Glossa Psycholinguistics

1.1 The importance of the passive

Passive constructions have long played a central role in linguistic theorizing and psycholinguistic research. On the one hand, passives have been used to argue for the autonomy of syntax (see, e.g., Alexiadou et al., 2018, for a historical review), since they reverse canonical participant role-ordering (i.e., PATIENT ACTION AGENT), yet still maintain the language’s canonical word order in terms of syntactic (as opposed to semantic) roles (e.g., SUBJECT VERB OBJECT for English). This line of argument dates as far back as at least Chomsky (1957), but also holds for non-Chomskyan “simpler syntax” accounts (e.g., Culicover & Jackendoff, 2005; Pollard & Sag, 1994) that posit a “syntactic level of representation [that] includes syntactic category information but not semantic information…or lexical content” (Branigan & Pickering, 2017, p. 8). Indeed, some frameworks go so far as to assume that the passive construction enjoys no independent existence, and that passives are derived from other sentence types or from basic principles:

Constructions such as…[the] passive remain only as taxonomic artifacts, collections of phenomena explained through the interaction of the principles of UG, with the values of the parameters fixed. (Chomsky, 1993, p. 4).

On the other hand, passives have been used as evidence for a tight link between syntax and function/semantics. For example, passives are used more often when only one of the noun phrases is animate, when the verb is of the theme-experiencer type (e.g., John was startled by a loud noise…; Ferreira, 1994) and when it allows an easily-retrievable (given) referent to be in subject position (…but he was reassured by his travelling companion; MacDonald, 2013).

In terms of the semantics of the (English) passive, one particularly influential proposal (Pinker et al., 1987; see also Pinker, 1989; Pullum, 2014) is as follows:

“[B] (mapped onto the surface [passive] subject) is in a state or circumstance characterized by [A] (mapped onto the by-object or an understood argument) having acted upon it”. (p.249).

Thus, a passive such as The pedestrian was run over by a bus is prototypical, while a passive such as *$5 was cost by the book is ungrammatical, since it is not possible to construe $5 as having been “acted upon” – even metaphorically – by the book. Theoretically speaking, the notion of a semantically-based passive construction is perhaps most strongly associated with cognitive-linguistic approaches, which view semantically-meaningful constructions in general (i.e., not just the passive) as a central tenet (e.g., Croft & Cruse, 2004; Goldberg, 1995; Hilpert, 2014). However, functional/semantics-based approaches to the passive exist in other theoretical frameworks too; and not only in the related approach of Lexical Functional Grammar (e.g., Pinker, 1989). For example, Aissen (1999) sets out a theory of subject choice in Optimality Theory, while Paolazzi et al. (2022) locate the (apparent) semantic difficulty with experiencer-theme passives (e.g., John was seen by Bill) in event structure, rather than syntax per se (specifically, the difficulty arises when coercing a state into receiving an eventive interpretation). Similarly, Nguyen and Pearl (2021) situate this effect – focussing on children – in lexical semantics.

While the theoretical implications remain debatable – and debated – the finding of something like Pinker’s semantic constraint on the (English) passive appears to be well supported. In the past decade, a decent number of adult studies have tested this proposal, and generally found evidence in its favour (Ambridge et al., 2021; Bidgood et al., 2020; Jones et al., 2021; though for partial null findings, see Darmasetiyawan et al., 2012; Messenger et al., 2012; for a meta-analysis of child studies, see Nguyen & Pearl, 2021).

Most relevant to the present work is the English study of Ambridge, Bidgood, Pine, Rowland and Freudenthal (2016): A group of native adult speakers were asked to rate each of 72 verbs for the extent to which each exhibits each of 10 semantic properties related to Pinker et al.’s (1987) notion of “affectedness” (e.g., A is doing something to B; B changes state or circumstances; the action adversely (negatively) affects B). These ratings were then combined using Principal Components Analysis to create a composite “affectedness” score for each verb. (Note that here and throughout this paper, we are using the term “affectedness” simply as a shorthand for the cluster of semantic properties that are characteristic of the passive construction, even though some are only tangentially related to affectedness per se; an issue to which we return in more detail in 2.1). The researchers then obtained, from a separate group of adult speakers, grammatical acceptability judgments of each verb in passive and, as a control, active sentences (e.g., Bob was kicked by Wendy; Wendy kicked Bob). For passive sentences, a large and significant correlation between semantic affectedness and sentence grammatical acceptability was observed. For example, the verbs hit, kick, push and shake all received close to the maximum possible semantic affectedness score, and close to the maximum possible acceptability scores in passive sentences (e.g., Bob was hit/kicked/pushed/shaken by Wendy). Conversely, the verbs like, remember, know and believe received low semantic affectedness scores, and lower acceptability judgment scores in passive sentences (e.g., Bob was liked/remembered/known/believed by Wendy).

