Experiencer troubles: A reappraisal of the predicate-based asymmetry in child passives

ABSTRACT

Children’s understanding of passives of certain mental state predicates appears to lag behind passives of so-called actional predicates, an asymmetry that has posed a major empirical challenge for theories of passive acquisition. This paper argues against the dominant view in the literature that treats the predicate-based asymmetry as theoretically irrelevant. We instead propose a novel account that locates the problem in the syntax of experiencer constructions. Synthesizing theoretical and developmental evidence, we build a case for an early misanalysis of transitive subject-experiencer constructions as unaccusatives – structures that, by design, cannot passivize.

1. Introduction

The passive is one of the best studied constructions within research on syntactic development. Since seminal work in the 70s (Bever 1970, de Villiers & de Villiers 1973) that found in children poor comprehension of passive sentences relative to active counterparts, researchers from widely varying theoretical orientations have pursued the question of what children do and do not know about this construction (see Deen 2011 for overview). Despite this wealth of empirical research, there is to date no consensus as to whether passive syntax is delayed in acquisition, or till what age.

The earliest work found that young children misinterpret passive sentences, often assigning the role of agent to the NP in subject position. Subsequent work in the 80s (e.g., Maratsos et al. 1985) revealed that some passives are harder for children than others, depending on properties of the verb from which the passive participle is derived. English-acquiring children understand verbal passives involving an “actional” verb such as hug by around age 4-5, yet they do not display adult-level performance with verbal passives involving a “non-actional” verb such as love. Non-actional passives seem to come in only around or after age 6. The finding has been replicated in a variety of studies, using diverse experimental methodologies (e.g., Sudhalter & Braine 1985, Gordon & Chafetz 1990, Hirsch & Wexler 2006; see Nguyen & Pearl 2021 for a meta-analysis), making the predicate-based asymmetry one of the most well-documented aspects of passive acquisition.

What lies at the heart of this asymmetry? Is it a genuine one—that is, reflective of a stage in the child’s language development—or is it an artifact of experimental or other factors? Researchers have approached the problem from both angles, but to date no consensus has emerged. Some treat the asymmetry as genuine, but their accounts of it fail to go beyond stipulating lexical restrictions on child passives. In the other camp are those who assume that the asymmetry is a mere artifact. But here again, researchers disagree over what this artifact is masking: competence or the lack thereof.

In this paper, we offer a novel account of the predicate-based asymmetry. The empirical basis of our proposal comes from a recent meta-analysis by Nguyen & Pearl (2021), which helps localize the problem to passives of subject-experiencer predicates. We then argue that the predicate-based asymmetry should be explained in terms of difficulties with the syntax of experiencers, rather than in terms of restrictions on passivization. More specifically, we propose that children initially hypothesize an unaccusative syntax for subject experiencer constructions, which renders them unpassivizable.

We begin, in Section 2, by giving a characterization of the predicate-based asymmetry. In the second half of Section 2 as well as in Section 3, we critically review previously offered approaches to this asymmetry. We discuss accounts that stipulate lexical-semantic restrictions on children’s passives and why they fail to reach a level of explanatory adequacy needed to advance our understanding of the phenomenon. We then turn to various accounts of passive acquisition that do not take the asymmetry to be modulated by grammatical capacity, and point out empirical and theoretical shortcomings of each. In Section 4, we present the details of our proposal and spell out specific predictions that may be submitted to direct experimental testing in future work. Section 5 discusses a different account that also takes the asymmetry seriously, but proposes a semantic solution to the problem. While the account as it stands does not make the right predictions, we suggest a reinterpretation of their core insight that integrates well with our own proposal.

2. The actional/non-actional asymmetry

The asymmetry in children’s comprehension between ACTIONAL passives (i.e., passives with verbs that denote a (physical) action such as kick) and NON-ACTIONAL passives (i.e., passives with verbs that denote mental or perceptual states, such as love or see) was first observed in Maratsos et al. (1985), and has since been one of the most stable findings in the literature on passive acquisition. In brief, children’s comprehension of non-actional passives tends to lag behind their comprehension of actional passives by 2-3 years.

Maratsos et al. (1985) noted that there is considerable variation, both within and across languages, with respect to the types of predicates that allow passivization. In English, all transitive verbs that (in the active) have an agentive subject and a patientive direct object allow passivization. But for the class of transitives that do not follow this template, there are some that passivize (e.g., love) and some that do not (e.g., lack). Other languages may be more (e.g., Russian) or less (e.g., Gaelic) restrictive than English. Given the cross-linguistic variation and seemingly idiosyncratic restrictions within a single language, it makes sense to ask whether the learner starts out: (i) by analyzing passivization as a general rule that applies to all subject-object pairs, only later identifying the language-specific restrictions, or (ii) by being conservative, initially semantically restricting their passives to those that are near-universally passivizable (e.g., agent-patient verbs) and only later extending it to other passivizable verbs in the language.

Maratsos et al. (1985) used two experiments to test these possibilities. In their first experiment, children were shown some characters and then heard a sentence which was followed by the question who did it?. The sentence presented to the children included either an actional verb, or a non-actional verb, and was either in passive or active voice. The 4- and 5-year-old participants performed significantly better on actional than non-actional passives, with scores of 67 and 40%, respectively. Both of these scores are lower than scores on their active counterparts (91 and 88%, respectively), but the crucial result is the distinction between actional versus non-actional passives. The second experiment used a picture-selection task in which children were presented with two pictures and were asked to point to the picture that best matched the sentence. Again, children had trouble with passives derived from non-actional verbs, and these problems persisted even in 7-year-olds. This actional/non-actional asymmetry has since been replicated in various studies in English (e.g., Sudhalter & Braine 1985, Gordon & Chafetz 1990, Hirsch & Wexler 2006) and has been found to hold in other languages as well (e.g., Dutch [Verrips 1996], Japanese [Sano et al. 2001], and Russian [Babyonyshev & Brun 2004]).

These experimental results led Maratsos et al. (1985) to conclude that children’s passives are initially semantically constrained; children only permit passives of predicates of a certain semantic type, namely predicates of actional verbs (physical action verbs in their words). This happens to be the set of verbs that passivizes in every language that has a passive construction (Keenan & Dryer 2007).¹ This account then takes the actional/non-actional distinction to reflect a genuine constraint in acquisition. More specifically, it ascribes children’s poor performance on non-actional (or mental/perceptual state) passives to a semantic restriction on passivization. Maratsos et al. (1985), following Hopper & Thompson (1980), define this restriction in terms of a notion of semantic transitivity, a structural-semantic gradient composed of properties of the verb and its arguments such as kinesis, agency, and affectedness of the object. Actional verbs such as kick are on the high end of the scale, as they involve an action and an agent, whereas a verb such as love is low on the transitivity scale, as it is stative, not an action, and has no agent or affected object. Despite having passive syntax, the account goes, English-acquiring children fail on certain types of passives because they have yet to extend the set of passivizable verbs to those lower on the semantic transitivity scale.

Despite the intuitive appeal of notions such as semantic transitivity, it is difficult to use this gradient to define, in a non-circular manner, natural classes of verbs that show uniform syntactic or semantic behavior. The absence of precise characterizations of the relevant semantic features has given rise to two problems, one methodological and one conceptual. First, there is no consensus across acquisition studies as to which verbs count as “actional” or “non-actional”, which makes a comparison difficult.² Second, it remains ill-understood what makes certain semantic features grammatically relevant. Why is it that certain sets of semantic features, but not others, modulate the syntactic property of “being passivizable”?³ Likewise, how does a learner figure out what features to look for when defining the necessary semantic restriction?

