Overt pronouns in null subject languages: experimental investigation of Kashubian, Polish, and Silesian

2.1 Variable binding and double embedding

In this section, we review existing claims about how overt subject pronouns differ from null subjects in two environments: variable binding and double embedding.

Whether overt subjects can undergo variable binding is often referred to as the Montalbetti effect, a.k.a. the Overt Pronoun Constraint (Montalbetti 1984), stated in (2) and illustrated in (3) for Spanish, a CNSL, and in (4) for Brazilian Portuguese, a PNSL. While in Spanish the bound variable interpretation requires the use of a null subject, in Brazilian Portuguese it is available with an overt pronoun.^[2]

With a non-quantificational matrix subject, as in (5), Montalbetti (1984: 85) reports that the coreferential construal becomes possible, with both null and overt subjects of embedded clauses potentially being able to corefer with the matrix subject. However, Montalbetti also suggests that null subjects might nonetheless be preferred in this context. Although Montalbetti attributes this particular preference for null subjects to pragmatics, Holmberg (2010a: 91) takes the null/overt subject facts in embedded clauses to be a defining property in the distinction between CNSLs and PNSLs.

In particular, Holmberg (2010a) capitalizes on the observation that in CNSLs such as Arabic, Greek, Spanish, Turkish, and Italian, the embedded subject cannot be overt unless it is contrastive or a shifted topic. On the other hand, in PNSLs such as Brazilian Portuguese, Finnish, and Marathi there is no such requirement, with the crucial difference being that the null subject has to be coreferential with the matrix subject and free reference is not possible in this case, unlike for the overt pronoun. However, it has been suggested that Bulgarian actually allows for both an ambiguous interpretation of an embedded null subject and a similarly ambiguous interpretation of an overt pronoun in the same position (6). Nevertheless, Genevska-Hanke (2022), who points out this observation, classifies Bulgarian as a CNSL, treating it as counter-evidence to the claim that CNSLs disallow coreference between overt subjects of embedded clauses and matrix subjects.

Furthermore, contrary to a widespread view, such a coreferential reading of overt pronouns has been reported as possible even in Italian, as (7) shows (see also Frascarelli and Jiménez-Fernández 2019 and references therein for other Romance varieties).

Contrasts for null and overt subjects seem to be clearer in the second environment – double embedding structures. Languages such as Brazilian Portuguese (a PNSL) have been reported to require that a null subject be coreferential with the closest c-commanding antecedent (8a) – though see Barbosa et al. (2005); Fernandes et al. (2018) for some counter-evidence – whereas an overt pronoun in this language (see (8b)) can corefer with either of the higher subjects. As for languages such as European Portuguese (a CNSL, see (9)), a null embedded subject can corefer with either of the higher subjects, or can have a discourse antecedent. The literature is less clear about what interpretations overt subject pronouns have in European Portuguese, but given the above discussion about variable binding, a reasonable expectation is that they cannot be coreferential with a local antecedent. Data parallel to (8) are reported for another PNSL, Finnish, as well as for a discourse pro-drop language, Mandarin Chinese, by Modesto (2008: 383). Holmberg and Sheehan (2010) propose to account for the PNSL facts by taking them to involve control between the antecedent and the null subject, modeled as syntactic binding by Agree between the unvalued [D]-feature of the null subject pronoun and the antecedent. The unavailability of coreference with the highest subject is a case of defective intervention on their analysis. It is not clear what mechanisms derive the interpretation of the overt pronominal subjects, though.

To sum up, there is a general intuition that CNSLs and PNSLs differ in the binding possibilities for overt and null subjects. To the extent that these differences are systematic, it suggests that binding is a basic differentiator between these two types of languages and raises interesting questions about the information structure and even the morphosyntax of overt pronouns in the two types of languages. However, as can be seen from our review above, (potential) counter-examples have been raised occasionally. Existing examples are also based on a relatively small sample of languages, with the most detailed discussion centered on Romance languages. We add to the cross-linguistic landscape by seeking to establish relevant generalizations in three West Slavic null subject languages, K, P, and S. To ensure that our generalizations are sound and replicable, our study makes use of formal experiments. But before discussing our experiments, we first review basic facts for these three languages. This is especially important since K and S are less well-studied than P is.

