A multivariate analysis of causative do and causative make in Middle English

1 Introduction

It has been argued at length in the literature that Middle English is a period in which English underwent drastic changes (see, for instance, the contributions in Blake 1992). Some of these changes have been thoroughly studied – see for instance the extensive literature on changes in word order (see amongst many others Kroch et al. 2000; Bech 2001; van Kemenade 2012). There are, however, some innovations involving other grammatical subsystems that have received less attention. In this regard, a subsystem which went through a substantial reorganization in Middle English is the one that concerns periphrastic (or analytic) causative constructions – that is, two-part configurations formed by a causative verb and an infinitival complement that express a causal relation in which the occurrence of the effect is entailed, such as he made his son kick the ball (Gilquin 2010). Specifically, the slot occupied by the causative verb is where we observe the greatest degree of variation. As previous studies have shown (Denison 1993; Lowrey 2002; Moretti 2021; Royster 1922), some verbs that were used causatively in Old English (for example hatan ‘order, name’) fell out of use, while many others (for example cause, make, have) joined the causative network in Middle English, to the point that Denison (1985: 53) describes it as a subsystem that ‘expands and flourishes’. This paper zooms in on the use of two Middle English causative verbs that share several similarities in their basic meaning, namely make and do, which are exemplified in 1-2.

While do is no longer used as a causative verb (cf. Oxford English Dictionary Online 2000, OED, do, v. 29), make has established itself as the ‘prototypical’ causative verb in Present-day English (Dixon 2000: 36). At the beginning of the Middle English period, however, the situation was different. Although not particularly frequent, do is attested in causative constructions already in Old English, while make was a rare verb altogether (Royster 1922). As make made its appearance in the causative subsystem in the transitional period between Old and Middle English, we observe some degree of overlap in terms of the contexts in which both do and make occur, to the point that these verbs have been described as being semantically ‘identical’ (Kuhn 1977: 5). There is, however, only a handful of studies that have looked at the dynamics that underlie the relationship between causative do and causative make in Middle English, with the result that not much is known about their use and their interaction. Ellegård (1953: 108), for instance, reports that one of the causes of the disappearance of causative do is the rise of causative make. Hollmann (2003: 61) notes that, already from the outset, causative make appears with infinitives that require an agentive subject, while Lowrey (2012) observes that causative do shifted from being used in non-agentive contexts to being more frequent in agentive ones. By means of a Multiple Correspondence Analysis (MCA), this paper aims to shed light on the use of causative do and causative make in Middle English, focusing in particular on the semantic domains that are distinctively associated with these verbs. The results of the statistical analysis will show that there is initially a rather neat distinction between make and do, as the former is more frequent with infinitives that require an agentive subject, while the latter occurs with non-agentive subjects. However, already towards the end of the thirteenth century, we see that make has taken over and replaced do in non-agentive contexts as well. In addition, the results of the MCA will help us understand (i) why make began to be used in causative constructions, and (ii) how productive causative do and causative make are in Middle English. Methodologically, the results will show that MCA can be a useful tool to investigate productivity, particularly if they are combined with type frequency counts.

This paper is structured as follows. Section 2 presents causative do and causative make in Middle English and discusses some diachronic data that represents the springboard for the quantitative investigation. Section 3 describes the rationale behind MCA and the methodology adopted in the present study. Section 4 reports the results of the statistical analysis, which are then discussed in Section 5. Section 6 concludes the paper.

2 Causative do and causative make in Middle English

Previous studies have suggested that causative do and causative make are largely interchangeable in Middle English, since they share several semantic and syntactic features (Kuhn 1977). Indeed, they appear in similar contexts, as they are attested with infinitives that require inanimate and non-agentive subjects, see examples 1-2, as well as animate and agentive subjects, see examples 3-4.

