Overall, previous experimental native Spanish PAS findings show no clear division of labor as in native Italian: null pronouns select subject antecedents, but overt pronouns are “flexible” (non-subject and subject antecedents). Each experimental study is unique in terms of, e.g., the type of method/stimuli/design, which could explain the different results across studies. Consequently, we present a thorough review of each study to detect possible limitations that will be later implemented in our corpus study. Note that we review both offline and online PAS studies in adult Spanish monolinguals and adult L2 learners, thereby excluding other populations (see Tables 1, 2 in the online Supplementary material for additional details).² Finally, no single corpus study has targeted PAS structures, so we review some corpus evidence on AR in general as their findings may shed light on PAS.

The bulk of experimental studies on AR have investigated the PAS with two potential antecedents (subject/non-subject) and two anaphoric forms (overt/null pronominal subject) in either inter- or intra-sentential configurations. So, what we know about the PAS comes mostly from a series of similarly-designed experiments that do not question whether (i) the PAS may represent an oversimplified way of resolving anaphora in native (and L2) Spanish; (ii) PAS scenarios may be more complex than traditionally assumed (i.e., they can contain more than two antecedents in other syntactic positions); (iii) the antecedents may be realized by other forms other than null/overt pronouns (i.e., NPs for example). Unlike experiments, corpus data can shed light on these questions since they contain natural language production (where AR configurations are neither controlled nor constrained) and offer contextually rich scenarios with anaphors and antecedents embedded in their entire discourse. Unlike experiments, corpus data can shed light on these questions since they (i) contain natural language production where AR configurations are neither controlled nor constrained; (ii) offer contextually rich scenarios with anaphors and antecedents embedded in their entire discourse. This led to RQ1a and RQ1b.

Experimental PAS studies have typically restricted their focus to two anaphoric forms (overt/null pronominal subjects). Corpus studies have reported the use of other anaphoric forms (e.g., repeated Ns and NPs) in several AR scenarios, so NPs may be also possible Refererential Expression (RE) forms in PAS.⁷

Experimental studies report Spanish null pronouns to bias toward a preverbal subject antecedent, whereas overt pronouns show a more “flexible” behavior. This contrasts with native Italian where overt/null pronouns show a clear division of labor. Additionally, experimental studies have not typically included NPs as a possible RE form.

The implicit assumption in the experimental literature is that purely configurational factors (null➔subject vs. overt➔non-subject) overlap with discursive information-status factors (null➔topic continuity vs. overt➔topic shift). RQ4/H4 (when contrasted to RQ3/H3) will determine the extent to which the overlap assumption is correct. This motivates theoretical questions having to do with likely deficits at the syntax-discourse interface.

Despite English being a non-null subject language, corpus data (Quesada and Lozano, 2020) have shown that English monolinguals allow null pronouns in very specific contexts: topic continuity and coordination at around 77% (e.g., Lucy_i walked for an hour and Ø_i had a picnic), but never in non-coordinate configurations. So, it could be argued that L2 Spanish learners’ production of null pronouns in topic continuity could be due to L1 transfer rather than actual acquisition, which leads to the following exploratory research question.

Previous PAS experimental studies are often contradictory depending on the sentential configuration: inter- vs. intra-sentential; main-subordinate vs. subordinate-main orders (cf. the tables in the online Supplementary material). RQ6 is an exploratory question to explore whether the sentential PAS configuration modulates the choice of RE in naturalistic corpus production.

Corpus Escrito del Español L2 (CEDEL2) (Lozano, 2022) is a multi-L1 corpus of L2 Spanish learners coming from 11 different L1 backgrounds, plus a Spanish monolingual control subcorpus. CEDEL2 (version 2) currently holds 1,105,936 words, 4,399 participants, and 14 task topics. It is freely available/downloadable at http://cedel2.learnercorpora.com.

Data are collected via online forms⁸ and participants complete three forms: (i) linguistic background; (ii) standardized placement test (just for learners) (University of Wisconsin, 1998); and (iii) written/spoken text.

