The nature of Romance clitics has been much debated in Generative Linguistics since the seminal works by Kayne (1975) and Zwicky (1977), but experimental evidence regarding how they are processed is scarce. The status of clitics and particularly Romance clitics are an important subject of research in generative linguistics due to their intermediate/mixed behavior between independent pronouns and affixed agreement morphemes. In the case of the Spanish object-clitics we studied, they agree with their antecedent in number [Anna vió la novela_fem.sg/las novelas_fem.pl y la_fem.sg/las_fem.pl compró; “Anna saw the novel_fem.sg/s_fem.pl and (she) bought it_fem.sg/them_fem.pl”], and gender [Anna vió el libro_masc.sg/los libros_masc.pl y lo_masc.sg/los_masc.pl compró; “Anna saw the book/s and (she) bought it_masc.sg/them_masc.pl”], unlike verbal inflection that agrees in person and number. These object-clitics correspond to the object arguments of the sentences’ main verb comprar (“to buy”). Hence, like pronouns, Spanish object clitics agree in gender and not person, satisfy verbal subcategorization properties and behave as arguments of the verb. However, like agreement (inflectional) morphemes, clitics are unstressed and affixed to the verb. In generative linguistics, there are two main competing approaches accounting for the nature of clitics:

Kayne (1975) originally proposed that clitics were syntactically independent elements in what we will refer to as the Clitics as Pronouns Hypothesis: clitics are pronoun noun phrases (NPs) generated at argument position that attach to the verb in the course of the derivation. In this view, NP-clitic dependencies are a case of referential co-dependency and the clitic occupies the argument position (Torrego, 1988; Uriagereka, 1995; Sportiche, 1998; Anagnostopoulou, 2003; Marchis and Alexiadou, 2013, among others). In a variant of this hypothesis, the clitic is generated in its surface position, while the argument position is filled by the empty pronominal pro (Strozer, 1976; Rivas, 1977; Jaeggli, 1982; Borer, 1984; among others). On the other hand, according to what we will refer to as the Clitics as Agreement Hypothesis, pronominal clitics are agreement morphemes, part of Inflection and not generated in argument position (e.g., Jaeggli, 1986; Suñer, 1988; Fernández Soriano, 1989; Monachesi, 2005 among others).

Our main objective is to contribute to better understanding the nature of clitics by testing whether and to what extent the behavioral and electrophysiological pattern found during clitic processing resembles that reported previously in the literature for verb agreement, or whether it aligns better with the processing patterns of pronominal concord. To that end, we explore (i) whether, in behavioral measures, antecedent-clitic dependencies are prone to number attraction effects similar to those found in subject-verb agreement, or whether they are more resilient to these effects as antecedent-pronoun dependencies are (see further discussion about this issue in next section); and (ii) whether, in electrophysiological measures, they elicit the same electrophysiological indexes of attraction as those previously reported for subject-verb agreement.

The study of the contexts where attraction phenomena occur during language production has shed light on the main factors involved in agreement processing: in sentence preambles such as The key to the cabinet(s)…, speakers produce more number agreement errors completing preambles containing an attractor noun that does not match (i.e., cabinets) in number with the agreement controller (i.e., the head noun key), than when the attractor matches (Bock and Miller, 1991; see Bock and Middleton, 2011; Franck, 2011 for exhaustive reviews of attraction effects in various types of agreement dependencies). Research on attraction effects in language comprehension is much more scarce than in production, and it has considered almost exclusively subject-verb agreement (in English: Nicol et al., 1997; Pearlmutter et al., 1999; Pearlmutter, 2000; Wagers et al., 2009; Shen et al., 2013; in Dutch: Kaan, 2002; Chen et al., 2007; Severens et al., 2008; in Spanish: Acuña-Fariña et al., 2014; Lago et al., 2015; in French: Franck et al., 2015). However, recent studies have also explored antecedent-reflexive pronoun concord (Dillon et al., 2013, 2014, 2016; Jäger et al., 2015; Patil et al., 2016; Parker and Phillips, 2017; for a thorough literature review on attraction effects in subject-verb and antecedent-pronoun dependencies, see Jäger et al., 2017).

Early studies adopted the feature percolation hypothesis postulated to account for attraction effects in language production. According to this account, attraction effects in both production and comprehension occur because the number features of the attractor noun can erroneously percolate over the number features of the agreement controller, which results in an erroneous number representation of the agreement controller (e.g., Nicol et al., 1997; Pearlmutter et al., 1999; Pearlmutter, 2000).