The first problem of terminological confusion is evident in the acquisition work that followed. A study by Demuth, Moloi and Machobane (2010), for instance, reports the absence of an actional/non-actional asymmetry in children’s passives in Sesotho. Their set of non-actional verbs, however, includes verbs such as help, look for, expel, and leave behind, which have markedly different syntactic and semantic properties from Maratsos et al.’s (1985) original set. For starters, they involve an agentive subject, unlike verbs such as love or remember whose subjects are experiencers. We can see this in the contrast between look for and love across two common tests of agentivity: (i) the acceptability of modifying an event with adverbs indicating volitionality in (1) (see, e.g., Roeper 1987, Talmy 1976), and (ii) the acceptability of imperative formation with the verb (2) (see discussion in Dik 1978, Lehmann 1993, K. Klein & Kutscher 2002).⁴

Table

A recent meta-analysis by Nguyen & Pearl (2021) helps resolve the terminological confusion. Their aim was to make precise which of the many lexical-semantic features that could characterize ‘non-actional’ predicates actually pose difficulties for children. They used a set of descriptive semantic notions (e.g., actional, stative, volitional, affected, agency) to annotate the verbs tested in the acquisition literature. This provided a set of five semantic profiles (in terms of combinations of these semantic features) for the verbs tested in the literature. In addition, they identified from the literature the age at which children succeed on each verb, which allowed them to estimate the age of acquisition for each of their semantic profiles. Using this method, they found that what sets apart the two latest-acquired profiles from all others is the +SUBJ-EXPERIENCER feature. That is, the latest-acquired verbs involve predicates whose subjects (in the active) are experiencers, such as love and remember. Recall that in the original Maratsos et al. 1985 study, the class of ‘non-actional’ items involved precisely such verbs. Unlike subject-experiencer verbs, non-actional predicates such as annoy or frighten, which involve experiencer objects, do not exhibit an acquisition delay.

Nguyen & Pearl’s (2021) own proposal is that passive-acquisition is modulated by specific conglomerates of semantic features. In a sense, then, their account is a restatement of Maratsos et al.’s (1985) hypothesis that children’s passives are semantically constrained. In fact, their semantic properties are similar to the semantic properties that make up a predicate’s transitivity as defined by Hopper & Thompson (1980). Their profile 1 verbs, which are acquired earliest, are agent-patient verbs that denote volitional actions with an affected object—that is, verbs that are high on Hopper & Thompson’s (1980) semantic transitivity scale. The crucial novelty is that they do not just find a binary opposition between physical action verbs and non-actional or psych verbs, but rather demonstrate a relationship between children’s performance and finer-grained semantic profiles.

At the same time, the semantic features invoked by Nguyen & Pearl (2021) fare no better than those in Maratsos et al. (1985) in helping us pick out classes of predicates that behave in a uniform way syntactically or semantically, distinct from other classes of predicates. In fact, it is not even clear to what extent some of the features are semantic in the first place. The proposed +SUBJ-EXPERIENCER feature, for instance, is an amalgam of a syntactic notion of subjecthood and a semantic notion of being an experiencer. What this means is that neither account—Maratsos et al.’s (1985) original one or Nguyen & Pearl’s refinement—helps us understand why these semantic features matter for child passives and not others. The predicates hug and love differ along as many dimensions as make hug and love distinct concepts, but why are syntactic processes sensitive to features such as affectedness, but not, for instance, physical contact (certainly an important cognitive concept)? For Nguyen & Pearl’s lexical-semantic profiles, it is further unexplained what unifies the particular semantic composites beyond the explanandum itself. There does not seem to be anything motivating fusing some of these features except the fact that children have difficulties with passive participles of verbs that have those features. What then is the nature of the semantic restriction in the language-acquiring child? How do children define the relevant semantic profiles, and as such, the set of passivizable verbs and how do they eventually go beyond the initial set? Even if we grant that semantic features such as actionality and affectedness are special (perhaps the child is predisposed to look for them), what evidence does the child need to extend her passives beyond the prototypical case?

Despite this criticism, we think that Nguyen & Pearl (2021) makes a crucial contribution in helping us arrive at a more precise descriptive characterization of the asymmetry: the delay is specific to subject-experiencer verbs, and this, we think, is an important clue towards the nature of the problem in acquisition as we will spell out in section 4. But before we do so, we consider accounts of passive acquisition that take the asymmetry between actional and non-actional verbs to be artifactual. We first discuss accounts that take performance on non-actional verbs to be the true indicator of children’s competence, and illustrate how they account for children’s good performance on passives of actional verbs. We will then turn to accounts that take performance on actional verbs to be the true indicator of children’s competence and explain how they account for children’s poor performance on passives of non-actional verbs. While these proposals don’t face the sort of conceptual problems that beset the semantic-restriction accounts, we will discuss other empirical and theoretical problems that arise in relation to the predicate-based asymmetry.

3. Explaining away the asymmetry

Unlike the proposals previously discussed, most previous accounts of passive acquisition take either the non-actionals or the actionals as the true indicator of competence and attempt to explain away the other. In the following sections, we discuss representative accounts that fall into each camp.

3.1. Performance on non-actionals reflect late maturation of passive syntax

Theories that take children’s performance on non-actional passives to reflect their true competence need to say something about what aspect of the passive poses difficulties for the language-acquiring child and how they move from a state of not having the grammatical knowledge to represent passives to a subsequent state in which they do. Most relevantly for our purposes, these accounts need to answer the question why children perform relatively well on passives of actional verbs, well before demonstrating good performance on non-actional passives.

The two accounts we consider are Wexler’s (2004) Universal Phase Requirement (UPR) and Orfitelli’s (2012) Argument Intervention Hypothesis (AIH), both of which locate the problem in the syntax of passives. On both accounts, the problem relates to the subject nominal of passive verbs occurring in a non-canonical position. Child grammar is restricted in such a way that movement of the subject of a passive verb from its original (object) position is blocked (for different reasons in the two theories).⁵ Furthermore, both accounts hypothesize that the capacity to represent passive syntax is in part driven by aspects of brain development that are not input-contingent. That is, children need to mature in order for the relevant syntactic means to become available, which does not happen before the ages of 6 or 7. Only when the required syntactic means are in place is the child able to perform well on passives.

Yet, children only under-perform on non-actional passives. They succeed on passives of actional verbs well before age 6, a fact that demands an explanation. Both UPR and AIH explain children’s success on actional passives in terms of the child relying on an alternative strategy in interpretation. That is, when confronted with an input string that would be analyzed by adults using a structure that is not available to the child, they analyze the input string using an alternative structure that is available. In the case of passives, this alternative structure is the adjectival passive, which is often surface-identical to verbal passives in English.⁶ Consider, for instance, (3a), which is ambiguous as to whether it represents a verbal passive (as in [3b]) or an adjectival one (as in [3c]).

Table

Crucially, adjectival passives do not involve movement of the internal argument to subject position; rather, the surface subject is projected as an external argument. This structure, then, is predicted to be unproblematic for both syntactic accounts. Importantly, this strategy applies less successfully with non-actional passives, which are not well-formed in an adjectival structure (Borer & Wexler 1987). Contrast, for instance, the two examples in (4). Whereas the actional verb root break in (4a) makes a perfectly good adjectival passive participle, the verb root hear in (4b) is somewhat odd.

Table

As a result, on both UPR and AIH, children’s comprehension of non-actional passives reveals their genuine inability to construct the relevant structure, whereas actional passive comprehension involves deducing and applying the adjectival strategy, which derives a roughly adult-like interpretation albeit through distinct syntactic means.

There are intrinsic conceptual disadvantages to explaining away, rather than explaining, a set of behavior, but this two-pronged approach faces certain empirical problems, as well, as the adjectival strategy does not straightforwardly account for the asymmetry. The main problem is that the inapplicability of the adjectival passive parse is not restricted to the relevant non-actional verbs. Activity predicates such as kick (5) also make for bad adjectival passives (Hirsch & Wexler 2006). One proposal about the oddness of (5a/b) comes from Kratzer (2000) (further developed in Anagnostopoulou 2003 and Embick 2004), who postulates a principled limitation on productive adjectival passive formation: roots that do not imply a caused state of the internal argument do not support adjectival passive interpretations. Thus, broken, which implies that the present state of the object was caused by the breaking event, makes for a good adjectival passive. In contrast, kick, which does not imply any such new state, does not. Crucially and problematically, children perform well on passive sentences such as (5c), which, on this account, should not have a suitable adjectival parse.

Table

Hirsch & Wexler (2006), which builds on Wexler (2004), recognize this problem, and make the empirical prediction that performance should vary per verb, depending on how good the corresponding adjectival passive of that verb is. They predict that children should not perform well on passives derived from actional verbs such as hold, as they do not imply a (resulting) target state. Furthermore, there are non-actional verbs, such as remember, that can imply a target state and these accounts make the additional prediction that children should perform relatively well on passives of such verbs. This empirical prediction, however, is not borne out. In fact, it is quite the opposite: held is among the earliest-acquired passives, whereas remembered is among the latest-acquired ones (Nguyen & Pearl 2021).