We selected an L1 English-L2 Spanish (plus a comparable Spanish monolingual control) sample (Table 1) based on the following criteria: (i) the participant’s age range was 18 ~ 40, since Working Memory, which may affect AR, appears to decay after the age of 40 (Bel et al., 2016b); (ii) learners’ proficiency-level range was intermediate~advanced; and (iii) only two composition titles were targeted (cf. 2.3 below).⁹ Two hundred two texts met these criteria but we finally selected those that had at least one instance of a PAS (N = 75). We originally departed from two intermediate groups: lower intermediates (placement score: 21 ~ 28 raw score, 49% ~ 65%) and upper intermediates (29 ~ 35, 67–81%). Since they did not significantly differ in our analyses, we decided to analyze both groups as a single group of intermediates to simplify the between-group statistical analyses and interpretations. Learners had an equivalent age of exposure (AoE) to L2 Spanish and their length of instruction (LoI) in Spanish and length of stay abroad (LoSA) in a Spanish-speaking country increased with proficiency.

We selected two task tittles (Talk about a famous person and Summarize a film you have recently seen), since they are narratives that contain (i) abundant [+human] 3^rd person antecedent-anaphor chains; and (ii) PAS constructions, which were more frequent in the second task than in the first task and which offered different characters in discourse suitable for the topic continuity/shift purpose of this study.

The corpus sample was manually annotated (i.e., tagged) with UAM Corpus Tool (O’Donnell, 2009), version 6.2j (February 2023).¹⁰ We firstly tagged each text to indicate the group category (intermediate, lower advanced, upper advanced, and monolingual), which allows between-group comparisons for the same linguistic feature, as will be explained below. We designed another tagset to count the frequency of two AR scenarios (PAS vs. other AR). Each RE in subject position was assigned either the PAS tag (when the RE was preceded by a subject/non-subject antecedents) or other (when the RE was preceded by an AR scenario other than PAS). Figure 2 shows the fine-grained, linguistically-informed tagset to annotate PAS.¹¹ It allows for multiple and intricate statistical analyses among tags, as will become obvious later. It is inspired by previous corpus studies on AR (Lozano, 2016; Quesada and Lozano, 2020), although we introduced new features.

Every 3rd person human subject that followed the syntactic configuration of the PAS was manually tagged. First, the PAS-type system included: (i) standard PAS with two antecedents, as in (13), and (ii) complex PAS with more than two antecedents, as in (14a-c). For example, the tag used to annotate complex PAS in (14a) is s1_nons2_nons3, which indicates we have 3 potential antecedents: the first one is in subject position (s1) and the other two in non-subject position realized via a complex NP: PP (nons3) within an NP (nons2). In (14b), the tag s1_nons2andnons3 indicates that there is a singular antecedent in subject position (s1, which happens to be a null pronouns) followed by two NP coordinated antecedents in non-subject position (s2&s3) embedded within a PP. Notice that, due to the complexity of the antecedents’ region, the anaphor is a complex NP for disambiguation purposes. Other complex PAS contained plural REs, as in (14c), but we excluded them from our current analysis since it has been shown that the truly problematic cases of AR are 3rd person singular and not plural (Lozano, 2009).

Naaven_i se ha enamorado de Tiana_j y Ø _i quiere pedirle matrimonio. [Monolingual: ES_WR_24_3_IZG.txt].¹²

a.La chica_i se enamora del amante_j de su madrek hasta que al final Ø _i acaba teniendo … [Monolingual: ES_WR_30_3_JVM].

The anaphor-form system includes the RE form (null/overt pronouns and NPs) in subject position, as shown in bold in (15). The anaphor-number system includes the RE number (singular/plural), which served us to exclude plural REs in the analyses, as justified above.