More recently, it has been proposed that attraction effects are best accounted for by means of a similarity-based interference model (Badecker and Kuminiak, 2007; Wagers et al., 2009; see Dillon et al., 2013; Jäger et al., 2017; for computational simulations of the model) inspired in the ACT-R model (Lewis and Vasishth, 2005). According to this model, dependency relations are established by retrieving from memory the agreement dependents. When the agreeing element (e.g., a verb or a clitic) is encoded, it engages a cue-based retrieval mechanism to search for a matching controller in memory. But this retrieval mechanism is susceptible to similarity-based interference from other items in memory. Hence, when a distracting element that carries similar features (e.g., semantic, structural features) as the controller is present in the sentence, interference occurs because the distracting element might be misidentified as the controller. Importantly, this model predicts attraction effects to be only present or to be larger during the processing of ungrammatical than grammatical sentences. Wagers et al. (2009) suggested two options for cue-retrieval mechanisms to account for these asymmetric effects: (a) encountering the agreeing element engages retrieval mechanisms that retrieve number-matching NPs but (almost) never retrieve partially matching ones (i.e., a number mismatching attractor in grammatical sentences); or (b) the correct agreeing element form is predicted after encountering the controller NP and the cue-based reanalysis process ensues almost exclusively when ungrammaticality is detected.

However, several studies report the presence of number attraction effects in both grammatical and ungrammatical sentences. In these studies, sentence acceptability, self-paced reading for comprehension and eye-tracking measures showed that participants are slower reading or accepting grammatical sentences with a singular subject and a plural attractor (e.g., The author of the speeches was… vs. The author of the speech was…) than accepting sentences where both NPs were singular (Nicol et al., 1997; Pearlmutter et al., 1999; Pearlmutter, 2000; Acuña-Fariña et al., 2014). In contrast, for ungrammatical sentences, mismatching attractors have been shown to elicit faster reading times as compared to matching ones in self-paced reading tasks (Pearlmutter et al., 1999; Wagers et al., 2009; Franck et al., 2015; Lago et al., 2015) and eye-tracking measures (Dillon et al., 2013). That is, attraction effects interfere in the processing of the agreement controller in grammatical sentences but facilitate it in ungrammatical ones. However, Wagers et al. (2009) identified a confound variable that might have led to the interference attraction effects reported in grammatical sentences: since in all these studies attractors and agreeing verbs where adjacent, the interference effects observed at the verb might be due to carry-over effects of the slower times needed to process the morphologically marked plural rather than unmarked singular attractors.

Nevertheless, in a recent study, Franck et al. (2015) showed both facilitation and interference attraction effects in grammatical sentences were the attractor and the verb were not adjacent and they suggested that experimental design factors might affect the direction of the effect. In a self-paced reading for comprehension task in French only including grammatical sentences (Experiment 1), they reported attraction facilitation effects. In contrast, in a speeded acceptability judgment task, participants showed attraction interference effects (slower acceptability judgments) when judging both grammatical and ungrammatical sentences containing number mismatching attractors, as compared to matching ones. Franck et al. (2015) interpreted their results as evidence that different behavioral tasks tap into different processes: while self-paced reading taps structure building processes, grammaticality judgment taps into later processes of agreement computation. Either way, the fact that attraction effects were detected in both grammatical and ungrammatical sentences might support feature percolation accounts.

However, many recent studies found reliable attraction effects in ungrammatical sentences but not in grammatical ones, favoring similarity-based interference accounts (Wagers et al., 2009; Dillon et al., 2013; Lago et al., 2015). This grammatical vs. ungrammatical asymmetry of attraction effects was interpreted as the main evidence that attraction effects are mainly due to similarity-based interference effects during the retrieval of the cues necessary to build dependency relations (e.g., Lewis and Vasishth, 2005), and not due to a faulty representation of the agreement controller, as suggested by the feature percolation account. As reviewed in the next section, electrophysiological evidence of attraction effects replicated the grammatical asymmetry of attraction effects (e.g., Kaan, 2002; Shen et al., 2013; Tanner et al., 2014, 2016).

Morphological markedness plays a crucial role during agreement attraction in comprehension: attraction effects are either only found in singular, but not plural agreement (Nicol et al., 1997; Wagers et al., 2009, in acceptability and self-paced reading data), or are larger in singular than plural agreement (Acuña-Fariña et al., 2014, in eye-tracking measures). These findings replicate the number markedness effects also reported in production studies (Bock and Miller, 1991; Bock and Cutting, 1992; Bock and Eberhard, 1993; Eberhard, 1997), suggesting that morphologically marked plural distractors are stronger attractors than non-marked singular ones in both modalities. Thus, attraction effects might sometimes be obscured and delayed due to carry-over effects of plural attractors when the attractor and the agreeing element are adjacent. However, those carry-over effects do not last long: they can be avoided by including a word between the attractor and the verb (e.g., Wagers et al., 2009) and even when the attractor and the verb are adjacent, attraction effects are detected at the region following the verb (Pearlmutter et al., 1999).

