Twenty eight native Basque-Spanish bilinguals (12 male) aged 18-40 (SD = 5.92) participated in this experiment. Their dominant language was Basque and they acquired it from birth¹.

The experimental stimuli consisted of 60 experimental items. As sentences with postnominal relative clauses are less frequent in Basque, we performed a naturalness judgment test on our stimuli. Twelve native Basque speakers who didn't take part in the experiment rated the sentences on a scale from 1 (=very unnatural, unusual) to 7 (=very natural, usual). The rating revealed a medium naturalness, with the average value of 3.85 (SD = 1.64) for the experimental items, and 4.16 (SD = 1.68) for the control items. We fitted linear mixed effects models, with Matrix Number (Singular, Plural) and Disambiguation (ergative singular, absolutive plural) conditions as fixed effects and random by-subject and by-item intercepts, separately for the experimental and the control items. Effect coding was used for all the independent variables. The analysis on the experimental items showed a main effect of Matrix Number, such that the sentences with singular matrix number were rated as more natural [Estimate: 0.36, SE = 0.10, t = 3.70]. The analysis on the control items showed no significant effects (all ts < 1.14).

The critical, ambiguous noun was located within an embedded relative clause, which was interpolated between a matrix subject and a matrix verb. In Basque, the sequence -ak in Basque denotes either singular ergative case subjects of transitive constructions (in which case it combines a singular determiner morpheme -a and an ergative marker -k), plural absolutive case subjects of intransitive constructions, or plural absolutive case objects of transitive constructions (in which case it combines a plural determiner morpheme -ak and a zero morpheme). The number of the matrix subject/verb was unambiguously singular (-a) or unambiguously plural (-ek). Relative clauses were disambiguated only at the auxiliary, either toward ergative (50%) singular or absolutive plural (50%), but the disambiguating region was not analyzed in the current paper² (Table 1). While we investigate post nominal relative clauses, we note that the prenominal relative clauses are more frequent in Basque (Laka, 1996), and that previous unpublished research on post nominal relatives shows less cost (and thus preference) for subject relatives (Yetano et al., 2010). However, this advantage shows only at the disambiguating region, so we don't expect it to affect the critical noun.

Sixty additional control items were created where the critical ambiguous noun did not interrupt a matrix subject-verb dependency but was preceded by a clause containing a singular or plural-marked adjunct (Table 1). In these sentences, the manipulated non-subject noun is an inanimate, making it even less subject-like. Although, ideally, these control sentences should be created from the same lexical items as the experimental ones, it was not possible in Basque to restructure them so that the clause containing the ambiguous noun doesn't interrupt a subject-verb dependency. Therefore, the control sentences were a separate set of lexical items and were analyzed separately.

The experimental and control items were Latin Squared into 4 lists. 80 fillers were added to each list, whose structure, length, and lexical content were varied. Both in experimental and control items, the critical ambiguous noun was animate, and nouns that are not usually used in either singular or plural form were avoided (e.g., wives). The sentences were recorded in standard Basque by a male native speaker.

Four images appeared on the screen with each item: two versions (singular and plural) of the critical, and two versions (singular and plural) of the distractor noun, as shown in Figure 1. For the fillers, one of the inanimate nouns from the sentence was depicted. For the 20 nouns shared by control and experimental items, the same images were used. The number of occurrences of each image across the experiment was balanced, as well as the position of singular/plural and critical/distractor images. The pictures were obtained from a free online source (www.freepik.com) and modified. They were 15 x 15 cm big and limited by a border.

Unlike in the classical visual world paradigm used in sentence processing (e.g., Tanenhaus et al., 1995), we presented the images 500 ms before the critical word onset, aiming to eliminate looking behavior related to the nouns prior to the critical one. Note that this time window length has previously been found to provide an adequate preview of the picture prior to the auditory stimulus onset (e.g., Apfelbaum et al., 2021; Kapnoula et al., 2015). The participants were tested individually and instructed to listen to the sentences carefully through the headphones. They were told to look at the fixation cross, and then wherever they wanted upon the appearance of the images.

The session lasted approximately 45 mins and consisted of two experimental blocks (100 items each), preceded by 5 practice trials. The yes/no comprehension questions appeared after 50% items, equally distributed over experimental, control, and filler items, and were focused on the clause containing the ambiguous noun (e.g., Does the boy like the girl? for experimental, and Did the girl invite her friends for a dinner? for control items in Table 1). On the remaining items, the participants were asked to listen to the ambiguous noun (isolated from the sentence) once again and click on the image (from the 4-image screen presented with the sentence) that corresponds to it. This task was added in order to make sure that participants paid attention to the visual stimuli and disambiguated well.

Right eye movements were recorded using the SR EyeLink 1,000 eye-tracker with the 500 Hz sampling rate and a chinrest. We performed 9-point calibration before the experiment, between the two blocks, and additionally if fixation stability was lost.