Complicating the picture somewhat, a similar – though significantly smaller – correlation was observed for active sentences. That is, participants also displayed a small preference for sentences such as Wendy hit/kicked/pushed/shook Bob over sentences such as Wendy liked/remembered/knew/believed Bob. There are two possible explanations for this semantic affectedness effect for actives; though the two are not mutually exclusive. The first is that the active transitive construction is – like the passive – also prototypically associated with a cluster of semantic properties related to affectedness (e.g., Ambridge et al., 2014; Hopper & Thompson, 1980; 1984; Ibbotson et al., 2012; Næss, 2007; Talmy, 1985). For example, Ibbotson et al. (2012, p. 1270) characterize the semantics of the active transitive construction as an “agent intentionally instigating an action that directly results in the patient being affected” (Ibbotson et al., 2012, p. 1270). If this possibility is correct, then we would expect to see semantic affectedness effects observed for acceptability judgments with both passives and actives; and possibly even similar-size effects for the two constructions (i.e., no interaction of semantic affectedness by sentence-type passive/active).

That said, at least assuming a construction-based/functionalist approach, languages clearly “need” a two-participant construction that does not (necessarily) denote high affectedness (e.g., to convey a scenario in which Wendy pushed Bob, but so gently that he barely noticed). Thus, for every language that shows a semantic affectedness effect for passives, we would expect to see a smaller, and close to zero, semantic affectedness effect for at least one of (a) actives and (b) pseudo-passives (defined in more detail below).

The second possible explanation for this semantic affectedness effect for actives is that it is an experimental artefact resulting from some kind of general dispreference and/or processing difficulty for verbs that score low on affectedness. For example, higher-affectedness verbs (e.g., push, bite, hit) tend to show higher scores on measures of concreteness, imageability or perceptual strength (e.g., Connell & Lynott, 2012) than lower-affectedness verbs (e.g., know, miss, remember), which tend to be more abstract, and generally refer to mental states rather than actions. If this possibility is correct, then even if the active transitive construction has no particular semantics, we would expect to see an (apparent) semantic affectedness effect for actives, merely by dint of the fact that low-affectedness verbs receive lower grammatical acceptability ratings across the board. To rule out this possibility, we would need to see, for every language that shows a semantic affectedness effect for passives, a smaller semantic affectedness effect for at least one of (a) actives and (b) pseudo-passives (defined in more detail below).

The original English study of Ambridge et al. (2016) did find evidence for an interaction of affectedness by construction type, such that the (continuous) semantic affectedness effect was greater for passives than actives, even when controlling for the frequency of each verb in passive and active constructions (no pseudo-passive construction was included). Likewise, in the present study, we conclude that the passive construction has the meaning of affectedness only when we see a significantly larger affectedness effect for passive than pseudo-passive and/or active constructions.

1.2 Replications of Ambridge et al. (2016, English) in Indonesian, Mandarin and Balinese

Subsequently, three replication studies of Ambridge et al.’s (2016) English study have been conducted for different languages. First, for Indonesian, Aryawibawa and Ambridge (2018) found semantic affectedness effects for both passives (1) and actives (2), but not for a noncanonical “pseudo-passive” construction (3) that shares passive word order, but not the morphosyntactic hallmarks of the passive construction, specifically the verbal passive prefix di- and the preposition oleh, ‘by’.

Second, for Mandarin, Liu and Ambridge (2021) found semantic affectedness effects for passives (4) and BA-actives ((5); a dedicated affectedness construction), but not for standard actives (6), or noncanonical “pseudo-passives”/“notional passives” (7), which again share passive word order, but lack the passive morphosyntactic marker, bei (le marks completive aspect).

Third, for Balinese, Darmasetiyawan and Ambridge (2022) found semantic affectedness effects for actives (8) and three different types of passives (-a, ka- and ma- passives; (9–11)), but again not for noncanonical “pseudo-passives”/“basic passives” (12), which lack passive morphosyntax (-a/ka-/ma-), but do include the preposition teken, ‘by’.

