An alternative view of the English alternative embedded passive

5.1 Regression analysis: language-internal constraints

Among the participants who most likely have the AEP in their grammars, visual inspection of the acceptability ratings shows that ratings appear to be continuously distributed (Figure 1). Given that so few verbs are attested in the AEP, this is somewhat surprising. While there are clearly unacceptable verbs as well as clearly acceptable verbs, this is not as bifurcated as we may have expected. Rather, at least some participants appear to accept a wide variety of matrix verbs, including rare and previously unattested matrix verbs.

Figure 1:

Average rating of each verb by country.

In the same way, there is no clear evidence of a strong implicational hierarchy governing acceptability. Figure 2 shows the percentage of participants rating need, want, and like as acceptable matrix verbs given that they accept a less- or non-attested verb. Here I use the same threshold of an average rating of 3.67 or higher to define acceptance as I do for need and the AEP more generally. As seen in Figure 2, the distribution of acceptance of matrix need/want/like is continuous, with a ceiling effect at 100 % acceptance. Were there a strict implicational hierarchy, we would expect to see a vertical line at 100 % acceptance at least for need and most likely for want and like as well. It is important to note as well that at least one participant accepted every potential matrix verb.

Figure 2:

Acceptance of matrix need/want/like given acceptance of other matrix verbs.

Each of the language-internal constraints appears to influence acceptability ratings among the participants taken to have the AEP in their grammars. These effects are quite similar across the UK and US. For example, there is a clear productivity effect in which more productive verbs are rated as more acceptable. As seen in Figure 3, the trendlines for the two nations have near-identical slopes.

Figure 3:

Acceptability ratings by log-scaled productivity.

There is a similar overlap with respect to syntactic type. For both the UK and US, Raising verbs are rated as more acceptable than Control and ECM verbs (Figure 4). The two nations diverge in ratings for ECM verbs, but ratings of Raising and Control verbs are quite alike.

Figure 4:

Acceptability ratings by matrix verb syntactic type.

Likewise, there is a clear effect of semantic type. Verbs belonging to groups related to volition and sentiment (A) and necessity (F) are rated higher than those in most other groups (Figure 5). Again, the US and UK appear to have similar effects. Although the nations have a similar overall pattern, the nations’ ratings for any given semantic group overlap to a lesser degree than they do for syntactic type. In particular, US participants rate verbs in groups related to assertion (D), and arranging for future events (H, I) higher than UK participants do. It should be noted that group D contained a comparatively large proportion of ECM verbs that were compatible with stative and eventive passives. In this sense, Figure 5 reflects the observation in Figure 4 of US participants rating ECM verbs higher than UK participants. Figure 5 also shows that several semantic groups are rated similarly to one another. This held strongly after accounting for other factors. For this reason, I recoded the semantic likeness factor into four combined groups: volition and necessity (A/F), permitting or preventing a future event (B/C), assertion or working toward a future event (D/E), and arranging for future events and forget (G/H/I).

Figure 5:

Acceptability ratings by matrix verb semantic likeness group.

Linear mixed-effects regression was conducted separately on ratings from US and UK participants who accepted matrix need. These regressions focused solely on language-internal factors; syntactic type, log-scaled productivity, and semantic likeness were included as fixed effects alongside interactions between semantic likeness and syntactic type/productivity. Random effects included participant and test item as random intercepts, as well as a random slope of the order in which items were presented to individual participants. The step function was used to arrive at a best-fit model through a step-down process. Models for both nations kept all effects except the interaction between semantic likeness and log-scaled productivity. Table 2 summarizes the model for US participants, while Table 3 summarizes the model for UK participants. All factors included in the final models significantly influence matrix verb acceptability. As seen, the models are nearly identical, and paint the same overall picture. The semantic group comprising verbs related to volition and necessity is strongly favored over all other semantic groups, while Raising verbs are strongly favored over Control and ECM verbs. More productive verbs in the canonical EP are more acceptable in the AEP. The interaction between semantic group and syntactic category effectively serves to avoid doubly penalizing non-favored categories. This additionally means that the effect of syntactic type only applies to one semantic group. Overall, the language-internal constraints suggest that there is a broad class of available matrix verbs to the AEP, with acceptability within this class constrained by verb syntax and productivity. Note that while the models are quite similar, the model for US participants consistently has slightly smaller effects for disfavored semantic and syntactic categories, as well as a slightly smaller productivity effect. This may reflect a tendency to accept novel matrix verbs more often than the UK participants.

