1.2.1. Syntactic analysis

Biberauer & Roberts (Reference Biberauer and Roberts2015, Reference Biberauer, Roberts, Ledgeway and Roberts2017) and Roberts (Reference Roberts2019) relate the NSP properties to two functional heads: (i) T(ense), which concerns the ‘richness’ of agreement and (ii) D(eterminer), which concerns the ‘richness’ of the article/clitic system. Formally, ‘richness’ correlates with the presence of φ-features (agreement features such as person, number, and gender) and a D-feature. On T, a D-feature entails a full specification of person and number, while on pronouns, it is linked to definiteness.

NSs are derived via incorporation into T: When the goal’s (pronoun’s) features are included in those of the probe (T), the pronoun can delete under feature identity with T, via copy deletion in a chain headed by T (reverberating Platzack Reference Platzack2004). In contrast, when the pronoun’s features are not subsumed under T’s features, goal incorporation (namely, an NS) is rendered impossible. If T bears a D-feature, as suggested for consistent NS languages like Italian, only DP (Determiner Phrase) pronouns, which bear a D-feature, can incorporate into T, while φP pronouns, which lack a D-feature, cannot (this distinction between pronoun types is proposed in Déchaine & Wiltschko Reference Déchaine and Wiltschko2002 et seq.). In contrast, if T bears no D-feature, as suggested for non-NS languages like English and partial NS languages like Hebrew, DP pronouns cannot incorporate into T, while φP pronouns can. Adopting Holmberg’s (Reference Holmberg, Biberauer, Holmberg, Roberts and Sheehan2010) proposal, Roberts (Reference Roberts2019) argues that, in partial NS languages, embedded third-person referential φPs raise to Spec.TP (specifier of the Tense Phrase) to inherit their reference from a higher definite DP [possibly via C (complementizer); see Landau Reference Landau2015], while first/second-person referential φPs – whether embedded or not – raise to Spec.TP to inherit their reference from a speaker/addressee feature in the left periphery. In partial NS languages, then, null referential pronouns are never intrinsically definite.

This analysis seems compatible with partial NS languages involving no omission of first/second-person subjects in root clauses (e.g. Marathi, Icelandic, and Bavarian German) but not with partial NS languages allowing such an omission (e.g. Hebrew, Finnish, and possibly Brazilian Portuguese and Russian).Footnote ⁹ I thus propose an alternative analysis for the latter type of partial NS languages, henceforth dubbed inconsistent NS languages. I suggest that in such languages, both T and referential first/second-person pronouns bear a D-feature, on par with consistent NS languages. In contrast, third-person referential NSs lack a D-feature and an inherent reference and can only assume a definite interpretation when embedded, via movement to Spec.TP and control of a locally c-commanding antecedent (in line with Holmberg Reference Holmberg, Biberauer, Holmberg, Roberts and Sheehan2010), which enables the valuing of the embedded T’s D-feature.Footnote ¹⁰ In root clauses, third-person NSs cannot assume a definite interpretation, and thus T’s D-feature remains unvalued and the derivation crashes.

Hebrew digresses from this analysis, as its present tense does not pattern with past/future tenses regarding subject omission, nor regarding agreement marking. Thus, I propose that present-tense T bears no D-feature in Hebrew as it lacks a person feature (Shlonsky Reference Shlonsky1997).Footnote ¹¹ Hence, only φPs can incorporate into present-tense T, while DPs cannot, as T lacks a D-feature.

Table 1 summarizes the relevant characteristics of the language types discussed above, based on Roberts (Reference Roberts2019) along with the proposed modifications.Footnote ¹²

Table 1. Features on tense (T) and pronouns in different language types, based on Roberts (Reference Roberts2019), combined with the current proposal concerning inconsistent NS languages and expletive pronouns¹

Note. Pro=pronoun; NS=null subject; T=tense; φ=agreement; D=determiner (definiteness).

¹ While Roberts (Reference Roberts2019) parametrizes the Extended Projection Principle (EPP; Chomsky Reference Chomsky1981a), arguing that T only bears an EPP-feature in some languages, in the current study the universality of the EPP-feature on finite T is opted for, à la Chomsky (Reference Chomsky, Martin, Michaels and Uriagereka2000). Hence, expletive NSs are included in the proposed analysis.

² In Hebrew, present-tense T lacks a D–feature.

³ In Hebrew present tense, expletives lack a D–feature.

The other NSP properties are also related to the absence/presence of D on T. Relying on Chomsky’s (Reference Chomsky2013) Labeling Algorithm, Roberts (Reference Roberts2019) proposes that a T bearing a D-feature can independently label the constituent it forms with vP (little v phrase), enabling the subject to remain in post-verbal position. However, if T lacks a D-feature, it cannot label the constituent it forms with vP, requiring either the subject’s raising to Spec.TP or expletive insertion. Complementizer-trace effects follow from the subject’s inability to remain in VP-internal (verb phrase internal) position, since extraction is rendered illicit, with local extraction from Spec.TP being dispreferred to non-local extraction from VP-internal position due to anti-locality (e.g. Rizzi & Shlonsky Reference Rizzi, Shlonsky, Sauerland and Gärtner2007). Extraction from Spec.TP should be possible, however, if this position’s ‘criterial’ nature changes once C is operated on, via e.g. that-deletion in English, or que-for-qui substitution in French (Roberts & Holmberg Reference Roberts, Holmberg, Biberauer, Holmberg, Roberts and Sheehan2010).

