1.4.1 Relations between templates

In approaches that decompose the verb in the syntax, transitivity alternations are commonly derived through layering (Alexiadou et al. 2015). On this view, alternating verbs are not derived from one another (that is, the transitive is not derived from the intransitive, nor vice versa). Rather, the two alternants share the same core structure and differ only in higher structural layers associated with voice and the external argument. In these theories, alternations are therefore not a grammatical primitive or a rule linking two templates, but an epiphenomenon of two independently derived structures that share a common core. I use the term alternation descriptively throughout, to refer to pairs of related intransitive and transitive verbs, without implying that one alternant is derived from the other or that the two are linked by a grammatical rule. The predictions regarding relations between templates are derived from claims about templates that share a core structure but minimally differ in their higher layers, the same mechanism these approaches use to explain why some template combinations alternate and others do not.

Two of the most influential approaches that assume layering to explain Hebrew verbal template alternations are Doron (2003) and Kastner (2020).

In Doron (2003) CaCaC, CiCeC, and hiCCiC each correspond to a syntactic configuration specified for the presence or absence of the two agency heads, the “causative” head γ, and the “intensive” head ι ([–γ, –ι], [–γ, +ι] and [+γ, –ι] respectively). These heads Merge above the root node and constitute part of the verb’s syntax. CaCaC ([–γ, –ι]) is termed “the simple template”, CiCeC ([–γ, +ι]) is termed “the intensive template”, and hiCCiC ([+γ, –ι]) is termed “the causative template”. Crucially, niCCaC and hitCaCeC are derived by layering the “middle” voice head μ on top of CaCaC and CiCeC respectively, and hiCCiC is derived by layering the “causative” head γ on top of CaCaC. Therefore, for Doron, the template pairs structurally related through layering are niCCaC-CaCaC, hitCaCeC-CiCeC, CaCaC-hiCCiC, and CaCaC-CiCeC (the intensive alternation, e.g., ŠaLaX ‘send’ ŠiLeaX ‘send away’).

Kastner (2020) states a similar layering claim in different theoretical terms. He proposes that templates are phonological realizations of configurations that include different values of the EPP feature [D] on the Voice head, which determines the presence or absence of an external argument in the specifier of Voice. When Voice carries a [+D] feature, the verb is in the hiCCiC template; when it carries a [–D] feature, the verb is in either niCCaC or hitCaCeC; and when it is unspecified, it is in either CaCaC or CiCeC. hitCaCeC and CiCeC minimally differ from niCCaC and CaCaC since they include the modifier $\sqrt{\text{ACTION}}$ on top of v. For Kastner, the structurally related template pairs are therefore niCCaC-CaCaC, hitCaCeC-CiCeC, CaCaC-hiCCiC and niCCaC-hiCCiC. For both Doron (2003) and Kastner (2020), since layering is the grammatical mechanism for relating these patterns, it is predicted that speakers will prefer pairs of templates that share the same syntactic core but differ in their specification of Voice⁴: presence of the middle head μ for Doron, and specification of the [D] features for Kastner. This yields the following structurally related pairs: niCCaC-CaCaC, hitCaCeC-CiCeC, CaCaC-hiCCiC, niCCaC-hiCCiC (Kastner), and CaCaC-CiCeC (Doron, in the intensive alternation).

Unlike Doron and Kastner, Arad (2005) puts more weight on the idiosyncrasy of the template system. For Arad, the template is a realization of the verbalizer head v, and in non-alternating verbs a specific template is specified under this head. She distinguishes between the syntax of alternating and non-alternating verbs, and claims that in the causative alternation, the root selects a conjugation class consisting of a pair of templates (e.g., CaCaC-hiCCiC). The correct template for the syntactic context is then selected based on the “flavor” of v (inchoative or causative). Arad proposes that there are six conjugation classes based on observed template pairs in alternating verbs: niCCaC-CaCaC (e.g., niFTaX-PaTaX – ‘open’), CaCaC-hiCCiC (KaFaʔ-hiKPiʔ – ‘freeze’), niCCaC-hiCCiC (niRGaʕ-hiRGiʕa – ‘relax’), hitCaCeC-CiCeC (hitXaMeM-XiMeM – ‘warm’), hitCaCeC-hiCCiC (hitBaHeR-hiVHiR – ‘clear’), and hiCCiC-hiCCiC (hiLBiN-hiLBiN – ‘whiten’). Additionally, four less-frequent template pairs are observed in Hebrew, not mentioned by Arad as conjugation classes: CaCaC-CiCeC (SaMaX-SiMe’aX – ‘be happy’, ‘cheer up’), niCCaC-CiCeC (niRPaʔ-RiPeʔ – ‘cure’), CaCaC-CaCaC (ACaR-ACaR – ‘stop’), and CiCeC-hiCCiC (ZiNeK-hiZNiK – ‘jump away’). Since conjugation classes lexically specify the possible template combinations for the grammar, it is predicted that speakers will prefer alternations that form a conjugation class over those that do not, i.e., that they will prefer alternations attested in the language over novel combinations.