Similarities can also be observed in the different syntactic patterns in which causative do and causative make occur, which are the following: V – INF, V – NP – INF, V – NP – to-INF, V – to-INF, V – for to-INF, and V – NP – for to-INF, where NP, which stands for noun phrase, refers to the subject of the infinitive verb (or causee, see Song 1996: 20).^[1] Although the V – INF pattern is no longer used in Present-day English causative constructions, it was possible in Old and Middle English to have causative constructions with the subject of the infinitive unexpressed (for example *he made build the castle). It is worth pointing out that not every construction is equally frequent in the data set. The data retrieved from the Penn-Helsinki Parsed Corpus of Middle English (PPCME2, Kroch and Taylor 2000) and Parsed Corpus of Middle English Poetry (PCMEP, Zimmermann 2015) returned a total of 746 hits (209 with do, 537 with make). Overall, the most frequent patterns are V – NP – INF and V – NP – to-INF, while V – INF, V – to-INF, V – for to-INF and V – NP – for to-INF are robustly underrepresented, as can be seen in Table 1. It is for the paucity of some constructions that I opted in the statistical analysis for a categorisation that distinguishes only between bare and to-infinitives, which means that I collapsed the constructions V – INF and V – NP – INF into one group, while V – NP – to-INF, V – to-INF, V – for to-INF and V – NP – for to-INF into another group. This classification, however, should not be taken as an indication that these different syntactic configurations are semantically equivalent. Several studies have in fact shown that while similar, bare and to-infinitive complements exhibit subtle semantic and pragmatic differences (see Fischer 1992, 1995; Hollmann 2003; Los 2005), which however became largely blurred towards the end of the Middle English period (Hollmann 2003: 160). Furthermore, another point of discussion in the literature concerns whether the patterns NP – V – NP – INF and NP – V – INF are variants of the same form (see Denison 1993: 165), or two distinct syntactic and constructional entities, with different origins and different properties (see Timofeeva 2013 and Lowrey 2015).^[2]

Figure 1 illustrates the frequencies of do and make. The left chart is concerned with token frequency, that is, the number of instances of the constructions occurring in each sub-period. The graph indicates that from the very outset, make is nearly as frequent as do, and its use increased over time, despite a modest decline in texts of the late fifteenth century. Causative do, on the other hand, underwent a robust decrease in token frequency between the thirteenth and the fourteenth century, although we can observe a moderate raise in Middle English texts of the fifteenth century.^[3] The right chart reports the variation in the type frequency, that is, the number of different infinitive verbs attested with causative do and causative make. As can be seen, make occurs with a larger number of verb types in every sub-period except for 1250–1350. Overall, there is a substantial overlap between token and type frequency for both verbs. The only exception concerns do in the last sub-period, when we observe an increase in token frequency and a decrease in type frequency. The interaction between a moderate high token frequency and low type frequency has been shown to be characteristic of poorly productive constructions (De Smet 2020), which lines up well with the observation made by Ellegård (1953: 39) that causative do was not very productive in late Middle English, being almost exclusively used in fixed phrases.

Figure 1:

Token and type frequencies of causative do and causative make in the data set.

An important aspect that distinguishes causative do from causative make is their distribution across different dialects. Previous studies (Ecay 2015; Ellegård 1953; Kuhn 1977) have argued that causative do was more common in the East, while causative make is more frequently attested in the West. The results reported in Figure 2 largely confirm this kind of diatopic distribution. Causative do is the main causative verb in the Eastern Midlands dialect, but only until the first half of the fourteenth century, when make takes over. Make, on the other hand, is more common in the Western Midlands dialect from the first period of interest and is, overall, the most frequent causative in every dialect from the fifteenth century onwards.^[4]

Figure 2:

Frequency ×100,000 words of causative do and causative make across different dialects in Middle English.

3 Method

Causative do and causative make are analysed here by means of a Multiple Correspondence Analysis (MCA). Correspondence analyses are an exploratory statistical method which take a bottom-up approach to uncover frequency-based associations that allow the identification of systematic relations and patterns between variables when ‘there are incomplete a priori expectations as to the nature of those relationships’ (Soares Costas et al. 2013: 1). The input of MCA is a large frequency table which, on the basis of distance-based clustering technique, is transformed to produce a graphical illustration in which every variable included in the study is visualised as a point. The results are therefore a geometrical representation of each row and each column in a multidimensional Euclidean space. The interpretation of the results is carried out on the basis of the relative positions of the points and their distribution along the dimensions. This means that the more similar the categories become in distribution, the closer they are represented on the map. In interpreting the data, it is important to bear in mind that the distances between the variables are not mathematically defined, but the degree of ‘closeness of points on the map with regards to their angle from the origin and points in the same quadrant can be used as guidelines to interpret relationshipships between row and column variables’ (Sourial et al. 2010: 643).

