1.1 Voice cross-references an argument and interacts with A-bar extraction

Verbs in Tagalog typically carry what has been called voice morphology, which always cross-references the argument of the clause marked by ang [aŋ] (Schachter & Otanes 1983). For example, the sentences in (1) involve a man, a fish, and a store in a buying event, and the thematic relations assigned by the verb to its arguments remain constant throughout. Notice that the verb form changes and the case markers of the arguments vary depending on the verb form. In (1a), the verb exhibits agent voice (AV). The agent lalaki ‘man’ is marked nominative; the patient isda ‘fish’, genitive, with ng [naŋ]; and the location tindahan ‘store’, dative, with sa.² In (1b), the verb exhibits patient voice (PV). The patient is now marked nominative; the agent, now genitive; and the location, still dative. In (1c), the verb exhibits locative voice (LV).³ The location ‘store’ is now marked nominative, and the other arguments, genitive. Throughout, the argument cross-referenced by the verb is in bold; and the agent, underlined.

As in many other Austronesian languages, the argument cross-referenced by voice interacts with A-bar extraction in Tagalog. This interaction has been called many names in the syntactic literature, from “subject-only restriction” (Keenan & Comrie 1977) to “absolutive restriction on A-bar extraction” (Aldridge 2008). We refer to this interaction throughout as the “extraction restriction.” In (2) is the received generalization that emerges upon surveying the Tagalog syntax literature. We refer to the received view as the strict version of the extraction restriction—to contrast with what we will be referring to as the less strict version presented in section 1.2.

What (2) says is that an argument that is extracted must match the argument that is cross-referenced by voice (i.e., the argument that is marked with ang). In (3), we show this interaction. When the verb exhibits AV, the agent babae ‘woman,’ which is cross-referenced by AV, is extractable and can serve as the pivot of a cleft, as in (3a). Meanwhile, the patient baro ‘dress,’ which is not cross-referenced by AV, is not extractable and cannot serve as the pivot of a cleft, as in (3b). In contrast, when the verb exhibits PV, the patient ‘dress,’ which is cross-referenced by PV, is extractable and can serve as the pivot of a cleft, as in (3c). The agent ‘woman,’ which is not cross-referenced by PV, is not extractable and cannot serve as the pivot of a cleft, as in (3d).

Throughout, we use the term voice-match to describe extractions where the extracted argument is cross-referenced by voice, as in sentences like (3a) and (3c). We use the term voice-mismatch to describe extractions where the extracted argument is not cross-referenced by voice, as in sentences like (3b) and (3d). In (4) we reframe the strict version of the extraction restriction in terms of voice-match.

For ease of exposition, we exemplified the voice-match restriction using AV and PV. The restriction is reported to extend to the other voices (Rackowski & Richards 2005).

1.2 The puzzle: Voice-mismatch is allowed if the agent is extracted

Contra the generalization in (2)/(4), others have observed that the interaction is not as clear-cut. For some speakers, voice-mismatch is allowed if the extracted argument is the agent (Ceña & Nolasco 2011; 2012; Tanaka 2016; Hsieh 2019; Bondoc 2021).⁴ In (5), we provide the less strict version. Note that (4) is a proper subset of (5).

Ceña & Nolasco (2011: 181) observed the following contrasts in Tagalog. In (6), the verb exhibits AV. The agent bantay, the argument cross-referenced by voice, can be relativized, as in (6a). The patient basi cannot be relativized, as in (6b). The source tao also cannot be relativized, as in (6c). These patterns are consistent with (2)/(4) thus far.

The data reported by Ceña & Nolasco diverge from the generalization in (2)/(4) when the verb exhibits non-AV voices. In (7), the verb exhibits PV. The patient tao, the argument cross-referenced by voice, can be relativized, as in (7a). The location palengke cannot be relativized, as in (7b)—at least, not using the same relativization strategy (Hsieh 2018). These patterns are consistent with (2)/(4) thus far. However, (7c) is inconsistent with it: the agent bantay can also be relativized.

In (8), when the verb exhibits applicative morphology, the recipient tao, the argument cross-referenced by voice, can be relativized, as in (8a). The patient pitaka cannot be relativized, as in (8b). These patterns are consistent with (2)/(4) thus far. However, (8c) is inconsistent with it: the agent bantay can also be relativized.

Table 1 summarizes the empirical picture by the thematic role of the extracted argument and the version of the extraction restriction. A ‘✔’ means that the extraction is licit; a ‘(✔)’, licit for some speakers; and ‘✖’, not licit.

Before proceeding, we make fully available Ceña & Nolasco’s description of the extraction restriction. Ceña & Nolasco also observed that those who do allow the agent to be relativized in voice-mismatch contexts only do so in head-initial relative clauses. This is shown in (7c), repeated below as (9a). Voice-mismatches are never allowed in head-final relative clauses, as shown in (9b).

In the judgment studies reported in sections 4.2 and 4.3 (and in our own work with speakers), we replicate the contrasts between agent-extraction in head-initial and head-final RCs, as shown in (9). We leave it to future work to see why these RCs behave differently. One possible reason for their difference in behavior vis-à-vis extraction could be due to structural differences. However, it is an open question as to whether these RCs have different derivations (Aldridge 2017) or share a single one (Law 2016).

1.3 The current study

We re-examine the nature of the interaction between voice and A-bar extraction in Tagalog. We have outlined the two competing generalizations of the restriction above, and we repeat these in (10). We framed this interaction in terms of voice-(mis)match to preview the terminologies that we will be using in the experiments that follow.

The present study asks two questions. First, we ask whether voice-mismatch extractions are indeed allowed—in particular, those with agent-extractions. Under the strict version of the restriction, only voice-match extractions are allowed; voice-mismatch extractions are not allowed. Thus, the mismatch penalty, the contrast between voice-match and voice-mismatch extractions should not vary by voice. On the other hand, under the less strict version, the mismatch penalty incurred in AV should be greater than that in non-AV voices because of the acceptability of agent-extractions even in voice-mismatch contexts. Second, if they are allowed when the extracted argument is the agent, we ask whether this permissiveness is a general property shared by all A-bar dependencies or whether this is specific to relative clauses. This is an empirical question because the studies in section 1.2 that observe the less strict version mostly only involved relative clauses.

The present study makes two main contributions. Our first contribution is empirico-theoretical. We show that voice-mismatch agent-extractions do exist “in the wild” and are indeed accepted by some speakers in an experimental setting. These types of extractions are important to formal syntacticians because they challenge current theories of extraction that rely on the patient/applied argument moving to a structural position higher than the agent to determine voice, as in (11) and in (12), and Attract Closest to derive the extraction restriction (Aldridge 2002; Rackowski & Richards 2005; Aldridge 2012). These proposals undergenerate; they cannot generate agent-extractions under non-AV voices for those who do allow them.

These types of extractions are important to (developmental) psycholinguists because they show how two types of Tagalog speakers can emerge despite being exposed to the same or similar linguistic input. Ultimately, we argue that what underlies this variation is grammar competition à la Han et al. (2007) that arises from the agent-first bias (Ferreira 2003; Bornkessel & Schlesewsky 2006) interacting with reanalysis (Frazier & Rayner 1982; 1987), and a noisy channel assumption in acquisition (Perkins et al. 2022).

Our second contribution is methodological. We applied the type of computational modeling that Dillon et al. (2017) employed to model English speakers’ behavior in agreement attraction, a completely unrelated phenomenon in an unrelated language. We constructed quantitative models that demonstrate that there are two separable types of Tagalog speakers with respect to extraction. This serves as further proof of concept for how a distributional analysis of judgment ratings can provide evidence for multiple types of speakers in a language community. Quantitative models that estimate participants’ judgment processes are particularly useful to researchers who argue for the existence of two varieties to account for variation in behavior with respect to a particular linguistic phenomenon. For example, this type of modeling could also be applied to the scopal variation in Korean with respect to negation and quantified objects. Even though Han et al. (2007) and Zeijlstra (2024) burden different parts of the Korean grammar, both proposals account for scopal variation by positing two varieties. The methodology employed in the present study can provide evidence for or against the tenability of having two varieties of Korean, and it can also help adjudicate between the two proposals if we model the speakers’ behavior in corners of the grammar where the predictions of the two proposals diverge.