5.2 Regression analysis: language-external constraints

To consider whether language-external factors correlate with matrix verb acceptability requires analyzing all participant ratings. This involves including participant demographic data in the analysis of ratings from UK participants, while including participant demographic data and population-level factors in the analysis of ratings from US participants.

The model of UK participant ratings began with the language-internal factors and random effects identified as significant predictors in Section 5.1. To this, I added fixed effects of participant education level, ethnicity, gender, and age. Because acceptability of the AEP was strongest with a single semantic group, I also included interactions between semantic group and each language-external factor. The step function was again used to arrive at a best-fit model. This yielded the same significant language-internal factors as in Section 5.1, but no other factors were included in the final model. As such, the only difference between the model of ratings for UK participants who accept matrix need and the model of all UK ratings is that the inclusion of ratings from participants who do not accept matrix need yields smaller effect sizes. At least in this data set, the only language-external influence on AEP acceptability will be spatial.

The initial model of US participant ratings included the same fixed effects, interaction terms, and random effects as the initial UK model. To this I added the population-level factors of by-zip code German heritage, Scottish heritage, Irish heritage, Scots–Irish heritage, population density, and educational attainment of a Bachelor’s degree or higher, as well as their interactions with semantic group. After using the step function to obtain the best-fit model, the demographic factors, German heritage, Scots–Irish heritage, community educational attainment, and interaction terms for all of these language-external factors except gender were kept in the final model (Table 4).

Like the model of all UK responses, in the US ratings model the language-internal factors pattern in the same way as when only participants who accept matrix need are analyzed. However, the effect sizes are smaller with all participants included. Male participants were more accepting of test items overall, but this may be a task effect. We find both a main effect of individual education level and an interaction with semantic group in which participants with higher educational attainment accept non-volition/necessity matrix verbs more than participants with a high school education. We see something similar with ethnicity. While there is little main effect of ethnicity, we find that overall, non-White ethnicity interacts with semantic group such that non-White participants accept non-volition/necessity matrix verbs more than White participants. I suggest that these interaction patterns demonstrate that Whites with lower educational attainment tend to be AEP users more than other groups. Overall, this group has a more robust semantic constraint on matrix verb acceptability than other groups. This interpretation is supported by the population-level effects. German and Scots–Irish ancestry correlates positively with ratings of volition/necessity matrix verbs, while university level educational attainment correlates negatively with ratings of these verbs. In the interaction terms, these effects are effectively undone. This suggests that communities with high German and Scots–Irish heritage populations (i.e., White communities) and low educational attainment accept AEP sentences in the broad semantic class shown in Section 5.1 to be highly rated and subject to language-internal constraints. That is, participants from such communities appear to be AEP users. It is important to note, however, that the language-external effects are far weaker than the language-internal effects. The primary constraints on matrix verb acceptability are linguistic.

5.3 Geospatial analysis

An interim summary of the regression analyses is that while matrix verb ratings are distributed continuously, there are clear constraints wherein frequent Raising verbs from a broad class of volition/necessity verbs are most acceptable in the AEP. Constraints on acceptability are nearly identical in the US and UK among participants who accept matrix need. In the US, these matrix verbs are most accepted among White participants with lower levels of educational attainment, evidenced by them growing up in communities with low educational attainment and high German and Scots–Irish ancestry. The correlation with Scots–Irish ancestry, in conjunction with the similar constraints in the US and UK, appears to be evidence in favor of the AEP appearing in American English through Scots–Irish migration.

Here, I contribute additional evidence in favor of this connection by examining geographical patterns in acceptance of potential AEP matrix verbs. Ratings for each verb were averaged by the zip code/postcode that the participant(s) grew up in. I first test whether ratings of individual verbs are distributed evenly or not across the two nations using global Moran’s I. This is a measure of global spatial autocorrelation, the degree to which the entire set of observed values are correlated across all sample locations. Like other geospatial measures, global Moran’s I relies on a spatial weights matrix that defines the relation between the different locations sampled. There are different kinds of spatial weights matrices, and geospatial statistics are sensitive to the kind of spatial weights matrix selected. Unfortunately, as Grieve (2016) and Wood (2019) discuss, there is not a firm guidance for determining how to decide which kind of spatial weights matrix to use. I use a nearest neighbors spatial weights matrix for both nations (Grieve 2018). Under this approach, the relation between zip codes/postcodes is defined as a value between 0 and 1 based on the distance between a zip code/postcode and the k nearest zip codes/postcodes to it. How many zip codes/postcodes are included as neighbors influences the smoothing of the data, and like the choice of spatial weights matrix, there is not a firm criterion for selecting this (Grieve 2018). For the UK, I use k = 15, while for the US I use k = 50. I found that larger values for k smoothed the data too much, while smaller values left it too fragmented to interpret clearly.