The MCA of causative do and causative make constructed in this study relies on the lexical features of the infinitives that complement the causative verbs. The approach taken in the present study subscribes to the main tenets of Construction Grammar (Goldberg 1995; Hilpert 2014; Traugott and Trousdale 2013), which means that syntactic patterns are inherently meaningful (cf. Goldberg 1995) and that the meanings of constructions are closely linked to the lexical items that typically occur within a construction (Goldberg 1995; Hilpert and Perek 2015). In order to carry out the statistical analysis, I relied on the following steps. The first step concerned in the exhaustive retrieval of all the instances of do and make in causative constructions from PPCME2 and PCMEP. As mentioned in the previous section, this process returned a total of 746 occurrences.^[5] The second step consisted in the identification of the semantic classes to which every infinitive belongs, since it is assumed that semantic classes are conducive to the identification of different semantic domains (cf. Levshina et al. 2013). The operationalization of the semantic classes is not a straightforward task, since meaning cannot be directly captured in the same way that, for instance, morphosyntactic or phonetic properties can. In this study, I opted for a middle-grained classification in which a total of 16 semantic classes, which are illustrated in table 2, were identified. The choice of the semantic classes stems from the study conducted by Levshina et al. (2013), which provided a contrastive analysis of causative constructions in English and Dutch. On the basis of semantic criteria, Levshina et al. (2013) analysed the infinitives occurring with a number of English and Dutch causative verbs and grouped them within 15 semantic classes. The same procedure was used in this study as well, the only difference being the addition of the class V.Destroy to refer to predicates like kill and destroy. A more fine-grained classification, such as the one proposed by Levin (1993) in which more than 40 verb classes are identified, would result in low-frequency classes that would not be particularly instructive for the purposes of this study.

It has to be noted that some verbs may be polysemous and have different meaning in different contexts. The verb comen ‘come’, for instance, may denote motion, as in example 5, as well as a change of state, as in example 6. In order to capture such differences in meaning, every instance has been analysed in context.

The third and last step involved the exploration of the data set using a Multiple Correspondence Analysis. In order to track the development across the entire Middle English period, an analysis for each sub-period (M1: 1150–1250, M2: 1250–1350, M3: 1350–1420, M4: 1420–1500) was performed. It is well acknowledged that there are other statistical techniques, such as variability-based neighbour clustering (Hilpert 2013), that allow a more accurate periodization of the data. However, this method relies on the precise year in which a text was produced, while determining the exact date of composition for Middle English texts is often rather complicated. For this reason, I follow the Helsinki dating system (Kytö 1996), which has been used by the compilers of PPCME2 and PCMEP.

4 Results

The semantic domains of causative do and causative make are visualised through four plots, one for each sub-period. Each plot has two dimensions: Dimension 1, which is represented on the x-axis, returns the largest amount of variation in the data and, therefore, has the most explanatory power, while Dimension 2, represented on the y-axis, captures the second largest amount of variation in the data. In interpreting the results, it is important to look not only at the closeness of the variables, but also at their distance from the centre of the map, which is located at the intersection of the vertical and the horizontal lines. Such intersection is the meeting point of the average row and column profiles, which means that the closer data points are to this intersection along the axes, the less ‘different’ they are, compared to the average profile (Deshors 2017).

The map in Figure 3 reports the results for the M1 sub-period (1150–1250). The position of the semantic classes of the infinitives with respect to where the causative constructions appear in the map helps us understand how they relate to one another and how strong their correlation is. The first observation concerns the distribution of the four constructions. As can be seen, they appear in three of the four quadrants, which means that they exhibit at least some degree of contrast. The first dimension contrasts make – INF, make – to-INF and do – INF, which appear on the left-hand side of the map, with do – to-INF, which occurs on the right-hand side of the plot. Secondly, we see that make is strongly associated with a higher and more diverse number of semantic classes than do. The top left quadrant, in which both make – INF and make – to-INF appear, contains a large amount of semantic classes that pertain to different semantic domains. We can see predicates that express motion, change of state, physical manipulation, creation as well as verbs of change of possession, verbs of destruction and light verbs. Note also that the V.Psych class, which appears in the top right panel, occurs near the origin and is rather close to the make – INF construction. Causative do-constructions, on the other hand, are associated to a smaller number of semantic classes that, furthermore, belong to similar semantic spaces. Specifically, do – to-INF has the strongest connection with V.Intel, with V.GetInfo being the only semantic class found in its proximity, while do – INF is found near verbs that refer to social interactions, physiological processes and states and mental influence.