The remainder of the paper is structured as follows. Section 2 presents naturally occurring examples of agent-extractions under non-AV voices and reviews previous experimental findings by Pizarro-Guevara & Wagers (2020). Both are inconsistent with a strict view of the restriction and is consistent with a less strict view. Section 3 reanalyzes experiment 1 of Pizarro-Guevara & Wagers using a Bayesian approach to statistical inference in order to incorporate their insights into the statistical analyses of the judgment studies presented after. Section 4 presents a series of acceptability judgment experiments, which were designed to be a replication and an extension of Pizarro-Guevara & Wagers, and updates the empirical landscape of the extraction restriction. Section 5 presents computational evidence that there are two separable types of Tagalog speakers vis-à-vis extraction. Section 6 concludes by proposing a way in which we can think of how the two types of Tagalog speakers emerge despite being exposed to the same or similar linguistic input. We view these two types of speakers as a form of grammar competition (Han et al. 2007) that arises via the idea that the agent-first bias could affect how child learners parse input strings under noisy conditions during acquisition.

2.1 Voice-mismatches “in the wild”

One of our main contributions to the discussion on the nature of the extraction restriction is to provide examples of these voice-mismatch extractions “in the wild.” The examples we provide below are naturally occurring and are from various sources online, ranging from newspapers and tabloids, from Wikipedia entries to the Bible.

The examples in (13) involve agent-extraction when the verb exhibits PV. The examples in (13a)–(13b) involve relative clauses; those in (13c)–(13d), topicalizations; and those in (13e)–(13f), ay-inversions. The examples in (13c)–(13f) suggest that the acceptability of agent-extractions under PV is not specific to relative clauses.

The examples in (14) involve agent-extraction when the verb exhibits applicative morphology. The example in (14a) involves a relative clause; that in (14b), topicalization; and those in (14c)–(14d), ay-inversion. The examples in (14b)–(14d) suggest that acceptability of subject-extractions under applicatives is not specific to relative clauses.

2.2 Experimental evidence from judgment studies

Pizarro-Guevara & Wagers (2020) investigated the extraction restriction experimentally by looking at whether and how comprehenders used voice when processing A-bar dependencies in real-time. They hypothesized that voice allows comprehenders to commit to an interpretation as early as the verb. They reasoned that at this point in the utterance, comprehenders already have information about the event structure via the lexical semantics of the verb, and about the thematic role of the moved element via voice. Thus, at the verb, comprehenders should already have enough information to evaluate the thematic fit of the moved argument.

Their hypothesis crucially hinged on the assumption that there was a one-to-one mapping between voice and the identity of the moved argument. In other words, their hypothesis assumed the strict version of the restriction. To evaluate whether this assumption was tenable, they conducted a series of acceptability judgment studies and tested three types of dependencies: ay-inverted sentences, wh-questions, and relative clauses. They manipulated whether the verb exhibited AV or PV (Voice: av, pv), and whether the extracted argument was cross-referenced by voice or not (Match: ±). The logic of their design is as follows. By comparing the participants’ ratings of agent-extractions in AV (+match), as in (15a), against patient-extractions in AV (–match), as in (15b), we can estimate the mismatch penalty incurred when extracting patients not cross-referenced by voice. By comparing their ratings of patient-extractions in PV (+match), as in (15c), against agent-extractions in PV (–match), as in (15d), we can estimate the mismatch penalty incurred when extracting agents not cross-referenced by voice.

Figure 1¹⁵ visualizes the distribution of their ratings by Voice, Match, and dependency type. They had three main findings. First, across dependency type, participants gave higher ratings to sentences in voice-match contexts, that is, when the extracted argument is cross-referenced by the voice morphology on the verb. We see this in the left panel: there is a higher proportion of blue and bluish bars compared to the gray(ish) bars. This finding is consistent with both versions of the extraction restriction.

Figure 1: Distribution of ratings in exp. 1 of Pizarro-Guevara & Wagers (2020) by Voice, Match and dependency type. The left panel corresponds to voice-match extractions; the right panel corresponds to voice-mismatch extractions. Blue corresponds to a rating of 7 ‘acceptable’, while gray corresponds to a rating of 1 ‘unacceptable.’

Second, across dependency type, participants gave lower ratings to AV-sentences in voice-mismatch contexts, as in (15b). We see this in the right panel: The top bar for each dependency type has a higher proportion of gray(ish) bars compared to the blue and bluish bars. This finding is again consistent with both versions of the restriction.

Third, participants were more variable with how they rated PV-sentences in voice-mismatch contexts, as in (15d). We see this in the right panel: The bottom bars for the ay and rc panels have a higher proportion of blue and bluish bars compared to gray(ish) bars—relative to the second finding. These proportions are not quite as high as the first finding, however. We see this pattern in the whq panel as well, but to a lesser extent. This finding is only consistent with the less strict version.

3.1 Analysis

We fitted Bayesian ordinal mixed effects regression models with a probit link function in R (R Core Team 2020) using brms (Bürkner 2017), a front end of the Stan language for Bayesian estimation of model parameters (Gelman et al. 2015). For each dependency, we used the participants’ ratings as the dependent measure and fitted two types of models: crossed and nested. In the crossed model, we included the main effects of Voice, Match, and their interaction as fixed effects. In the nested model, we separated the fixed effects parameters for Match in the av and pv conditions. The random effects structure included random intercepts for participants and items, and Voice, Match, and their interaction as random slopes for participants and items. Table 2 provides the contrast coding for all the fixed effects in both models for all the dependencies.

As a starting point, we used Normal(0,5)¹⁶ as the priors for all the fixed effects and the intercept. This means that the parameter has a mean of 0 and a standard deviation of 5. These are uninformative priors as they do not place strong constraints on the model’s predictions, and incorporate very little knowledge about what makes a plausible ratings distribution. We used LKJ(2) as the prior for our correlation matrix. For each model, we ran four Monte Carlo Markov chains in parallel, with 40,000 samples each. The first 8,000 were always discarded as part of warm-up. For all the models reported below, the R-hat statistic was at 1.0. No divergences were observed.

In addition, we quantified how much evidence we have for an effect by conducting Bayes factor analyses. We compared the marginal likelihood of the full model against the marginal likelihood of a null model. In the crossed models, the full model included Voice, Match and their interaction, while the null model excluded the interaction. Comparing the marginal likelihoods of these two models quantifies how much evidence we have in favor of/against the mismatch penalty being larger in AV than in PV. In the nested models, the full model included Voice, Match in av, and Match in pv. There are two null models: one that excluded Match in av and one that excluded Match in pv. Comparing the marginal likelihoods of the full model and the null model that excludes Match in av quantifies how much evidence we have in favor of/against the presence of a mismatch penalty in AV. Similarly, comparing the marginal likelihoods of the full model and the null model that excludes Match in pv quantifies how much evidence we have in favor of/against the presence of a mismatch penalty in PV.

Because BF analyses are highly sensitive to prior specifications (Nicenboim et al. 2020), we also conducted a sensitivity analysis by considering a range of priors, Normal(0,3) and Normal(0,1), for the interaction term in the crossed models and the separated Match parameter in the nested models. We used these increasingly informative priors for the parameters of interest, but kept the rest of the priors as originally described above.

3.2 Results

See our OSF page for two tables: (i) the mean posterior distribution, along with the 95% Bayesian credible intervals (CrI), which indicate where the plausible parameter values for these fixed effects parameters lie, given the data and the priors; and (ii) the sensitivity analysis, where we calculated the Bayes factor 100 times and report the median over these values, along with a 95% interval representing the 2.5 and 97.5 percentiles. For present purposes, what is important is that a Bayes factor smaller than 1 favors the null model, indicating evidence that the effect of interest is absent. A value larger than 1 favors the full model, indicating evidence that the effect of interest is present. A value larger than 10 is considered to be strong evidence (Jeffreys 1998).