A significant p-value using global Moran’s I indicates that values are likely not distributed randomly in space. That is, significant values suggest that there is geographical clustering of data. Because I conduct this test on each potential matrix verb for both nations, I use a Bonferroni-corrected p-value of 0.000208 (0.05/240) as a significance threshold.^[5] Even taking this correction into account, the ratings of most verbs were found by global Moran’s I to be distributed non-randomly. The test was significant for 81 verbs in the UK data (70 %) and 104 verbs in the US data (90 %). There was good agreement between the nations; 72 of the verbs had significant results for the UK (89 % of significant results) and the US (69 % of significant results). Of matrix verbs attested in Sections 1 and 2, 10 of 11 (91 %) had significant results in both nations. The one exception, enjoy had a significant result in the US but not the UK.

The global Moran’s I results on the one hand are not particularly informative, because so many results were significant. However, this is not as problematic as it may seem. We found above that the distribution of ratings for all verbs was continuous, that all verbs were rated highly by at least one participant, and that effect sizes in regression analyses were larger among participants who accept matrix need. This suggests that most potential matrix verbs had at least some degree of acceptability among AEP users. Thus, any positive ratings may be geographically marked. At the same time, global Moran’s I does not show that positive ratings are clustered non-randomly. Rather, it shows that there is non-random distribution of any ratings. Even if AEP users are not rating a particular verb highly, there may be locations in which participants rated the verb particularly low for acceptability. Likewise, the statistic may indicate that ratings are dispersed in a regular pattern rather than being clustered in a single region. While a useful start for showing that ratings for most verbs have non-random spatial distributions, more detail is clearly needed.

To obtain this added detail, I use the Getis-Ord G _i * statistic as a form of “hotspot” analysis (Grieve 2016; Wood 2019). This statistic measures local spatial autocorrelation: how different a value in one area is from those in surrounding areas. Because this statistic is applied to each measured location, it can be used to determine which regions prefer or disprefer a matrix verb compared to the rest of a nation. To calculate this, I again use a nearest neighbors spatial weight matrix and the same k-values as with global Moran’s I. Getis-Ord G _i * results are easily reported on a map. In the following figures, bright red areas will significantly favor a given matrix verb, while dark blue areas will significantly disfavor it. Upon mapping the Getis-Ord G _i * analysis, I decide whether a hotspot is found through visual inspection by looking for clusters of zip codes/postcodes which significantly favor the verb compared to the surrounding area. We can see an example of this in Figure 6. This figure maps results of the Getis-Ord G _i * analysis across the US and UK for the three widely attested AEP matrix verbs need, want, and like. In the US, we see that the verbs are significantly more favorably rated across the Midwest and in the Pacific Northwest. Likewise, they are significantly less favorably rated in New England and California. We also see that the hotspots are less clearly defined for like than need, with want in between the two. These results replicate the surveys reported in Murray et al. (1996) and Murray and Simon (1999, 2002, as well as the more recent and methodologically similar surveys conducted by Wood et al. (2020) and Wood et al. (2022) quite well with respect to both hot and cold spots. In the UK, we see a similar pattern with the three verbs. Need yields the strongest and largest hotspot, covering Northern Ireland and Scotland in their entirety, and extending into Tyneside in the North East of England. Want and like are also significantly favored in Scotland and Northern Ireland, but Tyneside is less of a hotspot. For all three verbs, there is a strong cold spot in the Greater London area. That Scotland as a whole is a hotspot replicates the fieldwork in Smith et al. (2019), while the overall geographic pattern for need replicates Strelluf’s (2020, 2022 work using data collected from Twitter.

Figure 6:

Hotspot analysis of need, want, and like in the US and UK.