Thus, accounts explaining away the predicate-based asymmetry in terms of the adjectival strategy do not succeed. The adjectival strategy fails to make the right empirical cut.⁷ It incorrectly predicts poor performance on some actional passives, while at the same time incorrectly predicting good performance on some non-actional passives. Given that the adjectival strategy does not distinguish between actional and non-actional passives in the right way, we are back to the question we started out with: why do children perform better on actional than non-actional passives? Next, we turn to those accounts that take children’s success on actional passives to demonstrate command of passive syntax in children. As we will see, such accounts face their own set of problems.

3.2. Performance on non-actionals reflects extra-grammatical factors

Theories that take children’s adultlike performance on actional passives to reflect their overall passive competence also have to say something about the verb-based asymmetry. In contrast to explaining good performance on actional passives, however, these accounts have to explain why passives of certain non-actional verbs pose difficulties for the child. Here we discuss three kinds of explanations in the literature (though they are not incompatible with each other). One kind places the blame on the asymmetric depictability of non-actionals compared to actionals. Another argues that the problem has to do with pragmatic factors that again place an asymmetric burden on non-actionals. A third type of account suggests that a parsing bias to treat the first-encountered DP as an agent impedes comprehension. In the following section, we review representative work that falls into each of these categories.

Messenger et al. (2012) argue that picture-matching tasks, where children must process the sentence and distinguish between two very similar pictures, may pose asymmetric difficulties when it comes to subject-experiencer predicates. They write: “In concrete terms, it is relatively easy to distinguish a picture in which a penguin is hitting from a picture in which it is being hit; however, it may be much more difficult to distinguish a picture in which a penguin is seeing (matching an experiencer-theme verb passive such as “the pirate was seen by the penguin”) from a picture in which it is not seeing (but rather is being seen)” (Messenger et al. 2012:571).

If, indeed, adult interpretation of non-actional passive sentences was impeded by depictability issues that did not equally affect the actional passive sentences, methods that rely less on picture-based comprehension should reveal better performance. Messenger et al. (2012), therefore, use a structural priming method to probe children’s knowledge of passive syntax. Their logic is as follows: if passives of non-actional verbs successfully prime passives of actionals, children must have given the non-actional prime sentence a syntactic parse that can appropriately apply also to the actional passive. Across two experiments, they found similar rates of priming across verb types, with no effects or interactions of verb-type. The authors take these findings to corroborate their hypothesis that the verb-based asymmetry is an experimental artifact.

While Messenger et al. (2012) suggest a plausible alternative explanation for children’s failures with non-actional passives in picture-selection tasks, they have a minimal amount to offer that throws light on why the verb-based asymmetry should have emerged in Truth-Value Judgment Tasks (TVJT) (Gordon & Chafetz 1990, Fox & Grodzinsky 1998, Deen et al. 2018, Nguyen & Pearl 2021). This remains a big disincentive to accepting their broader conclusion. Furthermore, their own methods are not beyond criticism. At present, we cannot be sure that the priming effect was due to surface similarities between the subject-experiencer passive prime and the actional passives produced, like the presence of a by-phrase. There is in fact evidence that intransitive, active locatives (e.g., the foreigner was loitering by the broken traffic light) prime passives (Bock & Loebell 1990). What is missing then in Messenger et al.’s (2012) experiment is a crucial control, demonstrating lack of priming by partial structures in children, for instance by showing that non-passive sentences such as in (6), with a nominal-modifying by-phrase, fail to prime actional passives.

Table

A second issue has to do with the subject-experiencer prime sentences used in the experiment. Messenger et al.’s (2012) prime sentences were all in the progressive aspect, for example, (7a-b). However, a well-known property of subject-experiencer predicates such as love is that they are stative and thus incompatible with the progressive (see [8]).⁸ The sentence in (7b) would then be infelicitous, unless the mismatch between the verbal aktionsart and aspect is somehow repaired. One way to do so is to coerce an eventive reading for love, that is, treat it as denoting a loving-activity (e.g., in Messenger et al.’s [2012] stimuli, an act of hugging), of which a sheep is the agent. This entails that the passive in (7b) is not derived from a subject experiencer construction, but rather from an activity predicate which projects a transitive structure with an agent argument.⁹,¹⁰

Table

Thus, while the methodological issues raised by Messenger et al. (2012) are reasonable, it is not clear to us that their own methods adequately demonstrate competence with these structures.

The second type of full competence account evokes pragmatic factors to explain away the asymmetry. Crain, Thornton & Murasugi (2009) and Crain & Fodor (1993) observed that long passives are pragmatically marked and appropriate only when the by-phrase is used to distinguish between one or more potential alternatives. Building on these observations, O’Brien, Grolla & Lillo-Martin (2006) argued that children’s poor performance on passives may have resulted from the use of experimental materials that did not meet the relevant felicity conditions. In their own study, they hypothesized that providing contexts that properly motivate the use of the by-phrase should lead to greater comprehension, even on non-actional verbs. Indeed, when three- and four-year-old children were presented with stories that made salient two potential agents/experiencers in a TVJT, they found that the actional/non-actional asymmetry was less pronounced. In their Experiment 1, for instance, 4-year-olds’ accuracy rates were 93% on actional long passives versus 83% on non-actional long passives.

Conversely, attempts to replicate O’Brien, Grolla & Lillo-Martin’s (2006) results have subsequently failed (Nguyen & Snyder 2017, Deen et al. 2018). Additionally, even if their findings were reliable, what remains underdeveloped on this pragmatic account is why this use condition on the passive by-phrase—which crucially applies to all passives—has an amplified effect on non-actional passives. Put differently, even if a more felicitous experimental set-up improves performance, there still seems to be a verb-based asymmetry: infelicity only seems to be problematic in the comprehension of non-actional passives.¹¹

A different explanation is offered by Y.T. Huang et al. (2013), who propose that difficulties with passives stem from a bias in incremental processing. The idea—a variation on an old idea proposed in Bever (1970)—is that children, and perhaps adults as well, first predict a canonical mapping between thematic roles and nominals; in other words, they hypothesize that the first nominal they encounter is an agent. Crucially, children are not able to revise this initial incorrect parse even as they get further evidence that the sentence they are hearing is a passive. Y.T. Huang et al.’s (2013) evidence comes from Chinese bei-passives, which differ in important ways from English verbal passives (see C.T.J. Huang 1999 for overview), but the parsing account has been endorsed for English by Deen et al. (2018). Deen et al. (2018) reasoned that if children’s difficulty with passives has to do with their prior expectations about thematic alignment, simply repeating the non-canonical sentence should give them a second chance to revise their expectations. They found that English 4-year-old children showed greater success on non-actional passives if the target sentence was repeated.¹²

The explanatory inadequacy of such processing accounts is discussed at length in Hirsch & Hartman (2006). We direct the reader to this work for details, while simply summarizing the thrust of their main argument. If non-canonical word orders lead to an insurmountable processing problem for children, we should find analogous difficulties in other structures involving movement, such as object wh-questions. This prediction is not borne out, for actionals or non-actionals. More generally, what holds for the pragmatic accounts holds for the processing accounts as well: why do these processing factors unevenly affect non-actional passives?

Thus, the various attempts that explain children’s poor performance on non-actional passives as stemming from extra-grammatical sources also fall short. Accounts that locate the issue in the lower imageability and thus comprehensibility of non-actionals fail to explain children’s difficulties with the same type of predicates in tasks that do not rely on picture selection. Pragmatic and processing accounts simply replace one question—why do children show asymmetric difficulties comprehending non-actional passives?—with another—why are children asymmetrically affected by pragmatic and processing factors with non-actional passives?

The state of the art, then, is one where existing proposals in both camps—those that take the asymmetry seriously and those that explain it away—are limited in empirical coverage or predictive power or both. This motivates the present proposal. We share the ambition with the lexical-semantic restriction accounts of accounting for the asymmetry in terms of the child’s linguistic development. However, like the maturational accounts, we maintain that the problem is syntactic in nature.