2.2 The null subject parameter in K, P, and S

In contrast with P, which has official status and is widely spoken, K and S are varieties spoken by much smaller communities. K is used mostly in northern Poland (Pomeranian Voivodeship) and S in southern Poland (Silesian and Opole Voivodeships) and northern Czech Republic (Moravian-Silesian Region). In the Polish 2021 Census, 87,600 and 457,900 speakers declared K and S respectively as the main language used at home (although some research shows a much bigger number for K; see Rogowska-Cybulska and Cybulski 2011 and references therein). K is classified as severely endangered in the Atlas of the World’s Languages in Danger (Salminen 2010) and has the official status of a regional language, while S is excluded from consideration in the Atlas and has been classified as a Polish dialect.^[3] Both K and S are characterised by a significant degree of dialectal variation. The judgment data presented in what follows come from speakers of Central K (Central Kashubia, Eastern Pomerania region of northern Poland) and Central S (the Katowice metropolitan area in the Upper Silesia region of southern Poland).

While the status of P as a CNSL has been clear (see Barbosa 2019; Franks 1995; Roberts 2019; Ruda 2022; Willim 1989, submitted for this classification, and see also Barski 2013; Bondaruk 2001; McShane 2009; Nillson 1982; Pisarkowa 1969; Wolna et al. 2022 for some previous discussions of null and overt pronominal subjects in P), the situation in K and S needs to be evaluated more carefully, not only because they are relatively understudied, but also because of microvariation in past tense agreement morphology (see Section 2.2.2). As mentioned in the introduction, a number of properties have been used to classify a language as a CNSL; we list six key properties in (10).

In this section, we consider all three languages in light of the properties in (10a–d), showing that they all pattern almost alike, as is expected if they are CNSLs. Property (10e) is difficult to verify due to the dispreference or complete unavailability of long-distance wh-extraction in these languages (see Cichocki 1983; Szczegielniak 1999 for discussions of P); but to the extent that such extraction is possible in K and P, the judgments suggest that the that-trace effect is not operative in these two languages; the S native speakers whom we have consulted do not accept long-distance wh-extraction at all. The final property (10f) is the subject of our experimental study, so we defer our discussion to Sections 3 and 4.

2.3 Interim summary: clarifying binding facts for subject pronouns with experiments

Overall, the grammatical features of K, P, and S reviewed above suggest that they fall straightforwardly within the category of CNSLs: null subjects of all person/number/gender combinations are available in matrix and embedded clauses, all person/number combinations are distinguished morphologically in all tenses (barring agreement drop in K and S), all three languages allow free subject inversion and couple third person agreement on the verb with the se marker to achieve the generic inclusive interpretation in out-of-the-blue contexts (but for modal impersonals in K). These criteria are standard in the NSL literature, even though the latter two may raise some concerns (see the references in Footnote 1 and Section 5 below).

With this background in mind, we move on to consider what the binding facts may be for K, P, and S subject pronouns in two contexts: variable binding and double embedding. Given the above properties, we expect these three languages to largely pattern with each other and other CNSL languages in how overt subject pronouns are interpreted. However, the phenomenon of agreement drop means that K and S are in a sense less consistent in allowing null subjects than P. This opens up a possibility that the languages might actually differ in how null and overt subject pronouns are interpreted, which we explore through formal experiments.

We chose formal experiments over eliciting informal judgments from native speaker consultants, because experiments produce quantifiable data that in principle can reveal more subtle variation between the three languages. Our experiments were presented in an auditory mode, with pre-recorded materials, because K and S are mostly spoken. A further advantage is that pre-recorded materials let us ensure that overt subject pronouns are pronounced consistently in an unstressed manner for all participants. This is important in light of prior observations that the Montalbetti effect can be neutralized in a CNSL like Spanish by stressing or focusing the pronoun.

More specifically, we used a two-alternative forced choice sentence interpretation task, in which participants first heard a sentence with either a null or overt subject pronoun. They then heard a second sentence that makes explicit a particular interpretation of the subject pronoun, and judged whether the second sentence accurately described the first sentence, by choosing between “True” and “False” options. Hence, while we did not use informal judgments, our experiments were closely modeled upon the informal introspective judgment task standardly used by syntacticians to investigate binding and coreference: for a given sentence that is in principle ambiguous, participants decide whether the coreferential (bound) reading is available.

