1Thanks to the great number of Bantu languages, their close genealogical relatedness and their rich morphology, Bantuists are naturally inclined to study grammar from a comparative perspective. Despite our rich tradition of comparative morphosyntactic research, the methodological challenges involved in cross-linguistic comparison have rarely been explicitly addressed in the Bantuist literature, with some exceptions such as Marten et al. (2007). This paper aims to make a methodological contribution to the initial step of comparative research, viz. the discovery of cross-linguistic variation and the organisation of language data. This initial step of data gathering and management is in principle common to all comparative Bantu studies, irrespective of their research goals and theoretical underpinnings. This paper is therefore intended to be theoretically neutral and generally relevant.
2Most importantly, it argues against the initial methodological step of reducing the analogue reality of linguistic variation by categorising properties, constructions or languages into a finite set of values or types and using the outcome of this initial categorisation as the input of language comparison. Such a reductionist approach has been used more or less implicitly in many or most comparative Bantu studies, and has been explicitly advocated in a number of recent publications, including Marten et al. (2007), where it is known as the parametric approach. I will here avoid the term parametric, because it can also be used to refer to generativist approaches to grammatical analysis which claim that cross-linguistic variation is inherently parametric, i.e. constrained by a discrete set of innate grammatical parameters. As has been said, the methodological issues discussed here are intended to be useful irrespective of one’s theoretical preferences. In contrast, I will use the terms parameter and (parameter) value to refer to dimensions of variation and their discrete segments in reductionist approaches. Names of parameters will be written in small caps and those of values in square brackets.
3Reductionist approaches have been used by scholars with very diverse research goals and theoretical persuasions. Their major methodological shortcomings, reduction and arbitrariness, have often been noticed in the literature, but appear to be considered as a necessary evil in the quest for cross-linguistic generalisations. In contrast, this paper presents a number of methodological alternatives, which are better at revealing patterns of cross-linguistic variation and at taking into account the full range of attested variation in the process of reaching generalisations. Some of these methods, the canonical approach and the use of comparative concepts, come from the field of linguistic typology. Another one, which I propose to call the scenario-based approach, is useful for comparative research within a group of genealogically related languages, and specifically well adapted for comparative Bantu studies.
4Section 2 is a critical evaluation of the use of parameters as a means for organising data and discovering variation. The next two sections illustrate alternative methodological tools, namely comparative concepts (Section 3) and different types of points of reference (Section 4). Section 5 argues that a scenario-based approach can be particularly useful for the comparative study of morphosyntactic variation within languages families such as Bantu. The issue of sampling is very briefly discussed in Section 6.
5In order to avoid getting lost in abstractions, I will illustrate the discussion with concrete examples. The paper’s intention is to compare methods for the discovery and organisation of language data, not theoretical approaches, research results or researchers. The case study in Section 2 was chosen because I find it perfectly executed and excellent in the clarity and relevance of its generalisations, making it easier to concentrate on issues of methodology. The studies cited in the other sections are rather chosen by default, because they are rare exceptions to the general preference for reductionist approaches.
6One way of dealing with morphosyntactic variation is to classify constructions (or languages) into discrete types. Domains of variation are then broken down into different parameters, which each have a finite number of possible values. Such an approach has the benefit of simplifying the sometimes exuberant range of variation with which the comparativist can be confronted, because each language or construction can be assigned a single value for every parameter. Aspects of formal variation that are not deemed essential — such as the exact shape of agreement markers, for instance — can be ignored. It is also easy to represent the values of two different parameters in a table and identify any correlations between them, which facilitates the discovery of generalisations. In contrast, it is well known that in many cases a fixed set of parameter values is not able to capture the full extent of observed variation, and therefore provides an incomplete overview. There are two types of work-around for this problem. The first is to reduce the observed variation by categorizing different phenomena under the same parameter value. This solution leads to problems of reproducibility and to potentially misleading generalisations, as I will argue below on the basis of an example.
