We begin with a brief introduction to Pangasinan clausal syntax. Pangasinan exhibits a so-called Philippine-type voice system, where one nominal argument is chosen to be the “pivot” in every clause and verbal morphology reflects this choice of pivot (Wolff 1973; Himmelmann 2002; 2005; Reid & Liao 2004; Ross 2002; Erlewine & Levin & Van Urk 2017). In the examples below, the pivot (in bold) is the agent Pedro in (7a), the theme ‘banana’ in (7b), and the instrument ‘his hand(s)’ in (7c). Following Philippinist tradition, we refer to these verb forms in (7a,b) as Actor Voice (AV) and Patient Voice (PV). There are also forms that take certain peripheral arguments as their pivots, such as the Instrumental Voice (IV) in (7c).

Nominal arguments may be introduced by case-marking articles, as in (7) above. Following Kroeger (1991) and many others, we refer to the case on pivots here as nominative, with non-pivot arguments bearing genitive case (that of nominal possessors) or oblique case.5

In examples such as (7), the case-marking articles, together with the choice of voice morphology, uniquely identify the semantic role of each postverbal argument.7 Word order cues are not necessary, and indeed postverbal word order in such clauses are free, as seen in the PV clause in (9):

Although both word orders are acceptable for the intended meaning in (9), the “agent < theme” word order in (9) is less marked. As a reviewer notes, this preference is common across Philippine languages (see e.g. Kroeger 1991: 111 on Tagalog). Despite this preference, it is important to note that speakers readily accept both orders in examples such as (9), in sharp contrast to examples where word order freezing applies.8

There are also situations where two arguments will bear the same case, such as the agent and non-specific theme in the Benefactive Voice (BV) clause in (10) which are both genitive. We observe that the relative order of the two genitive arguments cannot be changed; see (10b). This same word order restriction has been observed in a number of Philippine languages. For instance, the frozen word order of two genitive arguments is reported in Tagalog for arguments of psych verbs (Billings 2005: 307–308), in recent perfective aspect clauses (Kroeger 1991: 111), and in certain nominalizations (Shibatani 1988: 101–102).

We discuss word order freezing in Pangasinan in greater detail in the next section. Although most of our examples here will be in Patient Voice (PV), the overall generalization that postverbal word order is free up to freezing also extends to other voices for all examples that we have tested.

Pangasinan also allows for specific, third-person arguments to be introduced by a demonstrative article instead of a case-marking article and doubled in most cases by a corresponding pronoun. Here we adopt the term “apposition” for this construction from Benton (1971: 145,154). The examples in (11) show that a non-pivot agent — introduced by a genitive article in (9) above — can instead be introduced by a demonstrative article and obligatorily doubled by a genitive pronoun. The postverbal word order remains free.

We give the inventory of demonstrative articles in (12). Demonstrative stems encode a three-way distinction between near speaker ya, near addressee tan, and distal or unmarked man (Fernandez Yaptenco 1967: 68; Benton 1971: 47, 78).9 We encode these distinctions using two binary features for near-author [±dem.auth] and near-participants [±dem.part].10 When used as a prenominal article, they appear with the linker -y and the stem-final nasal is dropped: e.g. man + -y > may. In addition, demonstrative articles inflect for number, for which we adopt the binary singular feature [±sg] of Noyer (1992) and Harbour (2003; 2008).

Pronouns for non-pivot agents (genitive) and for pivots (nominative) are second-position clitics, as is common in Philippine languages (Reid & Liao 2004; Kaufman 2010; Erlewine & Levin 2021), and therefore encliticize to the verb in (11). We give the inventory of these clitic pronouns in (13).

Apposition can also target pivots, which are in nominative case, but with minor differences in the use of the pronoun. First, no corresponding pronoun appears with singular pivots, as in (14a). This naturally follows from a general gap in the pronominal inventory of the language (13); see also discussion in Erlewine & Lim & Branan (2022). Second, for plural pivots, the appearance of the third-plural nominative clitic pronoun =ira is optional, as in (14b).