We should also point out that what our experiments measure is how easy it is to access a particular interpretation. For this reason, for instance, we interpret our results as showing a Montalbetti effect if participants find it harder to obtain a bound variable reading with overt subject pronoun sentences than with null subject sentences, and consequently respond “False” more often to overt subject sentences than to null subject sentences. This departs somewhat from descriptions of the effect in the theoretical literature, in which the Montalbetti effect is the impossibility (ungrammaticality) of a bound variable reading with overt subject pronouns. The departure is necessary because experiments alone cannot definitively show whether a reading is possible or impossible. To illustrate this issue, suppose a participant gives a “False” response to overt subject sentences. This response might well indicate that the interpretation in question is ungrammatical and not available at all for overt subjects, but it might also be the case that the interpretation is available, but difficult to access for processing reasons, for example because it occurs less frequently in naturally occurring speech and texts. Deciding between the two possibilities crucially depends on one’s theoretical framework and assumptions about how behavior and grammar are related.

One might wonder whether there are other tasks that could produce results that allow us to make claims about the grammaticality of certain readings. Perhaps with the same pair of sentences, participants could be instead asked to judge the acceptability or possibility of the second sentence as a description of the first. Alternatively, perhaps participants could be presented with a scenario, and then decide whether it is more acceptable to describe the scenario with a sentence containing a null or overt subject pronoun. Or perhaps one could switch to more conventional fieldwork methods, by eliciting informal introspective judgments from native speaker consultants. However, the same limitation is equally applicable to these alternatives: an “unacceptable” or “impossible” response elicited through these tasks might just mean that the bound variable reading is difficult to access. In this respect, then, our judgment task is no worse than these alternatives.

Put differently, judgments obtained for the interpretation of null and overt subjects will always need to be handled carefully. Importantly, in the context of our experiments, we will assume that whatever patterns we observe should be accounted for in terms of grammatical factors (rather than processing ones), as is standardly assumed in the existing literature on null and overt subjects.

3.1 Design

Experiment 1 is intended to assess the Montalbetti effect in K, P, and S: whether K, P, and S speakers prefer a null subject (compared with an overt subject) to be bound by a quantificational subject (like a CNSL language like Spanish, as reported by Montalbetti 1984), or whether they are neutral about the binding of null and overt subjects.

More specifically, we created two conditions that differ in whether the embedded subject was null or an overt pronoun. Since this subject is in principle ambiguous, we made sure to disambiguate by having each sentence be accompanied by a second sentence (a “prompt”) that makes explicit the bound variable reading, as shown in (25). Participants were requested to judge whether the second sentence accurately described the first sentence. In the ideal case, if the language in question exhibited the Montalbetti effect, participants should reliably reply “True” if the first sentence had a null embedded subject and reply “False” if that sentence had an overt subject pronoun.

For thoroughness, we note that we included two other conditions, to study the availability of bound variable readings for overt and null possessors (e.g., in sentences equivalent to Every girl said that (her) mom is left-handed). We will not discuss this manipulation in this paper, as it does not necessarily concern the Montalbetti effect, which covers overt/null alternations (see (2)). One might choose to analyze null possessors instead as being absent in the syntax altogether, in which case there is no overt/null alternation to speak of.

3.2 Materials

For each language, we created 16 frames, each featuring a distinct combination of lexical items, and corresponding prompts. For each frame we created four versions, one per condition (null/overt subject/possessor), producing a total of 64 sentences. The sentences and prompts (64 pairs total) were sorted into four lists using a Latin Square design, such that each condition appeared four times and every frame appeared only once. All sentences and prompts for all three languages were recorded by native speaker readers, to be presented auditorily.

We then combined the lists with target items from Experiment 2 (see Section 4 below), effectively using one experiment’s materials as filler items for the other. We had also explored adding fillers that were not target items from either experiment, but received feedback from native speakers that doing so made the overall study too long and taxing.

3.3 Participants

We recruited 40 adult native speaker participants for each language, by word of mouth. Participants received gifts or vouchers as compensation. Because K and S speakers are almost always also proficient in P, we could not and did not impose any exclusion criterion based on their knowledge of P or any other language.