A difficulty in conducting this analysis for each potential matrix verb is that like with global Moran’s I, repeated measures risk drawing incorrect conclusions from false positives. Note, however, that the global Moran’s I results are helpful in part: we know for which matrix verbs the results are distributed non-randomly. For these verbs, we expect to see a hot/cold spot somewhere. To improve confidence in interpretation, I focus on these verbs, and additionally limit attention to hotspots in core areas which clearly demonstrate acceptance of the AEP with widely attested matrix verbs. As seen in Figure 6, for the US this is a roughly cohesive region stretching from Indiana to western Pennsylvania (cf. Wood et al. 2022), extending south to eastern Kentucky and West Virginia. For the UK, this is simply Northern Ireland and Scotland. The reasoning here is that although a hotspot in Los Angeles or London, for example, may well be indicative of a novel innovation with the AEP, because the construction is not found in these locations with widely attested verbs, such a hotspot is more likely a false positive. In contrast, a hotspot in Glasgow or Pittsburgh is consistent with these areas accepting the AEP with widely attested verbs.

Of the 116 test matrix verbs, 55 appeared to have a hotspot somewhere in the US and/or UK, and 37 of these were rated as significantly more acceptable in one or more of the core AEP regions than the rest of the US and/or UK (21 had a US core hotspot, 21 had a Northern Ireland hotspot, and 12 had a Scotland hotspot). These 37 matrix verbs include each of the 11 matrix verbs attested in examples (1, 4–14): need, want, like, love, deserve, hate, prefer, enjoy, risk, require, and demand. Of these attested verbs, all but risk have a hotspot in the US’s core region. In contrast, only six attested verbs have a hotspot in Northern Ireland, while seven do in Scotland. In this sense, the results from the core US region appear to encompass the bulk of what are clear true positive hotspots. Figure 7 illustrates the Getis-Ord G _i * results for four of the 21 verbs with hotspots in the core US region. There are three key takeaways. Firstly, the size of the US hotspot varies by verb. As seen, love has a far larger hotspot than the others illustrated here. Secondly, the center of the US hotspot varies. For hate, the hotspot is in the Indiana/Ohio region noted by Wood et al. (2022). However, for enjoy and request the (tiny) hotspot is in the region surrounding Pittsburgh. Finally, the degree to which the US hotspot is replicated in the UK also varies. While love has a clear hotspot in Northern Ireland and western Scotland, hate is limited to northeast Scotland and enjoy and request do not have UK hotspots.

Figure 7:

Hotspot analysis of love, hate, and enjoy, and request in the US and UK.

This variation in whether and where UK hotspots are located extends to other verbs as well. There are several verbs with hotspots somewhere in the core US area and part or all of Scotland, but not Northern Ireland (Figure 8).

Figure 8:

Hotspot analysis of deserve, require, demand, and expect in the US and UK.

At the same time, hotspots for prefer and stand appear somewhere in the core US region as well as Northern Ireland, but are largely absent from Scotland (Figure 9).

Figure 9:

Hotspot analysis of prefer and stand in the US and UK.

The location of hotspots for the 21 matrix verbs with hotspots in the US core region appear to follow an implicational hierarchy (Table 5). As seen, verbs with hotspots in Northern Ireland tend to also have hotspots in Scotland, and those with hotspots in Scotland tend to have hotspots in the US beyond the core region of Indiana/Ohio/western Pennsylvania.

It is important to note that these verbs only comprise a little over half of the 37 verbs with hotspots in a core AEP region. The remaining verbs (15/16) have hotspots primarily in Northern Ireland (Figure 10). The exception, avoid, has a small hotspot centered on Edinburgh, while three others (exemplified by seek and apply here) have hotspots around Glasgow in addition to Northern Ireland. Only three of these 16 verbs, exemplified by seek here, have a hotspot anywhere in the US. The majority of the 16 verbs without hotspots in the core US region, then, are like petition: a hotspot solely in Northern Ireland. One possible conclusion from these results is that Northern Ireland substantially diverges from the US and Scotland with respect to acceptable AEP matrix verbs. That one of these verbs includes risk, for which an attested example exists, may be evidence in favor of this position. However, it is perhaps equally likely that these examples are false positives generated by outlier rating(s) which appear on the northwestern edge of the UK map and which are relatively isolated from the rest of the map. Their existence is worth mention, but I set these, like hotspots elsewhere in the US, England, or Wales, aside as more likely false positives found over the course of many repeated measures.

Figure 10:

Hotspot analysis of seek, avoid, apply, and petition in the US and UK.