Note that demonstrative articles in Pangasinan do not encode case distinctions (Amurrio 1970: 46; McFarland 2008: 131–133). The demonstrative articles introducing the agents in (11) above and the theme pivots in (14) are therefore the same forms. When apposition applies simultaneously to both the agent and the pivot, then, there will be two postverbal arguments introduced by demonstratives instead of case-marking articles. We turn to such examples in the next section.

We now turn to the precise description of word order freezing effects in Pangasinan. All of our examples here will be PV clauses where both the pivot theme and non-pivot agent are introduced by demonstrative articles, with the agent doubled by a genitive clitic pronoun and the theme optionally doubled by a nominative clitic pronoun if available (see (13)). Such structures reflect the straightforward simultaneous application of agent apposition as in (11) and pivot apposition as in (14) above.

Consider example (15) below, which elaborates on example (2) from the introduction. Without context, all speakers judge the sentence as unambiguous, with the first demonstrative-marked nominal (‘woman’) as the agent and the second demonstrative-marked nominal (‘man’) as the pivot theme. We discuss the role of context and other manipulations below.

If instead the two postverbal phrases ‘woman’ and ‘man’ in (15) are reversed, the result unambiguously expresses that ‘The man saw the woman.’ In other words, the relative word order of the two postverbal arguments is frozen here, with “agent < theme.” Note that the requirement reflected in (15) is not for the agent to immediately follow the verb, nor for the pivot to immediately follow the agent. The linear position of demonstrative-marked arguments is still free with respect to other arguments and adjuncts, as seen here with ‘at school.’ Characterizing free word order as due to an operation of scrambling, as we do, we analyze the frozen “agent < theme” order as reflecting the base order. Scrambling of the theme over the agent fails to apply in this example, as the two are featurally identical, introduced by identical demonstrative articles.

However, as previewed in the introduction, if the two demonstrative-marked phrases differ in their number specifications, their relative order becomes free:

Word order flexibility is however not simply a special property of plural nominals. The word order of two plural demonstrative-marked arguments is again frozen, leading again to an obligatory “agent < theme” interpretation for (17):

The data so far suggests that the order of agent and theme is free as long as their articles are distinguishable in some way, as in (16), but otherwise limited to “agent < theme” order interpretations. However, unlike number contrasts, a difference in the deictic features of demonstratives is insufficient to block word order freezing, as seen through example (18) which allows only an “agent < theme” parse.

We also note that, in all of the examples so far, the nominals ‘man/men’ and ‘woman/women’ are equally likely as the agent and theme of the verb, to ‘see.’ In what follows, then, we will use various manipulations to bias the relative plausibility of a “theme < agent” construal. In such cases, we shall see that our seven speakers exhibit individual differences in the boundaries of their word order freezing effects.

First, we consider a manipulation in animacy in (19) below. For four of our seven speakers, example (19b) is judged as unnatural because it only has the nonsensical interpretation where a ball kicked a man. For these speakers, word order freezing applies even if there is a clear animacy contrast that may serve to disambiguate the roles of the arguments. However, our other three speakers allow for the plausible, intended interpretation, which requires a “theme < agent” parse of the two arguments. All speakers accept the intended interpretation in (19a), which is interpreted naturally with “agent < theme” order.

Next, we maintain the equal animacy of arguments but see if world knowledge alone can serve to resist word order freezing. In (20), we modify example (15) to specify that the man is blind, making him a less natural agent of ‘seeing.’ All speakers allow for the intended interpretation in (20a), where the agent precedes the pivot theme. Reversing the order of arguments, all but one of our speakers interpret (20b) as unambiguously expressing the less likely situation where a blind man sees a woman, and thus report it to be anomalous. Just one speaker allows the intended, more plausible interpretation, reflecting a “theme < agent” parse without freezing.