All research reviewed above studied subject-verb agreement dependencies. But do attraction effects also affect the processing, and more particularly the comprehension of antecedent-pronoun dependencies? In production, pronoun-antecedent agreement seems to be as sensitive to attraction effects as subject-verb agreement is, but the former is more sensitive to notional number factors (e.g., Bock et al., 1999, 2004), suggesting that pronominal dependencies may rely more on the retrieval of the semantic/lexical representation of the antecedent. In comprehension, early studies exploring the role of grammatical constraints in antecedent-reflexive pronoun gender agreement showed that they are resilient to interference from other possible antecedent candidates (Nicol and Swinney, 1989; Sturt, 2003; inter alia). More recently, these findings have been replicated in studies that compared the magnitude of attraction effects in antecedent-reflexive pronoun vs. subject-verb agreement dependencies. In a reading for comprehension eye-tracking experiment, Dillon et al. (2013) showed reliable attraction effects for subject-verb agreement (shorter total reading times and fewer regressions to the critical agreement region were obtained in sentences containing mismatching attractors as compared to sentences containing matching ones, but only in ungrammatical sentences, replicating the grammatical asymmetry of attraction). No signs of attraction effects were found for reflexive pronouns (e.g., The new executive who oversaw the middle manager/s apparently doubted himself/^∗themselves…). Dillon et al. (2013) interpreted the resilience of reflexive pronouns to attraction effects as evidence that subject-verb vs. antecedent-reflexive pronoun dependencies involve qualitatively different processes (see also Phillips et al., 2011). According to the authors, these different linguistic dependencies are sensitive to different linguistic features: (a) verbal agreement is a formal morphosyntactic mechanism to index the arguments of the verb, and feature retrieval is mainly driven by ranked morphological and structural cues (i.e., number feature and subjecthood cues, respectively); and (b) pronominal concord is a dependency between two NPs and therefore antecedent retrieval is driven by syntactic (structural) cues.

Interestingly, recent eye-tracking studies show that although (English) reflexives are more resilient to attraction, they are indeed susceptible to it. For instance, Patil et al. (2016) showed that when the role of structural-cues such as subjecthood is controlled (e.g., both the antecedent of the reflexive and the attractor were subjects), attraction effects occurred when the attractor mismatched in morphological cues such as gender. Parker and Phillips (2017) also showed that no attraction effects occurred when the attractor mismatched in a single feature (i.e., gender) with the antecedent, but they did when the attractor mismatched in two features (e.g., gender and animacy, number and animacy or number and gender). These authors suggested that both subject-verb and antecedent-reflexive pronoun agreement engage similar cue-based retrieval mechanisms. However, following Dillon et al. (2013), Parker and Phillips (2017) suggested that reflexive pronoun dependencies weight structural cues more strongly than morphological cues, which precludes the erroneous retrieval of non-licensed antecedent candidates (see also Dillon et al., 2014, 2016).

In sum, behavioral measures show that subject-verb agreement comprehension is prone to attraction effects (Nicol et al., 1997; Pearlmutter et al., 1999; Pearlmutter, 2000; Wagers et al., 2009; Acuña-Fariña et al., 2014; Franck et al., 2015; Lago et al., 2015), but antecedent-pronoun dependencies are more resilient to these effects (Dillon et al., 2013; Patil et al., 2016; Parker and Phillips, 2017; see also Jäger et al., 2017 for a thorough review and discussion). Although attraction effects have been also reported in grammatical sentences, they are stronger and more consistent in ungrammatical sentences (e.g., Pearlmutter et al., 1999; Wagers et al., 2009; Lago et al., 2015), supporting similarity-based accounts. Next, we review the main findings of ERP studies on agreement and number attraction effects.

In general, when processing syntactic violations in subject-verb, object-verb, or antecedent-pronoun dependencies, three types of electrophysiological correlates have been reported in the ERP literature: Left Anterior Negativity (LAN), N400 and a centro-parietal positivity (P600) (for a detailed description and interpretation of each component see i.e., Bornkessel and Schlesewsky, 2006).

Most studies observed biphasic patterns with negative components (LAN/N400) followed by a positive component (P600). Some studies reported a biphasic LAN – P600 pattern for subject-verb agreement violations (Kutas and Hillyard, 1983; Osterhout and Mobley, 1995; De Vincenzi et al., 2003; Silva-Pereyra and Carreiras, 2007, among others) as well as for determiner-noun or noun-adjective gender agreement violations (Gunter et al., 2000; Deutsch and Bentin, 2001; Barber and Carreiras, 2005; Martin-Loeches et al., 2006; Molinaro et al., 2008). Other studies reported a biphasic N400-P600 pattern for subject-verb and object-verb agreement violations (Coulson et al., 1998; Zawiszewski and Friederici, 2009; Díaz et al., 2011; Zawiszewski et al., 2011) as well as for antecedent-pronoun violations (Schmitt et al., 2002; Hammer et al., 2005, 2008; Lamers et al., 2006). Finally, some studies have also reported an isolated P600 component for subject-verb agreement violations (Osterhout et al., 1996; Nevins et al., 2007; Frenck-Mestre et al., 2008), for determiner-noun or noun-adjective gender agreement relations (Osterhout and Mobley, 1995; Osterhout et al., 1997; Foucart and Frenck-Mestre, 2011, 2012) and for antecedent-pronoun violations (Lamers et al., 2006, 2008; Silva-Pereyra et al., 2012; Xu et al., 2013; Rossi et al., 2014). As far as we know, no study has shown an isolated early negativity (N400 or LAN).