The differences between -a, ka- and ma- passives (essentially whether the agent is volitional, non-volitional or unimportant) are not relevant for our purposes here; we treat them all as bona-fide morphosyntactic (as opposed to pseudo-) passives.

Thus, a crosslinguistic pattern seems to be emerging, whereby semantic affectedness effects are always observed for genuine passives, but are attenuated – or absent altogether – for “pseudo-passives” (also called “noncanonical” or “basic” passives in the original papers) – topicalization constructions that share passive word order but lack passive morphosyntax (e.g., Keenan & Dryer, 2007) – and, to a lesser extent, actives (which generally seem to share some degree of affectedness semantics with passives).

In the present study, we extend this paradigm to Hebrew, and thereby to a fourth language family (Indonesian and Balinese are both Austronesian languages). We then conduct a Bayesian mixed-effects meta-analytic synthesis to investigate whether, across languages, semantic affectedness are (i) observed when looking only at passives (a relatively lenient criterion) and – crucially – (ii) bigger for passives than at least one of (a) pseudo-passives and (b) actives (a strict criterion; see Section 1).

2.2 Results and discussion (Study 1: Hebrew)

All data and analysis scripts are available at https://osf.io/hwb26/. Figure 1 plots the raw sentence acceptability ratings on the 5-point scale (Y axis) against the scaled and centred continuous semantic predictor (X axis) for the present Hebrew data (and – for comparative purposes – for Balinese, English, Hebrew and Mandarin). Inspection of this plot suggests that, consistent with some of these previous studies, a semantic affectedness effect is observed for passives, but not for actives.

Figure 1: Semantics effects for Hebrew (also Balinese, English, Indonesian and Mandarin).

In order to test this observation statistically, we ran a series of mixed effects models with verb and participant as random factors. The main analysis used Bayesian models (brms package; Bürkner, 2018), in R (R Core Team, 2022), with a cumulative link function [family = cumulative()], reflecting the fact that the acceptability judgement data are on an ordinal scale (this reflects a methodological improvement over the previous papers reviewed above, which treated the judgment data as continuous). Following Bürkner and Vuorre (2019), we used a wide, flat, Gaussian prior of (0,5), with all predictors centred and scaled (there is an argument for instead using a Dirichlet prior based on the anticipated distribution of ratings across the five response categories, but this is not implemented in brms). Following the suggestion of an anonymous reviewer, we adopted fully-maximal random effects structure (in the sense of Barr et al., 2013) for all models, in order to ensure maximum generalizability (particularly for the subsequent meta-analysis). That is, we included all random intercepts and slopes that are justified given the design. In any case, frequentist model selection using the MixedModels.jl package (Bates, 2016; Bates et al., 2016) in Julia (Bezanson et al., 2012) revealed that, for the main semantics-only model, a fully-maximal model was indeed optimal.

Response ~ Sentence_Type*Semantics + (1+ Sentence_Type*Semantics|Participant) + (1+Sentence_Type|verb)¹

where “Response” is participants’ ratings on the 5-point scale, “Sentence_Type” is the sentence being rated [active = 1, passive = 0] and “Semantics” is the scaled and centred continuous semantic affectedness predictor. (Note that in the actual R syntax and output tables reproduced below, Sentence_Type is abbreviated to S_Type). However, for the secondary semantics-and-frequency model, frequentist model comparison revealed that a slightly simpler model was preferred according to both the AICc and BIC criteria (Matuschek et al., 2017):

Response ~ Sentence_Type*Semantics + Sentence_Type*Frequency + (1+ Sentence_Type*Semantics|Participant) + (1+Sentence_Type|verb)

(where “Frequency” is the scaled and centred continuous chi-square verb bias predictor). However, as noted above, to ensure maximum generalizability we did not run this model, but the maximal model with a by-participant random slope for the interaction of Sentence_Type by frequency:

Response ~ Sentence_Type*Semantics + Sentence_Type*Frequency + (1+ Sentence_Type*Semantics + Sentence_Type*Frequency|Participant) + (1+Sentence_Type|verb))

All Bayesian models were run for 5,000 iterations with a 2,000 iteration warm-up, with 16 chains running on 16 cores, with adapt_delta set to 0.99.