Overall, the geospatial analysis shows that ratings of most verbs have some degree of geographical clustering in both the US and UK. Of these, particularly high ratings are clustered together in hotspots somewhere in one or both nations for roughly half of the verbs (55/116). An intentionally cautious approach to the hotspot analysis finds fewer hotspots, but nonetheless high ratings for far more matrix verbs than previously attested are clustered in core AEP regions. Under this approach, there are more verbs with hotspots in the US than in the UK. This may be an artifact of the study; because there were more participants from the US than the UK, the analysis of the US data may be finding a weaker signal than is possible to find in the UK data. Alternatively, these results may be showing that while the UK and US share a core set of AEP matrix verbs, the US has seen some innovation in the set of possible matrix verbs. It is possible that the UK and US have diverged further, as several verbs have hotspots in Northern Ireland alone. However, it is difficult to assess at this stage whether these are false positives, or evidence that Northern Ireland has a substantially different set of acceptable AEP matrix verbs.

6.1 Research aims

The regression and geospatial results show that far more verbs are available to the AEP than previously attested. Conservatively, there are hotspots showing spatial clustering of high acceptability ratings for at least 21 matrix verbs. Ratings of many more are distributed non-randomly, although this does not necessarily mean that these verbs are widely accepted in a clear geographic region. Beyond this, ratings of all verbs were distributed continuously among participants who rated matrix need highly, with no clear boundary between acceptable and unacceptable matrix verbs. Indeed, at least one participant rated each potential matrix verb highly. In a general sense, the results suggest that while some verbs are accepted by more participants than others, if a verb is available as a matrix verb in the canonical EP, there are people who will accept it in the AEP. That said, there is a clear class of verbs related to volition, sentiment, and necessity which are more likely to be acceptable. However, this is not a closed class. The community detection algorithm applied to the bipartite network model did not include verbs such as expect and stand in this broad class of verbs, yet they had clear hotspots of high ratings in the US and Scotland (expect) and the US and Northern Ireland (stand).

Like the distribution of potential matrix verb ratings, the distribution of participants rating need, want, or like highly given that they rate another verb highly is continuous. This suggests that there is no implicational hierarchy governing matrix verb acceptability. Rather, the regression models show that acceptability is constrained by lexical semantics, verbal syntax, and verb productivity. As mentioned, there is a strong effect of lexical semantics in which volition and necessity verbs are rated much more highly than other verbs. We also find that Raising verbs and verbs which are highly productive in the canonical EP, that is, highly frequent in COCA, are rated highly for acceptability. Note that these constraints derive the implicational hierarchy suggested in previous research. Need, want, like (and love) are all in the favored volition/necessity class. Of them, need is by far the most productive, followed by want and then like/love. Need is also a Raising verb; the others are Control verbs. Combining these facts together, the constraints found in this study point to need being the most acceptable AEP matrix verb, with productivity governing the ranking of the other primarily attested verbs. In this sense, the previously described implicational hierarchy is an artifact of an incomplete description of possible AEP matrix verbs.

The regression and geospatial analysis strongly support the hypothesis that the AEP appears in American Englishes as the result of Scots–Irish migration. We find that even with the expanded view of the AEP demonstrated here, the feature is effectively the same phenomenon in both nations. Among participants who rate matrix need highly, and thus appear to have the AEP in their grammars, the language-internal constraints on matrix verb acceptability are nearly identical between the nations. At the same time, the geospatial results show a clear relation between hotspots of high ratings in the US and UK; for a set of 21 matrix verbs, acceptability in the UK implies acceptability in the US. This finding greatly expands the number of AEP matrix verbs found to be accepted in both nations. We also find that acceptability ratings in the US correlate with community Scots–Irish ancestry. This is suggestive of a founder, or cultural inertia, effect maintaining the AEP in American regions that saw Scots–Irish migration during the colonizing period (Mufwene 1996; Zelinsky 2011). Note that if the AEP did indeed come to the US through Scots–Irish migration, that the AEP is strongly attested in both Northern Ireland and Scotland suggests that the AEP is an older Scots feature brought over during the plantation of Ulster (cf. Corrigan 2010, 2020a; Strelluf 2020).

6.2 Open issues

While the evidence clearly supports a conclusion that the AEP appeared in American Englishes via British Englishes, the geospatial analysis found some evidence that the US and UK may not share the same set of available matrix verbs. As noted, there were hotspots for some verbs within the US that were not present in the UK. At the same time, several matrix verbs had hotspots in Northern Ireland and nowhere else. More research is necessary to confirm whether these differences are divergent innovations, noisy results due to a weak signal, or simply false positives in the current study. Beyond these potential differences, the raw data appeared to show that ECM verbs which may select either stative or eventive passives were rated surprisingly highly, particularly among US participants. It is unclear whether this is because these participants forced a stative reading of the test items, in which case participants differed in task completion across the two nations, or if there may be an expanding use of the AEP in the US that is not being innovated in the UK.