The information-status system comprises topic-continuity and topic-shift contexts, as in (16a-b) respectively. The antecedent-function system included subject antecedent, non-subject antecedent, and subject/non-subject antecedent (for cases of complex PAS). This system allowed us to detect PAS scenarios with subject-antecedent biases, as in (16a), or non-subject antecedent biases, as in (16b).

b. Bella_i se da cuenta de que Jacob_j está enamorado de ella_i y ella_i también un poco de él_j [Monolingual: ES_WR_21_3_ICH].

In the syntactic-configuration, we tagged the type of intra-sentential and inter-sentential configurations, e.g., topic-continuity and coordination in (16b) and topic continuity and non-coordination, which can be of different types, e.g., subordination in (17) or new sentence in (18).

Finally, the anaphora-resolution system indicates the type of resolution: via morphosyntax or semantics. In this paper, we analyzed only the PAS that was morphosyntactically resolved. In order to avoid skewing our results, we excluded PAS that was semantically resolved (i.e., null pronouns in topic-shift scenarios like (19), which are ultimately resolved via directive verbs).

UAM Corpus Tool has an in-built statistical analysis software. Between-group (or between-system/tag) comparisons are based on the tags’ raw frequencies and statistical contrasts are chi square (χ²) tests, accompanied by their significance level (p) and their effect size (Cohen’s h).

Based on the linguistically-motivated tagging scheme (Figure 2), UAM Corpus Tool allows for multiple and sophisticated statistical contrasts between the different groups and the (sub) nodes and terminal nodes of the tagset. These contrasts were motivated by the linguistically-informed hypotheses from section 1.2. Following statistical recommendations for corpus data (Egbert et al., 2020), we purposely decided to use the χ² statistical contrasts provided by the software rather than submitting the data to more sophisticated statistical analyses (which involve transforming the data and abstracting away from the linguistic facts and interpretations):

In Figure 3, PAS scenarios (gray bars) were compared against other types of AR scenarios (black bars). Spanish monolinguals resolve anaphora via scenarios (68.2%, i.e., 296 REs out of a total of 434 tagged REs) other than standard PAS (21.2%) or complex PAS (10.6%). Thus, standard PAS only amounts to around 1/5th of the total possible AR scenarios. Learners show a similar pattern to monolinguals across all proficiency levels, though only the upper advanced group shows native-like behavior (standard and complex PAS χ² = 2.16, p = 0.1419 n.s., h = 0.204; other scenarios χ² = 0.15, p = 0.6964 n.s, h = 0.030). The lower-level learner groups significantly differ from Spanish monolinguals in other scenarios but not in standard and complex PAS scenarios (intermediates vs. monolinguals: standard and complex PAS χ² = 0.49, p = 0.4848 n.s., h = 0.087, other scenarios χ² = 8.80 p = 0.0030, h = 0.193; lower-advanced vs. monolinguals: standard and complex PAS χ² = 0.66, p = 0.4180 n.s., h = 0.106, other scenarios χ² = 12.44, p = 0.0004, h = 0.235).

Our findings support H1a (PAS represents one of the many possible mechanisms of AR in native and non-native Spanish) and H1b (PAS can contain more complex configurations than those traditionally reported in the literature). Corpus data reveal that the traditional assumption of standard PAS as a prototypical strategy to resolve anaphora has been overestimated in the experimental literature.