We also consider the effect of discourse congruence and manipulating the information-structural status of arguments. We asked four of our speakers13 to judge examples of the form in (15) in the context of different explicit Questions Under Discussion (QUDs) (Roberts 1996). In response to a question of the form ‘Who did the man see?’ in (21Q),14 only one speaker accepted example (15) above (repeated here as (21A)) as a felicitous answer.15 For (21A) to be congruent to the question in (21Q), the postverbal phrases must be interpreted with a “theme < agent” parse.

We summarize the points of speaker variation that we have observed here, reporting individual judgments for all seven of our speakers in (22) below. Our speakers sort into three groups with internally consistent patterns of behavior, which we label A, B, and C in (22). Group A speakers show the strictest form of word order freezing, whereas group B speakers allow for phrases with contrasting animacy to maintain free word order, and our one C speaker additionally allows for world knowledge and discourse context to distinguish arguments and resist freezing. We will argue that our proposal for scrambling and freezing in section 2 above can productively account for these three different grammars of Pangasinan in section 3.3 below.

We note that, formally, the addition of the explicit QUD in (21) makes the theme in the intended proposition the answer focus, so we hypothesize that the theme bears a [+foc] feature in contrast to the agent. As the table shows, the one speaker that accepted the dialogue in (21) was the one speaker that accepted (20) above, using world knowledge alone to disambiguate between the arguments. We hypothesize that the additional modifier in (20) supports this speaker’s accommodation of an implicit QUD such as ‘Who saw the blind man?’ or ‘Who/What did the woman see?’, thus reducing the case of this example to that of our explicit QUD example in (21).

Finally, we note that we have found one attested example in Benton (1971) of a sentence with two arguments introduced by identical demonstrative articles that appears to require a “causee < causer” parse; i.e. in “theme of causative < agent of causative” order. We reproduce this example in (23) with Benton’s translation. However, we have presented this example to two of our group A speakers and both reported that it unambiguously means that ‘the child made the young man cry,’ rejecting Benton’s translation.

We hypothesize that this example must have been produced by a speaker commanding the group C grammar, in a context that makes the two arguments distinguished in their information-structural status.

We now illustrate how our general proposal for word order scrambling and freezing, as in section 2 above, is able to account for the patterns of word order flexibility in Pangasinan and its individual variation.

We begin by describing a set of basic assumptions for the clause structure of Philippine-type languages. We assume that all arguments are introduced within a lower domain of the clause which we call vP. The lexical verb moves out of vP to a position to the left, leaving all material within vP linearized postverbally. Arguments with particular thematic roles are associated with particular positions within vP, in accordance with Baker’s UTAH (Baker 1988), and in particular the agent argument is always the highest argument within vP. This suggests that, without further manipulations, the agent will precede any other postverbal noun phrase (NP_X), as illustrated in (24). Here, art stands in for a prenominal article.

One of the noun phrase arguments within the clause will be identified as the pivot, receiving nominative case, and this choice will be reflected by voice morphology on the verb. Other nominals in vP may be oblique or express default genitive case (Erlewine & Levin & Van Urk 2020; Erlewine & Lim to appear). After case marking, a Non-Actor Voice (NAV) clause (including the PV clauses described above) will be linearized as in (25).

Next, we make two working assumptions regarding Pangasinan apposition. (Here we present these two properties of Pangasinan apposition as working assumptions, but we develop and motivate a syntax for apposition with both of these properties in Erlewine & Lim in prep.) First, nominals introduced by case-marking articles and demonstrative articles have different structural size — KP and DemP, respectively — and project the features that are encoded on their heads. Case-marking articles (K) in the extended projection of common nouns then bear case features but not ϕ-features, as in (26a), as case-marking articles for common nouns do not encode ϕ-feature contrasts (see (8)). In contrast, demonstratives bear the deictic features [±dem.auth, ±dem.part] as in (12) and ϕ-features and potentially other features of their nominal but not case features, as in (26b). Both project [+N], as extended projections of N.

As noted above, we use binary formal features for demonstratives and for ϕ-features, such as the [±sg] singular feature. We also assume that case values can be decomposed into binary features (see e.g. Jakobson 1936; Bierwisch 1967; Wunderlich 1997), of which we illustrate only [±gen] here.