Since Hebrew enables post-verbal subjects only with certain verb types, these subjects seem to be regulated by factors other than labeling. Hence, inversion in Hebrew is independent of the functional heads associated with the NSP. Roberts (Reference Roberts2019) suggests that, if only arbitrary NSs can remain in VP-internal position, extraction should be blocked since the variable bound by the wh-operator is indefinite. Indeed, I have proposed that third-person NSs are inherently indefinite in inconsistent NS languages and can assume definiteness only via movement to Spec.TP. Thus, extraction should initiate in Spec.TP, resulting in complementizer-trace effects.

This can account for the complementizer-trace effects in Finnish (and Russian), as well as for if-trace effects in Hebrew, but what about the lack of that-trace effects in Hebrew? It can be related to the clitic nature of še-, the Hebrew that-type complementizer (Shlonsky Reference Shlonsky1988). While Shlonsky argues that še- is a proclitic both syntactically and phonologically, I propose that it is an enclitic in the syntax, obligatorily attaching to the preceding predicate that selects a declarative complement, while being a proclitic in the phonology, as it unites with any phonological material following it (see Klavans Reference Klavans1985 for the suggestion that clitics’ structural and phonological hosts need not be identical).Footnote ¹³ Being an enclitic, še- cannot detach from the relevant declarative predicate, as can be seen in the ungrammaticality of the sentential subject in (5a) and of the topicalized complement clause in (6a), compared to the grammaticality of the English equivalents in (5b) and (6b), respectively. However, (5a) and (6a) can be salvaged by the insertion of the demonstrative pronoun ze ‘this’ before the initial clause. Ze can function as a nominal predicate interchangeable with propositional nouns such as ha-uvda ‘the fact’ (see respective (5c) and (6c)).

I thus suggest that še- bears an unvalued weak [uDec] feature that can only be valued [+Dec] upon cliticization to a verb carrying a [+declarative] feature. This can explain the ungrammaticality of (5a) and (6a): C_[+Dec] has to adjoin the main verb and cannot appear in isolation from it. Moreover, it can account for the lack of that-trace effects in Hebrew: Following the cliticization of še- to V, C is vacated and the ‘criterial’ nature of the embedded Spec.TP position changes, rendering extraction possible. Thus, the impossibility of if-trace sequences in Hebrew is related to the lack of a D-feature on third-person NSs, whereas the possibility of that-trace sequences is independent of the functional heads associated with the NSP and concerns instead the idiosyncratic nature of the complementizer še- ‘that’.Footnote ¹⁵

2.1.1. Research questions

The study addresses the following questions:

1. (i) To what extent is L1 transfer attested in intermediate Hebrew-speaking learners of English?

   1. a. Are expletive NSs accepted?

   2. b. Are referential NSs accepted in first/second persons and past tense but not in third person and/or present tense?

   3. c. Are post-verbal subjects accepted following unaccusative, but not transitive, verbs?

   4. d. Are complementizer-trace sequences accepted with complementizer that but not with complementizer if?

2. (ii) To what extent can feature bundles on functional categories (T, pronouns, complementizers) be reassembled and, in line with the meso-parametric approach, to what extent can functional categories associated with the NSP (T, pronouns) be reassembled? To what extent do learners pattern with English-speaker controls concerning constructions involving these functional categories?

   1. a. Are expletive NSs rejected to the same extent by both learners and controls?

   2. b. Are referential NSs rejected in all persons and tenses by both learners and controls?

   3. c. Are post-verbal subjects rejected with all verb types by both learners and controls?

   4. d. Are complementizer-trace sequences rejected with all complementizers by both learners and controls?

3. (iii) To what extent can (the relevant) interpretable features improve learners’ performance?

   1. a. Are referential NSs rejected more than expletive NSs?

   2. b. Are agentive post-verbal subjects (following transitive verbs) rejected more than non-agentive post-verbal subjects (following unaccusative verbs)?

   3. c. Are D-linked that-trace sequences rejected more than non-D-linked that-trace sequences?

2.1.2. Predictions

2.2.1. Participants

Learners were 104 high-school graduates, who had learnt English formally for 11 years [aged 18;0–19;0 (M_age = 18;2); 86 females]. They were recruited from a pre-college translation course delivered in 3 different semesters.Footnote ¹⁷ Ninety participants were Hebrew-speaking monolinguals, while 14 defined themselves as bilinguals (Hebrew–Russian/Hebrew–German/Hebrew–Spanish).Footnote ¹⁸

Learners’ English proficiency was established based on an initial 3-hour placement test.Footnote ¹⁹ Based on learners’ placement scores, as well as the acquaintance with them during the semester, their proficiency was determined as intermediate, comparable with the ‘independent user’ level (levels B1/B2, as defined by the Common European Framework of Reference for Languages; CEFR 2009: 123). Placement scores are provided in Table 1 of Appendix A at https://doi.org/10.17605/OSF.IO/3RVTA.

A total of 97 native English-speaker controls [aged 15;0–85;0 (M_age = 38;1); 70 females] were recruited from social networking sites (Facebook, LinkedIn, etc.; dialects varied: mainly American, British, and South African). Thirty-six controls were English-speaking monolinguals, while the remaining 61 defined themselves as multilinguals. Fourty-seven specified Hebrew as one of their L1s, and the remaining 14 specified Afrikaans, Yiddish, French, Spanish, Russian, and American Sign Language as one (or more) of their L1s.Footnote ²⁰

2.2.2. Materials and stimuli

Learners’ L2 competence was tested via a binary grammaticality judgment (GJ) and correction task. Two pen-and-paper test versions including the same questions in different order were created to examine four properties in the L2: (i) obligatoriness of expletive subjects, (ii) obligatoriness of referential pronominal subjects, (iii) impossibility of post-verbal subjects, and (iv) impossibility of complementizer-trace sequences.