Ay-inverted sentences. There is strong evidence for a mismatch penalty in AV. The evidence is between 2.58 × 10⁹ and 1.26 × 10¹⁰, suggesting that when the verb has AV, agent-extractions (+match) were rated higher than patient-extractions (–match). There is inconclusive evidence for a mismatch penalty in PV, ranging from anecdotal evidence for the null model to anecdotal evidence for the full model. The evidence is between .23 and 2.00, suggesting that when the verb has PV, we do not have enough evidence, given the data, to say whether patient-extractions (+match) were rated higher than agent-extractions (–match) or whether these were not rated differently from each other. There is strong evidence for a larger mismatch penalty in AV than in PV. The evidence is between 2.62 × 10⁴ and 8.96 × 10⁴.

Head-initial RCs. There is strong evidence for a mismatch penalty in AV. The evidence is between 8.81 × 10³ and 3.57 × 10⁴, suggesting that when the verb has AV, agent-extractions (+match) were rated higher than patient-extractions (–match). There is moderate evidence for no mismatch penalty in PV. The evidence is between .09 and .41, suggesting that when the verb has PV, patient-extractions (+match) were not rated differently from agent-extractions (–match). There is strong evidence for a larger mismatch penalty in AV than in PV. The evidence is between 11.22 and 15.90.

Wh-questions. There is strong evidence for a mismatch penalty in AV. The evidence is between 3.50 × 10⁴ and 9.33 × 10⁴, suggesting that when the verb has AV, agent-extractions (+match) were rated higher than patient-extractions (–match). There is also strong evidence for a mismatch penalty in PV. The evidence is between 771 and 2088, suggesting that when the verb has PV, patient-extractions (+match) were rated higher than agent-extractions (–match). There is anecdotal to moderate evidence for the mismatch penalties in AV and PV not being reliably different. The evidence is between .28 and .91.

3.3 Discussion

We reanalyzed experiment 1 of Pizarro-Guevara & Wagers (2020) using a Bayesian approach to statistical inference to estimate the magnitude of a mismatch penalty in AV and PV. We will use these values as informative priors for the parameters of interest in our replication, which will be presented in the next section. To estimate the mean of the interaction, we took the average of the mean point estimates of the interaction across the different prior specifications. For example, we estimated the mean of the interaction for head-initial RCs to be 1.58, which is the average of 1.74, 1.69, and 1.31. To estimate the standard deviation of the interaction, we took the average of the errors (i.e. the width of 95% CrI, divided by four). For example, we estimated the standard deviation of the interaction for head-initial RCs to be .56, which is the average of .59, .58, and .50. We followed the same procedure for calculating the prior specifications for the Match in av and Match in pv of the nested models.

In our replication, we will be incorporating their insights in our models as informative priors. For head-initial RCs, we will use Normal(1.58,. 56) for the interaction in the crossed model, and Normal(–1.90,. 09) and Normal(–.17,. 10) for Match in av and Match in pv in the nested models, respectively. For wh-questions, we will use Normal(.64,. 10) for for the interaction in the crossed model, and Normal(–2.26,. 10) and Normal(–1.64,. 08) for Match in av and Match in pv in the nested models, respectively.

Our reanalysis also quantified how much evidence there was in their experiment for a larger mismatch penalty in AV than in PV in order to see whether their empirical generalizations were warranted, given the data. Our reanalysis is congruent with their generalizations in broad strokes. We found strong evidence for a mismatch penalty in AV across dependency types. We also found that there was variable evidence for a mismatch penalty in PV. However, the finer details varied.

In ay-inverted sentences, the mismatch penalty in AV was larger than in PV. What we are uncertain about is whether it was a smaller mismatch penalty in PV or whether it was absent. In relative clauses, there was no mismatch penalty in PV and thus, the mismatch penalty in AV was larger than in PV. In wh-questions, there was evidence that the mismatch penalties in AV and PV did not differ.

Given their data, the empirical terrain that emerges is the following: only the extraction behavior in wh-questions is more consistent with the strict version of the extraction restriction. Their extraction behavior in ay-inverted sentences and head-initial RCs is more consistent with the less strict version of the restriction.

Their findings raise three important questions. First, are these findings replicable? Second, if they are, how can the less strict extraction restriction be modeled? Lastly, what does this tell us about the landscape of A-bar extraction in general? To answer the first question, we conducted our own judgment studies that sought to replicate and extend their findings. We present our studies comparing the extraction behavior of Tagalog speakers under AV and PV in section 4.2, and their extraction behavior under applicatives in section 4.3. To answer the second question, in lieu of presenting a formal analysis for the less strict version of the extraction restriction, we suggest a way to think about this variation as a form of grammar competition in learning, in the sense of Han et al. (2007), that arises via the idea that an agent-first bias could filter the linguistic input or otherwise affect how child learners parse input strings under noisy conditions during acquisition. While all three questions are equally important, we leave it to future work to develop a formal analysis for the second question and to answer the last question. We invite readers interested in a formal analysis of voice-mismatch extractions to see Hsieh (2023) for one way to model the phenomenon.

4.1 Participants and procedure

We recruited 30 participants from the University of the Philippines, Diliman and surrounding communities. Participants needed to (i) be 18 years or older; (ii) live in or around Metro Manila at time of testing; and (iii) use Tagalog every day.¹⁷ Prior to any analysis, we excluded one participant due to internet connection issues. The remaining 29 participants ranged from 18 to 42 years of age (M = 25, SD = 5). They received either a gift certificate from a local coffee shop or a bookstore, valued at 300 Philippine Pesos, for participating.

The same speakers participated in experiments 1 and 2, which were combined in one session and were randomized along with items involving extraction out of islands and complement clauses. The judgment studies were administered online via Ibex (Drummond 2016). Participants were asked to rate the acceptability of sentences using a 7-point scale, with 1 being talagang hindi katanggap-tanggap ‘really unacceptable’ and 7 being talagang katanggap-tanggap ‘really acceptable.’ They were explicitly instructed that a sentence is katanggap-tanggap if (i) as a speaker of Tagalog, they can imagine themselves saying or writing this; and (ii) they can imagine other Tagalog speakers saying or writing this, as well. On average, each session took about 30 to 40 minutes to complete.

4.2 Experiment 1: Comparing voice-mismatch extractions in AV and PV

We replicated the finding that there is a consistently large mismatch penalty in AV, irrespective of the dependency involved. We did not replicate the variable mismatch penalty in PV. Instead, we found a consistent mismatch penalty in PV, albeit a smaller one, in head-initial RCs and wh-questions. In other words, agents enjoyed a certain degree of extractability, congruent with the less strict version of the extraction restriction. This did not extend to head-final RCs, however. This is consistent with the observation made by Ceña & Nolasco (2011) and is more congruent with the strict version.

4.2.3 Results and discussion

In Figure 2, we visualize the distribution of participant ratings by Voice, Match, and dependency type. See our OSF page for the following: (i) the mean posterior distribution, along with the 95% Bayesian CrI; and (ii) the sensitivity analysis, where we report the median Bayes factor and a 95% interval representing the 2.5 and 97.5 percentiles, calculated over 100 estimates of the Bayes factor. What is important for present purposes is that a value larger than 1 favors the full model, indicating evidence that the effect of interest is present. A value larger than 10 is considered to be strong evidence.

Figure 2: Distribution of ratings in experiment 1 by Voice, Match and dependency type. The left panel corresponds to voice-match extractions; the right panel corresponds to voice-mismatch extractions. Blue corresponds to a rating of 7 ‘acceptable’, while gray corresponds to a rating of 1 ‘unacceptable.’

Head-initial RCs. There is strong evidence for a mismatch penalty in AV. The evidence is between 1.21 × 10⁶ and 4.46 × 10⁶, suggesting that when the verb has AV, agent-extractions (+match) were rated higher than patient-extractions (–match). There is also strong evidence for a mismatch penalty in PV. The evidence is between 81 and 296, suggesting that when the verb has PV, patient-extractions (+match) were also rated higher than agent-extractions (–match). At best, there is anecdotal evidence for a larger mismatch penalty in AV than in PV. The evidence is between .87 and 2.68.