Our K participants were speakers of the Central K dialect, living in Central Kashubia (Sierakowice and surrounding towns/villages), Eastern Pomerania region of northern Poland. The age of participants ranged from 23 to 70 years old, with the majority being above age 55. In a survey we administered, most participants reported acquiring K from birth, with parents using K at home. A majority of participants (26) reported using K actively at least 50 % of the time, with the other language being P.

Our S participants were speakers of the Central S dialect, living in the Silesian Voivodeship (the Katowice urban area) in the Upper Silesia region of southern Poland. Like K participants, S participants tended to be older adults (45–78 years of age), and generally acquired S from birth, from family and peers. Around half of participants reported using Central S actively at least 50 % of the time, with the other language being P or – in one case – German.

3.4 Procedure

This experiment was conducted in-person with research assistants who are native speakers of K, P, or S. After informed consent was obtained, the experimental session started with a short informal conversation in the target language, so that participants would feel more comfortable. In addition, doing so helped to mitigate the risk of the intrusion of P in the K and S sessions. Research assistants also explained the instructions in an informal manner in the target language.

We used PCIbex (Zehr and Schwarz 2018) for the display of materials and data collection. Participants were randomly assigned to one of the lists, and listened to recordings played from a laptop via PCIbex. They judged whether the prompt was a “True” or “False” description of the sentences by pressing keys on the keyboard. Participants could listen to each recording as many times as they preferred before giving a judgment.

3.5 Results

The results presented are for 40 P participants, 35 K participants, and 32 S participants; the K and S figures are lower due to unforeseen data loss, where responses were not successfully recorded on PCIbex, for presumably technical reasons. Figure 1 presents the percentage of “True” responses; recall that this measures the availability of a bound variable reading. For statistical analysis, we fitted a mixed-effects logistic regression model for each language’s data set, with pronoun type (overt vs. null) as predictors, and with random intercepts by participant. This was the most complex model that converged for all three data sets and did not produce warning messages indicating over-parametrization.

Figure 1:

Experiment 1 (variable binding) results; whiskers = standard errors.

P participants obtained the bound variable reading much more reliably with null subjects (98 %) than with overt subjects (69 %) (b = 4.67, z = 5.62, p < 0.01). In contrast, K and S participants obtained the bound variable reading much more easily for both types of subjects at similar rates (about 90 % for K and about 80 % for S). Numerically, there is a slight advantage for null subjects in both languages, but no significant difference was found for K (b = 1.59, z = 1.79, p = 0.07) or S (b = 0.42, z = 1.01, p = 0.31). Given our definition of the Montalbetti effect (Section 2.3), we take these results to indicate that there is cross-linguistic variation between K, P, and S, such that P shows the Montalbetti effect but K and S does not.

3.6 Discussion

Although P shows that bound variable readings are easier to obtain for null subjects than for overt subjects, as expected if P shows the Montalbetti effect, there is a data point that is potentially problematic: the relatively high rate of “True” responses for overt subjects in P (69 %). In existing descriptions of the Montalbetti effect, it is ungrammatical for overt subjects to have a bound variable reading. If so, the rate of “True” responses should have been much lower.

As an anonymous reviewer suggested, this result might indicate something deeper and surprising about binding possibilities in P, in contrast to another CNSL, like Spanish or Italian. Whether that is the case is an empirical question that we will need to leave to future research. But in the meantime, we argue that there might be simpler explanations for why P speakers show a high rate of “True” responses. For instance, we think there is likely to be a “True” bias attributable to our task. Consider a sentence with an embedded overt subject pronoun like “Every girl said that she is left-handed”. Participants were requested to judge whether in this sentence, each of the girls talked about herself. Although the grammar of P might not usually allow variable binding for the overt pronoun ona ‘she’, there is no other salient third person female individual that could serve as the pronoun’s antecedent. Participants therefore might feel more obliged to give a “True” response. Second, it is also possible that P speakers might make an exception to variable binding if the overt pronoun is stressed, as Montalbetti (1984) reported in the context of Spanish. While the pronoun was not stressed or focused in our stimuli, it is nonetheless possible that some participants perceived otherwise. In our task, participants listened to sentences and interpreted them. In the process of interpretation, participants might have to recall what they heard, but their recollection might not be an accurate reflection of the recordings played. And while participants could listen to the recordings multiple times, doing so would have been effortful and might have introduced further distortions.

