3. The crucial difference between the two analyses lies in the local ranking between *STRONG and *NULL, as shown in (26).

In the animacy faithfulness analysis, FAITH_ANIMACY favors human strong pronouns and inanimate null pronouns. As *NULL is ranked higher than *STRONG, it interacts more often with FAITH_ANIMACY. This correctly predicts an asymmetry for unfaithful outputs in competition contexts: null human pronouns are less acceptable than inanimate strong pronouns. However, this same ranking leads to incorrect predictions when it comes to without-competition contexts. Indeed, *STRONG is now ranked too low to compensate the effect of the strong-favoring constraints (LEX, *NULL and FAITH_ANIMACY with human antecedents), making human strong pronouns more acceptable than inanimate strong pronouns, contrary to the fact. The animacy faithfulness analysis thus faces a paradox: *NULL has to be ranked higher than *STRONG to penalize human null pronouns in competition-contexts, but *STRONG also has to be ranked higher than *NULL to equate the acceptability of human and inanimate antecedents in without-competition contexts.

The gender faithfulness analysis does not suffer this issue, because its faithfulness constraint FAITH_GENDER already embodies the asymmetry between inanimate and human antecedents: it penalizes only human null pronouns. Thus, *NULL can be ranked lower than *STRONG, so the latter can now compensate for the constraints favoring strong pronouns in without-competition contexts. This leads to the similar acceptability rates for human and inanimate strong pronouns in without-competition contexts observed in the experimental data.

In Section 4.1., we have already suggested that the gender faithfulness analysis is conceptually superior to the animacy faithfulness analysis. The SOT analysis provides further support in that the gender faithfulness analysis is also empirically more accurate for French. We assume that gender faithfulness holds in both French and Spanish, even if the animacy faithfulness analysis performs equally well in Spanish. This makes sense within OT, since all languages obey the same constraints but differ in terms of their ranking. Additionally, applying the same analysis highlights interesting differences between the two languages in terms of constraint ranking, as shown in (27).

The main difference between the two languages lies in the relative ranking of the structural constraints *STRONG and *NULL. In French, *STRONG is ranked higher than *NULL. Strong pronouns can thus be considered more marked structures than null pronouns in French PPs: when no other constraint is violated (LEX and FAITH_GENDER), null pronouns tend to be used. However, the faithfulness constraint FAITH_GENDER is ranked higher than the structural constraint *STRONG. This means that there is a stronger preference for marked and faithful outputs (i.e., strong pronouns) than for unmarked and unfaithful outputs (i.e., null pronouns). By contrast, in Spanish, *NULL is ranked higher than *STRONG, which makes null pronouns an overall more marked structure. All other things being equal, Spanish avoids null pronouns in PPs more than French. Given the well-known differences between French and Spanish with respect to the licensing of null subjects, the results from our study show that different syntactic contexts exhibit quite different restrictions as regards the choice of overt vs. null pronominal elements.

In our evaluation of the animacy-based approach, we focus on the analysis where both pronouns are specified for animacy since this corresponds to Zribi-Hertz’s (1984a; 1984b) original proposal. In addition, we have determined the R² score for two animacy-based alternatives: (i) the strong pronoun is specified as [+human] and the null pronoun is not specified for animacy, (ii) the null pronoun is specified as [–animate] and the strong pronoun is not specified for animacy. The scores for (i) are 0.97 (French) and 0.99 (Spanish), and the scores for (ii) are 0.64 (French) and 0.98 (Spanish). While analysis (ii) has to be discarded based on the low R² score for French and general conceptual disadvantages (it requires the stipulation of animacy features; the null pronoun carries more features than the strong pronoun), analysis (i) deserves more attention. After all, it captures the empirical results equally good as the gender-based analysis in terms of the R² score. However, it suffers from the conceptual disadvantage of requiring the stipulation of animacy features. Since this is not the case in the gender-based analysis as gender features are needed independently (e.g., for agreement), we favor the gender-based analysis over the animacy-based analysis where only the strong pronoun is specified for animacy. Although our decision between these two analyses ultimately relies on conceptual arguments, the determination of the R² score is nevertheless crucial. It is only after the evaluation of the analyses’ empirical power, and the outcome that they capture the data equally well, that conceptual arguments become decisive.