Wh-questions. There is strong evidence for a mismatch penalty in AV. The evidence is between 4.71 × 10⁵ and 1.68 × 10⁹, suggesting that when the verb has AV, agent-extractions (+match) were rated higher than patient-extractions (–match). There is also strong evidence for a mismatch penalty in PV. The evidence is between 6.02 × 10⁵ and 2.16 × 10⁹, suggesting that when the verb has PV, patient-extractions (+match) were also rated higher than agent-extractions (–match). There is strong evidence for a larger mismatch penalty in AV than in PV. The evidence is between 145 and 393.

Head-final RCs. There is strong evidence for a mismatch penalty in AV. The evidence is between 1.66 × 10⁵ and 6.55 × 10⁵, suggesting that when the verb has AV, agent-extractions (+match) were rated higher than patient-extractions (–match). There is also strong evidence for a mismatch penalty in PV. The evidence is between 5.61 × 10⁴ and 2.20 × 10⁵, suggesting that when the verb has PV, patient-extractions (+match) were also rated higher than agent-extractions (–match). There is moderate evidence for no difference in the mismatch penalties between AV and PV. The evidence is between .07 and .32.

Here our main goal was to see whether we can replicate the findings of Pizarro-Guevara & Wagers (2020). Let’s first focus on 2 of the 3 dependencies they examined, head-initial RCs and wh-questions. Like them, we found that voice-mismatch extractions in AV consistently incurred a large mismatch penalty, irrespective of the dependency involved. Unlike them, we found that voice-mismatch extractions also consistently incurred a mismatch penalty, albeit a smaller one, compared to voice-mismatch extractions in AV.

In head-initial RCs, our reanalysis and our replication found evidence for a larger mismatch penalty in AV than in PV. Where they vary is in terms of how large the mismatch penalty is in PV. Our reanalysis found moderate evidence for there being no mismatch penalty in PV, while our replication found strong evidence for a mismatch penalty in PV, albeit a smaller one compared to that in AV. One potential reason for the difference in effect sizes could be the priors used. Our reanalysis used uninformative priors, while our replication used more informative priors from our reanalysis. When the effect size is small, uninformative priors tend to favor the null model (Nicenboim et al. 2020). The estimated effect size of the mismatch penalty in PV in the original study is –.17 (c.f. the mismatch penalty in AV is between –1.98 and –1.76). Meanwhile, with more informative priors, the estimated effect size of the mismatch penalty in PV in our replication is between –1.43 and –.95 (c.f. the mismatch penalty in AV is between –2.40 and –2.21). Abstracting away from the underlying cause of the difference, the results of the original and of our replication are inconsistent with the strict restriction. Crucially, both are congruent with the less strict version.

In wh-questions, our reanalysis and our replication found strong evidence for a mismatch penalty in PV. Where they vary, however, is whether the size of this penalty is smaller or comparable to the size of the mismatch penalty in AV. Our reanalysis found no reliable difference between the mismatch penalties incurred in AV and in PV. Our replication, on the other hand, found strong evidence for a larger mismatch penalty in AV than in PV. One potential reason is the priors used. As discussed above, uninformative priors tend to favor the null model, which could be the reason why our reanalysis did not detect any reliable difference between the mismatch penalties incurred in AV and in PV. The difference could also be due to the differences in the shape of the experimental items. Our experimental items differed along two dimensions. First, Pizarro-Guevara & Wagers used matrix wh-questions, while ours used embedded wh-questions. Second, they used d-linked interrogatives (i.e., akin to English which X), while ours used bare interrogatives. We have not done any systematic investigations on the effect of embedding context and d-linking on extractability. We leave these questions open for now. What is important for present purposes is that in wh-questions, there is some evidence consistent with the strong version and some evidence consistent with the less strict version.

A secondary goal was to determine the extent to which agent-extractability in voice-mismatch contexts varies in head-initial and head-final RCs. Recall that Ceña & Nolasco (2011) observed that agent-extractions under PV are permissible in head-initial RCs by some speakers, but never in head-final RCs. This generalization is consistent with our data. In head-initial RCs, the mismatch penalty incurred in PV is smaller than that in AV. However, in head-final RCs, there is moderate to strong evidence to believe that the mismatch penalties in AV and PV are not reliably different.

4.3 Experiment 2: Comparing voice-mismatch extractions in applicatives

We found that agent-extractions incurred a mismatch penalty, albeit a smaller one relative to patient-extractions, when the verb has applicatives in head-initial RCs. These results are more congruent with the less strict version of the extraction restriction. It is unclear, however, if this permissibility extended to wh-questions and head-final RCs. At best, given the data, we only have anecdotal evidence that agent-extractions and patient-extractions under applicative incurred comparable mismatch penalties in wh-questions, and inconclusive evidence in head-final RCs.

4.3.3 Results and discussion

In Figure 3, we visualize the distribution of participant ratings by AgEx, Match, and dependency type. See our OSF page for the following: (i) the mean posterior distribution, along with the 95% Bayesian CrI; and (ii) the sensitivity analysis, where we report the median Bayes factor and a 95% interval representing the 2.5 and 97.5 percentiles, calculated over 100 estimates of the Bayes factor. What is important for present purposes is that a value larger than 1 favors the full model, indicating evidence that the effect of interest is present. A value larger than 10 is considered to be strong evidence.

Figure 3: Distribution of ratings in experiment 2 by AgEx, Match and dependency type. The left panel corresponds to voice-match extractions; the right panel corresponds to voice-mismatch extractions. Blue corresponds to a rating of 7 ‘acceptable’, while gray corresponds to a rating of 1 ‘unacceptable.’

Head-initial RCs. There is strong evidence for a voice-mismatch penalty when extracting the agent. The evidence is between 18.12 and 88.54, suggesting that extracting the agent when the verb had AV (+match) was rated higher than when the verb had applicative morphology (–match). There is also strong evidence for a voice-mismatch penalty when extracting the patient. The evidence is between 6.38 × 10⁴ and 3.13 × 10⁵, suggesting that when the verb had applicative morphology, extracting the applied object (+match) was rated higher than extracting the patient (–match). There is anecdotal to moderate evidence for a smaller mismatch penalty in agent-extractions under applicative than in patient-extractions. The evidence is between 1.94 and 6.17.

Wh-questions. There is strong evidence for a voice-mismatch penalty when extracting the agent. The evidence is between 2.34 × 10⁴ and 7.19 × 10⁵, suggesting that extracting the agent when the verb had AV (+match) was rated higher than when the verb had applicative morphology (–match). There is also strong evidence for a voice-mismatch penalty when extracting the patient. The evidence is between 5.68 × 10⁶ and 1.75 × 10⁸, suggesting when the verb had applicative morphology, extracting the applied object (+match) was rated higher than extracting the patient (–match). At best, there is anecdotal evidence for a smaller mismatch penalty penalty in agent-extractions under applicative than in patient-extractions. At worst, there is anecdotal evidence that their mismatch penalties are not different. The evidence is between .30 and 1.07. In other words, given the data, we do not know whether extracting the agent incurred a smaller mismatch penalty than extracting the patient when the verb had applicative morphology or whether there is no difference between their mismatch penalties.

Head-final RCs. There is strong evidence for a voice-mismatch penalty when extracting the agent. The evidence is between 1.21 × 10⁴ and 8.70 × 10⁴, suggesting that extracting the agent when the verb had AV (+match) was rated higher than when the verb had applicative morphology (–match). There is also strong evidence for a voice-mismatch penalty when extracting the patient. The evidence is between 257 and 1857, suggesting that when the verb had applicative morphology, extracting the applied object (+match) was rated higher than extracting the patient (–match). At best, there is anecdotal evidence for a smaller mismatch penalty penalty in agent-extractions under applicative than in patient-extractions. At worst, there is anecdotal evidence that their mismatch penalties are not different. The evidence is between .45 and 1.95. In other words, given the data, we do not know whether extracting the agent incurred a smaller mismatch penalty than extracting the patient when the verb had applicative morphology or whether there is no difference between their mismatch penalties.

Here our main goal was to examine the extraction behavior of Tagalog speakers when the verb exhibits applicative morphology. We found that agent-extractions incurred a mismatch penalty, albeit a smaller one relative to patient-extractions, when the verb has applicatives in head-initial RCs. This echoes the findings in experiment 1. It is unclear, however, if this permissibility extended to wh-questions and head-final RCs.

