Closest conjunct agreement with attributive adjectives

Abstract Contrary to most French grammars claiming that French only allows masculine agreement when mixed-gender nouns are conjoined, we show that closest conjunct agreement (CCA) does exist in contemporary French, as in other Romance languages, and is the preferred strategy for prenominal adjectives. Using data from a large corpus (FrWac) and an acceptability rating experiment, we show that (feminine) CCA is well accepted in contemporary French, and should be distinguished from attraction errors, despite the norm prescribing masculine agreement. We also show the role of the adjective position, i.e. prenominal or post-nominal, and humanness. CCA is the preferred strategy for prenominal adjectives, and non-human nouns favour CCA for post-nominal adjectives. Assuming a hierarchical structure for coordination, the closest noun is the highest in A-N order, whereas it is the lowest in N-A order. Thus CCA in prenominal position may be favoured by a shorter structural distance. One can also see CCA with a prenominal adjective as ‘early’ agreement. Regarding humanness, grammatical gender is interpreted as social gender with human nouns, and a masculine plural can refer to a mixed group. This ‘gender neutral’ plural may favour masculine agreement for human nouns, or the prescriptive norm is more influential for human nouns.


Introduction
Agreement with coordination is a complex issue and different languages may use different strategies (cf. Sadler, 1999;Wechsler and Zlatić, 2000;Kuhn et al., 2007;Borsley, 2009;Dalrymple and Hristov, 2010). When the coordination includes conjuncts with conflicting features, languages may follow resolution rules (cf. Givón, 1970;Dalrymple and Kaplan, 2000) or use closest conjunct agreement (cf. Corbett, 1991;Wechsler and Zlatić, 2003). A single language, such as Latin, may use different strategies in case of coordination of nouns with different genders: masculine resolution (1-a) (for human nouns) or agreement with the closest conjunct (1-b (Corbett, 1991, p.193) b. eadem alacritate ac studio same.F.SG.ABL ardor.F.SG.ABL and zeal.N.SG.ABL 'with the same ardor and zeal' (Caesar BG. 4.24;cited in Johnson 2013, p.6) Romance languages, such as Italian, Spanish and Portuguese, use the same masculine resolution strategy, but also allow for CCA. In example (2-a) (from Demonte and Perez-Jimenez, 2012), the determiner and prenominal adjective show CCA while the postnominal adjective is not marked for gender. CCA is observed for the post-nominal adjective in example (2-b) (Villavicencio, Sadler and Arnold, 2005), and for the prenominal adjective in Italian (2-c) (Benincà, 1988 Feminine agreement with nouns of mixed genders is also attested in classical French (Viennot, 2014). Aixiu  formation 'All citizens must contribute personally or through their representatives to its formation.' (de Gouges, Déclaration des droits de la femme, 1791) In (3a), the postnominal adjective nouvelle has wide scope over the two nouns, since ?Armez-vous d'un courage ('arm yourself with a courage') is weird without an adjective: un ('a') refers to a subtype of courage and de would be used without an adjective: Armez-vous de courage ('arm yourself with courage').

The masculine resolution rule in contemporary French
According to some authors (cf. Viennot, 2014), contemporary French has lost the possibility of CCA due to the power of the (masculine) prescriptive norm. Other authors suggest a more nuanced view: Grevisse and Goosse (2016, p.338) write: "si les noms sont de genre différents, l'épithète se met au genre indifférencié, c'est-àdire masculin". But they add: "La règle générale n'est pas toujours respectée [:::]. La tradition grammaticale, qui correspond à un certain sentiment des usagers, estime choquant pour l'oreille que le nom féminin soit dans le voisinage immédiat de l'adjectif." (5) (Grevisse and Goosse, 2016, p. 339) 4 and conclude: "il est préférable chaque fois que cela est possible d'accorder avec l'ensemble des noms". 5 3 'if the nouns are of different genders, the attribute must be with indistinct gender, i.e. masculine' 4 'the general rule is not always observed. [:::] The grammatical tradition, which corresponds to some speakers' feeling, finds it shocking for the ear if the feminine noun is immediately close to the adjective.' 5 'it is preferable every time it is possible to agree with the set of nouns'. Notice that in (5)  According to Curat (1999), examples (6-a), (6-c) and (6-e) obey the resolution rule but (6-a) is not acceptable. In (6-b), CCA is claimed to be dubious because it is not compatible with the resolution rule. In post-nominal position, Curat (1999) claims that the feminine adjective only modifies N2 (6-d), whereas the masculine adjective may modify the coordination as a whole (6-e Thus, the status of the masculine resolution rule is unclear in French, especially for prenominal attributive adjectives when the closest noun is feminine (5) (6-a). Furthermore, most authors rely on their own intuition or a handful of attested examples.
The aim of this article is to pursue a quantitative and empirical study of gender agreement of French attributive adjectives in case of coordination of nouns with different genders. We provide new contemporary data on gender agreement of plural adjectives with plural nouns, which shed some light on the acceptability of CCA, and on the factors that may favour it over resolution.