RQ2 explores the different RE forms in PAS scenarios, independently from the factors that constrain their choice. Spanish monolinguals produced mostly null pronominal subjects (66.1%), followed by NPs (23.2%) and overt pronominal subjects (10.7%) (Figure 4). Learners show a tendency toward the native norm as proficiency increases, yet only upper-advanced leaners (57.9% null, 26.3 overt, 15.8% NP) show a rather similar and non-significant pattern to the Spanish monolinguals (null pronouns: χ² = 1.30, p = 0.2551, h = 0.169; NPs: χ² = 1.55, p = 0.2135, h = 0.188), though a significant difference for overt pronouns (χ² = 7.80, p = 0.0052, h = 0.410). The lower-advanced group shows similar proportions for all three RE forms (34.5% null, 35.7% overt, 29.8% NP), which significantly differ from monolinguals for null (χ² = 19.16, p < 0.001, h = 0.642) and overt (χ² = 17.82, p < 0.001, h = 0.614), but are non-significant for NPs (χ² = 1.07, p = 0.3012, h = 0.149). Intermediates produce mainly overt REs (overt pronouns 39.1%; NPs 38.1%) and some null pronouns (22.8%), with the three RE production rates being significantly different from monolinguals (overt: χ² = 22.67, p < 0.001, h = 0.685; NP: χ² = 5.30, p = 0.0213, h = 0.324; null: χ² = 37.96, p < 0.001, h = 0.902).

These findings support H2. Whereas null pronouns are the tendency in Spanish monolinguals and in upper-advanced learners, the rest of learners differ from monolinguals and show more variability in RE forms. Null pronominal subjects are gradually acquired with proficiency level, whereas overt pronouns show the opposite pattern. Crucially, NPs are a frequent RE form to resolve anaphora in PAS scenarios for both learners and monolinguals. We turn next to the division of labor of such RE forms.

First, we focus on Spanish monolinguals’ production to clarify the division of labor in PAS scenarios and to settle the question of whether the alleged flexibility of overt pronouns is more apparent than real. Figure 5 shows a clear bias of null pronouns (93.6%) toward subject antecedents (13), which confirms the PAS and supports most previous research in Spanish. Overt pronouns (32.4%) show a timid bias toward non-subject antecedents, (16b), as previously reported in the literature but, crucially, if we include NPs as a possible RE form, NPs show a strong bias (64.7%) toward non-subject antecedents, (20). Thus, NPs play an important role in PAS scenarios and this could explain the apparently “flexible” bias found for overt pronouns previously reported.

Importantly, if we consider overt and NPs forms together (overtly realized REs), then a neater division of labor shows up (Figure 6): null pronouns are biased toward subject antecedents (93.6%) yet overtly realized REs are biased toward non-subject antecedents (97.1%). Thus, corpus data reveals that the division of labor of AR in native Spanish is more complex and more clear-cut than previously assumed since NPs play a key role. These findings explain the division of labor in native Spanish and therefore settle the dispute on the apparent flexibility of overt pronouns in PAS scenarios.

Let us now compare learners against monolinguals regarding the production of RE forms for subject vs. non-subject antecedents. As for subject-antecedent biases (Figure 7), Spanish monolinguals show a clear-cut bias as they produce almost exclusively null pronominal subjects (93.6%). Intermediates show equal variability across all three RE forms (null 35.2%, overt 35.2%, NP 29.6%), as illustrated in (21a, b, c) respectively, and their production is significantly different from monolinguals (null: χ² = 44.05, p < 0.001, h = 1.358; overt: χ² = 26.09, p < 0.001, h = 1.270; NP: χ² = 10.87, p = 0.0010, h = 0.638). From lower advanced to upper advanced we can see an increasing trend toward the native norm, particularly for null pronouns (lower advanced: null 47.8%, NP 26.1%, overt 26.1%; upper advanced: null 75.5%, overt 17.8, NP 6.7%), though, crucially, each advanced group significantly differs from the monolingual group: lower advanced vs. monolinguals (null: χ² = 28.75, p < 0.001, h = 1.101; overt: χ² = 18.20, p < 0.001, h = 1.072; NP: χ² = 8.07, p = 0.0045, h = 0.558); upper advanced vs. monolinguals (null: χ² = 7.00, p = 0.0081, h = 0.521; overt: χ² = 11.91, p = 0.0006, h = 0.870; except for NPs, where there are no significant differences χ² = 0.00, p = 0.9646, h = 0.009). In short, intermediates know that a null pronoun can select a subject antecedent, but they equally produce overt pronouns (as in their L1) and NPs. Nearly half of the productions of lower-advanced learners are null pronouns. Upper-advanced learners show native-like discriminations, but their productions are not fully native-like yet.