4.4 The empirical terrain: Summary of Experiments 1 and 2

We presented a series of acceptability judgment experiments, designed to replicate and extend the findings of Pizarro-Guevara & Wagers (2020). We compared the extraction behavior of Tagalog speakers when the verb exhibits AV and PV (experiment 1), and when the verb exhibits applicative morphology (experiment 2).

In experiment 1, head-initial RCs and wh-questions had comparable extraction profiles. The ratings distributions of voice-match extractions were more concentrated toward the higher values of the Likert scale. The ratings distribution of voice-mismatch extractions when the verb had AV (i.e., patient-extractions under AV), toward the lower values. The ratings distribution of voice-mismatch extractions when the verb had PV (i.e., agent-extractions under PV) was in a sense more intermediate. It was not quite like the distributions of the other two. This empirical terrain is more congruent with the less strict version of the extraction restriction. However, head-final RCs had a different extraction profile. The ratings distributions of voice-match extractions were more concentrated toward the higher values of the Likert scale, while the ratings distributions of voice-mismatch extractions were more concentrated toward the lower values. This empirical terrain is more congruent with the strong version of the extraction restriction.

In experiment 2, head-initial RCs replicated the extraction profile exhibited in experiment 1. The ratings distributions of voice-mismatch extractions were more concentrated toward the higher values; that of voice-mismatch patient-extractions, toward the lower values; and that of voice-match agent-extractions, more intermediate. Wh-questions and head-final RCs replicated almost all of the extraction profile they exhibited in experiment 1. What is unclear is their behavior in voice-mismatch agent-extractions. At best, there is anecdotal evidence for a smaller mismatch penalty penalty in agent-extractions under applicative than in patient-extractions. At worst, there is anecdotal evidence that their mismatch penalties are not different.

It is crucial to understand the nature of the ratings distribution that underlies the intermediate ratings distribution of most of the voice-mismatch agent-extractions. Intermediate mean ratings are compatible with (at least) two types of response distributions. In one scenario, participants classified these extractions categorically. The intermediate ratings distribution stemmed from a mixture of acceptable and unacceptable responses, creating an “illusory” intermediate ratings distribution. In another scenario, participants perceived these extractions as somewhere between acceptable and unacceptable. The intermediate ratings distribution therefore reflected an overall upward shift in acceptability relative to unacceptable extractions or an overall downward shift in acceptability relative to acceptable extractions, creating a true intermediate ratings distribution. The next section develops quantitative models to estimate the participants’ judgment process to gain a more nuanced understanding of the underlying distributions of their ratings.

5.1 Method

To determine which model best captured the response behavior of the participants, three types of distributions are needed as ingredients for the simulations: (i) a test distribution, which is the distribution of the linguistic object that we want to model; (ii) an acceptable reference distribution that gives us an estimate of the sampling space when a linguistic object is perceived as uncontroversially acceptable; and (iii) an unacceptable reference distribution that gives us an estimate of the sampling space when a linguistic object is perceived as uncontroversially unacceptable.

Recall that for our judgment studies we had 48 distractor items. The goal of these items was to compare Tagalog speakers’ extraction behavior out of two syntactic islands (i.e., whether-islands and complex NP-islands). We crossed whether the verb exhibited AV or PV (Voice: av, pv) and whether the extraction site was the complement clause or the island (Site: Complement, Island). This is, to our knowledge, the first experimental investigation of island effects in Tagalog. In Figure 5, we visualize the distribution of participant ratings by whether the extraction was out of an island. What is relevant is that extractions out of complement clauses were rated high, while extractions out of an island were rated low. For the purposes of modeling the participants’ judgment process, we used the ratings for extractions out of complement clauses as the grammatical reference distribution. We then used the ratings for extractions out of islands as the ungrammatical reference distributions. See our OSF page for the supplementary material with a description and the Bayesian analyses of these experiments.

Figure 5: Distribution of ratings in experiment 2 by Voice, Site and island type. The left panel corresponds to extractions out of a complement; the right panel corresponds to voice-mismatch extractions. Blue corresponds to a rating of 7 ‘acceptable’, while gray corresponds to a rating of 1 ‘unacceptable.’

We followed closely how Dillon et al. generated the distributional predictions of these response models. See our OSF page for more information about how these distributions were generated under the two response models. As a metric for evaluating how well a given model fits the empirical distribution, we calculated the difference in the models’ Bayesian Information Criterion (BIC). Negative BIC difference scores indicate an advantage for the discrete model over the gradient model, while positive BIC difference scores indicate an advantage for the gradient model over the discrete model.

5.2 Judgment process in experiment 1

We pooled the observations across participants within conditions. We fitted two types of discrete and gradient models for each dependency: (i) patient-extractions when the verb had AV; and (ii) agent-extractions when the verb had PV. We ran 40,000 Monte Carlo simulations of the models. Below we report the mean BIC, and in square brackets, the minimum and maximum BIC values.

Head-initial RCs. When the verb had AV, 30,918 of the 40,000 simulations had a positive BIC (Mean = 4.27 [–129.3, 38.71]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 77% of the time. When the verb had PV, 40,000 of the 40,000 simulations had a negative BIC (Mean = –108.49 [–212.94,–53.97]). This suggests that the discrete model provided a better fit than the gradient model for the empirical distribution 100% of the time.

Wh-questions. When the verb had AV, 30,176 of the 40,000 simulations had a positive BIC (Mean = 4.57 [–165.6, 36.95]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 75% of the time. When the verb had PV, 39,999 of the 40,000 simulations had a negative BIC (Mean = –38.95 [–90.79, 3.34]). This suggests that the discrete model provided a better fit than the gradient model for the empirical distribution a little less than 100% of the time.

Head-final RCs. When the verb had AV, 28,980 of the 40,000 simulations had a positive BIC (Mean = 4.77 [–183.32, 45.26]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 72% of the time. When the verb had PV, 20,747 of the 40,000 simulations had a negative BIC (Mean = –0.5 [–183.32, 45.26]). This suggests that the discrete model provided a better fit than the gradient model for the empirical distribution about 52% of the time.

To summarize, gradient distributions best characterize the mismatch penalty in AV. This suggests that patient-extractions are better thought of as an overall decrease in acceptability from the grammatical reference distribution/an overall increase in acceptability from the ungrammatical reference distribution. Meanwhile, discrete distributions best characterize the mismatch penalty in PV—at least in head-initial RCs and wh-questions. This suggests that agent-extractions are better thought of as a mixture of categorical responses, sampling from both acceptable and unacceptable reference distributions.

5.3 Judgment process in experiment 2

We pooled the observations across participants within conditions. We fitted two types of discrete and gradient models for each dependency: patient-extractions and agent-extractions when the verb had applicative morphology. We ran 40,000 Monte Carlo simulations of the models. Below we report the mean BIC, and in square brackets, the minimum and maximum BIC values.

Head-initial RCs. When the patient was extracted, 31,386 of the 40,000 simulations had a positive BIC (Mean = 3.34 [–98.84, 28.97]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 78% of the time. When the agent was extracted, 40,000 of the 40,000 simulations had a negative BIC (Mean = –128.83 [–245.13, –74.56]). This suggests that the discrete model provided a better fit than the gradient model for the empirical distribution 100% of the time.

Wh-questions. When the patient was extracted, 30,627 of the 40,000 simulations had a positive BIC (Mean = 4.17 [–125.48, 34.36]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 77% of the time. When the agent was extracted, 40,000 of the 40,000 simulations had a negative BIC (Mean = –65.44 [–137.48, –18.74]). This suggests that the discrete model provided a better fit than the gradient model for the empirical distribution 100% of the time.

Head-final RCs. When the patient was extracted, 32,360 of the 40,000 simulations had a positive BIC (Mean = 5.95 [–171.84, 41.23]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 81% of the time. When the agent was extracted, 30,757 of the 40,000 simulations had a positive BIC (Mean = 4.97 [–211.36, 40.1]). This suggests that the gradient model provided a better fit than the discrete model for the empirical distribution about 77% of the time.