CCA and Corbett's agreement hierarchy
Putting French in a more general perspective, we take into account Corbett's typological work. Looking at a wide variety of languages, Corbett considers that agreement is an asymmetric relation between a controller and a target, involving syntactic, semantic and pragmatic factors. In his view, CCA is not a 'repair' mechanism yielding acceptable but otherwise ungrammatical output (Peterson, 1986;Bhatt and Walkow, 2013), but a strategy available in many language families, such as Slavic or Bantu languages. He proposes that three factors may favour CCA for languages with different agreement strategies (see Corbett, 1983Corbett, , 1991Corbett, , 2006: • Controllers referring to inanimates • Targets before Controllers (postverbal subject vs. preverbal subject) • Agreement hierarchy attributive < predicate < relative pronoun < personal pronoun ← non-resolution ------------resolution → In his view, animate nouns may favour a more semantic agreement, for example plural agreement with two singular conjuncts, hence resolution. The interaction between CCA and directionality is compatible with a structural explanation (when the subject is postverbal, the closest noun is the highest noun, see below Section 4.2). It is compatible as well with an incremental processing account: only in backward agreement, when the target follows the controller, does the speaker know the features of all conjuncts before computing target agreement.
The agreement hierarchy may also be explained in terms of linear and structural distance. An attributive adjective (6-c) (6-b) is closer to the (controller) noun (and belongs to the same noun phrase) than a predicative one, which belongs to the verbal phrase and may be separated from the (controller) noun by a copula (5)  (7), hence a penalty for CCA for predicative adjectives. If these factors apply to French gender agreement, CCA should be more frequent for prenominal adjectives than for post-nominal adjectives. We also expect differences between human and non-human nouns, if human nouns favour resolution in general.
The article is organized as follows: Section 1 briefly reviews previous literature about adjective agreement and positions in French. The new empirical data are presented in Section 2 (a corpus study) and Section 3 (an acceptability experiment). We analyse the results, discuss the factors that interact with CCA, and compare gender and number agreement in Section 4.

French nouns and gender
Like most other Romance languages, French has two grammatical genders, feminine (fem) and masculine (masc). For non-human nouns, grammatical gender is usually considered arbitrary. For example, the noun chaise 'chair' (la chaise) has feminine gender, while the noun livre 'book' (le livre), on the other hand, has masculine gender.
As in many languages, grammatical gender is usually associated with social gender for human nouns: masculine nouns tend to refer to males and feminine nouns to females. For example, garçon ('boy') is masculine and fille ('girl') is feminine, even if many personal nouns have a common gender, such as journaliste, 'journalist.M/F'. 6 6 We leave aside some fixed gender nouns that can refer to both men and women (un génie 'a.M.SG genius', une victime 'a.F.SG victim'). See Bonami and Boyé (2019) for a discussion of the evolution.
For human nouns presenting gender alternation (chanteur/chanteuse 'singer.M/ singer.F'), we leave open the debate whether they are formed by parallel suffixation, by derivation of one noun from the other, or by inflection from a common lexeme (Spencer, 2002;Bonami and Boyé, 2019).
In what follows, we use a written corpus with various adjectives, but in our experiment we only use adjectives with an audible gender marking (petit/e) (Section 3).