Regarding the electrophysiological responses underlying number attraction effects, the available evidence is rather scarce and focused on subject-verb number agreement. To our knowledge, no study explored attraction effects in antecedent-clitic dependencies.

Electrophysiological indexes of attraction effects in subject-verb agreement are heterogeneous, but two main results have been observed: (a) electrophysiological indexes of agreement violation detection are less salient and harder to detect in sentences containing number mismatching attractors than matching attractors (Kaan, 2002; Chen et al., 2007; Severens et al., 2008; Shen et al., 2013; Tanner et al., 2014, 2016); and (b) the four studies that checked for asymmetrical attraction effects found an asymmetry: number mismatching attractors elicit a reduction of ERP components as compared to number matching ones in ungrammatical sentences, but not in grammatical ones (Kaan, 2002; Shen et al., 2013; Tanner et al., 2014, 2016).

Focusing on the studies that reported asymmetrical attraction effects, and thus support the cue-based retrieval account of agreement computation (Lewis and Vasishth, 2005; Wagers et al., 2009), Kaan (2002) investigated the effects of distance and number interference in subject agreement processing: Dutch participants performed an acceptability rating task in sentences containing subject and object NPs that either matched or mismatched in number. ERP responses following the critical verb revealed main grammaticality effects reflected by a bilateral negativity over central and posterior sites between 300 and 500 ms, and a P600 effect between 500–700 and 700–900 ms. A main number attraction effect was revealed by a significantly larger P600 component between 500 and 700 ms following subject agreement violations in sentences with only singular NPs (i.e., the control singular number matching condition) than in any other condition. Number mismatching attractors elicited a smaller P600 in singular subject agreement, but not in plural, replicating the number markedness effects (Eberhard, 1997; Nicol et al., 1997; Wagers et al., 2009). Finally, the modulation of the P600 related to attraction effects was asymmetrical, as it only occurred in ungrammatical sentences.

Tanner et al. (2014) provide behavioral and ERP evidence supporting the asymmetric pattern of attraction effects: English speaking participants showed a main P600 component elicited by subject-verb agreement violations, with attraction effects revealing a smaller P600 in sentences containing number mismatching attractors than number matching ones. These attraction effects were asymmetrical: Participants showed a reliable P600 effect and were less accurate judging ungrammatical sentences that contained number mismatching rather than number matching attractor NPs. In contrast, they showed no P600 effect and were similarly accurate while judging the acceptability of grammatical sentences containing number matching and mismatching attractor NPs. These results obtained both with and without an adverb intervening between the attractor noun and the auxiliary verb (The chemist with the test tube(s) (probably) is/^∗are…), suggesting that ERP indexes of attraction are resilient to carry-over effects of the plural attractor. In a recent study, Tanner et al. (2016) replicated this pattern of attraction effects revealing that number mismatching attractors reduce the magnitude of the P600 as compared to number matching ones in ungrammatical sentences.

Shen et al. (2013) used a comprehension task where participants listened to several narrations in English with a low proportion of violations. In sentences with no attractor NPs, singular subject agreement violations elicited a bilateral frontal negativity between 150 and 300 ms (interpreted as a LAN) followed by a P600 between 700 and 950 ms. In sentences with complex NPs (e.g., A catalog with color picture/s sit/^∗sits…), those containing number matching attractors elicited an atypical early posterior negativity between 150 and 300 ms, and no P600, while those containing number mismatching attractors elicited neither early posterior negativity nor P600 effects. Although the authors suggest that the posterior negativity resembles the timing and distribution of the N400, the distribution of the negativity related to morphosyntactic violations is rather frontal (and lateralized: LAN) and starts later on (300 ms after the stimulus onset) (see Molinaro et al., 2011, 2015; Tanner, 2015 for an extensive review and discussion). These different ERP patterns might be due to the naturalistic procedure used in this study (i.e., sentences were auditorily presented and embedded in discourse), as compared to the procedure used in most other studies. Regardless of the origin of the atypical early components in this study, the relevant fact for our discussion is that agreement attraction effects reached significance only in ungrammatical sentences (differences between sentences with number matching vs. mismatching attractor NPs), replicating the asymmetric pattern of attraction effects. Shen et al. (2013) interpreted these results as evidence that subject agreement is affected by the presence of number-bearing elements other than the subject itself, with number mismatching elements completely “masking” subject agreement violations.