The main semantics-only model and the secondary semantics-and-frequency models are shown in Tables 3 and 4, respectively. The credible intervals shown are 95% intervals for the main model terms (which are two-sided) and 90% intervals for the one-sided hypothesis tests (shown in bold and explained below). Posterior probabilities are shown in the column “B <> 0”. Although an advantage of the Bayesian approach is that it discourages dichotomous statistical thinking, readers who want to ask “Is this effect significant?” may interpret these values as “Bayesian p values” by mentally subtracting them from 1. Indeed (and perhaps somewhat ironically given the “statistics wars” between frequentist and Bayesian approaches), with a completely uniform prior and a one-sided test, Bayesian posterior probabilities (1–X) and frequentist p values are numerically identical (Marsman & Wagenmakers, 2017).

Since we used treatment coding (active = 1, passive = 0), the estimate for “Semantics” represents the estimated effect of the semantic affectedness predictor for passive sentences. The estimate for the “S_TypeActive:Semantics” interaction represents the difference between the effect of the semantics predictor for actives versus passives. This term is large and in the predicted (negative) direction for both models, indicating that the effect of semantics is greater for passive than active sentences. Note that this term represents the difference between the semantics effects for passives versus actives, not an estimate of the semantics effect for either passives or actives independently. In order to obtain these estimates, we used the “hypothesis” function of brms to perform a directional (i.e., one sided) test of the hypothesis that the effect of semantics is greater than zero for (a) actives and (b) passives (although this is just a one-sided version of the estimate for passives already described; note that the estimates are identical). Conceptually, this is equivalent to the more common practice of breaking down a significant interaction by running separate follow-up models for – in this case – (a) active sentences only and (b) passive sentences only. However, the present method of estimating the relevant contrast directly from the main model is preferable (e.g., Schad et al., 2020) for two reasons. First, estimation of variance and random effects terms is more stable compared to fitting individual models to subsets of the data (Schad, personal communication), since – in the latter case – each model is fitted to only half of the data, which decreases the reliability of parameter estimates (e.g., Kerkhoff & Nussbeck, 2019). Second, running separate models on subsets of the data increases the number of tested effects. Thus, unless we are rigorous about ignoring any effect that was not specifically predicted, we run the risk of turning up – and seeking to interpret – spurious chance findings. These values are shown in the bold rows of Tables 3, 4. This analysis revealed that the semantics slope is large and in the predicted direction for both the semantics-only (Table 3) and semantics-and-frequency (Table 4) models (i.e., whether or not frequency is included in the model). Indeed, the correlation between the frequency predictor and the semantic affectedness predictor was only r = 0.1, which explains why controlling out this factor makes little difference.

Following the same logic, this analysis also revealed that the effect of frequency (see Table 4) is (a) moderate for passives (row “Frequency” for the two-sided test; row “Frequency_Slope_for_Passive” for the one-sided test), (b) marginally smaller for actives than passives (row “S_TypeActive:Frequency), and (c) essentially zero for actives (“Frequency_Slope_for_Active”, one-sided test).

Before summing up, we pause to consider a potential objection; that our semantics manipulation is confounded with concreteness – or perhaps perceptual strength (e.g., Connell & Lynott, 2012) – which affects ease of processing. It is certainly true that our high-affectedness verbs (e.g., push, bite, hit) would also score highly on measures of concreteness, imageability or perceptual strength, while our low-affectedness verbs (e.g., know, miss, remember) would not. However, this shows the value of including active sentences (and, for some other languages, pseudo-passive sentences) as a control condition. For Hebrew, the correlation between affectedness and active sentence acceptability is not even in the direction one would expect if affectedness were a proxy for concreteness-related ease of processing. Thus, it does not seem that low-affectedness verbs show low acceptability in the passive merely because their low concreteness induces processing difficulties.

In summary, the findings for Hebrew were, if anything, even more clear-cut than for the languages previously studied: Verb-by-verb semantic affectedness ratings predict verbs’ acceptability in passive, but not active, sentences; and attempting to control for verbs’ relative frequency in the passive construction makes very little difference. To investigate the extent to which this holds true across languages, we now turn to our meta-analytic synthesis.

