Glossa Psycholinguistics

1. Introduction

Previous research has shown that a processing effect called agreement attraction does not uniformly impact all types of grammatical dependencies. For example, subject-verb agreement typically shows number attraction effects across languages, which consist of the facilitated processing of an ungrammatical verb in the presence of a preceding plural modifier – called an attractor – as in the sentence The key to the cabinets ARE on the table (Dillon et al., 2013; Hammerly et al., 2019; Jäger et al., 2020; but see Chromý et al., 2023a, 2023b). By contrast, anaphoric dependencies have a more variable behavior. Comprehension studies have mostly focused on reflexive anaphors like himself or herself, which sometimes show number attraction during reading (Jäger et al., 2020; Parker & Phillips, 2017; Patil et al., 2016), but other times do not (Dillon et al., 2013; Parker & Phillips, 2017; Sturt, 2003). A similarly variable attraction profile has been documented with object pronouns like him or her during reading comprehension (Badecker & Straub, 2002; Chow et al., 2014). The source of these variable attraction profiles is unclear, but they have sometimes been attributed to a differential use of syntactic constraints (Chow et al., 2014; Han et al., 2021) or to interpretation-related factors in production (Eberhard et al., 2005; Kandel & Phillips, 2022).

The variability observed with reflexives and pronouns motivates research into other types of anaphoric dependencies, as well as into less-studied languages. This brief report pursues this goal by investigating the processing of clitic pronouns in Bulgarian. To date, it is unknown whether Bulgarian object clitics show agreement attraction. Previous studies have only considered Greek and Spanish and employed different experimental methods. Particularly, Paspali and Marinis (2020) showed that Greek clitics are prone to gender attraction in several measures (self-paced listening times, speeded and offline judgments), in contrast to Santesteban et al. (2017), who found evidence of attraction with Spanish clitics in offline judgments and event-related brain responses, but not in self-paced reading times.

The contrast above could point to a cross-linguistic difference between Greek and Spanish clitics, an asymmetry in the processing of gender vs. number, and/or methodological factors. For example, the lack of attraction effects during reading in Santesteban et al. (2017) might have resulted from the properties of their materials (non-canonical configurations, such as clitic left-dislocated structures) and/or design. As acknowledged by the authors, the design of the experimental sentences did not allow controlling for spillover effects of the plural attractor, which can affect attraction during reading (Wagers et al., 2009). Finally, participants were required to provide grammaticality judgments after reading the sentences. This differs from most previous reading studies using self-paced and eye-tracking, which typically include comprehension questions to avoid focusing participants’ attention on grammaticality (Dillon et al., 2013; Lago et al., 2021; Paspali & Marinis, 2020; Wagers et al., 2009).

Our study addressed these potential issues and additionally tested a less-studied language: Bulgarian. Bulgarian clitics are syntactically, semantically, and prosodically deficient pronominal elements that are specified for gender (masculine, feminine, or neuter) in the third person singular, in addition to number and case (Franks, 2017, 2021). In contrast to other Slavic languages, case inflection in Bulgarian and Macedonian has been lost on nouns, but it is retained on clitics and tonic (strong) pronouns, with object clitics in the Accusative case and oblique clitics in the Dative case. In Bulgarian, clitic and tonic pronouns are in complementary distribution: clitics are the default option in antecedent-pronoun dependencies, while pronouns are used in marked contexts, such as clefts, contrastive configurations and as objects of prepositions. Table 1 presents the properties of Bulgarian object clitics (note that clitics are not specified for animacy).

Bulgarian clitics are licensed after any prosodically non-deficient elements, e.g., conjunctions (1), the future marker šte (‘will’), and the negative particle ne. Further, clitics are verb-adjacent and their canonical position is pre-verbal, unless post-syntactic (e.g., phonological) constraints reposition them post-verbally (Franks, 2017; Harizanov, 2014a; Pancheva, 2005). This distribution differs from the placement of clitics in Greek and Spanish, where the proclisis-enclisis alternation is governed by the finiteness of the verbal host of the clitic (Mavrogiorgos, 2013). The properties of Bulgarian clitics also show differences with those of the clitics in West Slavic and the rest of the South Slavic languages (there are no pronominal clitics in East Slavic).