Each version comprised 110 items: 60 target items and 50 filler items. Thirty target items involved constructions grammatical in the L1 but ungrammatical in the L2:

• - Five items incorporating expletive NSs parallel to It (see (7) below)

  

• - Five items incorporating expletive NSs parallel to There (cf. (8))

  

• - Five items incorporating first/second-person referential NSs in past tense (cf. (9))

  

• - Five items incorporating non-agentive post-verbal subjects (following unaccusative verbs; cf. (10))

  

• - Five items involving that-trace sequences where the subject trace/gap stands for a non-D-linked wh-phrase (cf. (11))Footnote ^21, Footnote ²²

  

• - Five items involving that-trace sequences where the subject trace/gap stands for a D-linked wh-phrase (cf. (12))

  

The remaining 30 target items involved constructions ungrammatical in both the L1 and L2:

• - Five items incorporating third-person referential NSs in past tense (cf. (13))

  

• - Five items incorporating first-person referential NSs in present tense (cf. (14))

  

• - Five items incorporating second-person referential NSs in present tense (cf. (15))

  

• - Five items incorporating third-person referential NSs in present tense (cf. (16))

  

• - Five items incorporating agentive post-verbal subjects (following transitive verbs; cf. (17))

  

• - Five items involving if-trace sequences where the subject trace/gap stands for a non-D-linked wh-phrase (cf. (18))

  

Filler items resembled target items but were mostly grammatical. They included the elements ‘missing’ from target items (i.e. overt expletive/referential subjects), demonstrated grammatical word order, and involved extraction from object (not subject) position, which did not result with complementizer-trace sequences. A small portion (10%) of the filler items involved ungrammatical sentences where the pronoun that appeared instead of the expletive it (cf. (19)).

The latter was used to ensure that participants accepting expletive NSs parallel to it (see (7) above) indeed accept a NS in those sentences, rather than misanalyze that as an expletive.

The items inspecting each construction were of equal length (number of words). Items were pseudo-randomly ordered such that no two consecutive items concerned the same construction. The full list of materials appears in Appendix B at https://doi.org/10.17605/OSF.IO/3RVTA.

2.2.3. Procedure

Instructions were based on Cowart (Reference Cowart1997). They were explained orally and were written in Hebrew on the actual test. Participants were required to read each of the target sentences and ask themselves if they seemed English-sounding or not. They were further asked whether these sentences could be said or written by a native English speaker. Participants were then requested to decide whether each sentence was possible/acceptable or impossible/unacceptable in English, and, in the case of the latter, to provide a correction of the sentence. Lastly, participants were requested not to worry about grammar rules they may have studied in school or elsewhere. The full instructions appear in Appendix B at https://doi.org/10.17605/OSF.IO/3RVTA.

The correction part was meant to determine whether learners judged the unacceptable sentences as such due to the experimental manipulation, but it elicited many unexpected corrections. Corrections were considered target-like when they did away with the problem deriving from the experimental manipulation (i.e. addition of an expletive/referential subject, change of subject position from post- to pre-verbal, or elimination of a complementizer-trace sequence), even when the result was ungrammatical for reasons independent from the experimental manipulation.

2.2.4. Data analysis

The sample size was determined based on the necessary statistical tests: Paired-samples t-tests and repeated-measures analyses of variance (ANOVAs). Since pairwise comparisons and interactions were in order, and since several multi-leveled factors were involved, a large sample was called for (at least 100 participants for some of the tests and even more for others; see e.g. Brysbaert Reference Brysbaert2019). Observations were thus collected from as many participants as possible, given the limitations imposed by resource constraints.

In what follows, only significant effects (p < 0.05) are reported. Moreover, all sphericity violations incurred were corrected using Greenhouse-Geisser estimates.

2.3.1. L1 transfer

Table 2 summarizes the results of learners and controls in the various target items bifurcated according to their (un)grammaticality in the learners’ L1.Footnote ²³ Figures 2 and 3 summarize the respective results of learners and controls in the various item types inspecting the four NSP properties (two item types per property): expletive NSs, referential NSs, post-verbal subjects, and complementizer-trace sequences. In all properties excluding expletive NSs, item type 1 involves a construction grammatical in the learners’ L1, while item type 2 includes a construction ungrammatical in the learners’ L1. In the case of expletive NSs, both item types involve constructions grammatical in the learners’ L1.

Table 2. Learners’ scores (rate of correct rejections of ungrammatical constructions in the L2) in the GJ task, compared to native-speaker controls: Mean accuracy levels (%) and standard deviations (in parentheses) in the item types inspecting the various NSP properties, split according to (un)grammaticality in the learners’ L1

Note. GJ = grammaticality judgment; NSP = Null Subject (Meso-)Parameter; NS = null subject; D = discourse; SD = standard deviation.

¹ In complementizer-trace sequences, item type 1 represents participants’ average performance in two item types: D-linked and non-D-linked that-trace sequences (as detailed below item type 1). Results from the two item types were united since D-linking did not affect participants’ performance, as reported below.