Consider now non-subject antecedent biases (Figure 8). Spanish monolinguals’ production clearly indicates that NPs (64.7%) (and not overt pronouns) are the privileged RE form to refer to a non-subject antecedent. Crucially, null pronouns are hardly an option for any group, so learners know from the outset that a null pronoun is not an adequate form to refer to a non-subject antecedent. It is therefore remarkable that no null pronouns are used in purely structural PAS configurations. As for learners, overt pronouns and NPs are highly produced, but learners are rather indeterminate about them, particularly intermediates, who show optionality in their production (47.2% overt vs. 52.8% NP), and the two advanced groups, who also show a rather indeterminate pattern where overt pronouns are slightly higher than NPs, as in (22 a, b): lower advanced (56.3% vs. 40.6%), upper advanced (54.6% vs. 40.9%).

b. …ella_i escribe algunas cartas a Michael_j. Pero Michael_j no responde. [Learner: EN_WR_38_9_30_3_JG].

Figure 8 visually shows that the learners’ pattern is either optional (intermediates) or somewhat opposite to the monolinguals’ (advanced groups). The low frequencies in production in all groups may explain why no significant differences are observed between each of the learner groups and the monolinguals (p > 0.05 in all cases, though p < 0.50 for each of the advanced groups vs. the monolinguals, which represent marginally non-significant differences).

These findings, taken together, support H3 since null pronouns show a strong bias toward subject antecedents (with learners showing an increasing sensitivity to this), whereas overt material (i.e., overt pronouns as well as NPs) shows a clear bias toward non-subject antecedents.

Recall that, at this point, we need to discriminate between purely structural PAS results (RQ 3, previous section) from purely information status/discursive PAS results (RQ4, this section). This will allow us to determine whether the traditional assumption of a correspondence/overlap between syntactic position (subject/non-subject) and information status (topic continuity/shift), as stated in section 1.2, is reflected in production data. Recall that RQ4 will additionally allow us to check for possible deficits at the syntax-discourse interface, as predicted by the IH.

Figure 9 shows the use of REs in topic-continuity contexts, where the production of null pronominal subjects is higher for all groups, although the percentages between groups vary considerably. There is a clear increase of nulls from the intermediate to the monolingual group: intermediate (38.8%), lower-adv (47.8%), upper-adv (76.2%), monolingual (95%). If we compare these results with Figure 7, we can observe a similar trend in the results and a similar statistical behavior. In particular, intermediates show again similar variability across all three RE forms (null 38.8%, overt, 30.6%, NP 30.6%), as shown in (23a-c) respectively and their production is significantly different from monolinguals (null: χ² = 40.39, p < 0.001, h = 1.346; overt: χ² = 21.30, p < 0.001, h = 1.173; NP: χ² = 12.83, p = 0.0003, h = 0.772). From lower advanced to upper advanced we can see again an increase toward the native norm, particularly for null pronouns (lower advanced: null 47.8%, NP 26.1%, overt 26.1%; upper advanced: null 76.2%, overt 19%, NP 4.8%), though, once again, each advanced group significantly differs from the monolingual group: lower advanced vs. monolinguals (null: χ² = 30.52, p < 0.001, h = 1.163; overt: χ² = 17.65, p < 0.001, h = 1.072; NP: χ² = 9.53, p = 0.0020, h = 0.621); upper advanced vs. monolinguals (null: χ² = 7.86, p = 0.0050, h = 0.568; overt: χ² = 12.40, p = 0.0004, h = 0.903; except for NPs again, where there are no significant differences (χ² = 0.00, p = 0.9563, h = 0.011).