Borer (2004) goes even further than Arad (2005) in highlighting the idiosyncrasies of the system, proposing that each root is listed with a specific template, thereby predicting an even less constrained selection of templates by speakers. She also argues that templates themselves encode transitivity, a claim I discuss in Section 1.4.2.

1.4.2 Transitivity

A recurring theme in theories of Hebrew verbal templates is the extent to which they are tied to transitivity. Borer (2004), discussing acquisition studies of Berman (1982; 1993; 1994), proposes that there are morphosyntactic rules that relate templates with the transitivity of the syntactic structure manifested in the [OM] feature denoting Object Marking, the ability to assign accusative case. In Borer’s system hiCCiC and CiCeC are realizations of [+OM], niCCaC and hitCaCeC are realizations of [–OM], and CaCaC is unspecified for transitivity. Since templates are listed for specific roots in Borer’s system, as discussed in Section 1.4.1, the [OM] feature does not directly determine the choice of template, but, according to Borer, may assist in lexical access by “delimiting the search” for the appropriate template for a specific root.

Similarly, Arad (2005) proposes that niCCaC and hitCaCeC are marked as [–tr] or [–ACC case], which confines their environment of insertion to intransitive verbs. This generates the correct pattern in the lexicon where all templates can be used in intransitive contexts, but normally only CaCaC, CiCeC, and hiCCiC can be used in transitive contexts.

In Kastner (2020) the templates’ relation to transitivity is through Voice, which determines the presence of an external argument, and, through Burzio’s (1986) generalization, also determines the verb’s ability to assign accusative case. Therefore, the prediction is that hiCCiC ([+D]) will have an external argument and will be able to assign accusative case, whereas niCCaC and hitCaCeC ([–D]) will not have an external argument and will not be able to assign accusative case. CaCaC is unspecified for [D] and therefore also for transitivity. Similarly, for Doron (2003), the middle head μ blocks an external argument, and therefore also transitivity for the middle templates niCCaC and hitCaCeC.

Since all of these theories associate verbal templates with transitivity, the prediction is that whatever the speaker knows about this association should surface in their preference for template in nonce verbs. Specifically, speakers will rule out transitives in niCCaC and hitCaCeC. Additionally, they might prefer transitives in CiCeC and hiCCiC over transitives in CaCaC, since CiCeC and hiCCiC are transitive whereas CaCaC is unspecified in Doron (2003), Borer (2004), and Kastner (2020).

Having reviewed the main theoretical proposals, I briefly outline what the lexicon of Hebrew actually shows about the relationship between transitivity and verbal templates. Although all alternating transitives are in either CiCeC, hiCCiC, or CaCaC, the opposite implication is not true. That is, there are verbs in CiCeC and hiCCiC that are not transitive (e.g., hiGiʕa – ‘arrive’, PiZeZ – ‘prance’), and at least one alternating verb in CiCeC that is not transitive (ZiNeK – ‘jump away’ is intransitive and alternates with hiZNiK). hiCCiC-hiCCiC alternations (hiLBiN – ‘whiten’) are also of this sort (although Borer 1991 claims that in these cases each alternant is independently derived from the corresponding adjective). Additionally, there are cases in which both alternants are syntactically transitive and assign accusative case, e.g., LaVaŠ-hiLBiŠ (‘wear’-‘dress’). In sum, it seems that the only generalization about the relation between transitivity and template is that niCCaC and hitCaCeC are not transitive.

The experiments presented in this paper investigate which aspects of template selection in the causative alternation are based on grammatical generalizations and which ones rely on listed idiosyncratic information about specific items. I examine this question using alternating nonce verbs in production and judgment, with the rationale that grammatical generalizations will be applied to nonce verbs, whereas listed information will be manifested as variability in template selection. Before turning to the present study, I briefly review previous work that has used nonce verbs in Hebrew.