Agreement strategies for French attributive adjectives
Attributive adjectives in French can be in pre-or post-nominal position, depending on their semantic class (cf. Wilmet, 1981;Bouchard, 1998;Miller, Pullum and Zwicky, 1997). For example, indefinites and cardinals are only found in prenominal position (8-a)- (8-b), while size, color or relational ones occur post-nominally (8-c) (8-d (9) can alternate between preand post-nominal positions with roughly the same meaning (Abeillé and Godard, 1999;Thuilier, Fox and Crabbé, 2012;Laenzlinger, 2005).  Bonami and Boyé (2005) argue that for singulars, on top of feminine and masculine, a third form is needed for prenonimal masculine singular adjectives (vieux, vieil, vieille). We leave aside this issue here since we only consider plural forms. 8 As in English, all adjectives must occur after the noun when they have a phrasal dependent: un bon repas 'a good meal'/un repas bon pour moi/*un bon pour moi repas 'a meal good for me', hence the [WEIGHT light] constraint proposed by (Abeillé and Godard, 1999) or the [LEX ] constraint by (Sadler and Arnold, 1994) Riegel, Pellat and Rioul, 2018), French would be unique among Romance languages in only allowing resolution in case of conflicting genders: the coordination of a masculine and a feminine noun is supposed to be resolved to masculine, regardless of whether the nouns are animate (10) or not (4).

10)
Un père et une mère excellents a.M.SG father and a.F.SG mother.F.SG excellent.M.PL 'an excellent father and mother' (Corbett 1991, p.279) Masculine agreement is sometimes considered as default agreement (Grevisse and Goosse, 2016), since it is used for expressions without grammatical gender: it is the gender used with sentential or verbal subjects (11-a) and with the expletive pronoun (11-b).

11) a.
Bien The difference between default agreement and resolution rules has been debated (cf. Nevins and Weisser, 2019), since in languages with three genders, default agreement is usually neuter while resolution is usually masculine. We use the term resolution rule in this article referring to masculine agreement when there is a conflict of features. 9 In Section 2 and Section 3, we use coordination with one determiner shared by two plural nouns D A N1 et N2 or D N1 et N2 A. We choose plural nouns so that there is no interaction with number agreement (see An andAbeillé (2017, 2019) for a discussion of CCA for number agreement with French determiners and adjectives). It is assumed that with a shared determiner, the two conjuncts are semantically related and form a conceptual unit (Wälchli, 2005;Le Bruyn and de Swart, 2014). The adjective thus tends to be interpreted as having scope over the coordinated nouns. We do not discuss adjectival scope much further, but only include examples where the adjective makes sense with scope over the whole coordination in the following sections. Our empirical data come from a corpus study on the one hand, and an acceptability rating experiment on the other hand.

Attributive Adjective agreement in FrWac
We chose FrWac (French Web as a Corpus, Baroni et al., 2009) because it is a large (1.6 billion word) corpus of contemporary French, including informal uses, annotated for parts of speech. In this corpus, we found 32,769 tokens for D A N1 et N2 and 59,818 tokens for D N1 et N2 A. We randomly took 2,500 items for each structure and annotated automatically the gender information of nouns and adjectives with the Lexique database (New et al., 2001). We only selected plural nouns (among the two sets of 2500, it turns out that 1081 were plural for prenominal A and 1,000 were plural for post-nominal A). Restricting our data to nouns with different genders, we checked each item manually and removed the examples where the A N combination is a compound (12-a) 10 and when the A has scope on only one conjunct or has a syncretic form (12-b Table 2 reports the occurrences of masc/fem adjectives with the two plural nouns of different genders. In both prenominal and post-nominal positions, when the closest noun is masculine, masculine agreement may be triggered by resolution or by CCA. The adjectives are always masculine in these cases (13-a), (13-b).