Another possibility, suggested by an anonymous reviewer, is the reported tendency for Polish to use overt pronominal subjects more frequently in colloquial spoken language than in written language (Nillson 1982; Pisarkowa 1969). The greater frequency of overt pronominal subjects in spoken P might reflect the disambiguating and emotive functions overt pronouns have, in addition to their different information structural functions (McShane 1999, 2009; Nillson 1982; Pisarkowa 1969 and references therein; see also Pešková 2019 for Spanish and Czech). Setting aside why there is such a difference in use, this general frequency of overt pronouns in spoken P might have made participants more willing to entertain and accept a bound variable reading, even when such a reading is not grammatically licensed.

Similarly, one might also wonder whether high rates of “True” responses and the smaller contrasts in K and S also reflect a more general “True” bias in K and S participants: since there is no standard variety of K and S, participants might be much more willing than P participants to accept any plausible description of the target sentence, even if the description involves an ungrammatical interpretation.

Such a bias seems unlikely to us, especially when we consider participant responses to Experiment 2; recall that each participant completed Experiments 1 and 2 in a single sitting. In Experiment 2, participants also had to give “True” or “False” responses to another set of constructions with overt and null subject pronouns. If K and S participants had a general bias to give “True” responses to overt subject sentences (or even any type of sentence), there should be similarly high rates of “True” responses for overt subject sentences in Experiment 2. That is actually not the case; “True” response rates are generally lower in Experiment 2 for K and S. In fact, contrasts between (certain) K and S conditions in Experiment 2 are also larger and comparable with equivalent contrasts in P (For more details, see Section 4.4.). In other words, even though K and S are not standardized like P is, K and S participants were not indiscriminately giving out “True” responses; they can be as sensitive as P participants are to certain contrasts. In the context of Experiment 1, then, the fact that K and S participants (but not P participants) gave similar responses to the null and overt subject conditions is unlikely to be a null result due to experiment limitations, but rather a genuine case of cross-linguistic variation.

To sum up the results of Experiment 1, a bound variable reading was easier to obtain for null subjects than overt subjects in P. This difference is consistent with P exhibiting a Montalbetti effect and being a CNSL. In contrast, there was no clear difference for both types of subjects in K or S, despite K and S being otherwise typologically very similar and genetically closely related to P.

4.1 Design

This experiment probed a different kind of binding and coreference involving null and overt subject pronouns: when these subjects appear in doubly-embedded contexts, whether they can corefer with the local antecedent or a non-local one. Given existing descriptions, we expect that in a PNSL, null and overt subject pronouns can both corefer with (be bound by) a local antecedent, but only overt subjects can corefer with a non-local antecedent. It is less clear from the literature to what extent overt subjects differ from null subjects in a CNSL, but a reasonable expectation is for null subjects (and not overt subject pronouns) to tend to corefer with a local antecedent, as we discussed in Section 2.

To evaluate these claims about coreference, we again used the same sentence interpretation task: participants listened to a target sentence and a prompt that described that sentence. We used a 2 × 2 design that crossed two factors: whether the sentence had an overt or null doubly-embedded subject, and whether the prompt named the subject of the main clause or the embedded clause, that is, whether the doubly-embedded subject was intended to be bound by a non-local or local antecedent. To the extent that the subject can be bound by the intended antecedent in the target sentence, participants should respond “True” to the prompt sentence. Example (26) provides an example of these four conditions for P. Again, note that this task is intended to mirror coreference judgments that syntacticians typically use in informal introspection.

4.2 Materials

Sixteen frames were constructed. For each frame, we created four versions of sentences and prompts, one version per condition. As in Experiment 1, the resulting 64 pairs of sentences and prompts were recorded and sorted into four lists using a Latin Square design, such that each condition appeared four times and each frame appeared only once.

The items for this experiment were combined with those of Experiment 1, as described above.

4.3 Participants and procedure

Since we combined both Experiments 1 and 2 in the same session, the participant recruitment and procedure were the same as those for Experiment 1.