There are two more studies that explored number attraction effects in subject-verb agreement, but they did not analyze whether these were asymmetric. In an acceptability rating task, Severens et al. (2008) explored in Dutch whether the ambiguity of the determiner of the controller NP affects number attraction. In number match conditions, an atypical ERP pattern related to morphological agreement violations was found, as subject agreement violations only elicited an N400, not followed by a P600. This was interpreted to reflect a blatant violation of the expected verb form during a first, syntactically shallow process that cannot be repaired by further analysis, resulting in the absence of a P600. In violations involving number mismatching conditions, only a P600 was elicited, which was interpreted as reflecting a deeper syntactic processing triggered by the strong conflict between a shallow syntactic analysis that suggests the first noun (singular) to be the controller and a combinatorial analysis that suggest the noun (plural) agreeing in number with the verb (i.e., the attractor) to be the controller. In other words, the agreement attraction effects were argued to prevent the generation of a N400 component correlated to the ungrammatical verb. Similar findings were reported by Chen et al. (2007) for English singular subject agreement, although this study reported a LAN instead of a N400. Here, a biphasic LAN-P600 pattern was observed in matching conditions, while only a P600 (but no LAN) was reported in mismatching conditions (i.e., The price of the cars ^∗were…”).

In summary, the electrophysiological indexes of attraction are mainly reflected by a reduction of main ERP components related to agreement violation detection. The most consistent finding is the reduction of the later P600 component, found in three out of six studies (Kaan, 2002; Tanner et al., 2014, 2016). The other studies showed a reduction of diverse early components: a posterior early negativity (Shen et al., 2013), an N400 (Severens et al., 2008), or a LAN (Chen et al., 2007), but two showed atypical ERP components in the baseline number matching conditions, which might pose problems to the generalizability of attraction effect to other types of dependencies. Further research needs to bring some light on the origin of such heterogeneous patterns of attraction effects in subject-verb agreement. However, it is worth noting that all the studies that explored it found an asymmetrical pattern of attraction effects (Kaan, 2002; Shen et al., 2013; Tanner et al., 2014, 2016), which supports the similarity-based interference account of attraction (Lewis and Vasishth, 2005; Wagers et al., 2009).

The grammaticality judgment accuracy data replicated two of the most common findings in agreement: (a) An attraction effect: participants produced more grammaticality judgment errors when sentences contained an attractor that mismatched the number of the antecedent NP as compared to sentences containing a number matching attractor (Nicol et al., 1997; Franck et al., 2015); (b) A markedness effect: attraction effects obtained with singular but not with plural antecedent NPs (Bock and Miller, 1991; Eberhard, 1997; Pearlmutter et al., 1999; Pearlmutter, 2000; Wagers et al., 2009). Plural attractors disrupted participants’ grammaticality judgment accuracy both when accepting grammatical sentences and when rejecting ungrammatical ones. This replicates the finding of attraction effects in the judgment of grammatical sentences by Nicol et al. (1997: Experiment 2), in contrasts with the results of Franck et al. (2015: Experiment 3).

Reading time results are less conclusive. This is because, despite the inclusion of spill-over effects in the model, the main number attraction effects seem to reflect carry-over effects of the larger difficulty of processing the number of the plural attractor presented just before the clitic (Pearlmutter et al., 1999; Wagers et al., 2009): in sentences containing singular antecedent NPs, the presence of plural attractors, as compared to singular ones, slowed down participants’ reading times. This effect persisted at the clitic region and the following ones, both in grammatical and ungrammatical sentences. In contrast, in sentences containing plural antecedent NPs, reading times at the attractor noun region were slower for matching plural than mismatching singular attractors. These slow down effects occurred only in grammatical sentences, and persisted only until the following clitic region. The fact that these effects appear in sentences with singular and plural antecedent NPs at the regions where the attractor is presented suggests that part, if not all, of the attractor number effects are due to the greater reading and processing cost of morphologically marked plural attractors. Consequently, we argue that these effects are not bona fide agreement attraction effects.

Importantly, similar grammaticality effects obtained while reading sentences containing singular or plural antecedent object NPs. In both cases, the presence of clitics mismatching in number with their antecedent NP (ungrammatical) led to slower reading times than those obtained for matching antecedent-clitic pairs (grammatical). These differences arose at the clitic region in sentences with singular dependencies, and at the following region in sentences with plural dependencies.⁴ But in both cases, at the two-final regions ungrammatical sentences were read faster than grammatical ones. This is probably because, once participants detected the ungrammaticality of the sentence at previous regions (CW or CW+1), they simply speeded up reading the sentence to complete the grammaticality judgment task.