² All referential NSs ungrammatical in the learners’ L1 were rejected consistently, regardless of tense/person. Thus, these numbers represent the average score of 4 item types: First/second/third-person referential NSs in present tense, along with third-person referential NSs in past tense.

Figure 2. Learners’ scores (rate of correct rejections of ungrammatical constructions in the L2) in the GJ task, in the item types inspecting the various NSP properties.

Figure 3. Native-speaker controls’ scores (rate of correct rejections of ungrammatical constructions) in the GJ task, in the item types inspecting the various NSP properties.

Overall, learners demonstrated the highest accuracy rates concerning referential NSs – both grammatical (item type 1) and ungrammatical in their L1 (item type 2) – and concerning post-verbal subjects following transitive verbs (item type 2), which are ungrammatical in their L1. Some variability was detected concerning null it-expletives (expletive NSs; item type 1) and post-verbal subjects following unaccusative verbs (item type 2), both of which are grammatical in these learners’ L1. Finally, learners grappled with null there-expletives (expletive NSs; item type 2) and that-trace sequences (complementizer-trace sequences; item type 1), which are grammatical in their L1, as well as with if-trace sequences (complementizer-trace sequences; item type 2), which are ungrammatical in their L1.

Since an analysis involving all possible comparisons increases the risk of false-positive results, learners’ average scores in each set of constructions were compared (grammatical vs. ungrammatical in their L1; see the rows shaded in grey in Table 2). A paired-samples t-test revealed a significant difference between constructions grammatical and ungrammatical in the learners’ L1 [t(103) = –8.813, p < 0.001], indicating that learners demonstrated significantly lower accuracy in constructions grammatical in their L1, compared to those ungrammatical in their L1, suggesting L1 transfer.Footnote ²⁴ Learners thus showed an increased tendency to judge as grammatical those structures that are grammatical in their L1. Surprisingly, the native-speaker controls demonstrated the same pattern in the constructions under inspection, although the learners’ L1 is not expected to affect their performance at all. Thus, controls demonstrated significantly lower accuracy in constructions grammatical in the learners’ L1, compared to those ungrammatical in the learners’ L1 [t(96) = –6.512, p < 0.001], suggesting that mistakes relating to the constructions grammatical in the learners’ L1 are harder to detect using the GJ task – for learners and controls alike.

To further explore the differences between item types inspecting the same property, a repeated-measures ANOVA was conducted for each participant group: learners and controls. Each ANOVA involved two within-subject factors: property (four levels: expletive NSs, referential NSs, post-verbal subjects, complementizer-trace sequences) and item type (two levels: 1, 2) as well as a between-subject factor: question order (two levels: 1, 2).

Both ANOVAs revealed a main effect for property [learners: F(3,306) = 158.875, p < 0.001; controls: F(3,153) = 22.794, p < 0.001]. Contrasts indicated that all properties differed from one another for the learners, whereas the controls demonstrated differences between all properties besides: (i) referential NSs and post-verbal subjects, and (ii) expletive NSs and complementizer-trace sequences. The property*item type interaction turned out significant too, in both groups [learners: F(3,306) = 23.048, p < 0.001; controls: F(3,153) = 8.335, p = 0.002].

To further explore the interaction in each group, paired-samples t-tests were conducted for each property to compare the two item types used to test that property. In both groups, participants were significantly more accurate concerning it-expletive NSs, compared to there-expletive NSs [learners: t(103) = 5.075, p < 0.001; controls: t(96) = 2.163, p = 0.035], and significantly less accurate concerning that-trace sequences, compared to if-trace sequences [learners: t(103) = –4.393, p < 0.001; controls: t(96) = –4.915, p < 0.001]. Moreover, learners proved less accurate concerning post-verbal subjects following unaccusative verbs, compared to those following transitive verbs [t(103) = –3.904, p < 0.001], while controls reached significantly higher accuracy concerning post-verbal subjects following unaccusative verbs, compared to those following transitive verbs [t(96) = 2.434, p = 0.018].

Table 3 and Figure 4 present the distribution of learners and controls’ mistakes in all item types in which learners’ mistakes were substantial.

Table 3. Percentages of learners and controls making mistakes in the GJ task in constructions involving more mistakes on the learners’ part

Note.GJ = grammaticality judgment; D = discourse.

Figure 4. Percentages of learners and controls making mistakes in the GJ task in constructions involving more mistakes on the learners’ part.

2.3.2. Feature reassembly and meso-parameters

Table 4 and Figure 5 summarize the scores of learners and controls in the four NSP properties, averaging the scores of item types inspecting the same property.

Table 4. Learners’ scores (rate of correct rejections of ungrammatical constructions) in the GJ task, compared to native-speaker controls: Mean accuracy levels (%) and standard deviations (in parentheses) in the different NSP properties, undivided by item type

Note. GJ = grammaticality judgment; NS = null subject; SD = standard deviation.

Figure 5. Learners’ scores (rate of correct rejections of ungrammatical constructions) in the GJ task, compared to native-speaker controls, in the different NSP properties, undivided by item type.

A repeated-measures ANOVA involving subject group as a two-level between-subject factor (learners, controls) and property as a four-level within-subject factor (expletive NSs, referential NSs, post-verbal subjects, complementizer-trace sequences) was conducted. Main effects were revealed for subject group [F(1,155) = 47.596, p < 0.001), indicating that learners were significantly less accurate than controls, and for property [F(3,465) = 132.127, p < 0.001], whose contrasts revealed that all properties differed from one another besides referential NSs and post-verbal subjects. The subject group*property interaction turned out significant as well [F(3,465) = 48.678, p < 0.001]. Separate one-way ANOVAs for each of the properties, incorporating the between-subject factor of subject group revealed that learners were significantly less accurate than controls only concerning complementizer-trace sequences [F(1,158) = 94.515, p < 0.001].