3.1 Individual languages model

Before proceeding to this all-languages model, however, we first built a separate model for each language. This allows us to “correct” for some small differences between the analyses in the published papers such as different priors, different ways of estimating separate effects for different sentence types, different instantiations of the frequency predictor and (for Balinese) the use of a 10-point, rather than a 5-point, scale (the ratings are re-scaled for the analyses presented here). We also used a cumulative link function throughout, reflecting the fact that the acceptability judgement data are on an ordinal scale (a methodological improvement over the original analyses, which treated the judgment data as continuous). As already noted above, we used maximal random-effects structure throughout, in order to ensure maximum generalizability. Treatment coding was used throughout with Passive designated as the reference category (i.e., 0). This allows us to test whether, as compared with Passives, the effect of Semantics (or Frequency) is smaller for (a) Actives and (b) Pseudo Passives; via the individual interaction terms S_TypeActive:Semantics and S_TypePseudo_Passive:Semantics, respectively (for Frequency, S_TypeActive:Frequency and S_TypePseudo_Passive:Frequency). In order to allow for comparable analyses across languages and sentence types, we collapsed Balinese -a, ka- and ma- passives as Passive (a supplementary model retaining the different sentence types is available on the OSF project site).

As for the Hebrew models described earlier, we again ran separate semantics-only and semantics-and-frequency models. The logic of including frequency here is to control for its effect, in order to allow us to investigate the effect of semantic affectedness above and beyond any effect of frequency (at least in principle; see Westfall & Yarkoni, 2016, for caveats). In order to be maximally conservative in this regard, we always include frequency in each semantics-and-frequency model, regardless of whether or not it reaches any particular statistical criterion. However, the semantics-and-frequency models must be interpreted with caution, since – in principle – there is likely to be collinearity between the semantic affectedness and frequency predictors (i.e., speakers tend to use verbs in only semantically compatible constructions). In practice, this correlation is worryingly high only for Mandarin (r = 0.48) and, to some extent, English (r = 0.26). Only a marginal degree of collinearity is present for Hebrew (r = 0.10), and none for Indonesian (r = –0.01), with the caveat that the frequency counts for Indonesian are sparse and likely rather unreliable.

The findings of these analyses are summarized in Figure 2 (for full models, see Tables 5, 6). Figure 2 summarizes the hypothesis tests which – analogous to those performed for Hebrew above – estimate the effect (i.e., slope) of semantics for each sentence type individually. Despite the differences noted above, it is clear that Figure 2 generally recapitulates the findings outlined in the original papers, with non-zero (“significant”) effects of semantics observed for all bona-fide passives (i.e., not pseudo-passives) in all languages. Indeed, for every language, the confidence interval for the effect of semantics for passive sentences always excluded zero, and the B<>0 value was always 1 (i.e., all posterior samples were greater than zero), except for Indonesian, which was close (CI = [–0.01,0.55]; B<>0 = 0.95; CI = [0,0.55]; B<>0 = 0.95; for the analyses excluding and including frequency, respectively).

Figure 2: Semantics (and frequency) effects by language and sentence type.

For pseudo-passives (i.e., Balinese “basic” passives, Indonesian “noncanonical” passives and Mandarin “notional” passives), the slope was always in the opposite direction to that for passives (see Figure 1). Indeed, as shown in Tables 5, 6, the interaction term S_TypePseudo_Passive:Semantics was always negative and non-zero (i.e., “significant”), indicating that the effect of semantics was always smaller for pseudo- than genuine passives.

For actives, the estimate of the semantics slope was – with the exception of Balinese – always (a) very close to zero and (b) “significantly” smaller for actives than passives (see the interaction term S_TypeActive:Semantics), though only marginally so for Indonesian. However, Balinese bucked the trend by showing a “significantly” larger effect for actives than passive.

Before moving on to the all-languages meta-analytic synthesis, it is important to clarify the extent to which the findings of the present by-languages reanalyses differ from – or are similar to – those presented in the original papers (though differences in the statistical modeling approaches mean that there is rarely a one-to-one correspondence per se).

• For Balinese, Darmasetiyawan and Ambridge (2022) found no evidence that the effect of semantics differed between actives (the reference category) and any type of passive (-a, ka- and ma-; the three were analyzed separately). In contrast, the present reanalysis found that the effect of semantics was (counter to expectations) bigger for actives than for (pooled) passives. The present reanalysis also found that the effect of semantics was (as expected) bigger for (pooled) passives than for pseudo-passives. No corresponding analysis was included in Darmasetiyawan and Ambridge (2022) (since actives were always the reference category). Looking at individual sentence types, Darmasetiyawan and Ambridge (2022) found semantic effects for actives and -a, ka- and ma- passives (the three were analyzed separately); but not pseudo passives. These findings were replicated in the present analyses, both when pooling a, ka- and ma- passives, and (in the supplementary analysis available at the OSF project site) analysing each separately.