2.1.1 Methods

2.1.2 Results

2.1.2.1 Template preference

All 25 participants produced well-formed Hebrew verbs on the majority of trials and answered the majority of filler items correctly (correct agreement and verbal template). Accordingly, none were excluded. At the item level, 27 responses were excluded since they were not in one of the five templates in Hebrew (e.g., GoCiV), the root differed from the root in the stimulus intransitive verb, or the participant did not provide a response. The percentages of transitive responses in each template for each intransitive nonce verb in each template are presented in Table 2.

Table 2: Rate of response (transitive) templates by stimulus (intransitive) template.

		Response (transitive)
		CaCaC	CiCeC	hiCCiC	hitCaCeC	niCCaC
Stimulus (intransitive)	CaCaC	4%	64%	32%	0%	0%
	niCCaC	32%	45%	23%	0%	0%
	hitCaCeC	9%	66%	25%	0%	0%
	hiCCiC	24%	50%	26%	0%	0%

As summarized in Table 2, there was variability in the selection of transitive templates across each of the conditions, and there was no case in which the selected template was consistent across participants and sentences. The most common response template for all four conditions was CiCeC, which accounted for 45% to 66% of responses for all four intransitive templates. Overall, CiCeC was the most frequent response with 56% of responses, followed by hiCCiC with 26%, and CaCaC with 17%. Among the 45% of CiCeC responses to intransitive niCCaC items, about a quarter (11% of all responses) were cases in which the ni- prefix was reanalyzed by participants as a root consonant, as evidenced by the vowel template characteristic of CiCeC with quadriliteral roots. For example, for an intransitive niLTaG (the triliteral root LTG in niCCaC), a participant might produce the transitive NiLTeG, apparently reanalyzing the initial /n/ of niCCaC as part of a quadriliteral root NLTG in CiCeC, analogous to the real verb NiRMeL ‘normalize’.

Speakers’ preferences among the transitive templates were assessed with the model described in Section 2.1.1.3. The model’s output is reported in Table 3, and overall template preferences (averaged over Intransitive) and pairwise comparisons with Tukey correction are reported in Table 4. CiCeC was preferred over both hiCCiC (p < .0001) and CaCaC (p < .0001). The contrast between hiCCiC and CaCaC was not significant (p = .27). This is illustrated in Figure 1. At the level of individual participants, CiCeC was the most common response for 20 of the 25 participants. Of those, 13 exhibited a CiCeC > hiCCiC > CaCaC preference. That is, the dominance of CiCeC held for most individual participants, but the relative ranking of hiCCiC and CaCaC varied. The only invariable preference observed in the data was the virtual avoidance of niCCaC and hitCaCeC as transitive alternants, with only one response each.

Table 3: Multinomial mixed-effects logistic regression model predicting choice of transitive template (Transitive ~ Intransitive + (1+ Intransitive | Participant)). Estimates are odds ratios. The reference level is hiCCiC for the transitive template and CaCaC for the intransitive template.

	Transitive: CaCaC (vs. hiCCiC)			Transitive: CiCeC (vs. hiCCiC)
Predictors	Odds Ratios	CI	p	Odds Ratios	CI	p
Intercept (Intransitive CaCaC)	0.13	0.05–0.30	<0.001	2.43	1.29–4.55	0.006
Intransitive hiCCiC	6.55	1.62–26.51	0.008	1.37	0.46–4.09	0.573
Intransitive hitCaCeC	1.15	0.32–4.11	0.835	1.44	0.70–2.96	0.324
Intransitive niCCaC	10.50	3.89–28.34	<0.001	1.03	0.55–1.92	0.919
N Participant	25
Observations	1171

Table 4: Pairwise comparisons among transitive templates, averaged over the intransitive templates, with Tukey correction for multiple comparisons. Estimates are on the log-odds scale and given as the first template minus the second.

Contrast	Estimate	SE	z	p
hiCCiC – CaCaC	0.10	0.06	1.54	.274
hiCCiC – CiCeC	–0.43	0.09	–4.76	<.0001
CaCaC – CiCeC	–0.52	0.08	–6.99	<.0001

Figure 1: Mean proportion of trials on which each template was selected as the transitive form in Experiment 1. Bars show means across participants; error bars show 95% confidence intervals; dots show individual participant proportions. Significance brackets indicate pairwise differences from a multinomial mixed model (*** p < .001; n.s. = not significant after Tukey’s correction).