2.3.3. Interpretability

Table 5 summarizes the scores of learners and native-speaker controls in three of the four NSP properties that could determine the effect of interpretability upon performance: NSs, post-verbal subjects, and that-trace sequences. For each property, two constructions were compared: One construction involved only uninterpretable features (dubbed ‘uninterpretable’), while the other involved both uninterpretable and (relevant) interpretable features (dubbed ‘interpretable’).Footnote ²⁵

Table 5. Learners’ scores (rate of correct rejections of ungrammatical constructions) in the GJ task, compared to native-speaker controls: Mean accuracy levels (%) and standard deviations (in parentheses) in the different NSP property pairs involving interpretable and uninterpretable members

Note. GJ = grammaticality judgment; NSP = Null Subject (Meso-)Parameter; NS = null subject; D = discourse.

A repeated-measures ANOVA was conducted for each of the participant groups, involving two within-subject factors: property (three levels: NSs, post-verbal subjects, that-trace sequences) and interpretability (two levels: interpretable, uninterpretable). Both ANOVAs revealed a main effect for property [learners: F(2,206) = 329.353, p < 0.001; controls: F(2,104) = 24.656, p < 0.001], and contrasts revealed that, in both groups, all three properties differed significantly from each other, regardless of interpretability. interpretability turned out significant only for the learner group [F(1,103) = 14.059, p < 0.001], indicating that properties involving (the relevant) interpretable features generally demonstrated higher accuracy rates than properties involving only uninterpretable features, regardless of property. However, the interaction between property and interpretability was significant in both groups [learners: F(2,206) = 7.427, p = 0.004; controls: F(2,104) = 10.975, p < 0.001], indicating that interpretability affected differently the various properties.

Paired samples t-tests revealed that both learners and controls were significantly more accurate concerning interpretable (referential) NSs compared to uninterpretable (expletive) NSs [learners: t(103) = 9.254, p < 0.001; controls: t(52) = 6.191, p < 0.001]. Regarding post-verbal subjects, the two groups differed from one another: Learners were significantly more accurate concerning interpretable (agentive), compared to uninterpretable (non-agentive), post-verbal subjects [t(103) = 3.904, p < 0.001], whereas controls were significantly more accurate concerning uninterpretable, compared to interpretable, post-verbal subjects [t(55) = –2.434, p = 0.018]. Neither group revealed an interpretability-related difference concerning that-trace sequences.

2.4.1. L1 transfer

Transfer was inconsistently attested in the intermediate L2 learners. It was seemingly observed in expletive NSs, which are grammatical in the learners’ L1 and were accepted in 12.9% of null it-expletive items, and in 29.2% of null there-expletive items. Hebrew-speaking learners’ performance replicated that of learners speaking consistent NS languages (e.g. Greek; Prentza Reference Prentza2010), although the latter accepted more ungrammatical expletive NSs (intermediate learners’ mean accuracy: 2.7/5; advanced learners’ mean accuracy: 3.7/5).Footnote ²⁶

Only 13% of the Hebrew-speaking learners accepted null it-expletives in at least two of the five items inspecting this construction, and their average placement score was 8.8 points lower than that of learners making zero–one mistakes. More learners (38%) made two or more mistakes concerning null there-expletives, and their average placement score was 3.9 points lower than the average placement score of learners making zero–one mistakes. Transfer may thus be attributed to lower English proficiency.

Controls also accepted expletive NSs in 13.2% of null it-expletive items, and in 25.7% of null there-expletive items. Seventeen percent of the controls accepted null it-expletives in at least two of the five items inspecting this construction, and 32% accepted null there-expletives in at least two items. Controls thus appeared to pattern with learners concerning expletive NSs, casting doubt on the role of transfer in learners’ mistakes.

Transfer was absent from referential NSs, which were rejected regardless of their L1 (un)grammaticality: Learners accepted only 2.2% and 2.6% of referential NSs grammatical and ungrammatical in their L1, respectively. These learners thus outperformed their Finnish-speaking peers (lower-intermediate learners’ acceptance rate: 46%; upper-intermediate learners: 11.7%; Alsaedi Reference Alsaedi2017) and even advanced learners speaking consistent NS languages (e.g. Arabic, 4.6%; Alsaedi Reference Alsaedi2017). Controls patterned with learners, rarely accepting referential NSs: 3.4%–4.2% of referential NSs grammatical and ungrammatical in the learners’ L1.

Post-verbal subjects did demonstrate transfer to some degree as learners accepted those grammatical in their L1 (following unaccusative verbs) significantly more than those ungrammatical in their L1 (following transitive verbs): in 8.3% versus 1% of the cases. This significant difference replicates Prentza’s (Reference Prentza2010) findings concerning intermediate and advanced Greek-speaking learners of English, although those learners accepted more ungrammatical post-verbal subjects (intermediate learners’ mean accuracy: 3.2/5 following unaccusatives vs. 3.8/5 following transitives; advanced learners: 4.1/5 vs. 4.5/5).Footnote ²⁷

Only 9% of the learners under study accepted post-verbal subjects in at least two of the five items inspecting this construction, so transfer was absent from most learners. The average placement score of learners demonstrating transfer was 3.6 points lower than that of learners making zero–one mistakes. Thus, transfer might again concern learners’ lower English proficiency. Controls scarcely accepted post-verbal subjects: 0.7%–3.2% of post-verbal subjects grammatical and ungrammatical in the learners’ L1, and none of them made more than a single mistake in this construction. Hence, transfer was attested concerning this construction, but only in a small portion of the learners.