• For English, Ambridge et al. (2016, Study 3) found a significant interaction, such that the effect of semantics was bigger for passive than active sentences; just as in the present reanalysis. The present reanalysis found a non-zero (“significant”) semantics effect for passive sentences but not for active sentences; no corresponding analysis was included in Ambridge et al. (2016, Study 3).

• For Indonesian, the present reanalysis yields an identical pattern of findings to the analysis reported in Aryawibawa and Ambridge (2018); even the numerical estimates and B values are broadly similar.

• For Mandarin, Liu and Ambridge (2021) found a significant interaction such that, compared to (“normal”; i.e., non-BA-) actives (the reference category), the effect of semantics was bigger for passives, bigger for BA-actives, and smaller for pseudo-passives. The present study found an interaction such that compared to passives (the reference category), the effect of semantics was smaller for actives and pseudo-passives, and bigger for BA-actives. Thus, although the two analyses are not directly comparable, both found a bigger effect for passives than (“normal”) actives. Looking at individual sentence types, both the original analysis and the present reanalysis found positive effects of semantics for passives and BA-actives, and a negative effect of semantics for pseudo-passives. The present study additionally yielded an estimate of approximately zero for the effect of semantics for (“normal”) active sentences; no corresponding analysis was included in Liu and Ambridge (2021).

In summary, there are no major discrepancies between the reanalyses reported here and those in the original studies, though the original analyses do show some important omissions. For this reason, and because the present analyses contain several improvements (most notably, the use of ordinal regression models), they should be considered as more definitive estimates of the relevant effects.

To sum up, then, the individual by-languages analyses generally found (a) evidence for a non-zero effect of passives in the predicted direction, and (b) evidence that this effect was bigger for passives than (always) pseudo-passives and (mostly) actives.

3.2 All-languages (meta-analytic synthesis) model

In order to investigate whether this pattern holds across languages, we built final semantics-only and semantics-and-frequency models including Language as a random intercept (and with maximal random slopes); i.e.,

All-languages semantics-only-model: Response ~ S_Type*Semantics + (1+ S_Type*Semantics|Participant) + (1+S_Type*Semantics|verb) + (1+S_Type*Semantics|Language)

All-languages semantics-and-frequency model: Response ~ S_Type*Semantics + S_Type*Frequency + (1+ S_Type*Semantics + S_Type*Frequency|Participant) + (1+S_Type|verb) + (1+ S_Type*Semantics + S_Type*Frequency|Language)

In order to allow these models to converge within the maximum time allowed by the server (7 days), it was necessary to reduce adapt delta from 0.99 to 0.95 and the number of iterations from [2000,5000] to [1500,3000] (all Rhat values remained at 1.0, indicating no convergence problems).

In order to allow for comparable analyses across languages and sentence types, we (a) collapsed Balinese -a, ka- and ma- passives as “Passive”, (b) collapsed Balinese “basic” passives, Indonesian “noncanonical” passives and Mandarin “notional” passives as “Pseudo_Passive” and (c) excluded Mandarin BA-Actives. The reason for this latter decision is that (garden variety) active sentences are included as a control sentence type that – at least on a strict standard (see Section 1) – allows us to rule out the possibility that apparent semantic affectedness effects are spurious, and reflect a general dispreference for verbs that show noncanonical linking (e.g., The woman feared [cf. frightened] the man), regardless of sentence type. We can rule out (or, at least, provide evidence against) this possibility by showing that semantic affectedness effects do not hold for (garden variety) actives (or, at least, are smaller for such actives than for passives). That is, the active functions as a control sentence type, precisely because it does NOT show the semantics of affectedness (or, at least, to a lesser extent than does the passive). Thus, a special type of active that has dedicated semantics of affectedness (i.e., the Mandarin BA-Active) clearly does not function as a control sentence type in this way.

Finally, note also that Balinese was excluded from the semantics-and-frequency model, since frequency counts were unavailable for Balinese. In order to tease apart the effects of (a) adding frequency and (b) excluding Balinese, we also ran a supplementary semantics-only model excluding Balinese. This model (available at the project OSF site) was more similar to the semantics-and-frequency model than to the semantics-only model that included Balinese. Thus, it seems that (as for the individual by-language models) adding frequency as a predictor has very little effect on the estimates of semantics, and that apparent differences between the all-languages semantics-only and all-languages semantics-and-frequency models are largely spurious, and reflect the exclusion of Balinese.