To summarize, when the verb has applicative morphology, gradient distributions best characterize the mismatch penalty involving patient-extractions. This suggests that patient-extractions are better thought of as an overall decrease in acceptability from the grammatical reference distribution/an overall increase in acceptability from the ungrammatical reference distribution. Meanwhile, discrete distributions best characterize the mismatch penalty involving agent-extractions—at least in head-initial RCs and wh-questions. This suggests that agent-extractions in head-initial RCs and wh-questions are better thought of as a mixture of categorical responses, sampling from both acceptable and unacceptable reference distributions.

5.4 Discussion

In this section, we developed quantitative models of the participants’ judgment process to estimate the underlying distribution of their voice-mismatch distribution ratings across experiments 1 and 2. In Table 4 we provide a summary of our simulations.¹⁸ The percentages reported indicate which model provided a better fit for the empirical distribution. When it is inconclusive, it is framed in terms of the discrete model.

Head-initial RCs and wh-questions. We found that the ratings distributions of voice-mismatch agent-extractions are better thought of as having discrete distributions. Speakers were either accepting or rejecting them in any given trial. For comparison, the ratings distributions of voice-mismatch patient-extractions are better thought of as having gradient distributions. Speakers were treating them uniformly, either as an overall downward shift in acceptability from the grammatical reference distribution or overall upward shift in acceptability from the ungrammatical reference distribution.

Head-final RCs. We found that voice-mismatch extractions—both agent- and patient-extractions—when the verb exhibits applicatives are better thought of as having gradient distributions. Voice-mismatch patient-extractions when the verb exhibits PV also had a gradient distribution. Speakers were treating them uniformly, either as an overall downward shift in acceptability from the grammatical distribution or overall upward shift in acceptability from the ungrammatical distribution. However, the simulations could not decide on whether the ratings distribution of voice-mismatch agent-extractions when the verb has PV is better thought of as having a gradient or discrete distribution.

One interpretation consistent with these findings is that there are two separable types of Tagalog speakers with respect to extraction: Tagalog A (those who conform to the strict version of the restriction) and Tagalog B (those who conform to the less strict version). We schematize these two types of speakers in Figure 6, along with a yet-to-be attested type of speaker, Tagalog C. Again, note that Tagalog A is a proper subset of Tagalog B, which in turn, is a proper subset of Tagalog C.

Figure 6: Visualization of the two Tagalogs with respect to extraction.

In this section, we have essentially identified a piece of grammatical knowledge (i.e., restrictions on extraction) that is systematic within an individual speaker but varies unpredictably across a population of Tagalog speakers. When these simulations are taken together with our own work with speakers and the previous descriptions about the variation in extraction restriction that Ceña & Nolasco (2011; 2012), Tanaka (2016), Hsieh (2019), Bondoc (2021) and others have observed, a hypothesis where there are two types of Tagalog extractors seems very tenable. We return to this in the general discussion.

6.1 Agent-first garden-paths Tagalog learners

In section 5, we interpreted the results from our simulations as being consistent with having two separable types of Tagalog speakers. A speaker of Tagalog A only allows voice-match extractions, conforming to the strict version of the extraction restriction. On the other hand, a speaker of Tagalog B allows what a speaker of Tagalog A allows, plus agent-extraction in voice-mismatch contexts. See Figure 6 for a visualization of these two types of Tagalogs.

A natural question to ask is why two types of Tagalog extractors develop based on the same or similar linguistic input. We suggest that the type of extractor a speaker of Tagalog ultimately becomes is a form of grammar competition in learning, in the sense of Han et al. (2007), that arises via the idea that an agent-first bias could affect how child learners parse input strings under noisy conditions. Because the core of our proposal draws heavily from language acquisition and sentence processing, we first give an overview of the intuition behind each ingredient. We will then discuss how the two types of Tagalog extractors emerge from the interaction of these ingredients.

The first ingredient is grammar competition in learning (Han et al. 2007). The main intuition behind this is that sometimes the primary linguistic data that a child in a language L is exposed to are compatible with at least two hypotheses, H₁ and H₂. When linguistic experience cannot adjudicate between the two, some learners acquire a grammar based on H₁ and others, based on H₂.

The second ingredient is the idea of a noisy-channel in acquisition. Production and perception errors are quite common in everyday language use. Noisy-channel models of sentence processing (e.g., Levy 2008) start with the assumption that some of the linguistic input we as comprehenders receive is noise. Thus, our task is to determine which parts of the linguistic input are considered noise and which parts are considered signal. Before children have fully acquired an adult-like grammar, they may parse some of the noise in the input as signal (Perkins et al. 2022). In other words, early in their language development, children may have a combination of signal that is correctly parsed as signal and noise that is parsed incorrectly as signal.

The third ingredient is the agent-first preference, a general bias to order the agent before any other arguments. In real-time comprehension, this manifests as a pressure to identify the most agent-like argument as quickly as possible. In other words, all else equal, assume that the first noun is the thematic agent (Ferreira 2003; Bornkessel & Schlesewsky 2006). This is an independently-motivated pressure that rears its head in many areas of the Tagalog grammar (Pizarro-Guevara & Garcia 2024). More on this in the subsection below.

Lastly, the fourth ingredient is a cognitively costly process called reanalysis (Frazier & Rayner 1982; 1987). In real-time comprehension, we make commitments to what the intended parse is even in the absence of fully disambiguating information. Sometimes, we are garden-pathed, and pursue an incorrect parse. Later on, when we are confronted with evidence against an incorrectly pursued parse, we must reanalyze and find a different analysis that is compatible with the current evidence we have. We know that children find abandoning their first analysis quite challenging, a phenomenon called the kindergarten-path effect (Trueswell et al. 1999). This has been argued to be a general cognitive limitation (Phillips & Ehrenhofer 2015).

Now let us consider how these four ingredients interact to give us the two types of Tagalog extractors. There are at least two hypotheses Tagalog children can consider with respect to what arguments can be extracted. These are provided in (22):

When the extracted argument is cross-referenced by voice, the linguistic output of these two hypotheses converge. Their output diverges only in contexts where the extracted argument is not cross-referenced by voice. However, these types of voice-mismatch extractions are rare in the input. For example, in an ongoing corpus analysis of Tagalog conversations, based on 105 minutes of recording and 22,663 words, Nagaya (2019) found 0 instances of voice-mismatch extractions. We do know that they exist, as we have seen examples of voice-mismatches “in the wild” in Section 2. We also know that child and adult speakers produce them in elicited production experiments (Tanaka 2016). With little or no evidence favoring one hypothesis over the other, different child learners may adopt different hypotheses. Some child learners might adopt H₁, ultimately leading to their becoming Tagalog A speakers. Some child learners might adopt H₂, ultimately leading to their becoming Tagalog B speakers. This is grammar competition in action.

Why would H₂ even be a part of the hypothesis space? Framed differently, why would some child learners give agents privileged status? We maintain that the agent-first pressure biases how comprehenders parse input strings under noisy conditions during acquisition. First consider how child learners arrive at the generalization that extractions need to, at the very least, be voice-matched. Recall that this corner of the extraction restriction is shared by both types of extractors. By hypothesis, when they were child learners, they must have correctly parsed the signal as signal. Perhaps this is unsurprising since this is presumably the type of input that they had strong positive evidence for. We argue that correctly parsing signal as signal is shared by learners who eventually become speakers of both Tagalog A and B.

How might child learners add the extra layer needed to allow agent-extractions in voice-mismatch contexts? Under a noisy channel assumption in acquisition, child learners also incorrectly parse noise in the input as signal. Our main claim is that the misparsing of noise to signal is caused by the agent-first bias. Consider child learners that encounter what would eventually be a head-initial RC, where the head noun functions as the patient of the verb inside the RC, as in (23a). This is the linguistic input that they are confronted with. Upon encountering the head noun katulong and the linker na, they may have realized that they are likely in a head-initial RC configuration. At this point in time, the head noun katulong is still role-ambiguous. They might be biased to interpret this as the agent of the upcoming predicate in the RC, per agent-first.

Upon encountering the RC verb sinipa, they might realize that their initial parse is inconsistent with the input, assuming voice-matching, and thus, it needs revision. In (23b), they are able to recover from their incorrect parse and reanalyze the role of the head noun. Thus, they are able to develop a veridical representation of the input. From these types of tokens, over time, child learners learn that when the verb has PV, patients are extractable (i.e., a voice-match extraction). We argue that successful reanalysis is shared by child learners who eventually become speakers of Tagalog A and B.