4. Proposal: A problem with syntax of subject-experiencer constructions

Due to Nguyen & Pearl’s (2021) meta-analysis, we have in hand a generalization about the predicate-based asymmetry in passive acquisition: children’s difficulties lie with passives of verbs with experiencer subjects (e.g., love, remember). Predicates with agentive or cause subjects, even when ‘non-actional’ in a general sense (e.g., scare, anger), are acquired earlier. Our proposal builds on this generalization. We propose that at the point where children reliably succeed on actional passives, they have command of passive syntax, but they may not have command of the syntax of experiencer constructions.¹³ In particular, we suggest that children initially misanalyze English subject-experiencer constructions—which project the experiencer as an external argument—as constructions involving an internally generated experiencer. As we describe in more detail in the following sections, experiencer verbs show variation across languages in the syntactic realization of the experiencer argument, such that expressions denoting identical concepts may appear with the experiencer projected externally (love in English) or internally (q’var ‘love’ in Georgian). Crucially, the latter configurations tend to be unaccusative structures, which systematically lack verbal passives (cf. *the child is being fallen). If children initially assign an unaccusative syntax to English subject experiencer constructions, their unpassivizability directly follows.

We first spell out our theoretical assumptions about passive syntax. We assume a syntactic analysis on which the defining feature of verbal passives is existential closure of the external argument position (as opposed to saturation by a syntactically realized nominal, as in the active). This approach naturally captures the absence cross-linguistically of verbal passives of unaccusatives. We then turn to experiencer constructions and their typology, building on the tripartite classification proposed originally in Belletti & Rizzi (1998). Figuring out where the experiencer argument originates has been a major theoretical challenge for linguists, and we argue that it poses a challenge also to the learner. Our main suggestion is that learners may not initially get it right and their misanalysis manifests in the passivization problems previously described. We consider further consequences of such an early misanalysis and discuss the kinds of evidence that may help the child move toward the adult grammar.

4.1. Core assumptions about passive syntax

Following work in the tradition of Kratzer (1996), we assume that a functional head in the verbal projection, Voice, plays a crucial role in the active-passive alternation. This Voice head is responsible for the realization of the external argument (we use the cover term “initiator” here to encompass various external argument roles, including agents, causes and experiencers). Kratzerian Voice comes in different varieties, but most important for our purposes are Active and Passive Voice. Both active and passive Voice heads semantically introduce an initiator theta-role and relate it to the event variable provided by the verb. However, only the active Voice supports the syntactic realization of a nominal that saturates this argument slot, by selecting for a DP to be merged in its specifier. In contrast, the initiator argument slot is existentially closed in verbal passives.¹⁴ The passive Voice is further distinguished from unaccusative Voice, which lacks an initiator both syntactically and semantically; the latter thus has no semantic content and could be considered a sort of expletive Voice (Schäfer 2008).

On this perspective, the defining characteristic of a passive is that it suppresses the syntactic projection of an external argument, and facilitates existential closure of the relevant argument slot. All other properties, whether or not the internal argument is obligatorily promoted, whether or not the structure can appear with a by-phrase, whether or not unergatives can participate in the structure, are language-specific parameters. This view on passives helps explain an empirical generalization that plays a crucial role in our proposal, namely that across languages, unaccusatives do not passivize (Perlmutter & Postal 1984). Because the semantics of passive Voice makes explicit reference to an external argument role, it cannot appear with predicates that do not project such arguments.

Suppose that the inventory of Voice heads is available to the child, for instance, as part of her innate endowment. If so, a passive-acquiring child’s learning task is reduced to identifying the relevant morphology and the various idiosyncratic syntactic properties of her language. We are not suggesting that this makes the task trivial or quick. The English-acquiring child is confronted with a range of complications: ambiguous participial morphology (perfect vs. passive), syntactic homophones (e.g., adjectival passives), embedding verbal element (be vs. get) all make learning difficult. Moreover, in English, forming a passive still requires the syntactic means to promote the internal argument to subject position, which could take time to develop. However, given children’s success on passives with a variety of verb classes around ages 4-5, we believe that the syntax of verbal passives is in place by that stage, including knowledge of English-specific morphosyntactic properties. Unlike the syntax of passivization, however, children may have continued difficulties with the syntax of subject-experiencer constructions, which in turn has consequences for their analysis of passives of such predicates. In the following section, we provide initial motivation for this idea by outlining the complexities surrounding experiencer constructions.

4.2. A cline of experiencer-predicates

Experiencer predicates—predicates that express psychological or perceptual states—pose well-known problems for the “linking” of syntactic positions to semantic or thematic roles. As implied by the nomenclature, these verbs typically involve an EXPERIENCER, the individual experiencing the mental state. Additionally, there is often a THEME argument, representing the content, cause or object of that mental state.¹⁵ Cross-linguistically and language-internally, psych verbs map these two roles onto different syntactic positions. Belletti & Rizzi (1998), and subsequent work, distinguish three main classes of experiencer-predicates, given in (9) along with examples from a range of languages:¹⁶

Table

Class 1 predicates behave like regular transitives: the experiencer subject receives Nominative case, they readily passivize, and take the transitive have-auxiliary in languages that have auxiliary selection (Belletti & Rizzi 1998). Class 2 and Class 3 verbs, however, are associated with special syntactic properties. Across languages, Class 2 and Class 3 experiencer constructions tend to display unusual control and binding patterns, quirky case marking, restrictions on extraction, among other properties—so-called ‘psych effects’, which have been the subject of study for decades (see, e.g., Lakoff 1970, Kuno 1971, Pesetsky 1987, Belletti & Rizzi 1998, Kim & Larson 1989, Grimshaw 1990, Bouchard 1995, Pesetsky 1996, Reinhart 2016, McGinnis 2001, Landau 2010). A particularly elegant proposal by Landau (2010) derives the unusual grammatical behavior of these predicates from the underlying locative syntax of experiencer arguments. Building on earlier intuitions (Kuno 1971, Speas 1990) that experiencers are associated with mental locations, Landau (2010) argues that Class 2/3 experiencer arguments are generated within locative Prepositional Phrases (PPs) and assigned oblique case by the prepositional head, which can sometimes be null (e.g. in the Class 2 examples in [9]).

Class 2 and Class 3 verbs can be further distinguished on the basis of their transitivity. Class 2 verbs come in two types: eventive and stative. In English, Class 2 predicates such as depress are most natural as statives, as shown by their incompatibility with the progressive in (10a), but others such as scare in (10b) are compatible with an eventive reading and are quite acceptable in the progressive.

Table

Eventive Class 2 verbs are transitives, projecting an external argument, typically the cause of the mental state (and an internal experiencer argument). Stative Class 2 verbs and all Class 3 verbs, however, are unaccusative, wherein both the experiencer and the theme are base-generated internally, similarly to the internal arguments of double object constructions. These predicates pass classic unaccusativity diagnostics, including, crucially, the inability to passivize (Belletti & Rizzi 1998, Pesetsky 1996, Landau 2010).

Table

But there is language-level and predicate-level variation in which of the internal arguments surface as the sentential subject. In some languages, the experiencer may move overtly to the subject position, resulting in the famous so-called ‘quirky’ subject constructions that contain a subject that bears inherent, non-nominative case instead of canonical nominative case typically assigned to subjects. The subject Gianni in (13a) for instance, as well as Raman in (13c) bear non-canonical oblique rather than nominative case.

Table

This option is restricted or unavailable altogether in some languages. For instance, whereas Icelandic, Faroese, and Greek allow both ACC and DAT experiencers to front, Italian, Spanish, and Dutch allow fronting only of dative experiencers. Finally, there are languages such as English and French that do not license non-nominative subjects at all.

In a number of cases, the availability of quirky subjects leads to ‘dual’ word orders, with the experiencer sometimes being realized as subject, but sometimes remaining in the base-generated internal argument position. Thus, the Italian (13a) alternates with (14), with the theme argument (questo ‘this’) surfacing as the subject. Notice that such permutability is not available for Class 1 predicates in Italian, whose experiencer argument is externally generated (15a-b).

Table

Given that across languages, both Class 1 experiencers (externally-generated) and Class 2/3 experiencers (internally-generated) can surface as subjects, a division of the class of psych-predicates into “subject-experiencer” predicates and “object experiencer” predicates is too coarse and often fails to slice the pie in an illuminating way. Henceforth, we will use the Class 1/2/3 terminology whenever appropriate.