Finally, transfer was apparently most evident in complementizer-trace sequences as that-trace sequences, which are grammatical in the learners’ L1, were wrongly accepted significantly more than if-trace sequences, which are ungrammatical in the learners’ L1: 60.3% versus 43.7%. Moreover, 72%–74% of learners accepted that-trace sequences in at least two of the five items inspecting these constructions, and their average placement score was 6 points lower than that of learners making zero–one mistakes. Transfer is thus associated with lower proficiency yet again. Nevertheless, transfer does not predict low accuracy concerning if-trace sequences, which should be rejected consistently, given their ungrammaticality in both the L1 and L2. Contra this prediction, 59% of learners accepted if-trace sequences in two or more of the five items inspecting this construction. Indeed, previous studies show that learners struggle with complementizer-trace sequences regardless of their L1 (e.g. White Reference White1985, Slavkov Reference Slavkov2009). Moreover, concerning that-trace sequences, the current learners seem to parallel those coming from consistent NS backgrounds, like Greek-speaking intermediate learners (mean accuracy: 2/5), although the former have consistently outperformed the latter in the other constructions under inspection. Learners’ poor performance regarding complementizer-trace sequences thus likely concerns additional causes, as shall be discussed below.

Controls also accepted significantly more that-trace sequences than if-trace sequences: in 20.9% and 8.2% of the cases, respectively, mirroring the learners’ difference between the two. However, only 20%–32% of controls accepted that-trace sequences in at least two of the five items inspecting these constructions, and only 11% of them accepted if-trace sequences in at least two items. Moreover, there is still an accuracy difference of about 35%–40% between learners and controls, which is probably the clearest indication for L1 transfer.

Nonetheless, we still have to address learners’ acceptance of if-trace sequences, which should be rejected under L1 transfer. I suggest that their acceptance be attributed to a methodological issue. Participants – both learners and controls – appeared to struggle with the correction of the if-trace items they determined as ungrammatical, using a variety of strategies. Such variance was absent from all other target items and, more relevantly, from the that-trace items, which were corrected by all participants using a single strategy: complementizer deletion.

Hence, participants’ uncertainty concerning the correction of if-trace sequences might have resulted in their acceptance, meant to avoid the correction part of the task. Learners demonstrated four correction strategies, as exemplified with item (18), repeated as (20a): (i) complementizer deletion (20b); (ii) addition of a resumptive pronoun (20c); (iii) changing the sentential structure, turning the main clause into a yes/no question and the embedded clause into a wh-question (20d); and (iv) changing if into that (20e). Only (i)–(iii) were considered target-like (although the deletion of if often sounded unnatural, especially without separating commas), while (iv) was coded as non-target-like. Controls also used (i)–(iii), with (iii) involving even more ‘drastic’ modifications, e.g. changing the main clause from interrogative to declarative (20f).

The various strategies used to correct the if-trace items attest to the non-triviality of their correction, which can explain the lower accuracy observed in all participants concerning these items: Their correction is less straightforward compared to the other items. The fact that, despite their greater complexity, if-trace sequences yielded higher accuracy rates than that-trace ones, can be attributed to their islandhood in the case of the controls, as they constitute weak wh-island configurations in English, which have surely rendered them easier to reject. As for the learners, the higher accuracy concerning if-trace sequences can be attributed to transfer from their L1, where only that-trace, but not if-trace sequences, are licit.

2.4.2. Feature reassembly and meso-parameters

Learners patterned with English-speaker controls concerning expletive NSs, referential NSs, and post-verbal subjects. This indicates feature reassembly on past/future-tense T and on third-person pronouns, in coherence with the meso-parametric approach: If T is reassembled as ‘weak’ (i.e. lacking a D-feature) and all pronouns are reassembled as DPs, pronoun incorporation into T (namely, subject omission) is rendered impossible. Moreover, a ‘weak’ T cannot label the constituent it forms with vP, forcing the subject to raise to Spec.TP in order to satisfy the EPP-feature and leading learners to reject post-verbal subjects.

Hebrew-speaking learners thus seem to outperform learners with consistent NS backgrounds, even at higher L2 proficiency levels (e.g. Greek; Prentza Reference Prentza2010, Reference Prentza2014a,Reference Prentzab). This might owe to the number of functional categories requiring reassembly. While both consistent and inconsistent NS languages require an added D-feature to a single pronoun type: Arbitrary pronouns (which were not discussed in the current study) in the former, and third-person null pronouns in the latter, eliminating the D-feature from T is necessary in only two tenses in an inconsistent NS L1 like Hebrew as opposed to all three tenses in consistent NS L1s.