13) a.
les In combinations like D A N1 f etN2 m and D N1 m et N2 f A, we found two possibilities for A agreement: the resolution rule (14-a), (14-b) or CCA (14-c), (14-d). However, with a closest feminine noun, resolution is very rare for prenominal A (5%), while it is at 55% for postnominal A. 10 Grandes Ecoles is a collocation which refers to selective French higher education establishments which are different from the universities. 11 A reviewer suggested an alternative analysis with two adjectives and ellipsis: les différentes ressources et différents services (the.PL different.F.PL resource.F.PL and different.m.PL service.m.PL); les ressources différentes et services différents (the.PL resource.F.PL different.F.PL and service.m.PL different.m.PL). But unlike number mismatch, ellipsis with gender mismatch between present and deleted form is an open issue (Merchant, 2014;Aparicio, Franich and Xiang, 2015). Furthermore, such an analysis would not capture the difference we find between prenominal and post-nominal adjectives. See section 3.3. 1  14) Table 2 shows that CCA exists in French and is not necessarily the same as resolution. It also shows a strong preference for CCA in prenominal position compared to post-nominal position (p < .001, Fisher's Exact Test). This may be explained by structural proximity: in cases of D A f N1 f et N2 m , the feminine closest conjunct is also the highest one (see Section 4.2 below).
Following previous literature (see Willer-Gold et al., 2017), we consider three possible strategies: closest conjunct agreement (CCA), (masculine) resolution, and first conjunct agreement (FCA). For example in Slovenian, on top of masculine resolution agreement, agreement is possible with the closest noun (neuter here) or with the first noun (feminine here): 15) Knjige in peresa so se podražila/e book.F.PL and pen.N.PL AUX.PL REFL become-more-expensive.N/F.PL 'Books and pens have become more expensive' (Slovenian, Willer-Gold et al., 2016, p.188) In a two gender system like French, FCA may predict the same agreement strategy as CCA, with a prenominal  Table 3 lists all potential cases of mismatch, with the three corresponding strategies. 12 In cases like A m N1 m et N2 f , FCA, CCA and resolution converge: masculine agreement is preferred over feminine. In the two cases A f N1 f et N2 m and N1 m et N2 f A f , CCA and resolution diverge. In the first case, CCA is preferred and also coincides with FCA since the trigger is the first conjunct. In the second case, (feminine) CCA does not coincide with FCA and it seems as acceptable as (masculine) resolution.
We can see that FCA is not an independent strategy in French, since On the other hand, CCA is an independent strategy since it is observed outside is the only factor that plays a role, as the feminine closest noun is the lowest one. We also discover that CCA has more weight than resolution since the preference for feminine in A N f et N m is greater than the preference for masculine in N m et N f A.
Overall  (1,7%). Looking now at postnominal adjectives, we found 291 examples compatible with CCA (78%) and 305 compatible with resolution Table 3. Adjective gender agreement in FrWac and the three agreement strategies In order to test the acceptability of such corpus data, we ran an acceptability rating experiment to see whether the corpus frequencies correspond to speaker's preferences. We also want to compare CCA and attraction errors (Bock and Miller, 1991), and to test the differences between human and non-human nouns, since grammatical gender has a social meaning for human nouns (Corbett, 1991). 13

Experimental data
Since formally collected judgements are more reliable than speakers' intuitions (cf. Wasow and Arnold, 2005;Gibson and Fedorenko, 2010;Sprouse, Schütze and Almeida, 2013), and more formal methods may reveal previously unobserved patterns in the data (cf. Keller, 2000;, we ran an acceptability judgement task. Most experimental studies deal with (number) subject-verb agreement (cf. Keung and Staub, 2018;Foppolo and Staub, 2020), while some deal with gender subject predicate agreement (Willer-Gold et al., 2017). As far as we know, this is the first experimental study of CCA in the nominal domain. We use a factorial design which treats adjective gender, humanness, and adjective position as three factors, each with two values (Am/ Af, human/non-human and pre/post). 14 As in our corpus study, we only use plural nouns, in order to avoid an interaction with number agreement. We also chose a gender neutral plural form for the determiner (de/des 'IND.PL'). We use 12 control items with attraction errors to compare their acceptability with that of CCA.

Materials
We built 24 experimental items, 12 with human (17) plural binomials and 12 with non-human (18). 15 The adjectives can appear in both pre-and post-nominal positions and they can have scope over the coordination in both positions. We chose human nouns with distinct masculine and feminine forms, in order to avoid interference from implicit expectations about social gender bias.
For each item, there are four conditions: masculine (Am) and feminine (Af) adjectives in prenominal (pre) position (17-a), (18-a), as well as in post-nominal (post) position (17-b), (18-b). We changed the order of binomials in (17-a) and 13 Our corpus data do not provide enough examples with human nouns for a statistical analysis since most of the examples involve inanimates. 14 In our experiment, we distinguish between humans and non-humans instead of using a more complex animacy hierarchy (cf. Corbett, 2000;Haspelmath, 2013;Zaenen et al., 2004), since because grammatical gender is related to social gender for human nouns, the relation between sex and grammatical gender tends to be looser in the case of animals. In French, for instance, a panda is always un panda (masculine) and a whale is always une baleine (feminine), regardless of their biological sex. 15 All items are in Appendix A. One non-human item was removed because of a manipulation error. This leaves us with 11 items with non-human nouns.
(17-b) so that the closest conjunct is always feminine in order to distinguish CCA from resolution. Thus, Am corresponds to resolution agreement and Af to CCA. With prenominal adjectives, the determiner is de (and not des) in order to force the adjective to have wide scope over the coordination -plural de is a variant of indefinite des only with a prenominal adjective (des/*de garçons vs des/de grands garçons) (Milner, 1978). With postnominal adjectives, the more natural reading is wide scope too since the determiner is shared between the two nouns.
We also included 12 control items, in two versions, one grammatical (19-a) and one with an agreement error (19-b), in order to test the differences between CCA and attraction errors (cf. Fayol, Largy and Lemaire, 1994;Keung and Staub, 2018