Figure 3:

MCA map concerning the sub-period M1 (1150–1250).

Moving on to the second sub-period (M2, 1250–1350), for which the results are illustrated in Figure 4, we observe a different situation. Firstly, the distribution of the four constructions along the first dimension contrasts do with make, since both do-constructions appear on the left-hand side of the map, while make-constructions are found on the right-hand side of the plot. Furthermore, we observe that not every semantic class is represented in the map, since only 13 out of 16 classes are part of the plot. This is due to the paucity of data that characterises the M2 sub-period, particularly of prose texts, which has been only partially mitigated by the addition of poetic texts (for a similar approach, see Zimmermann 2020). The main indication that can be drawn from Figure 4 is that, compared with the sub-period M1, the infinitives that occur in the left portion of the plot and are associated with causative do only belong to three semantic classes, namely V.Other, V.Create and V.Intel, with V.PhysManip and V.Body that have an intermediary position with regard to the horizontal dimension and are located midway between do – INF and make – INF. The three semantic classes associated to do are different from the ones we can observe in Figure 3. The only class that we see in both plots is V.Intel, while V.Other and V.Create were previously associated with make. The two constructions involving make, on the other hand, cluster with the vast majority of the semantic classes present in the plot.

Figure 4:

MCA map concerning the sub-period M2 (1250–1350).

Figure 5 illustrates the results for the third sub-period (M3, 1350–1420). The MCA map shows that now virtually every semantic class is associated with make. Make – INF occurs in the top right quadrant, while make – to-INF has an intermediary position and is located near the centre of the plot. Note that only one semantic class, V.Intel, clusters with do – INF, while do – to-INF is rather isolated in the bottom right quadrant. In analysing the position of the causative constructions in the map, perhaps the most unexpected result is the peripheral location of the construction do – to-INF. However, looking at the data, the odd position of this construction is explained by the low number of instances, only three, of do – to-INF in this period.^[6]

Figure 5:

MCA map concerning the sub-period M3 (1350–1420).

Lastly, Figure 6 illustrates the results concerning the last sub-period (M4, 1420–1500). The positions of make – INF and make – to-INF and the semantic classes to which they are more strongly associated with mirror, to a great extent, their localisation observed in the previous sub-period. As for do-constructions, both do – INF and do – to-INF occur in the top left quadrant. In that quadrant, we find infinitives referring to intellectual processes and, nearer the centre of the plot, predicates related to physiological processes/states and physical expression of emotions. Lastly, we also see that do – INF clusters with three semantic classes, specifically V.Exist, V.Other and V.ChPoss, which occur in the bottom left quadrant.

Figure 6:

MCA map concerning the sub-period M4 (1420–1500).

5 Discussion

The statistical analysis carried out in the previous section has uncovered some clear tendencies that characterise causative do and causative make in Middle English. Firstly, there is a neat distinction between these two verbs in texts of the first sub-period (1150–1250). Figure 3 shows that the predicates with which both causative do-constructions are more strongly associated are all verbs that require a non-agentive subject (for example V.MentInfl, V.Intel, V.Body). This is not surprising, since previous studies (Lowrey 2012; Moretti 2021) have shown that do in Old English was typically used in non-agentive contexts. On the other hand, causative make-constructions occur with infinitives whose subject is typically an agent (for example V.Create, V.Motion, V.PhysManip). This kind of distribution is indicative with respect to the emergence of make as a causative verb. In fact, it appears that make was recruited to the causative subsystem to be used in agentive contexts, where do appears to be rarely employed. Note that the typical agentive causative verb in Old English was hatan, which underwent a rapid decrease in frequency of use and died out in early Middle English (Lowrey 2013). Furthermore, Lowrey (2013) argues that the role of hatan was taken up by causative let. However, in light of the results of the MCA, it seems that the gap left by the disappearance of hatan was also covered up by make, since do was mainly confined to non-agentive contexts.