The second working assumption for Pangasinan apposition is that the nominals introduced by demonstratives are generated in regular argument positions, rather than for example being adjoined higher as a form of dislocation. Support for this view comes from binding. Binding possibilities in Pangasinan reflect the base positions of arguments, just as in Tagalog (see e.g. Kroeger 1991; 1993; Rackowski 2002): for instance, agents are able to bind into other arguments, but not vice versa, regardless of voice. As example (27) shows, an agent can bind a reflexive or reciprocal within the PV theme pivot. The use of pivot apposition in (28) does not affect these binding possibilities, indicating that these demonstrative-marked nominals are generated in (or very close to) the regular base positions for themes.16

Based on these assumptions, demonstrative-marked arguments in apposition will have the same base word order as that of their corresponding case-marked variants. (29) below schematically represents an example where both the pivot and non-pivot agent of a Non-Actor Voice clause undergo apposition; corresponding pronouns are not indicated here. Without scrambling, the demonstrative-marked agent will precede the demonstrative-marked pivot.

Finally, we follow Richards (1993; 2013) (see also discussion in Rackowski 2002: 22–27) in the view that postverbal word order flexibility in Philippine languages is the result of optional “scrambling” movements. Following our proposal in section 2, this involves positing optional scrambling probes of the form [probe:F] for some feature F ∈ Σ, where Σ is a designated set of features that can be targeted for scrambling. Scrambling probes project above vP but below the surface position of the verb. In (30), we illustrate a case where the scrambling probe [probe:F] matches the pivot and moves it over the agent, leading to a word order distinct from the base word order without scrambling in (29).

Following the logic in section 2, a constituent β can be scrambled over a higher constituent α if β has a formal feature in Σ that α does not have. The relative order of two constituents is then frozen in its base “α < β” order when there is no feature in Σ that matches β but not α. We then model the attested microvariation in scrambling through minor differences in the set of scrambling features Σ for our three different groups of speakers:

Using the scrambling feature specifications in (31), all speakers are able to change the word order of constituents that vary in category, ϕ-features, or case features. Additionally, as summarized in (22) above, group B speakers (using Σ_B) can use animacy to scramble constituents that are otherwise frozen for our A speakers, and our group C speaker (using Σ_C) can additionally use information-structural (IS) features based the discourse context (such as congruence to the QUD) to scramble constituents. We discuss the acquisition of these features in section 4.2 below.

We propose that all postverbal word order variation is due to movement by such scrambling probes.17 For concreteness, for all Pangasinan examples that we report above with non-default word orders, we describe the type of scrambling probe specification involved in its derivation under our proposal:

Note that we hypothesize that scrambling probes in Pangasinan may target the positive or negative value of each feature in Σ, rather than just a marked feature value or privative feature. This choice is empirically motivated for instance by the availability of scrambling a [+sg] theme over a [–sg] agent using [probe:+sg] in example (16a), as well as scrambling a [–sg] theme over a [+sg] agent using [probe:–sg] in example (16b). As a reviewer notes, this symmetry may be surprising given the typological prevalence of plural-probing processes (i.e. ‘omnivorous number’) and corresponding scarcity of specifically singular-probing processes. However, this assumed universal asymmetry between plural and singular has been challenged; see Murugesan (2021), Raghotham (2021), and Kumaran (2023) for recent arguments that some morphosyntactic processes do in fact make reference to a marked singular feature value.19

Finally, we reiterate that we reject the postulation of a purely optional feature [scr/Σ] that can be added arbitrarily to a constituent to designate it as a target of scrambling, as in Müller (1997; 1998) and Grewendorf & Sabel (1999) and much subsequent work. We also reject approaches to Philippine-type languages that allow for flexible postverbal word order through non-movement means, as in Fowlie (2013), Erlewine & Levin & Van Urk (2020), Erlewine & Lim (to appear), and Branan (2022). Our proposal explains the fact that all speakers are susceptible to some form of word order freezing, including our C speaker in the absence of contextual cues (see (15) and (17)), which these alternative accounts do not explain.