Given the differences in the characteristics of Bulgarian object clitics compared to clitics in Slavic and non-Slavic languages, as well as the issues surrounding the testing of clitics in previous work, our research question was whether Bulgarian clitics would show number attraction effects in comprehension. To our knowledge, our study is the first to test attraction with pronominal dependencies in Slavic and as such, it provides an important addition to existing attraction studies on subject-verb agreement in Russian (Slioussar et al., 2022) and Czech (Chromý et al., 2023a, 2023b).

2. Eye-tracking reading experiment

We conducted an eye-tracking-while-reading experiment to examine whether number attraction effects occurred with clitic object pronouns in Bulgarian. Based on previous findings with Spanish clitics (Santesteban et al., 2017), we might expect a lack of attraction effects with Bulgarian clitics. However, as pointed out above, our materials and method differed in several aspects from Santesteban et al. (2017), including the use of clitics in canonical positions, as well as the deployment of a reading-for-comprehension paradigm without a grammaticality judgment task. Due to these modifications, we expected to find number attraction with Bulgarian clitics.

3. Analysis

The data pre-processing steps were identical to those reported in Lago et al. (2021). The statistical analysis also closely followed Lago et al. (2021): Due to the brief report format, we only summarize the main aspects and any points of departure. We only analyzed reading measures that have consistently shown attraction in previous studies: first-pass regressions, regression-path, and total reading time (Acuña-Fariña et al., 2014; Dillon et al., 2013; Jäger et al., 2020; Parker & Phillips, 2017; Pearlmutter et al., 1999). The probability of first-pass regressions is an early processing measure denoting the likelihood of initiating a regression when first encountering a region. Total time is a global measure that denotes the amount of time spent in a region, including re-reading. Finally, regression-path time describes the amount of time spent since first entering a region from the left until leaving it to the right, including regressions to preceding regions.

Reading times were log-transformed. The analysis focused on the critical region – the clitic plus verb – and the spillover region – the following adverbial phrase(s) – because attraction effects in reading typically emerge after an agreement target has been encountered. We analyzed the critical and spillover regions jointly, using Region as a fixed effect (Cunnings & Sturt, 2018; Lago et al., 2021). Reading data were analyzed with frequentist mixed-effects logistic regression (first-pass regressions) and linear regression (regression-path and total times) in R (R Development Core Team, 2023). To control for the impact of testing three dependent measures on the Type 1 error rate (von der Malsburg & Angele, 2017), we applied the Bonferroni correction by dividing the .05 alpha level by three, which yielded an adjusted alpha level of 0.017.

Two statistical analyses were conducted for each measure. The first analysis – reported in Table 2 and in the text – used a model with the sum-coded factors Grammaticality (–0.5 grammatical / +0.5 ungrammatical), Attractor Number (–0.5 singular / +0.5 plural), and Region (–0.5 critical / +0.5 spillover), as well as their interactions. The second analysis – reported in the text – quantified the nested effect of attraction within grammatical and ungrammatical sentences separately, using the emmeans package (Lenth, 2024). This analysis was performed because previous research has shown that attraction mainly occurs as processing facilitation in ungrammatical sentences (Hammerly et al., 2019). Thus, the attraction effect in ungrammatical sentences was our critical diagnostic of attraction in Bulgarian clitics.

The random structure of the models included intercepts and slopes for the two theoretically relevant fixed effects – Grammaticality and Attractor Number – and their interaction. We also included a random intercept for trial, because there were two non-independent data points from each trial (e.g., one observation from the critical region and one from the spillover region).

4. Results

Mean accuracy in the comprehension questions was 80% (SD = 21%) in the experimental sentences and 90% (SD = 6%) in the filler sentences. Skipping rates were 3.29% and 3.06% in the critical and spillover regions, respectively. We focus the discussion of the reading data on the two effects of interest, grammaticality and attraction, but include the full output of the statistical models in Table 2. A figure with the reading times of the regions preceding and following the critical region is available in the Appendix.