However, learners’ distance from the controls concerning complementizer-trace items apparently fails to cohere with the meso-parametric approach. If definite third-person subjects cannot remain in post-verbal position, they must raise to Spec.TP and extraction from the latter position should be blocked under anti-locality. I propose to resort to L1 transfer to explain why extraction still occurs although subjects do not remain in post-verbal position. Concerning that-trace sequences, I suggest that learners misanalyze that as a clitic, due to its status in their L1.Footnote ²⁸ As such, it bears an unvalued weak [uDec] feature that can only be valued [+Dec] upon cliticization to a verb carrying a [+declarative] feature. This vacates C, enabling subject extraction and rendering that-trace sequences grammatical. If this is the case, learners did reassemble the features on T and on pronouns but failed to do so on the complementizer that, which is not one of the functional heads associated with the relevant meso-parameter. As for if-trace sequences, their non-islandhood in the learners’ L1 (Shlonsky Reference Shlonsky1990) should decrease accuracy among L2 learners, as learners transfer the non-islandhood of this construction from their L1. Clustering thus arguably holds since feature reassembly seems to have taken place concerning the elements involving the two functional heads associated with the NSP: D and T. However, feature reassembly on the declarative complementizer failed to occur, but this does not impinge on the clustering conclusion.

Let us turn to discuss the lower accuracy of both learners and controls concerning expletive, compared to referential, NSs, and concerning complementizer-trace sequences in general. The target items involving null it-expletives incorporated raising constructions (see (7), repeated below as (21)), whose interpretation has been reported to be complex for L2 learners (e.g. Yoshimura, Nakayama, Fujimori & Shimuzu Reference Yoshimura, Nakayama, Fujimori and Shimizu2016).

This comprehension deficit might have rendered these sentences more difficult to judge. Raising is also marked cross-linguistically (Givón Reference Givón2001) and has been shown to pose difficulty to L1 learners (e.g. Anderson Reference Anderson2006). This might have influenced both learners’ and controls’ performance.

As for there-expletives (see (8), repeated as (22)), these belong to a high, literary register, which appears unfamiliar to both learners and controls. Thus, the rarity of this construction could have affected participants’ performance.

Indeed, syntactic priming effects have been demonstrated in comprehension (e.g. Ledoux, Traxler & Swaab Reference Ledoux, Traxler and Swaab2007), and a sentence’s comprehensibility affects its acceptability (e.g. Bever Reference Bever and Hayes1970). Syntactic priming has also been argued to constitute a long-term phenomenon (e.g. Bock & Griffin Reference Bock and Griffin2000), and Favier & Huettig (Reference Favier and Huettig2021) report that lifelong literacy experience affects syntactic processing and increases error detection in spoken language in a GJ task. It is thus possible that judging rare constructions is harder than judging more frequent ones – for both learners and controls. Moreover, as noted by an anonymous reviewer, some of these constructions were syntactically ambiguous and could be interpreted as involving heavy NP shift (Kimball Reference Kimball1973), which would render expletive insertion unnecessary. Indeed, the item in (23) was the most accepted null there-expletive item – by both learners and controls (31% and 28%, respectively, out of all accepted there-expletive items).

Regarding complementizer-trace items, both that- and if-trace sequences involve subject extraction from embedded clauses, which can cause processing difficulties in both L1 and L2 (e.g. Juffs & Harrington Reference Juffs and Harrington1995; Dussias & Piñar Reference Dussias and Piñar2010). Moreover, these constructions’ grammatical equivalents are extremely rare in the input (Pearl & Sprouse Reference Pearl and Sprouse2013). In addition, none of the that-trace items involved in this study was prototypical (i.e. of the form Who do you think/say…?; Dąbrowska Reference Dąbrowska2008) and, thus, participants are unlikely to have ever encountered the grammatical equivalents of these items. Hence, these items’ rarity might have also affected participants’ judgments, as suggested above: If familiarity affects comprehensibility (Ledoux et al. Reference Ledoux, Traxler and Swaab2007) and comprehensibility improves acceptability (Bever Reference Bever and Hayes1970), then a construction’s rarity might decrease its comprehensibility and, consequently, its acceptability. Furthermore, these items were longer than all other items since they included background sentences, which might have influenced participants’ performance as well. The above-mentioned problematic correction of the if-trace items is also likely to have induced their acceptance. All these factors can explain the lower accuracy of both learners and controls in these items, while L1 transfer can account for the still lower accuracy of learners compared to controls.

One last issue should be raised concerning the nature of the GJ task. This task required intensive, word-by-word reading, which is unintuitive since readers tend to ‘skip’ function words and pay attention to larger, semantically informative, units (e.g. Greenberg, Healy, Koriat & Kreiner Reference Greenberg, Healy, Koriat and Kreiner2004). Indeed, all experimental items yielding lower accuracy among learners and controls involved functional elements: either missing expletives or ‘redundant’ complementizers. In contrast, ungrammaticalities concerned with referential NSs and post-verbal subjects were rarely missed by learners and controls, and both are semantically informative. If we examine specifically the complementizers, we can also explain the lower accuracy concerning that, compared to if, the semantic purport of the latter being ‘stronger’. It was thus easier for participants to ‘skip’ that, compared to if.

2.4.3. Interpretability

The Interpretability Hypothesis predicts learners’ better performance concerning agentive, compared to non-agentive, post-verbal subjects, since the former, but not the latter, involve the interpretable feature of agentivity, which is relevant for the rejection of post-verbal subjects. However, the fact that only 9% of learners demonstrated this effect misaligns with the hypothesis, under which this difference should be attested in most, if not all, learners, even at more advanced proficiency levels. Moreover, interpretability can account for learners’ better performance concerning referential, compared to expletive, NSs, but it cannot account for the same discrepancy attested in the native-speaker controls. Furthermore, the fact that expletive NSs were more easily missed can provide an alternative account to the discrepancy between referential and expletive NSs, especially since it was attested in both learners and controls. Finally, the hypothesis fails to predict learners’ nativelike performance concerning expletives, which involve no interpretable features. Nor does the hypothesis predict the lack of D-linking effect upon learners’ performance concerning that-trace sequences (a finding replicating Prentza Reference Prentza2010 while contradicting Tsimpli & Dimitrakopoulou Reference Tsimpli and Dimitrakopoulou2007). It thus fails to provide an exhaustive account for the findings.