Grammaticality judgment accuracy data indicate that attraction effects occur both in grammatical and ungrammatical sentences in antecedent-clitic dependencies, but no significant effects were found in reading times. Reading time measures revealed no attraction effects, and we argue this is because they were obscured by the carry-over effects of the processing of the preceding plural attractor NPs. Hence, accuracy data suggests that clitic dependencies are affected by the same factors as subject-verb agreement, which would favor the Clitics as Agreement Hypothesis (e.g., Suñer, 1988; Franco, 2000) according to which clitics are agreement morphemes. In contrast, reading time data suggest that the processing of clitic dependencies is not affected by the same factors as subject-verb agreement and that might be processed differently (as suggested by Phillips et al., 2011; Dillon et al., 2013), which could be interpreted as evidence favoring the Clitics as Pronouns Hypothesis, according to which clitics are pronouns generated at the argument position that moved to the verb (e.g., Kayne, 1975; Torrego, 1988; Uriagereka, 1995; Sportiche, 1998; Anagnostopoulou, 2003; Marchis and Alexiadou, 2013, among others). In order to shed more light on this issue, in Experiment 2 we use ERP methods with which, due to their finer temporal resolution than self-paced reading methods, we might be able to detect the presence of (any) attraction effects that overcome the carry-over effects of the processing of plural attractor nouns.

In Experiment 2, violations in sentences with singular attractors (e.g., …el paquete para el vecino ^∗los…) elicited a frontal negativity followed by a P600 component. These components have been previously reported for antecedent-clitic dependency violations, but not simultaneously: Silva-Pereyra et al. (2012) report an N400 for feminine gender violation and a P600 for masculine gender violation, while Rossi et al. (2014) report a P600 for both gender and number violations. Our biphasic ERP pattern replicates the one usually reported for agreement violations (see Molinaro et al., 2011) and other types of pronominal dependency violations such as reflexives or subject pronouns (Schmitt et al., 2002; Hammer et al., 2005, 2008; Lamers et al., 2006).

Regarding number attraction effects, violations involving singular antecedents and plural clitics with intervening plural attractors elicited a P600 component with no trace of a preceding negativity (e.g., el paquete para los vecinos ^∗los…). This pattern of results, together with those from grammaticality judgment accuracy data in Experiments 1 and 2, reveals greater difficulty detecting clitic number violations when a mismatching plural attractor intervenes. We interpret the absence of a negative component as signaling an attraction effect due to the mismatching attractor (replicating Chen et al., 2007; Severens et al., 2008; Shen et al., 2013). We did not find a reduction of the amplitude of the P600 component that could be interpreted as evidence for attraction effects, as in some studies on agreement (Kaan, 2002; Shen et al., 2013; Tanner et al., 2014).

Importantly, our results also revealed electrophysiological indexes of asymmetrical attraction effects: attraction effects only occurred in ungrammatical sentences, not in grammatical ones. In ungrammatical sentences, plural clitics with singular antecedents elicited a large and broadly distributed negativity when preceded by plural attractors, as compared to those preceded by number matching singular attractors. No equivalent differences were found for grammatical sentences. These results converge with grammaticality judgment accuracy and response time data in Experiment 2, where weaker number attraction effects obtained for grammatical sentences as compared to ungrammatical ones. Importantly, these asymmetrical effects suggest that they are in fact due to attraction and not to carry-over effects originated while reading the preceding plural attractors, as might have occurred in the self-paced reading task. If the effects shown in ungrammatical sentences were due to carry-over effects, they should also have been detected in grammatical ones.

In sum, behavioral and ERP results from Experiment 2 showed that antecedent-clitic dependencies are also subject to attraction effects and that these effects are detected in ungrammatical sentences only. Our ERP results identified frontal negative components as the main electrophysiological indexes of attraction effects.