4.3. The learning challenge for the English-acquiring child

There is a general learning problem presented by experiencer predicates with the linking of argument roles and syntactic positions: the same argument roles (e.g., EXPERIENCER) can seemingly be merged in different syntactic positions with different classes of experiencer predicates. Belletti & Rizzi (1998), however, suggests that there is still systematicity: across the three classes in (9), the experiencer is uniformly merged above the theme, with the divergent surface orders being the result of movement. Variants of this idea have also been adopted by Pesetsky (1996) and Landau (2010). If this assumption is on the right track, then the task of the learner is simplified, as they would not, they do not need to identify item-by-item what the linking rules are.

If we accept this, then the main challenge a learner faces lies in distinguishing between transitive (Class 1) and unaccusative psych-predicates (Class 3 and stative Class 2 verbs). The crucial difference between the two lies in whether the experiencer is realized as an internal or external argument, and, indeed, whether there is any external argument at all. How might one determine which class a predicate will fall into? Principled semantic distinctions are hard to draw. Verbs denoting near-identical concepts are mapped varyingly to Class 1, with the experiencer as an external argument, or Class 2/3, with the experiencer as an internal argument. Even more strikingly, sometimes one finds Class 1 and Class 3 alternations for the same verbs within the same language, with no obvious meaning change. We can observe this alternation in Faroese in (16), where the verb like appears in a transitive (Class 1) structure where the (experiencer) subject receives nominative case in (16a), yet in an unaccusative (Class 3) structure with a fronted dative subject in (16b).

Table

Given that it is not (always) possible to predict from meaning which configuration a predicate will appear in, how does the child figure out whether a given verb is associated with an internally or externally generated experiencer? Depending on the language, there may be morphosyntactic evidence that can guide the learner. For instance, in languages such as Faroese, the case morphology on the subject NP helps tell apart whether an experiencer subject is an internal (oblique case) or external (nominative case) argument. Other languages, such as Italian, distinguish unaccusative and transitive experiencer predicates in auxiliary selection (essere for the former and avere for the latter [Belletti & Rizzi 1998, Pesetsky 1996]). Word order evidence might also help: as demonstrated in (12)-(13), Italian distinguishes the two classes in the availability of word-order alternations, with Class 3, but not Class 1, permitting permutability of the two arguments (Belletti & Rizzi 1998).

No such transparent evidence exists in a language such as English, however. English fails to make distinctions in auxiliary selection. The language does mark case, but the distinctions are only visible in the pronominal paradigm and even there, there is rampant syncretism. Perhaps testament to its ‘poor’ case morphology, English-acquiring children take quite a while to figure out the case system of the language (Rispoli, 1998, 2005, Landau & Thornton 2011, Aravind 2022; cf. Schütze & Wexler 1996). As a result, even if case-morphology could in principle be informative, it is not clear that it would be informative to the young child who might still be trying to learn the ropes of that system.¹⁷

We believe that the task of distinguishing transitive versus unaccusative experiencer predicates is a cumbersome one in languages such as English, which lacks systematic and overt cues to the structure. We propose that a child lacking such cues initially defaults to an unaccusative structure, with an internally generated experiencer. That is, the child starts out from the simplest assumption: all experiencers are internal arguments.

4.4. Testing the predictions

If the child assigns an unaccusative structure to subject-experiencer predicates, those predicates are not passivizable as a result. While our account can thus account for the observed difficulty with passivization, it is important to consider what independent evidence could bear on its validity. For instance, there is evidence from both comprehension and production that subject-experiencer predicates pose difficulties for English-acquiring children. A study by Tinker, Beckwith & Dougherty (1989) showed that 4-year-old children systematically misinterpreted sentences such as Bugs Bunny feared Gumby as meaning Bugs Bunny frightened Gumby. That is, instead of assigning sentences with subject-experiencer verbs a transitive structure in which the subject is the (external) experiencer (Class 1), children seemed to assign a transitive structure in which the subject is the (externally-generated) cause and the object is the experiencer (eventive Class 2), that is, a structure with an internally generated experiencer argument. More recent studies replicate this finding. Using a modified TVJT, Hartshorne, Pogue & Snedeker (2015) found that 4-year-olds were at-chance when evaluating sentences involving Class 1 subject-experiencer verbs such as fear, hate, love, and like when both arguments were animate (e.g., The lion feared the monkey). In contrast, they were systematically above-chance with Class 2 verbs such as frighten, scare, and surprise. While Hartshorne, Pogue & Snedeker’s (2015) explanation differs from our own—they cite “privileged links” between causal events and transitive structures that predict easier acquisition of frighten-type verbs—their results accord with our hypothesis that children struggle with the syntax of subject-experiencer verbs.¹⁸ In particular, the chance behavior receives an explanation if children treat these constructions as reversible— the subject can either be the (fronted) experiencer, or the theme argument—as is often the case with Class 3 and unaccusative Class 2 predicates. If so, word order alone is not sufficient for the child to identify the experiencer. Indeed, when animacy cues could help identify the experiencer (e.g., The giraffe loves the beach ball), children’s performance improved to above-chance rates.

This interpretation of the comprehension findings gains support from the fact that children also spontaneously produce reversed structures. A diary study by Bowerman (1990) had found production errors of the sort in (17), where subject-experiencer predicates appear in configurations with a clearly internally generated experiencer argument. No errors of overapplying subject-experiencer syntax to object-experiencer predicates were found.

Table

While these observations are in line with our hypothesis, they might not provide conclusive enough evidence that children, as a rule, take experiencers to be internally-generated. After all, the observed difficulties could be the result of item-specific problems. We have to look elsewhere for more direct evidence. In fact, our hypothesis predicts specific non-adult-like behaviors that could be submitted to future empirical testing.

One such prediction that follows from our hypothesis is that children will over-accept adult-ungrammatical word orders with subject-experiencer predicates, while failing to show comparable permissiveness with other transitives. As mentioned, English does not show word order reversibility with Class 3 verbs that we see in a language like Italian, yet there are structures where both arguments of unaccusative experiencer predicates can remain in-situ, that is, in internal argument position (Cheung and Larson 2014). More specifically, when the theme argument of an unaccusative Class 2 or Class 3 predicate is clausal, both arguments (theme and experiencer) can remain in their internally generated position and an expletive can appear as the sentential subject, as in (18).

Table

This is not possible when the experiencer is an external argument, i.e., with subjects of Class 1 (subject-experiencer) verbs as in (19a),¹⁹ or agentive transitives as in (19b).

Table

If a child treats verbs such as love as unaccusatives, however, those verbs should pattern with unaccusative Class 2 and 3 verbs, which means children should find sentences like (19a) as acceptable as those in (18) (while still rejecting sentences like (19b)).

A more subtle test would involve patterns of verbal-compound formation in English. Grimshaw (1990) first observed that the internal (theme) argument of Class 1 predicates can occur as the incorporated object of verbal compounds such as in (20a-b). The same does not hold for the internal experiencer argument of Class 2 or 3 predicates. The examples in (20c-d) show that the experiencer argument (worker and student resp.) cannot be incorporated with the verb.

Table

Landau (2010), building on Baker (1997), argues that the restriction follows from the fact that (i) prepositions are disallowed in these compounds and (ii) all internally generated experiencers are headed by a preposition, overt or null. If this generalization is correct, and children take the experiencer argument of English subject-experiencer predicates as internally generated, they should find compounds such as (20a-b) ill-formed, much like (20c-d). Correlations between passivization difficulties and compound-formation difficulties with Class 1 predicates could thus also provide corroborating evidence for our hypothesis.

The empirical tests we suggest rely on collecting acceptability judgments from children. Acceptability judgments, while straightforwardly and reliably collected from adults (e.g., Sprouse, Schütze & Almeida 2013), have been proven to be more difficult to obtain from child participants before the age of 6 (Thornton 2021). Given that the difficulties with passivization of subject experiencer verbs seem to continue in 6-year-olds, and in some tasks even older children, acceptability judgments on expletive constructions and verbal compounds could potentially provide the right empirical support for our hypothesis.²⁰

A more promising avenue that does not rely on acceptability judgments is to indirectly test the underlying structure by looking at online processing signatures. Friedmann et al. (2008), using the Cross Modal Priming Paradigm (CMPP), for example, found that unaccusative predicates give rise to a reactivation of their NP-argument sometime after the offset of the verb. Koring, Mak & Reuland (2012) and Koring et al. (2018) replicated and extended these findings using the Visual World Paradigm. The distinctive processing signature of unaccusative sentences gives us a way of probing whether children’s subject-experiencer constructions—in contrast to those of adults—are underlyingly unaccusative.