The model also revealed that the intransitive template modulates the choice of transitive template. The contrast between CiCeC and hiCCiC was stable across intransitive templates, with no significant effect of Intransitive on this contrast (all ps > .32). The contrast between hiCCiC and CaCaC was modulated by Intransitive: the odds of selecting CaCaC over hiCCiC increased when the intransitive template was niCCaC (OR = 10.50, p < .001) or hiCCiC (OR = 6.55, p = .008); the shift at the hitCaCeC intransitive was not significant (OR = 1.15, p = .835). This is illustrated in Figure 2. Crucially, some significant shifts run counter to the inventory pattern, as presented in Section 2.1.2.3.

Figure 2: Mean proportion of trials on which each transitive template (CiCeC, hiCCiC, CaCaC) was selected in Experiment 1, plotted by intransitive template. Points show means across participants and error bars show 95% confidence intervals; lines connect the same transitive template across intransitive templates.

2.1.2.3 Comparing the results to template distribution in an inventory of Hebrew alternating verbs

To examine whether there is a relation between the distribution of responses in the experiment and the distribution of templates in verbal alternations in the language, I compared the results with an inventory of 443 Hebrew alternating verbs collected by Laks (2011; 2014a).⁶ Comparing the rate of transitive templates in the causative alternation overall, regardless of the intransitive template, the experiment’s results align closely with the inventory (frequency matching, Ernestus & Baayen 2003; Hayes et al. 2009, among others). As Figure 3 illustrates, CiCeC is the most prevalent transitive template in both the inventory and the experiment, followed by hiCCiC, with CaCaC being the least common. Neither niCCaC nor hitCaCeC is ever a transitive alternant in the inventory and are also virtually absent as transitive responses in the experiment.

Figure 3: Mean proportion of trials on which each transitive template was selected in Experiment 1 (Experiment), compared to the distribution of transitive templates in the Hebrew verbal inventory (Inventory). Bars show means across participants; error bars show 95% confidence intervals; dots show individual participant proportions.

However, when comparing the conditional distribution of transitive templates given an intransitive template, a different picture emerges. As demonstrated in Figure 4, speakers do not seem to know which template pairs are frequent in Hebrew, and do not systematically consider the intransitive template when selecting a transitive template. As presented in Section 2.1.2.1, the intransitive form does not modulate speakers’ preference for CiCeC over hiCCiC, even though in the inventory CiCeC is the most frequent transitive only for hitCaCeC intransitives. For instance, intransitives in CaCaC very infrequently alternate with CiCeC in the inventory, yet CiCeC was the predominant response in the experiment. Particularly striking is the case of intransitive hiCCiC, where the inventory shows that 100% of its transitive alternants are in hiCCiC, while the majority of responses in the experiment were in CiCeC.

Figure 4: Mean proportion of trials on which each transitive template was selected in Experiment 1 (Experiment), broken down by intransitive template, compared to the distribution of transitive templates in the Hebrew verbal inventory (Inventory). Experiment bars show means across participants; error bars show 95% confidence intervals; dots show individual participant proportions. Inventory bars show corpus proportions from Laks (2011; 2014a).

The preference for CaCaC over hiCCiC was modulated by intransitive in the experiment, but not consistently following the pattern of the inventory. The shift toward CaCaC in niCCaC intransitives is compatible with the inventory, where CaCaC and hiCCiC alternate with niCCaC at roughly equal rates. The shift in hiCCiC intransitives, however, runs counter to the inventory: hiCCiC intransitives always alternate with hiCCiC transitives, yet speakers shifted away from hiCCiC toward CaCaC.

Together, these comparisons indicate that while speakers’ aggregate template preferences resemble the inventory, their behavior in choosing a transitive template for a given intransitive template does not reflect the conditional distributions attested in the language.

2.2.1 Methods

2.2.2 Results

Six participants were excluded from the analysis: two assigned the same score to all sentences including fillers, three did not show a difference in rating between fillers with correct agreement and fillers with incorrect agreement (difference in average scores < 1), and one selected an incorrect answer for a comprehension question in the instructions. I therefore analyzed results from 69 participants, with each list containing a minimum of four participants (one list had six participants, one had seven). The average ratings for each template pair are presented in Table 6, and Figure 5 presents the distribution of ratings across combinations of transitive and intransitive templates.

Figure 5: Acceptability ratings by intransitive and transitive template. Each point represents a single response. Colors indicate the transitive template presented for each intransitive template.