4.4 Results

Figure 2 plots the results for all three languages. Across all three languages, with a non-local antecedent, participants tended to give a “True” response when the doubly-embedded subject was overt, although the rate of “True” responses is generally low (around 50 % for overt, lower for null subjects). With local antecedents, we observe variation between languages. Visually, P and S pattern together: participants are more likely to give a “True” response if the doubly-embedded subject was null than if it was overt. This was not the case for K: participants were equally likely to give a “True” response regardless of whether the doubly-embedded subject was null or overt.

Figure 2:

Experiment 2 (double embedding) results; whiskers = standard errors.

Statistical tests confirmed these observations of cross-linguistic differences. More specifically, for each language’s data set, we fitted a mixed-effects logistic regression model, with pronoun type (overt vs. null) and antecedent type (non-local vs. local) as predictors, and with random intercepts by frame (again, this was the most complex model that converged for all three data sets). The interaction between pronoun and antecedent types was statistically significant for P (b = 2.04, z = 5.60, p < 0.01) and S (b = 1.95, z = 4.55, p < 0.01), but not K (b = 0.91, z = 1.91, p = 0.06).

4.5 Discussion

We next discuss what these results mean for the classification of K, P, and S. We start by reviewing the responses for the two local antecedent conditions. P and S participants were more likely to allow null subjects than overt subject pronouns to corefer with these antecedents, while K did not show any clear preference. These responses are consistent with a scenario where P is a CNSL – as is usually assumed in the literature – and so is S, while K is a PNSL.

It is harder to draw clear conclusions from the responses for the non-local antecedent conditions. To the extent that any of the three languages are CNSLs, null subjects should be able to be bound by non-local antecedents, and thus one would expect a high rate of “True” responses for the non-local antecedent/null subject condition. However, the actual rate of “True” responses is much lower, around 25 %. Similarly, to the extent that any of the languages is a PNSL, overt subjects should be able to be bound by non-local antecedents, but again, the rate of “True” responses for the non-local antecedent/overt subject condition is much lower, around 50 %.

We suggest that the generally lower rates of “True” responses for non-local antecedents are an artifact of the experiment. For each trial, participants listened to a relatively long sentence with a doubly-embedded structure, processed it, and maintained its representation in memory while judging the prompt (recall that our materials are all auditory). In order to interpret the overt or null subject pronoun, participants would presumably build a dependency between it and what they believe is the intended antecedent. In this dependency-building process, the non-local antecedent arguably has a disadvantage relative to the local antecedent: it appeared earlier in the sentence, before the local antecedent did. As a result, it might be more difficult to retrieve the non-local antecedent (i.e. have it bind the subject pronoun), due to the relatively longer dependency, memory decay, or both. To the extent this dependency is more difficult to build and/or maintain, this would reduce the percentage of “True” responses for the non-local conditions. That said, this processing account still does not explain why “True” responses are more common for overt subjects than for null subjects. Some other explanation must be sought, perhaps in the differences in the information structural properties of null and overt subjects.

To sum up, while results for the non-local antecedent conditions are tricky to interpret, those for local antecedent conditions nonetheless allow us to form a tentative conclusion: P and S, in allowing null subjects to be bound by local antecedents at a higher rate, have the profile one might expect of a CNSL. K, in allowing both null and overt subjects to be bound by these antecedents at the same rates, provides a clear match with Modesto’s (2000) description of Brazilian Portuguese, a PNSL. Setting aside issues relating to interpretation, another noteworthy aspect of the results is the fact that the three languages again do not pattern uniformly with each other, despite being closely related typologically and historically.

5.1 Are variable binding and double embedding reliable bases for generalizations about pronoun interpretation?

One possible response to the inconsistency in the results is to maintain that overt subject pronouns in CNSLs do have marked interpretations relative to null subjects, but variable binding and/or double embedding does not provide a reliable environment for determining the binding possibilities of overt pronouns and null subjects. If that is the case, the associated experiment is confounded, and its results should be set aside.