A traditional argument for the link between freezing and confusability comes from data where morphological syncretism makes particular arguments inflectionally underspecified and concomitantly frozen. Such examples were made famous by Jakobson (1963: 269) for Russian and by Chomsky (1965: 126) for German. Both Russian and German have core case distinctions as well as some degree of word order flexibility; in particular, simple transitive sentences may appear in SVO or OVS order, with arguments distinguished by nominative and accusative case inflections. However, with particular noun phrases where the nominative and accusative case forms are syncretic, only the SVO word order construal is available without additional contextual manipulation.21

As a reviewer notes, at first glance, these facts appear to be challenging for the approach to scrambling and freezing effects that we put forward here. We assume that in the derivation of such examples, the subject is assigned nominative case and the object assigned accusative. If Σ includes case features such as [±acc], it should be possible to posit a probe [probe:+acc] to attract the object over the subject. The fact that these particular nominals then happen to be realized with forms that are ambiguous in terms of their case should not affect the possibility of such scrambling. The reported sensitivity of freezing to morphological syncretism thus appears to be challenge our approach to scrambling and freezing effects.

Such “freezing by syncretism” interactions can nonetheless be derived under our account, if we assume a derivational architecture where syncretism may bleed the featural distinctions available for scrambling. Here we appeal to rules of Impoverishment in Distributed Morphology: language-specific rules that can delete particular features in certain contexts, prior to Vocabulary Insertion (see e.g. Bonet 1991; Noyer 1992). For example, Krifka (2009: 164–165) suggests that the pervasive nominative/accusative case syncretism of feminine nominals in German grammar may be the result of an Impoverishment rule that neutralizes this featural distinction; see (35). We could similarly posit an Impoverishment rule as in (36) for Russian nouns of inflection class III, which includes both the ‘mother’ and ‘daughter’ of example (33).

If scrambling applies following the application of Impoverishment, it will naturally be susceptible to freezing by syncretism: features of the form [±acc] which are underlyingly syntactically active are not present at this point in the derivation.22

The idea that Impoverishment may precede scrambling naturally raises questions about the timing of scrambling. One possibility is for scrambling to take place postsyntactically; this would predict it to feed no or very few other syntactic processes. This prediction is borne out for postverbal scrambling in Pangasinan, which does not feed other hierarchy-sensitive syntactic processes such as pivot-oriented Ā-extraction processes (see Erlewine & Lim in prep.), as Richards (2013) also notes for Tagalog. A second possibility is that Impoverishment applies during cyclic Spell-Out of a lower portion of the clause (Uriagereka 1999; Chomsky 2000; 2001; a.o.) with scrambling applying to its output. This would allow for scrambling in narrow syntax — potentially with interpretational reflexes, for example for binding and scope — while still being susceptible to freezing by syncretism. This latter approach appears to be necessary for the German facts above, given Frey’s (2006) claim that non-information-structurally-motivated object movement to the prefield is fed by clause-medial A-scrambling.

Finally, we discuss the inventory of scrambling features and its determination. As noted above, the phenomenon of word order freezing appears to reflect a functional need to encode grammatical relations in a sufficiently unambiguous manner. Restating in terms of our proposal, there appears to be a functional basis for the inventory of scrambling features Σ: features that do not serve to disambiguate the semantic roles of arguments are not included in Σ. We suggest that this property of the feature set Σ can be explained as an epiphenomenon of the acquisition process involved in learning scrambling feature specifications.

We first describe a concrete and general procedure for the acquisition of scrambling features. We assume here that learners are conservative (see e.g. Snyder 2007; 2011 and Sugisaki & Snyder 2013) and thus require explicit evidence in order to learn an optional grammatical process such as scrambling using particular features. We illustrate this learning procedure using Pangasinan examples from above. To simplify our discussion, we start at a state where the learner hypothesizes the correct base order of arguments, i.e. “agent < theme” for the examples below. Suppose then that the learner encounters an example such as (37), the scrambled variant from example (1) above. Recall that the genitive clitic pronoun in such structures (here, to) cross-references the ϕ-features of the agent, in this case indicating that the agent is the singular ‘woman.’