Finally, divergences in acquisition trajectories cross-linguistically could also bear on our proposal. As discussed in section 4.3, one reason English children might have an especially difficult time identifying the syntax of subject-experiencer predicates is that disambiguating evidence is hard to come by. But the situation may be less dire in other languages. For starters, there are many languages that have both Class 1 and Class 3 experiencer predicates, with reliable morphological and syntactic cues distinguishing the two. In principle, the learning task should be easier for children acquiring such languages. There is suggestive evidence pointing in this direction. Volpato, Verin & Cardinaletti (2016) tested 3-to-5-year-olds’ knowledge of Italian actional and subject-experiencer passives, and found that children successfully produced both kinds in an elicited production task. The force of this line of argument is mitigated by the small number of cross-linguistic studies that test the verb-type asymmetry and the fact that some of the existing studies do seem to replicate the asymmetry. Lower performance on subject-experiencer passives have been found in Spanish (Oliva & Wexler 2018) and European Portuguese (Agostinho 2020), two languages that do distinguish Class 1 and Class 3 experiencer predicates. A direct comparison between these studies cannot be made because whereas the Italian-speaking children were tested in an elicited production task, the Spanish- and European-Portuguese speaking children performed a comprehension task. Still, one reason we might find a genuine difference in performance between Italian- versus Spanish- or European Portuguese-speaking children is a difference between the languages in the availability of surface cues for unaccusativity. Whereas Italian distinguishes between the verb types in terms of auxiliary selection (have vs. be), Spanish and European Portuguese do not. That is, there might be more surface evidence for unaccusativity available in Italian as compared to the other two languages. It is also worth noting that there were striking item effects in Agostinho’s (2020) study on European Portuguese. Two non-actional verbs were tested—ver ‘see’ and ouvir ‘hear’—but the verb-type asymmetry seems to be largely driven by poor performance on the latter. By age 5, there were no longer any significant differences between either of the two actional verbs tested (pentear ‘comb’ and pintar ‘paint’) and the perception verb ver. Perhaps item effects are exactly what one would expect in this situation: suppose the child’s initial hypothesis is that all experiencers are merged internally. Then, given the absence of any semantic cues that single out externally generated experiencers, the child would need to alter this hypothesis on an item-by-item basis for each subject experiencer verb. As such, we might expect that subject-experiencer verbs that are more frequent (in active voice) come in earlier. For these reasons, we hesitate to draw any firm conclusions for or against our proposal on the basis of cross-linguistic data at this point. A more detailed analysis of the languages to generate predictions, as well as more focused cross-linguistic empirical studies are necessary to state and test the predictions of our account.

4.5. Becoming adult-like

We have argued that English-acquiring children initially misanalyze Class 1 predicates as unaccusatives. We argued that this misanalysis stems from difficulties determining whether the experiencer argument is generated internally or externally: there are no obvious semantic correlates, and in a language such as English, syntactic evidence is limited. Still, eventually the English learner manages to retreat from their wrong hypothesis. How do they accomplish this and eventually become adult-like? We suggest that syntactic evidence, albeit sparse, might provide the necessary evidence over time to help the child eschew the unaccusative hypothesis.

One clear piece of evidence, of course, is the passive itself. Unaccusative structures do not passivize, so hearing a verb such as love in the passive should serve as evidence that the child’s assumption about the argument-structure of such predicates is wrong-headed. Though passivizability is strong evidence against the unaccusative hypothesis, the relevant data might be hard to come by in the input. Passives of subject-experiencer predicates are very rare. Gordon & Chafetz (1990) presented a study of the types of passives in children’s input and found that 93% of verbal passives were actional passives. The remaining 7% consists of both object and subject experiencer verbs (non-actional verbs), so only a subset of those are presumably subject experiencer verbs. Nguyen & Pearl (2018) state that subject-experiencer passives are virtually non-existent in the input, finding only 2 instances (both with the verb love) out of 1636 uses of subject-experiencer verbs. The sparsity of such crucial evidence might explain the late acquisition of subject experiencer predicates. A handful of cases of subject-experiencer passives might be sufficient to rid the child’s grammar of the unaccusative analysis, provided the child analyzes the structure as such, but the child might have to wait quite a while to encounter one.²¹

A related piece of evidence could come from deverbal adjective formation with the morpheme -able: Class 1 experiencer-predicates form good bases for -able attachment (21a), whereas unaccusative Class 2/3 do not (21b) (Trips & Stein 2008, Alexiadou 2018).

Table

Many have observed a tight relationship between adjective formation with -able and passivization (Chapin 1967, Horn 1980, Williams 1981, Kayne 1984, Fabb 1984, Roeper & van Hout 1999, Oltra-Massuet 2013, Alexiadou 2018). The generalization, across languages, seems to be that those verbs that can feed passivization can serve as the base for -able adjectivization (*arrivable, *happenable vs. huggable, pullable). Oltra-Massuet (2013), building on earlier work (Chapin 1967, Kayne 1984, Williams 1981, Fabb 1984, Roeper & van Hout 1999), accounts for this link by taking the internal structure of -able-adjective to contain the passive. That is, the adjectivizer takes as its input a structure that contains a passive Voice head, which in turn implies an external argument slot (see, also, Alexiadou 2018). On this account, the examples in (20b) are ill-formed for the same reason the associated verbs cannot passivize: they are tied to an unaccusative structure without an external argument. A child who knows the restrictions on -able-adjectivization might hear expressions such as lovable and forgettable and use them to rule out the unaccusative hypothesis. At the same time, the evidence from -able adjective formation is even more limited than that from passivization. They are subject to further, often subtle, constraints (e.g., affectedness of object), and certain passivizable subject-experiencer verbs do not make good bases for -able (e.g., *seeable).²²

Thus, surface evidence ruling out an unaccusative analysis of subject-experiencer predicates exists, but is scarce. Beyond their low frequency, another issue with passives and -able-adjectivization is that the evidence may not be of practical help to a child whose grammar is constrained in a way that does not permit such constructions. In order for a subject experiencer passive to count as evidence against an unaccusative analysis, the child needs to analyze that passive as such. But can they do this? Many passive acquisition experiments indicate that children presented with subject experiencer passives seem to (incorrectly) analyze them as active sentences (i.e., Mommy is loved by the baby → Mommy loved the baby). If the child does the same in the wild—that is, if the child simply reanalyzes subject experiencer passives (not permitted by their grammar) as actives (permitted by their grammar, but not informative about transitivity)—input sentences involving the passive would not constitute evidence against an unaccusative analysis.

A more systematic shift may be triggered by a better understanding of the case system of English, in particular the recognition that obliques never surface as subjects. If, following Landau (2010), internally generated experiencers are uniformly assigned oblique case by a (null or overt) preposition, and only as-yet caseless nominals can move to subject position in English, then it follows that an experiencer argument in subject position cannot in fact be internally generated. Once the child knows that obliques never surface as subjects in English then, every active sentence with a subject experiencer verb can function as evidence that these verbs project a transitive (Class 1) structure rather than an unaccusative one - extending the domain of positive evidence.²³

We have offered a few suggestions of types of evidence the child might rely on in their journey toward reaching the adult grammar. We think, however, that before establishing the sufficient and necessary learning evidence, we have to pin down the nature of the learning problem by providing empirical support for the present account (see section 4.4).

5. An alternative account: An eventivity restriction?

The central claim in this paper is that the verb-based asymmetry in passivization should be explained not so much in terms of restrictions on passivization, but instead in terms of a misrepresentation of the syntax of subject experiencer verbs. At the point the child has full command of passive syntax, they still fail to passivize subject experiencer verbs because those have unaccusative syntax in their grammars, rendering them unpassivizable. That the asymmetry shows up in passive comprehension is thus epiphenomenal. Our account, like the lexical-semantic accounts discussed in section 2, takes the predicate-based asymmetry seriously rather than explaining it away. But like the maturational accounts discussed in section 3.1, we locate the problem in the syntactic derivation of certain constructions (albeit different ones from those accounts). Before concluding the paper, we contrast our proposal with a final set of accounts that also aims at providing a grammatical explanation for the verb-based asymmetry, but differs in the nature and scope of the explanation (Grillo 2008, Gehrke & Grillo 2009, Snyder & Hyams 2015). These proposals hypothesize that the derivation of non-actional passives requires an additional grammatical operation that is either not available to children until the age of 6 or 7 or for which the child does not have sufficient processing resources. Different from other grammatical accounts, and closer in spirit to our own, they do not take children’s grammar, after age 4-5, to differ from adults in terms of the availability of passive syntax itself.