Most template combinations, including non-existent and rare combinations, were judged as acceptable with a rating between 4 and 5. hitCaCeC was consistently judged as an unacceptable transitive regardless of the intransitive template. In cases where the intransitive and transitive templates were identical (e.g., CaCaC-CaCaC, hiCCiC-hiCCiC, and hitCaCeC-hitCaCeC), the ratings were the lowest compared to all other transitives for a given intransitive (apart from hitCaCeC transitives). Consistent with this pattern, the model testing template identity yielded a negative and significant coefficient for SameIT (p < .001). In other words, identity between the intransitive and the transitive template had a reliably negative effect on acceptability ratings, with some variability across participants and items for non-identical pairs (SameIT = 0), as illustrated in Figure 6 and in the analyses that follow. This result supports the conclusion that participants systematically dispreferred combinations where the transitive template preserved the intransitive template. The full model output is presented in Table 7.

Figure 6: Distribution of acceptability ratings by identity between intransitive and transitive template (SameIT).

Given the observed dispreference for identity pairs, I excluded these cases from the analysis about overall template preference. An additional advantage for the exclusion is that the experiment did not include items with a CiCeC intransitive, and therefore there were no CiCeC-CiCeC pairs tested. This exclusion helps prevent potential confounding by eliminating the possibility that the absence of CiCeC intransitive-transitive pairs could artificially inflate the average score for CiCeC transitives.

In Experiment 1, CiCeC was the preferred transitive template overall, hiCCiC was second, CaCaC the third, and hitCaCeC and niCCaC were essentially absent as transitives. Therefore, I tested whether the same preferences would also emerge in Experiment 2.

For preferences among the transitive templates (non-identical trials only), a likelihood-ratio test against a null model with the same random-effects structure but no fixed effect of Transitive revealed a significant effect of transitive template (χ²(3) = 83.29, p < .001). Fixed-effects estimates showed that all three alternative templates were rated significantly higher than the reference level hitCaCeC (see Table 8 for the full model output). Pairwise comparisons with Tukey correction (Table 9) confirmed that hiCCiC and CiCeC were both rated significantly higher than CaCaC (both p < .001) but did not differ significantly from each other (p = .96). This suggests a ranking of preferences: hiCCiC ≈ CiCeC > CaCaC > hitCaCeC. Figure 7 presents the average rating for each transitive template.

Figure 7: Mean acceptability ratings for each transitive template in Experiment 2 (SameIT = 0 trials only). Bars show means across participants; error bars show 95% confidence intervals; dots show individual participant mean ratings. Significance brackets are shown for adjacent pairs (*** p < .001; n.s.). All non-adjacent comparisons were significant at p < .001.

At the individual participant level, CiCeC and hiCCiC were rated similarly (a mean difference of 0.5 points or less) by 38 of the 69 participants, consistent with the group-level non-significant difference between the two. Of the remaining participants, 16 gave higher ratings to CiCeC and 15 gave higher ratings to hiCCiC, showing a roughly even split. The preference for CiCeC over CaCaC was present in 33 of the 69 participants, and the preference for hiCCiC over CaCaC in 36. The strongest and most consistent contrast between adjacent templates in the ranking was between CaCaC and hitCaCeC: 60 of the 69 participants rated CaCaC higher than hitCaCeC, mirroring the clear group-level avoidance of hitCaCeC as a transitive template.

The full group-level ranking (CiCeC ≈ hiCCiC > CaCaC > hitCaCeC) was observed in 16 participants. As with Experiment 1, the preferences held at the group level but with considerable variability across individuals.

Finally, the manner-faithfulness model revealed a significant main effect of Faithfulness and significant effects of two of the intransitive templates compared to the reference CaCaC: hiCCiC and niCCaC. Crucially, the main effect of BP was not significant, nor were the two-way interactions between BP and Faithfulness or BP and Intransitive, indicating that there is no evidence that BP uniformly modulates the preference for faithful template responses. A marginally significant three-way interaction between BP, Faithfulness, and Intransitive was observed for the intransitive niCCaC (p = .04), indicating that in this intransitive template the effect of BP on Faithfulness was negative, in the opposite direction to the hypothesis that participation in the stop-fricative alternation increases manner faithfulness. See Table 10 for the full model output.

3.1.1 Grammatical generalizations

I found that the intransitivity of niCCaC and hitCaCeC and the application of the stop-fricative alternation in the contexts tested in this study are known synchronically by speakers of Hebrew based on the finding that they can apply them to nonce verbs with minimal variability.