2.4.4. General discussion

The current study deals with an understudied L1 in the context of the L2 acquisition of the NSP: Hebrew, an inconsistent NS language. Compared to learners coming from a consistent NS background – often including advanced learners – the intermediate Hebrew-speaking learners involved in the present study prove more accurate concerning most examined constructions: expletive NSs, referential NSs, and post-verbal subjects. While this could stem from the larger number of functional categories requiring reassembly in consistent NS L1s, methodological differences should also be considered. For example, the present study used a binary untimed GJ task, whereas Alsaedi (Reference Alsaedi2017) involved timed grammaticality rating on a four-point scale, and Prentza (Reference Prentza2010) used a bi-modal paced task involving acceptability rating on a five-point scale. These conditions have probably rendered judgements more difficult, reducing participants’ accuracy.

Another important methodological issue concerns the fact that learners were tested on unpaced written language. This made it more difficult to detect missing or redundant functional elements, like expletive pronouns and complementizers, since those are generally ‘skipped’ during reading as attention is drawn to semantically informative elements (Greenberg et al. Reference Greenberg, Healy, Koriat and Kreiner2004). Moreover, it made it harder to test the Interpretability Hypothesis concerning the property of NSs, since uninterpretable (expletive) NSs were initially at a disadvantage relative to interpretable (referential) ones. Indeed, both learners and controls made more mistakes concerning the detection of expletive, compared to referential, NSs, as well as concerning complementizer-trace sequences. Participants’ performance concerning expletive NSs and complementizer-trace sequences should thus improve in a parallel task involving spoken language, or, alternatively, self-paced reading, where attention is drawn to every single element. Moreover, an unpaced bi-modal task involving both auditory and visual versions of each sentence should also assist in the detection of the elements missed by the participants in the present study. Such conditions would provide a more neutral setting to test the Interpretability Hypothesis concerning learners’ detection of NSs.

L1 transfer was observed concerning post-verbal subjects, albeit only in a small percentage of the learners, who accepted them significantly more following unaccusative verbs, where they are allowed in their L1, than following transitive verbs, where they are disallowed in their L1. Transfer was further observed concerning complementizer-trace sequences, as learners accepted that-trace sequences, which are licit in their L1, significantly more than if-trace sequences, which are illicit in their L1. The latter were nevertheless accepted in the L2, a finding that could not prima facie be explained by transfer. However, a methodological problem could explain learners’ low accuracy concerning if-trace sequences, as both learners and controls struggled with these items’ correction more than in any other item type. Moreover, transfer could still account for the difference between learners and controls, though it concerned the (non-)islandhood of if-interrogatives rather than the functional heads associated with the meso-parameter. Finally, transfer was absent from referential and expletive NSs, with learners and controls accepting them at similar rates. Expletive NS acceptance by both learners and controls was arguably attributed to processing difficulties, alongside the rarity and ambiguous interpretation (i.e. heavy NP shift) of these constructions.

When contrasting feature reassembly and meso-parameters on the one hand with interpretability on the other, the former seems to have the upper hand, as the latter cannot account for all findings. Interpretability did not improve learners’ performance consistently (e.g. in that-trace sequences), and when it did, it sometimes had a similar effect on controls (e.g. concerning referential vs. expletive NSs). In contrast, properties related to the relevant functional categories (D, T) by and large clustered together, while deviant patterns were explained on independent grounds. This kind of clustering is unexpected under micro-parametric approaches, which predict languages to vary arbitrarily, while clustering effects are not expected to occur, unless by chance. Such clustering further advocates access to UG, at least concerning the heads associated with the NSP. However, as pointed out by an anonymous reviewer, to make a strong claim about the validity of this account, a second learner group of higher proficiency should be incorporated, allowing us to see if and which structures continue to pose learnability problems.

Indeed, the participants were intermediate learners. I predict more proficient learners to demonstrate higher accuracy concerning that-trace sequences, provided they are exposed to constructions demonstrating the non-clitic nature of that in English (namely, finite sentential subjects such as That he won surprised Dan). Regarding the islandhood of if-interrogatives, the situation is more complex, since learners are expected to acquire the prohibition on any extraction out of English if-interrogatives, namely, to ‘notice’ the absence of such cases from the input. I believe that here explicit instruction would be crucial. Otherwise, learners’ lower accuracy concerning if-trace sequences, compared to controls, is likely to persist.

If the properties were in fact acquired as a cluster, the difference revealed between the learners and the controls can perhaps be attributed to disparities in input – in terms of both quality and quantity – as well as to their initial states. While L1 learners are mainly exposed to (implicit) positive evidence, and to a great deal of it, learners in L2 classrooms receive substantially less input, which often involves explicit negative evidence. This could explain, for example, the learners’ difficulty in reassembling features on the declarative complementizer based on input alone. Moreover, contra L1 learners, L2 learners already have a grammar when they set out to acquire an L2. This was evident in the effect of the L1, which, although limited, was still manifest among the L2 learners inspected.