The fact that no clear attraction effects obtained in reading times for antecedent-clitic dependencies (either at clitic position or following word regions) suggests that this type of dependencies are resilient to attraction effects, as previously revealed by Dillon et al. (2013) and Parker and Phillips (2017). However, in Experiment 2 these effects were detectable by methods with finer temporal resolution such as ERPs. Off-line grammaticality judgment measures showed attraction effects in both experiments, but asymmetrical attraction effects were only obtained in Experiment 2. Certain experimental design variables might have contributed to these differences. For instance, in Experiment 1 grammatical and ungrammatical sentences containing plural antecedents were included, so that a grammaticality judgment task could be performed with sentences containing plural clitics. In contrast, in Experiment 2 all sentences with plural clitics were ungrammatical. Although some researchers counterbalanced this by adding as fillers grammatical sentences containing plural controllers (e.g., Franck et al., 2015: in all three Experiments), many self-paced reading and ERP studies do not (Pearlmutter et al., 1999: Experiments 1 and 2; Shen et al., 2013; Tanner et al., 2014: in all three Experiments), or do not report it (Chen et al., 2007; Lago et al., 2015: in all four Experiments; Wagers et al., 2009: Experiments 2, 4, 5, and 6). The fact that in Experiment 2 a plural clitic was a perfectly reliable signal of an ungrammatical sentence might have lessened the capacity of attractors to elicit attraction. Although the confound between grammaticality and the processing of plural clitics might have had some effect, we believe this cannot be the determinant factor behind the results of Experiment 2. This is because we should expect similar electrophysiological response patterns related to grammaticality (or clitic number) effects in sentences containing number matching and mismatching attractors. However, we report different ERP patterns in grammatical vs. ungrammatical sentences: frontal negativity-P600 vs. only P600, respectively. If participants used a plural-clitic equal ungrammatical task-specific strategy, attraction effects might have diminished overall, augmenting the possibility to detect asymmetrical attraction effects. The impact these design differences might have had on the off-line results obtained in both experiments seem to be reflected in our on-line measures too, where ERP data provided finer grained timing effects than the self-paced reading data. Next, we discuss the main implications of these findings for current models of language processing.

The negative components (e.g., LAN/N400) reported above have been generally interpreted as indexing a greater difficulty to integrate the predicted critical word into the previous context (Friederici et al., 1993; Münte et al., 1993; Friederici, 2002; Rossi et al., 2005). In the case of antecedent-clitic dependencies (e.g., …el paquete para el/los vecino(s) ^∗los…), attraction effects modulate these negative components, revealing a greater difficulty to predict/integrate a plural clitic (^∗los) that disagrees in number with the singular antecedent (el paquete) when it was preceded by a singular attractor (el vecino), as compared to when it was preceded by a plural attractor (los vecinos).

These results can be accounted for under the cue-based retrieval and similarity-based interference accounts of dependency processing (Lewis and Vasishth, 2005; Wagers et al., 2009). In these models, when the number feature of the dependent element (in our case, the clitic) is encountered, a retrieval mechanism searches for a matching element stored in working memory (in this case, the antecedent NP). Accordingly, the grammaticality effect reported here ensues: when a plural clitic is encountered in a context where all possible antecedents are singular, the predictability and/or integration of the clitic is most difficult, eliciting a frontal negativity. But when a plural clitic is encountered after a plural attractor, the similarity-based interference of the plural attractor with the singular antecedent to be retrieved from memory leads to erroneously interpreting the plural attractor as the antecedent of the clitic, so that no frontal negativity is elicited.

Conversely, the asymmetric effect emerges because, in grammatical sentences, the number mismatching attractor is not retrieved from memory, either because the retrieval mechanism is not deployed or because it only retrieves antecedent candidates that fully match the features of the clitic. In other words, when the number of the antecedent and the dependent clitic match (i.e., in grammatical sentences), it is assumed that the number of the attractor noun is not retrieved, so that no ERP differences ensue for matching and mismatching attractors. In contrast, when the number of the antecedent and the clitic do not match (i.e., in ungrammatical sentences), a reanalysis process ensues. Thus, a larger frontal negativity is expected in sentences containing number matching attractors as compared to mismatching ones, because illusions of grammaticality only occur in sentences where plural attractor NPs match the number of the plural clitic and can be mistakenly retrieved as their antecedents (Wagers et al., 2009; Phillips et al., 2011). Our results fully support these predictions.

The absence of a frontal negativity in sentences containing mismatching plural attractors is an index of the presence of number attraction effects during antecedent-clitic dependency resolution. Additionally, the asymmetric effect revealed by the absence of ERP components indexing attraction in grammatical sentences (in contrast to the negative component elicited in ungrammatical ones), provides compelling evidence in support of cue-based retrieval models as accounts of attraction effects in comprehension (Wagers et al., 2009; Shen et al., 2013). Unfortunately, our data cannot adjudicate between the possibility that encountering a singular clitic out-competes retrieval of plural antecedent candidates so that the attractor is not retrieved, and the possibility that the dependency is correctly processed without the deployment of retrieval mechanisms. Importantly, the present results cannot be accommodated into feature percolation models because they assume attraction effects are driven by an erroneous number representation in the antecedent NP. Hence, they predict equivalent effects in grammatical and ungrammatical sentences (Nicol et al., 1997; Pearlmutter et al., 1999), contrary to our findings (see also, Wagers et al., 2009; Dillon et al., 2013; Lago et al., 2015). Electrophysiological evidence in Experiment 2 revealed a clear asymmetrical effect of attraction. In sum, the evidence here provides strong support for cue-based retrieval models of dependency resolution in language processing, and are incompatible with alternative feature percolation accounts.