In Experiment 1, there were virtually no instances of transitive verbs in these templates. In Experiment 2 (where niCCaC was not tested), hitCaCeC was rated the lowest compared to all other transitive templates. The dispreference for niCCaC and hitCaCeC as transitives mirrors their absence as alternating transitives in the language. This finding is consistent with several theoretical accounts, as summarized in Section 3.3.2.

Similarly, speakers know how to apply the stop-fricative alternation to novel items at the first root consonant, the morphophonological environment tested in Experiment 1, as can be seen by the scarcity of misapplication errors. As an anonymous reviewer pointed out, previous studies have found variation in the application of the stop-fricative alternation, but crucially, the locus of variation has been mostly in post-consonantal positions, following certain consonants (e.g., yiTFoR ‘sew.3sg.m.fut’; Adam 2002, for spontaneous speech data and analysis, Albert 2014, for elicited production, Temkin-Martínez 2010, for acceptability judgment, Temkin-Martínez 2013, and Temkin-Martínez & Müllner 2016, for elicited production of verbs and nonce-verbs). In Experiment 1 of the current study, the potential locus of variation was the first root consonant, target transitive verbs were all in third person singular past tense, and hitCaCeC forms were not produced by participants; consequently, all relevant morphophonological environments were word-initial or post-vocalic. The lack of variation in the application of the stop-fricative alternation is therefore consistent with previous literature.

Speakers’ consistent and reliable application of these generalizations to nonwords in these domains suggests that the intransitivity of niCCaC and hitCaCeC (or at least their inability to be transitive alternants) and knowledge about the stop-fricative alternation in the tested environments are both encoded in the synchronic grammar and considered in real-time production and judgment.

3.1.2 Item-specific listing

Conversely, speakers are inconsistent in their choice of transitive template, and are permissive when asked to rate template combinations. They are not constrained by specific templates for the expression of causation, nor by strict pairings between templates, nor even by the pairings attested in the lexicon. I also found no evidence that they adhere to paradigm uniformity with respect to the stop-fricative alternation. This implies that speakers cannot reliably generate the appropriate template based on grammatical generalizations, and therefore the template must be listed for each specific root and for both alternants.

Two caveats are worth highlighting. First, the finding that template selection in alternations is mostly not based on grammatical generalizations does not imply that the relation between alternants is not grammatically determined. While there is a grammatical relation between abstract alternants at the thematic, syntactic, and semantic levels, their morphological realization is idiosyncratic and listed, except for the grammatical constraint on niCCaC and hitCaCeC transitives. Second, the present work focused specifically on template selection in transitivity alternations, and the conclusions drawn here are limited to this domain. The Hebrew verbal system is morphologically rich and encompasses a range of systematic relationships that were not investigated in the current experiments. Whether similar patterns emerge in nonce verb experiments targeting other types of verbs and alternations remains an open question for future experimental work.

With these clarifications in mind, the consistency of these findings across tasks and modalities (including an acceptability judgment task in Experiment 2 which minimizes the effect of performance factors) strengthens the reliability of the result as a reflection of the grammatical system underlying verbal template selection in Hebrew.

3.3.1 Relations between templates

The findings indicate that participants produce and accept a broader range of template combinations than those that are frequent or even attested in the Hebrew lexicon. Thus, any theory that restricts the set of possible grammatical template combinations or assumes strict grammatical relationships between members of certain template combinations and not others is too strong in light of the current data. Specifically, the layering approaches of Doron (2003) and Kastner (2020) treat the following template pairs as structurally related (sharing a core structure and differing only in higher layers): niCCaC-CaCaC, hitCaCeC-CiCeC, and CaCaC-hiCCiC. Kastner (2020) additionally takes niCCaC-hiCCiC to be structurally related, and Doron (2003) additionally takes CaCaC-CiCeC to be structurally related (in the intensive alternation). See Section 1.4.1 for the details.

This is too restrictive given that speakers produce and accept other combinations as well. In fact, speakers prefer niCCaC-CiCeC alternations over niCCaC-CaCaC or niCCaC-hiCCiC, and CaCaC-CiCeC alternations over CaCaC-hiCCiC in the causative alternation.