These accounts start from a novel account of passive syntax/semantics (Gehrke & Grillo 2009), which treats passivization as an operation on event structure rather than argument structure. Functionally, passivation involves the “zooming in” on a consequent state subevent of a complex event. This is possible only for eventive predicates, such as achievements and accomplishments, whose internal argument undergoes a change of state and as a result becomes the bearer of a consequent state. Such predicates are taken to contain rich enough verbal structure to represent this transition. Specifically, a secondary verbal projection headed by a (silent) BECOME-operator occurs above the VP constituent denoting the consequent state. In the passive, the lower VP, which contains the internal argument and denotes the consequent state, is singled out and undergoes movement to the specifier of VoiceP.

Stative verbs (e.g., love) are simplex and lack the secondary verbal projection associated with BECOME. As such, Gehrke & Grillo’s (2009) analysis renders stative subject experiencer verbs unpassivizable. Yet, in a language such as English, such verbs do passivize. Gehrke & Grillo (2009) account for this by assuming that passivization in such cases is possible by shifting the state into an event (i.e., an achievement or accomplishment). The structure has to be enriched by adding a BECOME predicate and the associated verbal structure in order to enable passivization.

Gehrke & Grillo (2009) suggest that this additional step of semantically coercing the state into an achievement or accomplishment might exhaust children’s processing capacity, such that they fail to compute a passive derivation of stative verbs. In fact, they propose this to account for data from agrammatics who show the same actional/non-actional asymmetry in their comprehension of passives (Grodzinsky 1995). The idea is that the passive derivation, where movement of the relevant VP layer is triggered by a discourse feature, is burdensome in itself. Adding in the shifts required to enable passivization of stative verbs depletes the resources completely, causing the observed actional/non-actional asymmetry. This account, they note, can be extended to children to explain the same observed asymmetry. Snyder & Hyams (2015) adopt Gehrke & Grillo’s (2009) proposal wholesale, modulo certain technical adjustments,²⁴ and propose that the crucial ability to coerce a stative predicate to an eventive one is a maturational change the child goes through in development around age 6.²⁵

This account has several advantages over those already discussed. It aims to give a grammatical explanation for the predicate-based asymmetry and does so in a manner that makes testable predictions. It also helps make novel and interesting connections to the linguistic behavior of atypical populations. However, the account makes certain wrong predictions for both child and adult grammar. Eventivity, while relevant as we argue in the following section, is not the main determinant of children’s problems with non-actional passives. Children continue to fail, past age 4, on subject-experiencer predicates such as forget, an achievement and thus eventive (Nguyen & Pearl 2021). In adult language, stative passives, including passives of stative experiencer predicates, retain their stativity in their passive form. This is contrary to what is predicted on Gehrke & Grillo’s (2009) theory, on which the only way to passivize a state is to coerce an eventive structure. In general, verbal passive formation has a minimal effect on a predicate’s stativity or eventivity, and, therefore, we are reluctant to endorse this theory.

We do, however, think that the insight that eventivity plays a role in children’s ability to passivize experiencer verbs is important and integrates well with our own account. Recall that Class 2 experiencer verbs come in two types, eventive and stative, and display a correlation between stativity and unaccusativity (e.g., Pesetsky 1996, Tenny 1998, Pylkkänen 2000b, Landau 2010). That is, stative Class 2 verbs are unaccusative, whereas eventive Class 2 verbs are transitive (i.e., project an external argument). This is not accidental. Eventivity in these predicates is a direct result of their transitivity. Eventive Class 2 experiencer predicates are causative: they involve a verbalizer head, which both introduces a causal event, and projects a CAUSE external argument.²⁶

We suggest that the same stativity correlation holds for children’s experiencer verbs (including their misanalyzed subject experiencer verbs). Stative experiencer verbs are unaccusative and therefore do not passivize; eventive experiencer verbs project an external cause argument and do passivize. There are several pieces to knowing this correlation. The child has to know: (i) that causal events involve a causal verbalizing head bringing in eventivity, (ii) that cause arguments are systematically merged externally, and (iii) that a verbal passive implicates a Voice head associated with an external argument slot (see Section 4.1). Existing empirical evidence suggests that these pieces are in place by age 4-5. Children at that age succeed on passives of eventive Class 2 verbs such as frighten or scare (e.g., Nguyen & Pearl 2021).

Thus, eventivity does play an explanatory role, but in a different way than suggested. For adults, stative subject experiencer verbs passivize (without coercion), as the experiencer is an external argument. For children, however, stative subject experiencer verbs, having been misanalyzed as unaccusative, do not passivize.²⁷ For both populations, passive formation is permitted with experiencer verbs if the non-experiencer argument can be interpreted as the causer argument. The constraint on children’s passives then, is the same as in adults: passive formation is only possible if there is an external argument slot that can be existentially closed. For children, external arguments are restricted to agents and causes, whereas for adults, some experiencer arguments project externally. Children thus differ from adults in assigning unaccusative syntax to all experiencer verbs that cannot be construed as having an external agent or cause argument, including subject experiencer verbs. This reduces the set of verbs that can be passivized in child language as compared to adult language.

6. Conclusion

This paper has been concerned with a well-attested asymmetry in the acquisition of passives: children have difficulties with passives of certain verbs at a stage when they succeed with others. Theories of the phenomenon crucially differ in whether they opt to explain away this asymmetry or take it as reflecting a theoretically important distinction. Within the former camp, there are accounts that try to minimize the significance of children’s success and others that do so with children’s failures. We pointed out problems with both types of approaches. We believe that theories that can explain the asymmetry have an inherent advantage over the ones that explain it away. At the same time, theories on the market that do take the asymmetry seriously have their own share of problems, either making incorrect predictions or relying on stipulative lexical-semantic restrictions. Our proposal aims to explain the asymmetry by locating the problem in children’s initial hypotheses about the syntax of subject-experiencer predicates. We argued that an early misanalysis of transitive subject-experiencer verbs as unaccusatives results in their non-passivizability as a corollary. This proposal accounts for the existing data and makes specific predictions for future empirical work.

Disclosure Statement

No potential conflict of interest was reported by the authors.

Notes

¹ Keenan & Dryer (2007) call short passives derived from transitive, actional verbs ‘basic passives’, as those types of passives occur most frequently across languages. According to their Generalization 2, “[i]f a language has any passives it has ones characterized as basic … ; moreover, it may only have basic passives” (329).

² Note that Maratsos et al.’s (1985) original definition of mental state verbs was quite specific: they define mental state passives as passives where “the underlying subject and object are always experiencer and stimulus, respectively” (170), that is, subject experiencer verbs. The notion of “actional verbs”, which has been adopted in the acquisition literature, is much less specific.

³ One argument that passivization is indeed a syntactic operation is that passives can be formed of Exceptional Case Marking (ECM) verbs, which do not select the nominal that surfaces in their object position. For example, in the passive sentence in (i), the subject was never a semantic argument of the passivized verb expect, but rather, the external argument of win. Therefore, it seems difficult to define passivizability solely in terms of semantic properties.

(i) Dana was expected to win.

⁴ Note that a positive result in this test does not necessarily imply that the verb in question is interpreted as agentive in all contexts, but that it can occur in a context that supports the agentive interpretation.

⁵ Technical details and theoretical assumptions differ between the two accounts, but we will not go into the details here, as they are not relevant for our purposes. We are concerned with accounting for the predicate-based asymmetry, and, as such, focus on how the syntactic theories account for this phenomenon. For the same reasons, we also do not discuss other challenges faced by these accounts.

⁶ This idea has its genesis in Borer & Wexler 1987.

⁷ That the adjectival strategy does not seem to account for the predicate-based asymmetry does not mean that there is no earlier stage where UPR or AIH hold in which children have command of adjectival passives exclusively—and use it to understand verbal passives for which they lack the syntactic means to analyze them, with relative success. There is evidence that children start out with adjectival (stative) passives only (e.g., Horgan 1978, Israel, Johnson & Brooks 2000, Gavarró and Parramon 2017, Oliva & Wexler 2018)—our claim is that it fails to explain the predicate-based asymmetry.