Finally, in addition to the absence of frontal negative components as an electrophysiological index of attraction effects, we also reported that number violations elicited a P600 component both when sentences contained a matching singular attractor and when they contained a mismatching plural attractor. This in turn reveals that, despite the presence of attraction effects, participants could detect the ungrammaticality of sentences containing number violations. These findings contrast with those reported in Shen et al. (2013), where attraction effects led to the absence of associated ERP components, and in Kaan (2002) and Tanner et al. (2014), where attraction effects caused a reduction in the amplitude of the ungrammaticality/reanalysis related P600. In the former case, the differences in results likely originate from task differences: the grammaticality judgment task we used required participants to explicitly and consciously check and reanalyze sentences for well-formedness, which would encourage the appearance of the P600 even in sentences where number attraction effects occurred, while the comprehension task used by Shen et al. (2013) did not require participants to pay attention to grammaticality. However, the differences between our results and those of Kaan (2002) and Tanner et al. (2014), where attraction effects reduced the amplitude of the P600 component, are harder to explain based on task differences, given that all studies used a grammaticality judgment task. The main differences with regard to our study involves a smaller ratio of ungrammatical sentences (28.6% in our study vs. 50% in theirs); the type of phenomenon explored, where we studied attraction effects in antecedent-clitic dependencies and they did so in subject-verb agreement. These findings might tentatively be interpreted as evidence that antecedent-clitic dependencies tap into processes different from those involved in agreement.

Whether Romance clitics are pronouns or agreement morphemes is under debate. Although our experimental approach does not provide direct evidence supporting either type of syntactic analysis, we believe it offers indirect evidence that can be informative. Some studies compared the magnitude of attraction effects elicited in different types of dependencies (Dillon et al., 2013; Parker and Phillips, 2017) and observed that antecedent-reflexive pronoun dependencies were more resilient to attraction than subject-verb agreement. We suggest that these differences correlate with the distinct nature of these two types of dependencies. Although both types of dependencies rely on similar cue-based retrieval mechanisms, antecedent-pronoun dependencies involve a referential dependency between two nominal arguments and weight structural cues more strongly than morphological ones, precluding the erroneous retrieval of non-licensed antecedent candidates. In contrast, subject-verb agreement is a morphological mechanism used to index the arguments of sentences where morphological cues weigh more than structural ones, making the erroneous retrieval of non-licensed attractors possible.

Although a direct comparison of attraction effects between subject-verb and antecedent-clitic dependencies goes beyond the scope of this study, we argue that our results align better with results previously obtained for antecedent-reflexive pronoun than for subject-verb agreement dependencies. Subject-verb agreement resolution shows consistent attraction effects in self-paced reading studies and these effects have been shown to mainly modulate late positive P600 components in ERP experiments. In contrast, antecedent-clitic dependency resolution is resilient to attraction effects in self-paced reading (Experiment 1; replicating reflexive pronoun studies: Dillon et al., 2013; Parker and Phillips, 2017) and affected early frontal negative ERP components (Experiment 2). Hence, we tentatively interpret the observed resilience of antecedent-clitic dependencies to attraction effects and the fact that they modulate different electrophysiological components than in subject-verb agreement to indicate that antecedent-clitic and verb agreement dependencies constitute different types of linguistic dependencies. Thus, we interpret our indirect evidence to favor the Clitics as Pronouns Hypothesis originally proposed by Kayne (1975), which suggests that Spanish object-clitics are processed as pronominal elements.

Regarding the ERP patterns indexing attraction effects in antecedent-clitic dependencies, we observed that attraction effects lead to the absence of a frontal negativity, while in previous studies subject-verb agreement attraction effects modulated the later positive P600 component (Kaan, 2002; Tanner et al., 2014). Although we reckon that linking the modulation of different ERP components as evidence for different types of processes might be seen as speculative, we tentatively argue the difference in the type of associated components signals the different processes involved in the resolution of pronominal and agreement dependencies: attraction effects revealed by frontal negativities might be related to pronominal processing and active retrieval of the lexical representation of possible antecedents and signals difficulty of syntactic/semantic integration of the full arguments (i.e., difficulty of establishing antecedent-clitic pronominal dependencies) (Barkley et al., 2015). Thus, in ungrammatical sentences, no negative component appeared because the plural attractor NP might have been incorrectly identified as the plural clitic antecedent NP. In contrast, attraction effects revealed at later positivities might be mainly related to purely morphological processes like agreement and signal difficulty to integrate and reanalyze morphological features (e.g., in verb agreement dependencies) (Hagoort et al., 1993).

Certainly, further research making direct comparisons between antecedent-clitic and subject-verb agreement dependencies in relatively similar syntactic contexts (i.e., distance between the two agreeing elements, structural position of the elements, whether they are in their canonical position or not, etc.) will help to better identify the processes underlying both structures. Further research ought to provide a fuller and more systematic picture of the main electrophysiological indexes involved in the resolution of the different types on linguistic dependencies across a wider array of languages.