Even Arad’s (2005) far less restrictive conjugation classes are too strong. Speakers’ preferences are not limited to the six conjugation classes proposed by Arad, nor to pairs attested in Hebrew. For example, the speakers in the current study produce and give high ratings to CaCaC-CiCeC pairs (rare in Hebrew and not a conjugation class in Arad’s theory), and even hiCCiC-CiCeC pairs (unattested in Hebrew). The implication of these findings is that abstract relations between templates are not specified in the grammar, and therefore the two templates must be listed for each root, as proposed by Borer (2004).

3.3.2 Transitivity

Speakers know that niCCaC and hitCaCeC cannot be transitive. This aligns with Borer’s (2004) and Arad’s (2005) proposal that a verb’s ability to assign object marking is marked by the template, and that transitive verbs must be in CaCaC, CiCeC, or hiCCiC. This finding is also consistent with accounts in which niCCaC and hitCaCeC cannot project an external argument, since by Burzio’s (1986) generalization they would also be unable to assign accusative case. However, this is a stronger claim than is necessary to account for the current findings. In Kastner’s (2020) analysis, niCCaC and hitCaCeC are spellouts of [-D], indicating the absence of an external argument; in Doron (2003), the middle head μ blocks the external argument. In both accounts, the inability to assign accusative case follows from Burzio’s generalization. While these predictions are consistent with the findings of this study, only the claim regarding accusative case is necessary. The stronger claim about the absence of an external argument is not required, and indeed is not fully supported by the empirical evidence from the Hebrew lexicon.

There are verbs in niCCaC and hitCaCeC with an external argument: non-alternating verbs with PP objects, alternating reflexives, and simple agentive verbs in hitCaCeC. This claim is based on their failure to pass unaccusativity diagnostics, even when tested experimentally against unaccusatives in the same template (Plotnik et al. 2024). To reconcile this apparent contradiction, Kastner proposes that PP-verbs in niCCaC, which do not pass unaccusativity diagnostics, are figure reflexive. For this case, following Wood (2015), Kastner proposes that it is the p head rather than the v head that possesses the [–D] feature. This prohibits an external argument for the preposition while allowing an external argument for the verb, thereby enabling an agent to appear in the subject position. Both p[–D] and Voice[–D] are spelled out as niCCaC, and therefore formally there is no contradiction with the claim that Voice[–D] is unaccusative. For reflexives in hitCaCeC, which also do not pass unaccusativity diagnostics, Kastner argues against the diagnostics themselves, and therefore he claims that they cannot be used to determine the presence of an external argument in the syntax. However, experimental evidence suggests that unaccusativity diagnostics in Hebrew are relevant even when controlling for semantic and morphological factors (Plotnik et al. 2024). In addition, Ahdout (2021) notes that there are some clear instances of agentive verbs in hitCaCeC such as hitLONeN (‘complain’), hiŠtoLeL (‘act wildly’). These verbs are not reflexive, fail unaccusativity diagnostics, and pass agentivity diagnostics. Additionally, there are PP-verbs in niCCaC that Kastner categorizes as figure reflexives such as niLXaM (‘fight’), but it is not clear that semantically they have a self-directed motion component that defines figure reflexives. To sum up this argument, verbs in hitCaCeC and niCCaC indeed cannot have a direct object, and this generalization is part of speakers’ synchronic knowledge, as Borer (2004) proposed. The same prediction is derived by Doron (2003) and Kastner (2020), but in a version that is too strong: they suggest that verbal templates are a morphological realization of a “Burzio head” that marks both lack of external argument and inability to take a direct object, but hitCaCeC and niCCaC only mark the lack of a direct object, without prohibiting an external argument.

Since speakers generalize the intransitivity of niCCaC and hitCaCeC to nonce verbs in both production and judgment, it seems that their intransitivity is encoded in the grammar. This grammatical generalization seems to apply to a syntactic notion of objecthood. Verbs in niCCaC and hitCaCeC can select PP objects, i.e., they can be semantically transitive. They can also appear with a cognate object, even with the accusative case marker et, as demonstrated in (7)–(8):

The fact that verbs in hitCaCeC and niCCaC can be semantically transitive suggests that the generalization speakers make is not based on a simple ban against an internal argument. Moreover, the verbs’ ability to appear adjacent to an accusative-marked argument indicates that surface frequency alone is insufficient for learning the behavior of these templates. Rather, speakers must generalize over true instances of direct objects to acquire this rule.