The results of the full models in the combined critical-spillover region showed a main effect of grammaticality in regression-path and total times, with longer reading times in ungrammatical than in grammatical sentences in both measures. In total times, the grammaticality effect interacted with attractor number and region. This reflected the fact that the Grammaticality × Attraction interaction – the statistical marker of a larger attraction effect in ungrammatical sentences – was larger in the clitic region than in the spillover region.

The results of the nested models showed that plural attractors facilitated the reading of ungrammatical sentences through fewer first-pass regressions and faster regression-path and total reading times (Figure 1). Specifically, there was a significant attraction effect in the probability of first-pass regressions in ungrammatical sentences (estimate = –0.31, SE = 0.13, z = –2.37, p = .018), but not in grammatical sentences (estimate = –0.06, SE = 0.15, z = –0.39, p = .696). Similarly, the attraction effect was significant in the regression-path times of ungrammatical sentences (estimate = –0.06, SE = 0.02, t = –2.73, p = .006), but not of grammatical sentences (estimate = 0.00, SE = 0.02, t = 0.31, p = .750). The same pattern was observed in the total reading times (ungrammatical sentences: estimate = –0.07, SE = 0.02, t = –3.37, p < .001; grammatical sentences: estimate = 0.02, SE = 0.02, t = 0.89, p = .376). Taken together, these results show that Bulgarian participants successfully detected number agreement violations with clitics, but also that they experienced attraction in the presence of a plural attractor in ungrammatical sentences.

Figure 1: Descriptive summary of reading measures for the critical (clitic + V) and spillover region.

Note. Points show by-condition means and error bars show 95% confidence intervals calculated for each condition across participants and items.

5. Discussion

This brief report described an eye-tracking-while-reading experiment designed to test the existence of number attraction with object clitic pronouns in Bulgarian. Our goal was to provide evidence to broaden the empirical record regarding the processing of clitics, using a reading paradigm with fine-grained temporal resolution. Our findings showed attraction effects in ungrammatical sentences in all three measures of interest: first-pass regression probability, regression-path and total reading time. The time course (in terms of reading measures) and location (in terms of sentence regions) of attraction effects were similar to those reported for subject-verb agreement (Dillon et al., 2013; Lago et al., 2021; Wagers et al., 2009). In addition, there was evidence of Bulgarian participants detecting antecedent-clitic number violations, leading to grammaticality effects in regression-path and total times. Below we discuss our findings in the context of the variable attraction profiles of pronominal dependencies and the use of attraction effects as a diagnostic of the grammatical status of clitics.

Appendix. Reading times in the regions adjacent to the critical region

Figure S2 shows the reading times of the two regions preceding and following the critical region, in order to provide a more integral picture of the reading profiles in the experimental sentences. Sample sentence (regions indicated between parentheses):

‘While Maria roamed through the things in the attic |, (she) found the picture of the house | and reluctantly | hid it | in the dusty drawer | with old things.’

Figure S2: Empirical reading times of the two regions before and after the critical region.

Note. Points show condition means and error bars show 95% confidence intervals calculated for each condition across participants and items. The x-axis shows the different sentence regions, starting from the second sentence region: (she) found the picture of the house.

Abbreviations

ACC – Accusative

CL – Clitic

F – Feminine

M – Masculine

SG – Singular

PL – Plural

Data accessibility statement

Materials, data, and analysis code are publicly available at the Open Science Framework: https://osf.io/n5fv7/.

Ethics and consent

The study with human participants was performed in accordance with the Declaration of Helsinki. Informed consent was obtained in writing from all participants prior to testing.

Competing interests

The authors have no competing interests to declare.

Author contributions

Conceptualization: T.I.S. and S.L.; methodology: T.I.S. and S.L.; formal analysis: S.L.; investigation: T.I.S.; resources: T.I.S. and I.A.S.; data curation: T.I.S. and S.L.; writing – original draft preparation: T.I.S.; writing – review and editing: T.I.S., I.A.S. and S.L.; visualization: S.L.; project administration: T.I.S.; funding acquisition: T.I.S. and I.A.S. All authors have read and agreed to the published version of the manuscript.

ORCiD IDs

Tanya Ivanova-Sullivan: https://orcid.org/0000-0002-8604-5204

Irina A. Sekerina: https://orcid.org/0000-0003-3859-3000

Sol Lago: https://orcid.org/0000-0002-4966-1913