There is evidence in the RC-processing literature that is consistent with the idea that comprehenders temporarily misparse the head-noun as the agent, per agent-first, but successfully recover from that initial parse upon encountering more information (i.e., the verb and the disambiguating co-argument). Consider two minimally different examples from Pizarro-Guevara (2020) in (24); these are string-identical, save the verb. Many researchers have found that RCs involving a patient head-noun, as in (24b), are more difficult to process than RCs involving an agent head-noun, as in (24a) (Pizarro-Guevara 2014; Tanaka 2016; Bondoc et al. 2018; Pizarro-Guevara 2020). We can interpret this difficulty—at least, in part—as reanalysis cost. In (24b), comprehenders posited that the head noun is the agent, per agent-first. Upon encountering the verb (and coargument), their initial parse was confirmed. On the other hand, in (24b), comprehenders posited that the head noun is the agent, per agent-first. Upon encountering the verb (and coargument), they realized that their initial parse was incorrect and had to revise.

Opposite successful reanalysis, an alternative scenario involves children’s failure to reanalyze their initial parse. Sometimes, even after encountering the verb, child learners might be unable to let go of their initial parse—a common feature of child sentence processing that many researchers have documented before (Trueswell et al. 1999; Phillips & Ehrenhofer 2015). Consider again the linguistic input that child learners are confronted with in (23a), repeated here as (25a). Upon encountering ng bata, they may try to revise the parse again, given that they now have more evidence in the input. Or they may try to save their initial parse by assuming that their interlocutor had intended to say ang bata [ʔaŋ.ba.taʔ]—instead of ng bata [naŋ.ba.taʔ]. The distortion of the input seems plausible enough since ang and ng are phonetically very similar; the only difference is the onset. One advantage that this distortion affords the child learners is that they get to form a locally coherent parse with the string sinipa ang bata and crucially, the locally coherent parse—even though globally ungrammatical from the perspective of a strict version of the restriction—allows them to maintain their initial parse, where katulong is the agent. After accumulating enough of these misparsed noise, child learners might transform these tokens into bona fide signal. Over time, these child learners might learn that when the verb has PV, agents are also extractable (i.e., a voice-mismatch extraction). We argue that failing to reanalyze and then reclassifying noise as signal is a characteristic of child learners who eventually become speakers of Tagalog B.

There is evidence in the RC-processing literature that is consistent with the idea that children initially parse head-nouns as agents and that they sometimes fail to abandon their initial parse (i.e., failing to reanalyze despite more evidence in the input). Thematic reversals are the most common errors children make. In particular, they have a bias to interpret the head noun as the agent of the predicate inside the RC. For example, Tanaka et al. (2019) found that 5 to 6 year olds interpreted an RC where the head noun was the patient as an RC where the head noun was the agent three times more often than the other way around. In other words, when the head noun was a patient, children often interpreted this as an agent. Misinterpreting an agent head noun was significantly less common. These results are corroborated by an earlier study by Pizarro-Guevara (2014): the rate of thematic reversals 4 to 9 year olds made was higher on average when it was patient-to-agent (i.e., a patient head noun being misinterpreted as an agent head noun) than when it was agent-to-patient (i.e., an agent head noun being misinterpreted as a patient head noun). Furthermore, apart from the naturalistic data presented in section 2.1, there is evidence that speakers—both children and adults—do produce structures like (25b), as shown by Tanaka (2016) in RC-production studies. The production of such structures is consistent with the idea that learners may have failed to reanalyze, and that they may have reclassified noise as signal.

To summarize, we argued that the language development of a speaker of Tagalog A is characterized by the processes in (26). On the other hand, the language development of a speaker of Tagalog B is characterized by the processes in (27).

A natural question to ask now is how the proposal developed above can account for the participants’ behavior in head-final RCs. Recall that voice-mismatches are not allowed—even for those who do allow them in head-initial RCs. In head-initial RCs, schematized below in (28a), upon encountering the first noun, this noun is still role-ambiguous. Only upon encountering the verb do comprehenders know its role (assuming voice-matching), as discussed above. On the other hand, in head-final RCs, schematized in (28b), upon encountering the first noun, this noun is no longer role-ambiguous. Comprehenders will have already encountered the verb with voice morphology and the case marker of the noun, which they can then use to infer its role. We claim that this lack of temporal role-ambiguity is the reason why the effect of agent-first is attenuated.

There is evidence in the literature consistent with the idea that the garden-path effect caused by the agent-first bias is attenuated in head-final RCs, compared to its effect in head-initial RCs. For example, Pizarro-Guevara (2020) compared how head-initial and head-final RCs were processed, using items like (29). He found that in head-initial RCs, as in (29a), RCs with a patient head-noun were more difficult to process than those with an agent head-noun. However, in head-final RCs, as in (29b), there was no evidence that one was more difficult than the other.

In this section, we presented a way for us to think about how the two types of Tagalog extractors came to be by drawing insights from language acquisition and sentence processing. We placed the locus of variation on the agent-first bias. This may seem like an entirely ad hoc proposal. However, as we will see in the next subsection, the agent-first bias has a far-reaching influence in the grammar of Tagalog.

6.2 Agent-first is at the epicenter of Tagalog grammar

The agent-first bias has been argued to be rooted in an evolutionary bias to attend to agentive features and is central in general event perception and cognition (Sauppe et al. 2023). In the proposal above, it was at the epicenter. Here, we exemplify how it is an independently-motivated pressure that seems to have a far-reaching influence in Tagalog grammar (Pizarro-Guevara & Garcia 2024).

First, it influences the post-verbal word order in declarative sentences, along with two other pressures: the ang-last and heavy NP-shift, the tendency to have “heavier” elements follow “lighter” elements (Kroeger 1993). For example, when the verb has PV, the verb-agent-patient order, as in (30a), is reported to be more natural than the verb-patient-agent order, as in (30b). The former satisfies both agent-first and ang-last, holding heaviness constant, while the latter violates both. On the other hand, when the verb has AV, both verb-agent-patient, as in (30c), and verb-patient-agent, as in (30d), have been reported to be equally natural. This is corroborated by experimental evidence from judgment studies (Hsieh 2016), sentence completion (Bondoc & Schafer 2022), and the visual world paradigm (Sauppe 2016).

Second, agent-first also plays a role in Tagalog speakers’ interpretation of globally ambiguous RCs. Using a picture selection task, Pizarro-Guevara (2020) showed that when participants were presented with a globally ambiguous head-initial RC, as in (31a), they interpreted the head noun as the patient 30% of the time, and as the agent 70% of the time. In other words, there was a strong preference to interpret the head noun of a globally ambiguous head-initial RC as the agent. By contrast, when they were presented with a globally ambiguous head-final RC, as in (31b), they interpreted the head noun as the patient 50% of the time, and as the agent 50% of the time. Unlike in head-initial RCs, there was no agent-first bias in how they interpreted the head noun of a globally ambiguous head-final RC.

Finally, from a developmental perspective, the centrality of agent-first in Tagalog is also evident in the type of errors that children make when comprehending different types of sentences. In the previous subsection, we saw that children made thematic reversal errors. They were more likely to misinterpret a patient head-noun as the agent than misinterpret an agent head-noun as the patient (Pizarro-Guevara 2014; Tanaka 2016). Similar reversal errors are also found when comprehending verb-initial sentences with two arguments in post-verbal position, as indicated by their accuracies in picture-verification (Garcia et al. 2019) and in picture-selection tasks (Garcia et al. 2020).

Lastly, the agent-first bias is not unique to Tagalog. In fact, it can be observed cross-linguistically in both child and adult languages: Äiwoo (Sauppe et al. 2023), Basque (Erdocia et al. 2009), English (Ferreira 2003), German (Haupt et al. 2008), Hindi (Bickel et al. 2015), Mandarin (Huang et al. 2013), Ojibwe (Hammerly et al. 2022), Spanish (Gattei et al. 2015), and Turkish (Demiral et al. 2008), to name a few. These studies suggest that agent-first is a cross-linguistically robust pressure, and thus, it should be unsurprising for it to also influence/be leveraged by Tagalog learners and speakers.