Indeed, there are some factors that could make it difficult to reliably detect the Montalbetti effect in practice. As we mentioned in the introduction, Montalbetti (1984), among others, observes that for Spanish speakers, an overt subject pronoun can be bound by a quantificational main clause subject if the overt pronoun is focused (e.g., if stressed). In addition, when the binder is not quantificational, languages seem to vary as to whether an overt pronoun can be bound or not (Camacho 2013; Frascarelli 2018; Genevska-Hanke 2022; Montalbetti 1984). That said, we proactively took steps to eliminate these potential issues in Experiment 1: our binders were exclusively quantificational, and we made use of audio recordings, which let us make sure that the overt pronoun was not pronounced with stress that could be misinterpreted as contrastive focus. While we obviously still could not prevent our participants from perceiving that the overt pronoun received focus, we believe that the results of Experiment 1 are as reliable as they can be. Rejecting the variable binding results in favor of the double embedding results – in which case one might classify P and S, but not K, as CNSLs – would then be arbitrary.

5.2 A purely morphosyntactic approach toward the classification of CNSLs and PNSLs

Another response would be to drop the binding possibilities of overt pronouns as a criterion for classifying languages as CNSLs. Instead, the classification criteria would consist of only morphosyntactic ones like (27a–e), or even (27a–b) exclusively (if the issues raised in the references in Footnote 1 with respect to (27c–e) are insurmountable, which is beyond the scope of this paper to determine). In this analysis, K, P, S would all be classified as CNSLs.

There are several arguments in favor of this view. First, research on pronouns has suggested that overt pronouns in general can be further divided into smaller classes across and within languages, with different structural and interpretational properties. To pick just two influential proposals in this literature, Cardinaletti and Starke (1999), for instance, distinguish between ‘strong’ and ‘weak’ pronouns (which are in turn different from pronominal clitics), while Déchaine and Wiltschko (2002) propose that overt pronouns can be DPs, ϕPs, and NPs. While both proposals differ in technical details, they agree that overt pronouns can vary in prosody, structure, and importantly, semantics (see also, among others, Noguchi 1997; Patel-Grosz 2020; Patel-Grosz and Grosz 2017).^[7] From this perspective, it is unsurprising to find that overt pronouns do not behave uniformly in K, P, S, although we must leave a detailed analysis of overt pronouns in these languages to future research.^[8] We only note here that in light of structure-based constraints such as Cardinaletti and Starke’s (1999) Minimise Structure or Patel-Grosz and Grosz’s (2017) Minimize DP!, a language in which null and overt pronouns have equally economical structures (e.g., the same number of syntactic nodes) could be expected to evade (27f) irrespective of its pro-drop type.

Relatedly, there appears to be an implicit assumption in the literature on Consistent and Partial NSLs that overt subject pronouns are such that they must either allow variable binding and also corefer freely with the higher subjects in double embedding (in the case of PNSLs) or not do so at all (in the case of CNSLs). However, this assumption is called into question when we consider a wider range of languages. Mandarin Chinese overt subject pronouns, for instance, can be bound by (corefer with) either of the higher subjects in double embedding; in this respect Mandarin behaves like a PNSL (28). But Mandarin has also been reported to show a Montalbetti effect, in that an overt subject pronoun cannot undergo variable binding, much like in a CNSL (29) (Huang 1991; see also Aoun and Li 1990 for a more nuanced discussion). While we also do not have much to say about overt pronouns in Mandarin, we believe that these observations provide further evidence that overt pronouns across languages are a heterogeneous class of lexical items in their own right. Given their diversity, overt pronouns might not always stand in clear contrast with null subjects, contrary to what is usually assumed in the CNSL/PNSL literature.