Given how clear the Scots–Irish migration connection is, one question that emerges is why the AEP is not widely attested in other settler colonial varieties. Although the modern United States was the most common destination, Scottish and Scots–Irish migrants also settled in Canada and Australia (Corrigan 2020a, 2020b; Wilson 1997), for example. However, in the Electronic World Atlas of Varieties of English (eWAVE, Kortmann et al. 2020), the AEP is attested to be absent from Newfoundland English and Aboriginal Australian English, and is said to be impossible in (White) Australian English. In contrast, eWAVE does attest the AEP as “pervasive or obligatory” in Colloquial American English, Scottish English, and English dialects in the North of England, and as “neither pervasive nor extremely rare” in Irish English and Appalachian English. We may have expected Canada and Australia to pattern more like this. New Zealand provides an example which does seem to match the expected pattern, as the AEP is attested there in regions with a high proportion of Scottish settlers (Kortmann et al. 2020; Maclagan and Hay 2009). That Canada and Australia apparently differ from the US and New Zealand is surprising, because sociolinguists often suggest that dialect features brought by early settler populations linger (see Tagliamonte et al. 2005, 2010; cf. Mufwene 1996; Zelinsky 1992). While this lingering effect can fade (Stanford et al. 2012), Denis and D’Arcy (2019) show that it is particularly strong in settler colonial nations such as Canada. Assuming, then, that eWAVE is correct in attesting an absence of the AEP in Canada and Australia, there is a question of why this is the case.

It should be noted that while the US geospatial results replicated past survey work well (Wood et al. 2020, 2022), these researchers find clear effects of participant ethnicity and participant urbanness which are not as clearly replicated (see also Murray et al. 1996 with respect to ethnicity). In effect, Wood et al. find that rural Whites are more likely to accept the AEP. In this study, there was not a clear effect of ethnicity for the volition and necessity verbs, while urbanness was not tested in the same way. However, the correlation between AEP acceptability and community demographics points to the same effect. Participants from zip codes with high degrees of German and Scots–Irish ancestry rated the AEP as more acceptable; this effectively means that participants from communities with larger White populations found the feature more acceptable. The German ancestry correlation is perhaps surprising, but may reflect the AEP being found broadly in White communities across the Midwestern US. Similarly, because rural communities in the US have lower access to higher education (Sowl and Crain 2021), the negative correlation between ratings and community educational attainment may reflect rural communities favoring the AEP.

While beyond the scope of this paper, an important question is what implications, if any, the findings here carry for syntactic analyses of the AEP (Edelstein 2014; Strelluf 2022). For example, a central claim in Edelstein’s (2014) analysis is that the feature is a Raising construction. This appears to be consistent with the strong constraint of syntactic type found here; it would make sense for Raising verbs to be strongly favored in a Raising construction. Beyond this, simply expanding the set of matrix verbs may not bear on formal analyses. However, I suggest that the data here points not only to an expanded set of matrix verbs, but to an expanded set of canonical constructions which the AEP may alternate with. One attested matrix verb not tested here was [could] use, because it did not occur in COCA in the canonical EP. This absence may have arisen because the context would overlap with the semi-modal used to. For me, [could] use seems to embed a passive better in a context such as (19):

However, this preference is not unique to use; indeed, enjoy and prefer, verbs newly attested as AEP matrix verbs in this study, also fit well into the VERB being X’ed frame. It appears, then, that the AEP may alternate not only with the canonical EP and concealed passive (as studied in Strelluf 2022, for example) but with this frame as well. Future research may find considering the VERB being X’ed frame in addition to these fruitful both from a quantitative perspective and for updating a formal analysis.

6.3 Conclusions

Although it is typically attested with a narrow set of matrix verbs, the AEP has been well attested with additional ones. This study takes those attestations seriously, aiming to determine more exhaustively the full set of AEP matrix verbs. In doing so, I provide an alternative view of the feature with an updated description. Matrix verb availability is widespread, yet highly constrained. In this sense, the AEP is more like other constructions that embed passives than previously thought. By examining this question through comparison of the US and UK, we find additional evidence that the American AEP is the same as the British AEP, and that this likely means the AEP was brought from one nation to the other. These findings could not be obtained through attention only to widely attested matrix verbs. Rather, this study serves as an example of how attention to rarely attested or non-attested data can inform morphosyntactic and dialectological research.