6.3 A bifurcation in the results

Irrespective of dependency type, we found evidence for a mismatch penalty when (i) extracting an agent not cross-referenced by voice, and (ii) when extracting a patient not cross-referenced by voice. Where the dependencies differed is when we consider the difference in mismatch penalties. That is, the difference between (i) and (ii) above. In other words, there is a bifurcation in their difference of differences. We see that in head-initial RCs and for the most part, wh-questions, the mismatch penalty in agent-extractions is smaller than in patient-extractions. Meanwhile, in head-final RCs, the mismatch penalties in agent-extractions and patient-extractions are not different. We also see this bifurcation in our simulations: head-initial RCs and wh-questions pattern in one way, while head-final RCs pattern in another way.

A natural question to ask is why head-initial RCs and wh-questions would form a natural class and exclude head-final RCs. We offer two potential reasons—and they need not be mutually exclusive. The first reason is structural similarity. It is well-established in the literature that argument wh-questions in Tagalog are derived via a (pseudo-)clefting strategy, which means that the interrogative phrase serves as the predicate (i.e., as a predicate nominal) and the remaining material is a headless relative clause (Kroeger 1993; Aldridge 2002). Their similarity is schematized in (32):

As briefly mentioned in section 1.2, it is an open question whether head-initial and head-final RCs have different derivations (Aldridge 2017) or they share a single one (Law 2016). If Aldridge is on the right track, head-final RCs are derived differently from head-initial RCs (and by extension, argument wh-questions). Thus, head-initial RCs and wh-questions can form a natural class and exclude head-final RCs.

The second reason is processing similarity. As discussed in section 6.1, in head-initial RCs and wh-questions, Tagalog comprehenders are first confronted with the filler (i.e., the argument displaced from the position where it is interpreted), followed by the gap (i.e., the position where it is interpreted). In a sense, comprehenders engage in a forward search for where to interpret the displaced argument. In contrast, in head-final RCs, Tagalog comprehenders are confronted with the gap first, followed by the filler. In a sense, comprehenders engage in a backward search for where to interpret the filler. Viewed this way, we have another reason as to why head-initial RCs and wh-questions can form a natural class and exclude head-final RCs.

In the proposal above, we have alluded indirectly to the processing similarity in head-initial RCs and argument wh-questions. In these dependencies, they are confronted with the filler first, which accentuates the garden-path effect caused by the agent-first bias.

6.4 Implications for theories of extraction

Before we conclude, we briefly discuss the implications of our current findings for theories of A-bar extraction in Austronesian languages. In many of the formal accounts in the literature (Aldridge 2002; Rackowski & Richards 2005; Aldridge 2012), the argument cross-referenced by voice is syntactically privileged because it is the highest DP in the structure. We use “highest DP approach” as an umbrella term for this class of proposals. Even though the finer details of each proposal are different, they have the following features in common. First, AV and non-AV voices are structurally different, and we provided partial derivations in (11) and in (12), repeated here as (33) and (34), for AV and for PV (which we take as representative of non-AV voices), respectively. In PV, the patient vacates the VP and moves to the outermost specifier of vP, above the agent.

Second, the highest DP receives ang-marking. This captures the generalization that when the verb has AV, the agent is ang-marked, and when the verb has PV, the patient is. Third (and perhaps most importantly), voice-match extraction is accounted for using Attract Closest. The basic idea is that the highest DP is the closest goal for an A-bar probe. With this formal machinery in place, the strict version of the restriction is a natural consequence. If an A-bar probe like F were to search for a goal, as in (35), only the agent is visible to the probe since it is the highest DP in the structure and is thus, the closest goal. This effectively allows agent-extractions under AV (+match) and rules out patient-extractions under AV (–match). If an A-bar probe like F were to search for a goal, as in (36), the patient, which has moved to a position higher than the agent, has become the highest DP and has become the closest goal. This effectively allows patient-extractions under PV (+match) and rules out agent-extraction under PV (–match).

The highest DP approach, in its current form, undergenerates. It cannot generate agent-extractions under non-AV voices for those who do allow them. The present study has demonstrated that these speakers do exist. There are (at least) two ways in which this approach can be modified to allow these structures to be generated: (i) modify the calculus of how the distance between the goal and the probe is determined; or (ii) posit another operation that places the agent in a structurally more prominent position after the patient has vacated the VP. In a recent proposal that takes into account these voice-mismatch structures, Hsieh (2023) makes a case for (ii) and argues that agents that are not cross-referenced by voice undergo what he calls genitive inversion, an independently attested operation in the language, which places the agent in a structurally more prominent position. This operation, in turn, feeds A-bar extraction.

A more thorough consideration of Hsieh’s proposal is outside the immediate scope of this paper. A main feature of his system is how intricately tied agent-extraction under NAV is to what he calls genitive inversion. Thus, his system would predict that the extent to which speakers allow agent-extraction under non-AV voices is tightly correlated with the extent to which they allow genitive inversion in their Tagalog. We leave it to future work to investigate the extent to which this prediction obtains.

There are also proposals in the literature, which we subsume under the umbrella term “in situ approach,” where the argument cross-referenced by voice need not be the highest DP in the structure (Chen 2017; Aldridge 2018). The finer details of each proposal differ, but they share some core features. First, of particular relevance is that in NAV, the patient does not occupy a structural position higher than the agent. The patient remains in-situ, irrespective of voice. Second, whatever proposal-specific mechanism is employed to determine voice is the same mechanism used to account for voice-mismatch extractions. Framed differently, whatever argument is agreed with by the probe that determines voice in Aldridge’s system—or probes in Chen’s system—is the same argument that satisfies the feature responsible for extraction. In AV, this argument is the agent and thus, the agent is extractable. In PV, this argument is the patient and thus, the patient is extractable. One way to modify the in situ approach is to decouple the mechanisms responsible for voice determination and for implementing the extraction restriction. They need not be a single process. That is, they could be separate but concurrent processes instead. It is outside the immediate scope of this paper to provide a detailed treatment of these proposals and to fully explore the theoretical ramifications of decoupling the two mechanisms.

6.5 Conclusion

To summarize, we re-examined the extraction restriction in Tagalog. We situated our investigation in the context of the received view of the restriction and provided an alternative generalization. We provided naturally occurring data of voice-mismatch extractions. Second, we replicated the findings of Pizarro-Guevara & Wagers (2020) in broad strokes, and found that voice-mismatch extractions incurred greater mismatch penalties in AV than in PV. Third, we extended their findings and found that extracting the patient incurred greater mismatch penalties than extracting the agent when the verb had applicative morphology. Together, these results strongly suggest that the restriction is not as clear-cut as previously described. At the very least, the generalizations we present should figure in any theory of how voice interacts with A-bar extraction in the language.

We also developed quantitative models that estimated the participants’ judgment process and the results suggest that participants were treating agent-extractions under PV and applicatives categorically. This is a novel application of the type of computational modeling that Dillon et al. (2017) performed for an unrelated phenomenon in an unrelated language (i.e., agreement attraction in English). Our results can thus be viewed as further proof of concept for that method and how a distributional analysis of ratings can provide evidence of multiple types of speakers in a language community.

We identified a piece of grammatical knowledge (i.e., extraction) that is systematic within an individual speaker but varies unpredictably across a population of Tagalog speakers. We proposed that this is a form of grammar competition that arises via the idea that the agent-first bias, an independently motivated pressure in Tagalog that also has cross-linguistic basis, is at the epicenter of this variation. We advanced the claim that it affects how child learners parse input strings under noisy conditions during acquisition.

If we take seriously the claim that there are two types of Tagalog extractors, we should expect to find some evidence of their emergence in how children judge voice-mismatch extractions. There are studies that examine how children judge sentences involving extraction, mainly involving head-initial RCs (Pizarro-Guevara 2014; Bondoc et al. 2018; Tanaka et al. 2019). Future work can investigate how children judge sentences when the extracted argument is not cross-referenced by voice. A comparison of head-initial and head-final RCs would be particularly fruitful. Future work can also systematically investigate if there are any sociolinguistic correlates of this point of variation or whether this truly is stochastic in nature.