Procedures
These materials were included in an acceptability judgement paradigm on the Ibex on-line platform (Drummond, 2013). After providing informed consent, participants read each sentence on a computer screen and judged its acceptability on a range from 0 (not at all acceptable) to 10 (completely acceptable), which is the usual scale in the French school system. 16 After rating each sentence, participants had to answer a simple comprehension question, to ensure that they were attentive. Participants could only see one version for each item, the distribution of which was counterbalanced across participants. The order of experimental items was also randomized in each trial. In addition there were 20 filler items with four conditions for each item from an unrelated experiment. Experimental items, controls and fillers were distributed across three lists using a Latin square design so that each list contained 24 experimental sentences, 12 controls, 20 fillers (from an unrelated experiment) and 3 practice items.
The duration of the experiment was estimated at 10 minutes on average. 43 native speakers of French (21 to 82 years old, median = 34, 26 female, 10 male, 3 did not report their gender), recruited on the RISC website (http://www.risc. cnrs.fr/) volunteered to participate in the experiment. Three participants were removed because their accuracy for comprehension questions was less than 75% and one was removed because they rated the ungrammatical controls higher than the grammatical controls.

Results
The results are shown in Figure 1 and Appendix B. Error bars in all figures in this article correspond to 95% confidence intervals. In general, the experimental items were rated higher than the ungrammatical controls (with attraction errors) (mean = 2.73), but lower than the (very simple) grammatical controls (mean = 9.52). Feminine adjectives (mean = 6.73 in prenominal position, mean = 6.37 in post-nominal position) were also preferred over masculine adjectives (mean = 5.03 in prenominal position, mean = 5.89 in post-nominal position), and this preference was stronger in prenominal position. We did not find effects of gender or age of participants.

Effects of Directionality
We analysed the data with a mixed-effects ordinal regression model using the clmm() function in the ordinal package (Christensen, 2018). This is an appropriate statistical model for ratings that cannot be assumed to represent an interval scale, i.e., the values may not represent equally spaced points in subjects' subjective acceptability space. Fixed effects in this model were position (pre vs. post) and gender (Am/Af). We also included maximal random effects (random 16 This scale is more familiar to French speakers from France than a 1-7 Likert scale, and has been used successfully in previous linguistic experiments (Abeillé, An and Shiraïshi, 2018, Abeillé et  intercepts and random slopes) for items and subjects. The coefficient of random and fixed effects is presented in Appendix C1 (Table 4).
There were no significant main effects of gender and position. But the interaction between gender and position was significant (p = 0.006). CCA was particularly preferred in prenominal position, which is consistent with our corpus data (Section 2) and with Corbett (1991)'s typological observation that CCA is more acceptable when the target precedes the controller (cf. Introduction).

Effects of Humanness
In the A-N1f-et-N2m condition, we ran a similar mixed-effect ordinal regression model as in the previous section, with gender, humanness and their interaction as fixed effects and random slopes for subjects and gender as random slopes for items. The results (see Appendix C2 Table 5 for more details) showed a significant effect of A gender (p = 0.01). As shown in Figure 2, in prenominal position, feminine adjectives were preferred with both human nouns and nonhuman nouns. The difference between Am/Af with non-human nouns was bigger than with human nouns, but the interaction between humanness and A gender was not significant (p = 0.24).
However, in post-nominal position, masculine adjectives were preferred with human nouns, while feminine adjectives were preferred with non-human nouns. This interaction was significant (p = 0.004) in the mixed-effect ordinal regression model. We did not find significant main effects for A gender (p = 0.59). Across humanness conditions, masculine and feminine adjectives were equally acceptable. The main effect of Humanness is marginal (p = 0.07).
To sum up, (feminine) CCA was preferred with prenominal adjectives compared to post-nominal adjectives. This is consistent with our corpus data, but quite striking given the weight of the masculine resolution rule which is taught as the prescriptive norm in France. If we zoom in on the data, the interaction between gender agreement and humanness was significant only in post-nominal position. Masculine (resolution) agreement was more acceptable with human nouns, while feminine agreement (CCA) with non-human nouns (compared with human nouns).