Examples such as (37) are unambiguous in their interpretation and necessarily involve scrambling, thereby supporting the learner’s postulation of scrambling with a particular feature (here: [–sg]). We highlight the salient properties of example (38):

Examples with the properties in (38) serve as explicit evidence of scrambling for the learner, supporting the postulation of scrambling via an optional probe [probe:F], i.e. F ∈ Σ.

Suppose further that the learner encounters an example such as (39), repeated from (19b) above, uttered with the intended interpretation reflected by the translation below. Would such an example lead the learner to postulate a form of scrambling? This depends on whether the learner takes (39) to unambiguous express its intended interpretation (property (38a) above). If the learner interprets the string as a non-scrambled structure, expressing the less plausible interpretation ‘The ball kicked the man,’ or even considers both parses to be options, the example will not contribute to the postulation of a scrambling feature.23

However, if the learner does treat such an example as unambiguously expressing its intended meaning — perhaps, for instance, if the intended interpretation is also clear from other linguistic or non-linguistic support in the utterance context — they may then take the example to be explicit evidence for a particular form of scrambling that is possible in their target language. Only then would the learner consider identifying a feature that distinguishes β from α — perhaps the inanimate [–anim] or else a distinguishing information-structural feature, depending on the context — and consider adding it to Σ.

In reality, as we reported in (19b) above, some but not all Pangasinan speakers accept scrambling of the form in (39). The possibility of producing such an utterance by even a subset of speakers means that the linguistic input of a Pangasinan speaker may include examples of the form in (39), making the above discussion of learner reactions to (39) a possibility worth discussing. At the same time, learner experiences may vary as to whether their input includes (sufficient tokens of) unambiguous utterances of this form, leading to the attested microvariation in scrambling feature specifications across adult speakers.24

Recall furthermore that two postverbal arguments in Pangasinan that vary only in deictic features are subject to frozen word order, for all of our speakers. The relevant data point is repeated here from above in (40). Note that we predict that at least some Pangasinan speakers (group C) would be able to produce sentences of the form in (40) with its intended interpretation, given an appropriate discourse context that distinguishes the two arguments by information-structural features, and therefore for the input of learners to potentially include examples of this form.

The contrasting deictic feature values do not serve to disambiguate the arguments’ roles, and unlike in (39), there is not even a clear difference in semantic plausibility of the two possible parses. This makes it very unlikely — although not impossible — for examples of this form to be interpreted unambiguously as involving scrambling and thereby support the learning of scrambling by deictic feature values such as [+dem.auth].

Summarizing the consequences of the discussion above, the proposed procedure for the learning of the featural triggers of scrambling in Σ predicts that formal features that can serve to disambiguate the semantic roles of arguments — for instance, ϕ-features as in (37) or animacy features as in (39) — are much more likely to be learned as possible grammatical triggers of scrambling. We thereby explain the apparently functional basis for the inventory of scrambling features, without treating word order freezing itself as a dynamic response to argument confusability, as in the strong functionalist account above, and also allowing for individual variation in the categorial judgments of different speakers.

More generally, we assume that the inventory of formal features in grammar must itself be learned from the linguistic input (see e.g. Zeijlstra 2008; Cowper & Hall 2014; Koeneman & Zeijlstra 2014; Biberauer & Roberts 2017; Biberauer 2019). We propose that no learner will hypothesize the existence of an arbitrary syntactic annotation [scr/Σ], with no semantic import (Müller 1998: 43), in order to account for an input utterance with marked word order. The grammatical trigger for the displacement of a constituent must then be attributed to one of its independently detectable features, based to its distinguishing semantics or morphology. These conservative assumptions for the acquisition of individual grammars successfully account for the various observations that at first glance suggest a description of word order freezing as a response to argument confusability.