The second indication that can be drawn from the quantitative analysis is that already by the end of the thirteenth century, make had replaced do in non-agentive contexts. Figure 4 indicates that several infinitives that typically require a non-agentive subject, such as V.ChState, V.Psych, V.Aspect, are associated with make-constructions. This means that in the second sub-period make has become more frequent in both contexts and replaced do as the main semantically ‘neutral’ causative verb (see Kemmer and Verhagen 1994 on the notion of neutral causative verb). On the other hand, we see that causative do has become a peripheral element of the causative subsystem. The expansion of causative make had inevitably an impact on the semantic domains in which do was used, and the competition between them caused do to be restricted to particular ‘semantic niches’ (cf. Torres-Cacoullos and Walker 2009; Traugott and Trousdale 2013). This is particularly clear in the last sub-period (M4, 1420–1500). The MCA shows that three semantic classes, namely V.Other, V. ChPoss and V.Exist, appear in the proximity of do – INF (cf. Figure 6). However, if we look at the infinitives that make up these semantic classes, we see the V.Other class is represented only by one infinitive type, that is, make, see example 7, and the same holds for the V.ChPoss class, as we find only one infinitive type, that is, pay, as shown in 8.

In the first case, it is likely that speakers used causative do when the infinitive was make to avoid constructions in which make was both the causative verb and the infinitive complement. In the second case, it has been argued in the literature that these do-constructions involve iussive do, a specific late Middle English use of causative do found solely in administrative and legal texts which was used to declare that certain tasks needed to be carried out (see for example Stein 1990; Budts 2020).

Lastly, the results of the MCA indicate that make was more productive during the entire Middle English period. Previous studies have argued that the fundamental indicator of productivity is type frequency, that is, the number of different types attested with a construction, following the assumption that the more types occupy a specific slot in a construction, the more productive a construction is (Barðdal 2008; Bybee 1995; Bybee and Thompson 1997). Type frequency, however, only tells part of the story when it comes to the productivity of a construction, since it does not consider the degree of semantic similarity between the elements that fill a slot (Baayen 2009). The more diverse the types are semantically, the more productive a construction is (cf. the notion of variability in Suttle and Goldberg 2011). In this regard, MCA can be a useful tool, since they measure the degree of semantic similarity between the types attested in a specific slot of a construction. The fact that make is more strongly associated to a greater variety of infinitive classes throughout the entire Middle English period than do suggests a higher degree of productivity.

6 Conclusion

This paper set out to investigate the behaviour of causative make and causative do in Middle English, focusing in particular on the identification of distinctive semantic domains of the infinitives they occur with. The results of the Multiple Correspondence Analysis have shown that causative make was originally used mainly in agentive contexts, that is, with infinitives that require an agentive subject. This suggests that one reason for the emergence of make in causative constructions was to be used in agentive contexts, since the main agentive causative verb in Old English, that is, hatan, had died out and do was not frequent in this context. Another interesting indication provided by the MCA is that, although do was initially more frequent, make was more productive, as the infinitives it co-occurred with belong to more diverse semantic classes. The results indicate that make began to be used in non-agentive contexts early on, that is, already in the second sub-period (M2, 1250–1350), while do became to be restricted to semantic niches, particularly in late Middle English.

As for the methods used in this study, MCA maps have proved to be a useful tool to study the productivity of grammatical constructions. Firstly, they offer new insights into how similar the elements that occur in a specific slot of a construction are. In this regard, it represents a solution to one of the limitations of type-based counts stressed by Baayen (2009). Secondly, they are able to transform frequency data into graphical representations that map similarities and differences in terms of several interacting factors. The case study presented in this paper illustrates how this method can be useful in informing hypotheses about changes in the productivity of constructions by examining the entire set of items occurring in them. This makes this method a valuable addition to other techniques available to researchers working particularly in the framework of (diachronic) construction grammar that take a quantitative and corpus-based perspective.