⁸ One reason for this is that the progressive aspect makes reference to the run-time of an event (e.g., it specifies that the event-time is contained within the topic time or of reference [W. Klein 1994]), but states do not have a measurable run-time—they hold indefinitely.

⁹ Coercion as a repair mechanism is not tied to subject-experiencers, but is generally available to language users to rescue otherwise uninterpretable structures. There are various ways of implementing the operation, including via semantic type-shifters (de Swart 1998, Reinhart 2001), syntactic reanalysis (Pylkkänen 2000b), or both (Gehrke 2012). What is crucial to our point is that such a process is necessary for the interpretability of sentences like (7b), and that it results in a shift in the verbal aktionsart and, in turn, argument-structure (e.g., activity-predicates have agents).

¹⁰ Note also that, generally, core properties of experiencer constructions do not persist in agentive contexts (Landau 2010).

¹¹ Another pragmatic account is recently put forth by Liter & Lidz (2021). They argue that the discourse-structural properties of passives in neutral contexts conflict with the requirements of experiencer constructions. In particular, they claim that the argument of by is at-issue or focused in long-passives, whereas it is the experienced state—and not the experiencer—that is at-issue in experiencer constructions. The placement of an experiencer in a passive by-phrase is consequently odd. They argue, furthermore, that the oddness can be overcome if the experiencer is quantificational, as in “Mary is loved by everyone”. They test this idea in a TVJT, comparing non-actional actives and passives with referential and quantificational antecedents, and find across-the-board above-chance accuracy rates (roughly 70%). We are skeptical about the basic theoretical premise, as at-issueness is not a lexical-semantic property nor a general property of constructions. To wit, in a sentence with narrow focus on the subject, as in (i) following, clearly what is at issue is who, among a set of alternative candidates, was heard by Lisa. Liter & Lidz’s (2021) arguments go through only on the assumption that across experimental settings, the target sentences were necessarily pronounced with default accenting, where the nuclear pitch accent tends to fall on the rightmost constituent inside the VP (see Büring 2016). It is not clear that this assumption is warranted with most TVJT scenarios, where, crucially, the target sentences are not presented out-of-the-blue and the context is not neutral.

(i) HOMER_F was heard by Lisa.

Even if we put these theoretical issues aside, we hesitate to take Liter &Lidz’s (2021) findings as providing evidence in favor of a pragmatic explanation of the verb-based asymmetry, because their results are incomplete in crucial ways. First, there was no comparison with actional predicates, so we simply cannot tell whether children would have displayed a verb-based asymmetry. Second, the critical interaction effect that would lend support to their hypothesis (i.e., are children significantly more accurate on non-actional passives with quantificational by-phrases?) was never directly tested.

¹² It should be noted that in addition to two subject-experiencer predicates (understand, remember), Deen et al.’s (2018) non-actional set included object-experiencer verbs such as surprise and anger, which according to Nguyen & Pearl’s (2021) findings are largely unproblematic for children in the first place.

¹³ Our account is compatible with but not committed to a two-stage development such that passives of subject-experiencer verbs come in late, and all other passives come in earlier. There may be difficulties with other passives for other reasons. In this case, passive syntax may be even delayed till age 4 or 5, and earlier success on any passive participles might in fact be the result of applying an alternative (adjectival) strategy, leading to relatively good performance on passives of some verbs, but not others. Our focus, as has been the focus of the acquisition literature, is on the predicate-based asymmetry.

¹⁴ Or, the argument slot may remain open to be accessed by a by-phrase adjunct (see Bruening 2013, Legate 2014, Legate et al. 2020 for implementation).

¹⁵ Note that this is a rather broad definition of theme that includes a causer argument. The difference between themes and causers becomes more important in section 5, and we assume, following Pesetsky (1996), among others, that transitive Class 2 verbs project an external cause argument.

¹⁶ This is a slight simplification, as Class 2 verbs come in different types as we will explain shortly.

¹⁷ In spite of lack of surface evidence, children do distinguish intransitive unaccusatives and unergatives at a young age; therefore, one might wonder why distinguishing transitive versus unaccusative experiencer predicates should pose further issues. We think the difference is that there are principled distinctions between unaccusative and unergative predicates, whereas these seem to be absent for experiencer predicates. One such difference is the implied (conceptual) presence of a causer with unaccusatives (Reinhart 2016). Another cue to the unaccusative/unergative distinction that is perhaps utilized by the learner is animacy: the agent subjects of unergatives tend to be inanimate (Becker 2015, Becker & Schaeffer 2013). Notice that such cues are unhelpful for experiencer predicates. Experiencers are invariantly animate—only animate entities can experience perceptual or mental states.

¹⁸ One way to tease apart their hypothesis and ours is to compare children’s performance on predicates such as depress (e.g., The political situation depressed Mary), which have an internally generated experiencer, but unaccusative (i.e., non-causal) syntax. Hartshorne, Pogue & Snedeker (2015) should predict that children miscomprehend such structures, whereas our proposal makes no such prediction.

¹⁹ As mentioned earlier, Class 2 experiencers can be ambiguous between transitive (with a cause argument) and unaccusative variants (Pesetsky 1996, Landau 2010, Cheung & Larson 2014). The expletive construction with verbs such as frighten disambiguates to the unaccusative.

²⁰ As with any one experiment, if children do not provide the predicted judgments, this does not necessarily mean the hypothesis should be rejected. There could be many reasons why children would not perform the way we predict in such experiments. One of the reasons is that the expletive constructions we predict to be acceptable for the child do not appear in the input, as they are adult-ungrammatical, which might be reason in itself for the child to reject such sentences. Furthermore, there is disagreement as to what the source of the set of judgments in (16) is (cf., Grimshaw 1990, Landau 2010), and, moreover, speakers seem to differ in the extent they judge certain compounds to be (un)acceptable, which might interfere with the proposed experiment.

²¹ That is, encountering a subject experiencer passive may count as evidence that subject experiencer verbs project a transitive rather than unaccusative structure. This does not necessarily mean that the child immediately knows how to passivize such structures; hence, we do not necessarily predict a correlation between subject experiencer passive frequency and timing of acquisition.

²² Furthermore, the relevant evidence may also be quite rare in the input. A preliminary search of the Brown corpus turned up 0 instances of subject-experiencer -able adjectives.

²³ Of course, this only works if the child knows that the subject is in fact the experiencer—it would not rule out an unaccusative Class 2 analysis in which the subject argument is the theme—and this is precisely what the children seem to be doing in the (active) comprehension experiments (Hartshorne, Pogue & Snedeker 2015) as well as their own production data (Bowerman 1990). We are assuming here that the positive evidence is a sentence-meaning pair and so the child knows the subject is the experiencer (different from the comprehension experiments where it’s the child’s task to select a meaning among different options).

²⁴ They suggest that the effect of coercion does not involve adding an extra verbal layer hosting a BECOME operator, but rather, to make an existing, ‘stative’ verbalizer head ‘eventive’.

²⁵ In fact, Snyder & Hyams (2015) posit two distinct maturational shifts: one that enables children’s passives of actional verbs, and a second one that enables passives of non-actional verbs. The first maturational change shares certain core assumptions with the grammatical accounts discussed in section 3, and in particular with Orfitelli’s Argument Intervention Hypothesis (2012). Snyder & Hyams (2105) crucially, however, do not explain away the predicate-based asymmetry but rather propose another maturational shift.

²⁶ See Pylkkänen 1999, 2008, Marantz 2001, Schäfer 2008, Deal 2009, Harley 2013, Legate 2014, and others for arguments that a causal verbalizing head is responsible for bringing in eventivity. Cross-linguistically, the head that introduces causal semantics and the one that projects the external argument can come apart, but in English these two properties are packaged together (e.g., Pylkkänen 2008).

²⁷ An eventive reading may be coerced, in sufficiently supportive contexts. This might account for children’s relative success on progressive psych-passives such as in (10b) from Messenger et al. (2012), repeated here in (i). As we pointed out in section 3.2, this sentence can only be interpreted as a loving-activity of which the sheep is the agent.

(i) A girl is being loved by a sheep.

Note, however, that agentivity might in itself account for this: agentive psych-verbs universally permit verbal passives and do not display so-called psych-effects (Landau 2010).