Closest conjunct agreement is not an attraction error
Our experiment has shown quite sharp differences between closest conjunct agreement in conjoined NPs and attraction errors. CCA is in fact preferred over resolution for gender adjective-noun agreement, except for human nouns with post-nominal adjectives. On the other hand, sentences with attraction errors are judged almost three times lower than closest conjunct agreement.
Proximity plays a role in language processing in general (Bock and Miller, 1991;Deevy, 1999), but its effects in CCA are different from attraction errors (Keung and Staub, 2018), so these two phenomena should be distinguished. In examples with attraction errors, N1 is the syntactic head (see Figure 4 next section) and it contributes its gender and number features to the NP subject, whereas coordination is a specific syntactic structure with specific properties, which we will discuss in the following section.

CCA and the syntactic Structure of Coordination
From the theoretical point of view, CCA is puzzling since agreement usually obeys locality constraints (between a head and its subject, specifier or local modifier, or between a trigger and the highest c-commanded probe cf. Chomsky, 2001). Some authors have tried to reduce CCA to standard agreement with clausal coordination and ellipsis (Aoun, Benmamoun andSportiche, 1994, but see Munn, 1999). Most authors consider that it is related to the exceptional syntax of coordination, which does not have the syntactic features of a standard controller. Either the conjunction is a syntactic head (Kayne, 1994), but without gender features, or the coordination is unheaded (Ross, 1967;Borsley, 2005).
Looking more closely at the syntax, two asymmetric syntactic structures have been proposed for coordination in the literature (Figure 3). Even though they differ from each other as to whether the conjunction is the head (Structure (a)) or the structure is non headed (Structure (b)), they both consider that the first conjunct is in a structurally higher position than the second conjunct.
As shown in Figure 3a, in minimalist approaches, the conjunction is the head, the first conjunct occupies the specifier position and the second the complement position. If the conjunction (and) is the head, it may have its own number value (plural) but it does not have a gender value (Bhatt and Walkow, 2013). Assuming agreement can be triggered by the specifier, only the gender of the first conjunct is accessible for agreement. 17 This predicts first conjunct agreement but not CCA with the second conjunct.
In unification-based grammars, there is no Agree operation but feature matching through unification. Certain features, in particular person and gender, are not distributive (Dalrymple and Kaplan, 2000;Sag, 2005): they are not necessarily shared by the conjuncts and their value for the coordination as a whole is computed by special feature resolution rules and not by unification. 18 In HPSG, a hierarchical structure is assumed for coordination (Borsley, 2005, see Figure 3b), but it is non headed, and the conjunction is only the head of the conjunct it combines with. Since there is no head to assign morphosyntactic features, such as gender, number, person, several options are available for the coordination as a whole. 19 Villavicencio, Sadler and Arnold (2005) analyse gender and number agreement of attributive adjectives in Portuguese. In their analysis, agreement is always with the whole coordination phrase, which has three agreement features: CONCORD (for resolution), LEFT-AGR (with concord features from the left-most conjunct) and RIGHT-AGR (with concord features from the right-most conjunct). This analysis may well account for our French data, although not taking preferences into account.  Johannessen (1998) for similar cases of asymmetric coordination for case assignment. As shown by Borsley (2005), Structure (a) is problematic for labeling, since the coordination should have the (nominal, adjectival etc.) category of the conjuncts, and also for ternary coordination (apples, pears and oranges) since it needs multiple specifiers. 18 In LFG, Dalrymple and Kaplan (2000) assume a flat constituent-structure for coordination and a set value in functional-structure. Dalrymple and Hristov (2010) account for CCA in Serbo-Croatian with precedence rules in f -structure. Since the f -structure of the coordination is a set, the agreement rules may trigger their first or last member. They do not expect differences between backward and forward agreement. 19 Unlike structure (a), structure (b) is not defined as a binary phrase, it is well suited for ternary coordinations (étudiants, professeurs et étudiantes) with N1, N2 and N3' at the same level, as well as for coordinations with several conjunctions (et Paul et Marie 'both Paul and Marie') with N1' and N2' at the same level Mouret (2005). However, it needs specific feature percolation principles and does not obey X-bar theory.