Short fixations ( ≤ 80 ms) were merged with the preceding/following fixation when within 0.5 degrees of visual angle (e.g., Duñabeitia et al., 2009); otherwise, they were removed with an automatic procedure by Data Viewer (SR Research Ltd, 2019). We analyzed only the time window during the auditory presentation of the critical noun (see Ito and Knoeferle, 2023 for a procedure on linear mixed effects models over time windows in VWP data). The length of the critical noun time window ranged from 416 to 980 ms for the experimental, and from 203 to 1,066 ms for the control items. The samples were binned (20 ms) for the analysis. We calculated log-ratio of proportions of looks to the plural vs. singular image, using the formula: log (proportion of looks to plural image/proportion of looks to singular image), with positive values indicating more looks to the plural image (Ito and Knoeferle, 2023; Arai et al., 2007). Note that unlike typical VWP experiments and their corresponding analyses, we don't have a natural target image in the current experiment, so the choice of plural image was rather arbitrary. We used VWPre (Porretta et al., 2018) and lme4 packages (Bates et al., 2015) in R Statistical software (R Core Team, 2017, version 3.4.1) for data preprocessing and analysis, respectively.

The dependent variable was the log-ratio of proportions of looks to plural vs. singular, while the independent variable was the Matrix Number (Singular, Plural). We tested models for the exclusion of random by-item and by-subject slopes (by-item and by-subject random intercepts were included), using the “best-path” algorithm (Barr et al., 2013), starting with the maximal model. When encountering non-convergence problems, a different optimizer was specified. Effect coding was used for all the independent variables. We report the intercept, the estimate, standard error, and the t value in Table 2, together with the best fitting model. Experimental and control items were analyzed separately.

The accuracy was 81.03% on the picture-selection task, and 86.11% on the comprehension questions. Comparable means for the experimental (77.02%) and control items (79.52%) on the comprehension questions suggest that the participant understood the experimental and control sentences equally well.

For the experimental items, we found a significant Matrix Number effect, such that the preference toward a plural image (expressed by the log-ratio) was significantly higher in the plural matrix subject condition, or conversely, significantly lower in the singular matrix subject condition (Table 2, Figure 2). For the control items, we found no significant effect of Matrix Number (Table 2, Figure 2).

In this paper, we investigated whether the number feature of a matrix subject-verb dependency is actively maintained during its establishment. Specifically, we tested whether singular/plural matrix number, which we assumed to be actively maintained and thus available, could modulate the interpretation of a number-ambiguous noun interpolated between the matrix subject and–a mechanism that has not been proposed before. We hypothesized that the noun will be interpreted as singular more often in the singular matrix verb condition, and as plural more often in the plural matrix verb condition, due to maintenance of the singular and plural feature, respectively.

The analysis of the log ratios of proportions of looks to plural vs. singular images showed that the plural preference increased significantly with plural matrix subjects (or decreased significantly with singular matrix subjects). This implies that the ambiguous noun was interpreted as plural more often in the plural matrix subject condition. Importantly, this increase or decrease was absent from the control items, making it possible to attribute this effect to feature maintenance during subject-verb dependency establishment. Certainly, a design where control and experimental items are more similar and analyzed together would remove potential confounds. For example, an additional experiment where the manipulated singular/plural noun in both control and experimental items is inanimate would clear up whether number maintenance is aided by animacy.

Finally, it is important to underline that the matrix subject in the experimental sentences is also a filler for the embedded relative clause. Previous research has evidenced active feature maintenance in filler-gap dependencies, although not with embedded relative clauses (Ness and Meltzer-Asscher, 2019; Chow and Zhou, 2019). Therefore, future research should determine whether: 1. the maintained feature is part of a subject representation, or the verb representation, or both (as noted previously); the maintenance mechanism is triggered because the subject noun is a subject, or because it is a filler, or both. It should be noted, however, that (Ristic et al. 2022) found maintenance costs even when a non-relative clause interrupted the matrix subject-verb dependency, indicating that filler status cannot completely explain maintenance effects.

Active maintenance of the features necessary for the establishment of a subject-verb dependency can manifest either as an increased processing cost, due to encoding interference, or it can serve as an additional disambiguating information, as in the current experiment. We propose that this maintenance is a means of keeping the number feature predicted or preactivated prior to encountering the verb, with the final goal of ensuring a more efficient agreement validation at the verb. In line with this, studies focused on other subject-verb dependency-related phenomena have proposed that number is predicted in its computation Lago et al., (2015).

In sum, our findings suggest that number feature is maintained during long-distance subject-verb dependency establishment in real time, similarly to feature maintenance in filler-gap dependencies. While more data is needed to determine whether these results are specific to the Basque morphosyntactic system, we suggest that this finding reflects a predictive mechanism which, in case of long-distance subject-verb dependency, manifests as active maintenance of relevant features.