Finally, there is also a conceptual argument for dropping the interpretation of overt pronouns from the list of criteria in (27): while the properties in (27a)–(27d) clearly reflect the formal, morphosyntactic features of the language, the interpretation of overt pronouns is a matter of the interface between syntax and information structure, where the alignments clearly vary across CNSLs (see Jasinskaja and Simík forthcoming for a discussion of perspectives from Slavic) and where judgments and preferences are especially susceptible to influences from language processing (see, among many others, Fernandes et al. 2018; Filiaci et al. 2014; Sorace 2005; Sorace et al. 2009; Wolna et al. 2022). In this domain, we note that CNSLs can differ in the interpretation of overt subjects, even non-pronominal ones. The quantifier interaction data in (30) provide a concrete example: these Polish sentences are natural with a narrow scope interpretation of the subjects, i.e. the intended, common-sense interpretations involving different handsome men, policemen, and lottery winners (this interpretation is facilitated by placing stress on jakiś ‘some’/ktoś ‘someone’). In contrast, parallel structures are deviant in Greek, another CNSL, and are offered to support an A-bar analysis of preverbal subjects, in which such subjects always receive an obligatory wide scope interpretation (Alexiadou and Anagnostopoulou 1998: 505–506). Thus, if the logic of the argument holds, Polish (but not Greek) preverbal subjects can in fact be in an A-position, despite the pro-drop status of both languages, or, alternatively, the effect of A-bar movement on the scopal properties of the subject observed in Greek, as well as in Romance varieties (Barbosa 1995), is not universal for all CNSLs. In either case, overt subjects are not interpreted uniformly even across the CNSLs which (on the surface) pattern alike with respect to the property in (27f).

While investigating data of this type goes beyond the scope of this paper, they further underscore the need for much more detailed comparative research into subject interpretation in CNSLs before the correlations observed in the current literature can be lifted to the status of reliable generalizations expected to apply beyond the language types investigated so far. Clearly, various independent aspects of the grammars of particular languages can be expected to influence the interpretive options available to both overt and null pronouns, diminishing the categorizing potential of pronoun interpretation in diagnosing null subject language types (in addition to the structure of the pronouns themselves, potentially, e.g., free/fixed word order, expression of information structure, presence of CLLD, presence of articles; see, e.g., Barbosa 2019 and references therein for a discussion of interpretive differences between subjects in discourse pro-drop languages).

To the extent that typological classifications such as the one in (31) from Barbosa (2019: 487–488) can be theoretically informative, we suggest that if a language fulfills the description specified in (31) for CNSLs, this should be sufficient to classify it as such, regardless of its patterns and frequencies of use of overt pronouns (see also Genevska-Hanke 2022). In other words, a CNSL is a language in which pro-drop is licensed by rich agreement (in one way or another, see Barbosa 1995, 2011a, 2019; Camacho 2013; Cole 2010; Holmberg 2005, 2010a; Roberts 2010, 2019; Saab 2020, 2024; Sigurðsson 2011).

A consequence of this definition is that drawing the fine line between consistent and partial pro-drop can now be based on purely formal considerations: languages in which verbal inflection is derivationally responsible for the distribution of null subjects are CNSLs. This sets them apart from PNSLs, where several alternative derivational mechanisms have been proposed: null NP anaphora (Barbosa 2013, 2019) or control (Holmberg 2005, 2010a; Holmberg et al. 2009; Holmberg and Sheehan 2010); and see Frascarelli (2018); Frascarelli and Jiménez-Fernández (2019) for another view where CNSLs and PNSLs exist on a continuum.

Focusing on the mechanism rather than the surface distribution of null and overt subjects may in some cases be crucial for categorizing a language one way or another, as this provides flexibility to impose additional formal restrictions on pro-drop even in CNSLs. To illustrate this point, suppose we use only surface distribution as a criterion (contrary to what we propose). If that is the case, we might conclude that Italian, the generative poster child of CNSLs, is not a CNSL, since a null [2sg] subject cannot be freely dropped – it is blocked in the subjunctive in the language due to inflectional syncretism (Cardinaletti 1997; Cognola and Casalicchio 2018b; Roberts 2019). In contrast, taking into account the derivational mechanism generally available in a language lets us accommodate additional restrictions on pro-drop (such as inflectional syncretism), and in doing so, avoid an intuitively incorrect shift in the classification of languages like Italian. On this approach, inflection-based constraints can in fact provide evidence for the role of agreement in licensing pro-drop in a language, supporting its classification as a CNSL.

Importantly, viewing cross-linguistic differences in pro-drop in terms of derivational mechanisms rather than purely surface properties is a crucial component of classifying languages such as K and S. As Section 2 noted, there is a correlation between pro-drop and agreement drop. Since this correlation basically boils down to inflectional syncretism arising from agreement drop, it supports the view on which the derivation of null subjects in K and S is agreement-based (in addition to the availability of null subjects of all person/number/gender combinations in both matrix and embedded clauses), implying that they belong to the class of CNSLs.