Linear Distance vs Structural Distance
Comparing prenominal and post-nominal adjectives, if CCA is superficial and only sensitive to linear proximity (number of intervening words), the preference for CCA should be the same in the two positions. On the other hand, if we look at structural distance (number of intervening nodes), prenominal and post-nominal adjective agreement should be different. In case of CCA, the prenominal adjective agrees with the highest conjunct (highest conjunct agreement or HCA, cf. Aoun, Benmamoun and Sportiche, 1994;Munn, 1999), while the post-nominal adjective agrees with the lowest conjunct. Thus, CCA should be preferred in prenominal position since the structural distance (the number of syntactic nodes between the adjective and the controller noun) is shorter with the first conjunct (see discussion of Willer-Gold et al., 2017 for South Slavic languages), while linear distance (the number of words between the adjective and the closest noun) does not change between prenominal and post-nominal positions.
Both our experimental and corpus data show an effect of the syntactic position when the adjective agrees in gender with a coordinated NP. The preference for CCA is stronger in prenominal position than in post-nominal position for both human nouns and non-human nouns. This result can be explained by taking structural proximity into account: assuming a hierarchical structure for coordination (Figure 3), the closest conjunct is the highest one (and is structurally closest to the target counting the number of intervening nodes) with prenominal adjectives. Both linear and structure distance may also explain the difference between attributive and predicative agreement, since predicative agreement favours Resolution: the attributive adjective is closer to the (closest) noun than the predicative one which is separated from the subject by the copular verb (5) (7). In terms of structural distance, the attributive adjective belongs to the same noun phrase as the (controller) noun, while the predicative adjective belongs to the verbal phrase and is thus separated from the (controller) noun by an extra phrasal boundary.
This asymmetry is also consistent with the agreement hierarchy proposed by Corbett (1991) that is to say that CCA is preferred when the target precedes the controller (Introduction). This effect of directionality can also be explained in terms of processing difficulties (see Haskell andMacDonald, 2005 andKonieczny, 2004 for psycholinguistic evidence). Human language processing is incremental in nature (cf. Tanenhaus et al., 1995;Kaiser and Trueswell, 2004;Levy, 2008): we do not wait until we have heard an entire sentence to start disambiguating and understanding. In N-A order, the speaker can anticipate agreement with the whole coordination phrase. However, in A-N order, the speaker cannot anticipate the coordination and could agree with the first noun in a strategy that we may call 'early' agreement.
But neither linear nor structural proximity is the only factor. One must take into account the syntactic function of the items. In CCA, the second noun is the lowest one but can nevertheless trigger agreement. In attraction errors, the complement is in the same structural position but it does not trigger agreement (Figure 4). The key difference is that with coordination, neither of the two nouns is the head (Figure 3), while in attraction errors, N1 is the head and N2 is a complement. Gender is thus assigned directly by the head noun to the NP in the latter case, but not in coordination.

The role of humanness
Our acceptability rating experiment shows that the preference for CCA is sensitive not only to directionality (or structural distance), but also to humanness. However, the effect of humanness is only significant in post-nominal position when adjective agreement takes place after the whole coordination phrase has been seen. This effect can be explained by the role of interpretation for human-based gender. While it is arbitrary for non-human nouns, grammatical gender is usually non-arbitrary for human nouns, and associated with social gender (Section 1.2). Note that for human nouns, especially 'professional nouns', their interpretation involves a social meaning. Plural masculine is ambiguous between a gender-neutral reading (i.e., referring to persons of both sexes) and a gender specific reading (i.e. referring to men only) (Gygax et al., 2009). les habitants (the inhabitant.M.PL) can be used to refer to a group of men and women ; however, a noun phrase with feminine grammatical gender, such as les habitantes (the inhabitant.F.PL), exclusively picks out women.
We suppose that the preference for masculine (resolution) for humans can be explained by this interaction with social meaning. For human nouns, the interpretation of a coordination involving masculine and feminine can be a mixed gender reading, leading to a preference for a (neutral) masculine adjective.