3.3.3 Toward a formal analysis

The experiments were designed to adjudicate between claims about the transitivity of templates and claims about the relations between verbs in different templates based on existing theoretical accounts of the Hebrew verbal system. The findings constrain the hypothesis space for existing and future theories in both respects: theories must include a component that makes hitCaCeC and niCCaC verbs intransitive, and at the same time a list with specific root-template combinations. Multiple frameworks can be made compatible with these requirements, and committing to any single one would require taking a stance on questions that the current experiments are not designed to address; I therefore leave this choice for future research. In this section, I demonstrate how these findings can be derived in principle, focusing for concreteness on the theoretical accounts of the type presented in Section 1.4, which assume syntactic decomposition of the root and the functional elements that are realized as templates.⁸ I show that the model in Borer (2004) is largely compatible with the findings of the current paper, and propose a modification that makes it fully compatible. I demonstrate how the same idea can be implemented in the Distributed Morphology framework (Halle & Marantz 1993).

Borer’s (2004) model directly predicts the results of the current experiment. She proposes two components in the grammar: a syntactic feature responsible for (in)transitivity whose realization constrains template selection, and a post-syntactic list of root-template combinations. She explicitly argues that both are needed:

“While the syntactic structure would exclude some possible morphophonological binyanim from consideration, […] the presence or the absence of the [+OM] feature still leaves the final morphophonology rather underdetermined and subject to root-specific listed information.” (Borer 2004: 34)

This is largely in line with the present findings. The intransitivity of niCCaC and hitCaCeC is categorical and productive: speakers treat it as a grammatical constraint, consistent with the role of the [+OM] feature, which can be realized as CiCeC, hiCCiC, or CaCaC (the latter unspecified for [OM]). Within each set (transitive and intransitive), the choice of template in nonwords shows preferences rather than categorical behavior and does not track pair-specific generalizations, consistent with Borer’s claim that, for real words, the specific template is a matter of root-specific listed information. The one area where the findings diverge from Borer’s proposal is her claim that [–OM] is realized by niCCaC, hitCaCeC, or unspecified CaCaC, whereas in the lexicon (and by speakers’ judgment in Experiment 2) there are also CiCeC and hiCCiC intransitives. This could be resolved by removing the morphophonological restrictions on [–OM], making the realization of intransitive verbs unrestricted by template, and leaving only the restriction on the [+OM] feature.

Arad’s (2005) proposal that niCCaC and hitCaCeC are marked as [–tr] or [–ACC case], which confines their environment of insertion to intransitive verbs, is similar to Borer’s modified account, and is consistent with the findings about (in)transitivity. However, as discussed in Section 3.3.1, Arad’s theory about the relation between alternating templates, namely that alternations are constrained to six template combinations, is too restrictive to account for the experimental findings.

The general idea of a restriction on transitivity together with full listing of templates can also be implemented in the Distributed Morphology framework. For the syntactic representation of alternating verbs I assume, following Arad (2005), that the relation between alternants is attributed to the flavor of the verbalizing head v above the shared root (v_tr and v_intr), which mark the verb’s transitivity, replacing Borer’s [OM] feature. As Arad (2005) notes (footnote 42, p. 277), this kind of syntactic account, presented here for simplicity, does not treat transitivity alternants as derivationally related on the syntactic level, but as independently derived verbs with a shared root. Arad points out that the same morphological system could in principle be recast in configurational terms. However, recasting the findings in configurational terms is non-trivial: if the template is not the realization of Voice, it is unclear what the syntactic relation between the alternants is, which heads are relevant for spelling out the template, and whether they are local enough to the root for it to define the environment for template selection. I leave these questions open.

The realization of the verbalizer is determined by lexical insertion rules in List 2 (the Vocabulary), where there is conditional allomorphy between possible templates, conditioned by the root. Crucially, hitCaCeC and niCCaC are not in the list of allomorphs for v_tr, which directly captures their inability to act as transitive alternants without further stipulations. A sketch of the lexical insertion rules is presented in (9).

I emphasize that this sketch is intended as a proof of concept rather than a full analysis. The details of the syntactic representation and of the Vocabulary Items remain to be worked out in future research. For instance, roots that share consonants but are related only abstractly in meaning (Arad’s 2005 Multiple Contextual Meaning roots) are not distinguished by (9) when the two verbs share a transitivity value (e.g. P-N-H in PiNaH ‘clear’ and hiFNaH ‘direct’). This could be resolved by treating them as distinct roots in the syntax, or by a difference in their functional context other than transitivity (such as properties of the other arguments). Such factors may influence template selection more generally but were not tested in the present experiments, and whether they affect it synchronically remains a question for future work.