By contrast, Figure 10 shows the use of REs in topic-shift contexts. Again, these results show a similar trend to those in Figure 8: monolinguals produce mainly NPs (63.9%), followed by overt pronouns (30.6%). Lower-adv and upper-adv learners show a trend that is rather inverse (though less marked) to monolinguals’, by producing overt (58.1, 50%) followed by NPs (41.9, 41.7%), as in (24a, b). Intermediates produce more NPs (54.1%), closely followed by overt (45.9%). Once again, the rather low frequencies in production in all groups may be behind the non-significant differences between each of the learner groups and the monolinguals: non-significant differences (p > 0.05) in most contrasts; marginally non-significant differences (0.05 < p < 0.10) for NPs in the lower-advanced vs. monolinguals contrast (χ² = 3.23) and the upper-advanced vs. monolinguals contrast (χ² = 2.87); and only one significant difference for overt pronouns in the lower-advanced vs. monolinguals contrast (χ² = 5.13, p = 0.0234, h = 0.561).

The results in Figures 9, 10 thus show that syntactic position (subject/non-subject) overlaps with information status (topic continuity/shift) in such a way that null pronouns typically mark a continuation of topic of the subject antecedent, whereas overt material (NPs and overt pronouns) typically marks a shift in topic. These results empirically demonstrate that the traditional experimental assumption in section 1.2 is correct in corpus production data.

When it comes the syntax-discourse interface, recall that Sorace’s IH predicts deficits with AR even at advanced levels. This is confirmed in this study, but only partially since our results show that not all syntax-discourse PAS scenarios are equally problematic. In topic-continuity contexts, despite learners’ steady increase of null pronouns, the upper-advanced group (76.2%) still significantly differs from monolinguals (95%), but no significant differences were found in topic-shift scenarios with either NPs or overt pronouns. This differential effect is in line with Lozano’s (2016) Pragmatic Principles Violation Hypothesis (PPHV), originally proposed for general AR in L1 English-L2 Spanish but also confirmed in other scenarios: AR in L1 Greek-L2 Spanish (Lozano, 2018; Margaza and Gavarró, 2022); AR in L1 English-L2 Spanish and L1 Spanish-L2 English (Quesada, 2021); clitic pronouns in L1 English-L2 Spanish (García-Tejada, 2022); and pragmatic implicatures in L1 Chinese-L2 English (Feng, 2022). The PPVH postulates that learners typically obey the pragmatic Principle of Clarity (i.e., they attain native-like knowledge in topic-shift contexts by using full RE forms to avoid ambiguity) but often violate the Principle of Economy (i.e., they produce overt pronouns in topic continuity, which leads to redundancy).

To summarize, the results showed that the choice of REs depends both on (i) the syntactic position of its antecedent (null➔subject vs. NP/overt➔non-subject), and (ii) the information status of its antecedent (null➔topic continuity vs. NP/overt➔topic shift). H4 is confirmed as there is a correspondence between syntactic position and information structure in PAS. Finally, the PPVH is confirmed since the most advanced L2ers cannot attain full native-like competence in topic-continuity contexts, but they can in topic-shift contexts.

Recall that a null-subject language like Spanish allows null pronominal subjects in all syntactic configurations (coordination and non-coordination), whereas a non-null subject language like English allows them only in topic continuity and coordination. If L1 transfer plays a role in PAS, L1 English-L2 Spanish learners are expected to produce null pronouns mostly in contexts where English allows them.