Comparison with determiner agreement
As in other Romance languages, CCA is also attested for determiner agreement in French. Abeillé, An and Shiraïshi (2018) tested gender agreement. Most French plural determiners are not specified for gender (les, des), so we chose certains/ certaines ('certain') (Schnedecker, 2005). In Frantext (after year 1950: 31 millions words), we found as many tokens for certains Nmpl et Nfpl (20-a)    When the feminine N is the first conjunct (the closest to the determiner), we found that the feminine certaines is preferred, and that certains is rated as low as ungrammatical controls. We conclude that feminine agreement is a case of CCA and that it is the only strategy for gender agreement when the closest noun is feminine. Thus masculine agreement (when the closest noun is masculine) can also be analysed as a case of CCA. This is consistent with what we have found for prenominal adjectives.

Comparison with number agreement
As for number agreement in the nominal domain, An andAbeillé (2017, 2019) tested determiner (D) agreement with two coordinated singular nouns. In a corpus study (FrWac), we found that Ds (22a,22b) is more frequent (90% of occurrences) than Dp (22c,22d), even with non coreferent nouns (Ds in 71 % of occurrences with non coreferent nouns): We also found that syntactic position plays a role: Dp is more frequent in subject position (with a plural verb) than in object position. We also ran several acceptability studies, and found an effect of humanness: Ds is slightly more acceptable than Dp with non-human nouns (votre/vos nom et prénom 'your name and first name'), but Dp is more acceptable with human nouns (mes père et mère 'my father and mother'). We consider singular agreement as a case of CCA, and plural agreement as a case of resolution. So humanness favours resolution, as for gender agreement. This is different from other Romance languages, which seem to prefer singular agreement for determiners with two singular coordinated nouns (Le Bruyn and de Swart, 2014;Heycock and Zamparelli, 2005).
An (2020)  In an acceptability judgement task, she found that As is as acceptable as Ap, and there was no effect of humanness. This suggests that in the nominal domain, CCA is stronger for gender than for number in French, since for number both singular (CCA) and plural (resolution) are acceptable in prenominal position. 20

Conclusions
A number of conclusions can be drawn from this work. We were able to show that CCA plays an important role in attributive adjective agreement, using a large corpus of contemporary French and an acceptability experiment. Contrary to the prescriptive norm, feminine agreement may be acceptable for attributive adjectives with conflicting coordinated nouns. In this respect, French is not different from other Romance languages, such as Spanish (Demonte and 20 An (2020) also tested number agreement of predicative adjectives and showed that they favour (plural) Resolution more than attributive adjectives, which is similar to the difference observed between attributive and predicative adjectives for gender agreement, see ex. 5 and 7.
Our corpus data strongly suggest that contrary to most French grammar books, closest conjunct agreement is quite frequent for noun-adjective gender agreement in French, and even the most frequent option for prenominal adjectives. Assuming a hierarchical structure for coordination (Kayne, 1994;Borsley, 2005), we suggest that the preference for CCA with prenominal adjectives may be explained by structural distance, as well as by incremental processing.
We use an acceptability rating task which shows that (feminine) CCA should be distinguished from attraction errors (Bock and Miller, 1991;Franck, Vigliocco, and Nicol, 2002), which do occur in spontaneous production but which are not well accepted (Fayol, Largy and Lemaire, 1994). Our experimental data also show that the preference for CCA is sensitive not only to the adjective position but also to the semantic features of nouns. Human nouns may favour resolution (masculine) agreement, especially for post-nominal adjectives. This difference may be explained by the interpretability of gender with human nouns and the availability of a masculine gender neutral plural for human nouns (les habitants may include men and women, while les habitantes do not) (Gygax et al., 2009). It may also be explained by the role of the prescriptive norm, which tends to take examples with human nouns (and to teach French children about boys and girls more than about tables and books).
We conclude that CCA is an independent strategy sensitive to the target-trigger ordering and to humanness as well, which confirms the typological tendencies proposed by Corbett (1983Corbett ( , 1991Corbett ( , 2006, and the corpus study in other Romance languages (Villavicencio, Sadler and Arnold, 2005). We suggest it is the only acceptable option in prenominal position. We conclude that closest conjunct agreement may be part of the grammar of contemporary French and suggest that it should be taught as such. Further work should test predicative adjectives.
C Model analysis C.1 Effects of position C.2 Effects of humanness Table 5. Coefficients of the mixed-effects ordinal regression model testing effects of adjective's position. This regression test the effects of adjectives' position on gender agreement, with fixed effects A (Am/Af), position (pre/postnominal) and their interactions. There were also random intercepts, as well as A, position and their interactions as random slopes for subjects and items