In topic continuity and coordinate syntactic configurations (cf. (16b)), all groups produce mostly null pronominal subjects. Learners show a slight increasing trend toward the native norm, though only the upper-advanced group shows native-like knowledge (Figure 11): intermediates vs. monolinguals (χ² = 10.54, p = 0.0012, h = 1.159); lower-advanced vs. monolinguals (χ² = 7.78, p = 0.0053, h = 0.994); upper-advanced vs. monolinguals (χ² = 2.93, p = 0.0870 n.s, h = 0.613). By contrast, in topic continuity and non-coordinate configurations learners’ production of null subjects (cf. (17) and (18)) is much lower and is always significantly different from monolinguals’: intermediates vs. monolinguals (χ² = 34.27, p < 0.001, h = 1.687); lower-advanced vs. monolinguals (χ² = 23.97, p < 0.001, h = 1.463); upper-advanced vs. monolinguals (χ² = 5.98, p = 0.145, h = 0.715). Additional within-group comparisons¹³ show that Spanish monolinguals’ production of null pronouns in topic-continuity coordinate vs. non-coordinate configurations is not significantly different, as expected (χ² = 3.38, p > 0.05, n.s), whereas learners’ production is significantly different: intermediates (χ² = 16.29, p < 0.02); lower-advanced (χ² = 13.29, p < 0.02); upper-advanced (χ² = 5.52, p < 0.02). Results suggest that learners’ significantly higher use of null pronouns in coordinate than in non-coordinate configurations reflects L1 English influence. Interestingly, learners show a strong gradual trend toward the native norm (intermediate 15.1%, lower-adv 24%, upper-adv 60%), which suggests their sensitivity to the allowability of null pronouns in non-coordinate scenarios increases with proficiency, though their production rates (even at upper-advanced levels) are far from Spanish monolinguals’. This confirms learners’ transfer of null pronouns but an increasing sensitivity to their pragmatics.

Table 2 shows that that the production of intersentential configurations is two thirds (or higher) the production of intrasentential configurations for both L2ers and monolinguals, which indicates that the most natural sentential configuration for AR in PAS scenarios is intersentential, either independent sentences as in [sentence].[sentence] or coordinate sentences as in [sentence]&[sentence]. This clear-cut trend has been rather overlooked in the design of stimuli in previous experimental studies, where intrasentential configurations like [main [subordinate]] have been typically the focus of attention. Importantly, only the upper-advanced learners can attain native-like competence as they are not significantly different from Spanish monolinguals (χ² = 0.60, p = 0.4371, h = 0.116), whereas the intermediates (χ² = 5.57, p = 0.0182, h = 0.338), and the lower-advanced learners (χ² = 11.31, p = 0.0008, h = 0.506) significantly differ from monolinguals.

Recall that an additional question concerns the order of main and subordinate clauses. Table 2 shows the clausal order for the low-frequency intrasentential configurations: Main-subordinate is overwhelmingly more frequent than subordinate-main for both L2ers and monolinguals. Note that no inferential statistics are performed here due to the low frequencies.

In short, in natural production PAS scenarios are overwhelmingly intersentential and, when they happen to be intra-sentential, the most frequent clausal order is main-subordinate. This is so in native and non-native grammars. These findings provide clear tips for those researchers wishing to design experimental PAS configurations that intend to look as natural as possible.

A final consideration is whether the sentential configuration is a factor that modulates the choice of RE in PAS in native Spanish (Figure 12). Null pronominal subjects are clearly biased toward subject antecedents regardless of the type of sentence (100% intrasentential, 90.5% intersentential), whereas an interesting subdivision of labor is observed when the bias is toward non-subject antecedents: overt pronouns in intrasentential (85.7%) but NPs in intersentential (77.8%), as shown in (25a, b). In other words, topic continuity (subject bias) is marked via null pronouns irrespective of the sentential configuration, but topic shift (non-subject bias) is marked via overt pronouns intrasententially yet via NPs intersententially, which is a finding not reported in the previous literature. Sentential configuration is therefore an additional factor that modulates the division of labor of REs in PAS configurations in native Spanish. This issue merits further investigation in future studies containing larger frequencies of learner and native corpus data.

Marco_i está celoso y Ø_i no se adapta bien a esta nueva vida de Verónica_j cuando ella_j empieza a tomar a sus amigos como amantes. [Learner: EN_WR_42_21_8_3_LBK].

Ella_i le_j tiene mucho cariño, pero Ø_i se niega a desmentir sus votos para estar con él_j. Nacho_j se deja guiar por un idealismo optimista… [Learner: EN_WR_42_21_10_3_LBK].