7The second work-around is to gradually add parameters or parameter values whenever new grammatical phenomena emerge from the data. Such an autotypologising approach is advocated by Bickel & Nichols (2002), for instance, and applied to the conjoint/disjoint alternation in Bantu studies by Van der Wal (2017) to reach a comparative analysis of the phenomenon that is much more complete than what was previously available.
8Whether sticking to a predefined set of parameters and values, or adding them as more morphosyntactic patterns are encountered, reductionist approaches are very limited in their ability to create hypotheses about the variation that may be encountered beyond the patterns that have already been attested. Although they may push the researcher to look specifically for constructions that have previously unattested value combinations, they do not provide specific hypotheses about the range of variation that may exist beyond the perimeter of the already known parameters and their values. In Sections 4 and 5, we will see how the canonical and especially the scenario-based approaches provide stronger exploratory guidance. By way of an example, we will now look at a domain of cross-linguistic morphosyntactic variation that has typically been studied by means of reductionist approaches in Bantuist studies, viz. object indexation.
9A popular topic in comparative Bantu since at least the 1990s is the domain of object indexation (see, for instance, Bresnan & Moshi 1990; Marten et al. 2007; Riedel 2009; Marten & Kula 2012; Zeller 2014; Marlo 2015; Van der Wal 2020). Van der Wal (2020: 200) summarises the main dimensions of variation in this domain as follows:
| (1) |
Oft-discussed dimensions of variation in the domain of object indexation |
|
(a) |
Number of object markers allowed: one-two-multiple; |
|
(b) |
Behaviour in ditransitives: only the highest object can be marked (asymmetric) or either object can be marked (symmetric); |
|
(c) |
Nature of the object marker: syntactic agreement (doubling) or pronominal clitic (non-doubling); |
|
(d) |
Types of objects marked, specifically locative object markers, and animacy, definiteness, givenness (differential object marking); |
|
(e) |
Position of object marker: pre-stem or enclitic. |
10The topic has been studied by Bantuists working in different theoretical frameworks, but the method of gathering and organising data has generally been reductionist, notwithstanding the fact that authors regularly point out the shortcomings of this approach and their dissatisfaction with it. The following two quotes, one from a generativist publication, the other from a historical-comparative one, are illustrative. The first is from Zeller (2012: 232), cited by Van der Wal (2022: 7), and concerns the dimension of variation mentioned in (1c). The addition in square brackets is mine.
It is uncontroversial that object markers behave like agreement markers in some Bantu languages, but like pronominal clitics in others. However, most languages lie somewhere between the two opposite ends of the agreement-pronoun continuum, showing ‘mixed’ properties, which perhaps reflect intermediate stages of a grammaticalization process that turns pronominal object markers into agreement morphemes (…). Despite the existence of many Bantu languages in which object markers are neither prototypical agreement markers nor prototypical pronouns, the theoretical debate about object marking in Bantu is still characterized by a strict [i.e. reductionist] agreement marker versus pronoun-dichotomy. Because of this dichotomy, the strongest arguments for one type of analysis are often provided by those properties of object markers which cannot be explained easily by the competing analysis. The problem with this method is that reducing the theoretical analysis to these two possibilities may prevent a better understanding of the true nature of object marking.
11Zeller’s observation is a good summary of the central message of this section, viz. that it is impossible to make insightful generalisations over patterns of gradual, analogue variation when this variation is digitised from the outset, that is, when the result of a premature categorization is mistaken for data.
12The second illustrative quote is of a type very frequently encountered where the binning of languages or constructions is discussed. It comes from the main comparative publication that studies the dimension of variation shown in (1e), about the morphological nature of object indexes (Beaudoin-Lietz et al. 2003: 177‑178). The authors divide the Bantu languages into three types. In Type 1, object indexes are prefixed to the verb stem, in Type 2 they are postverbal (either enclitics or independent pronouns). Object indexes in languages of Type 3 can be either prefixed or postverbal. The full fragment discusses the map that shows the geographical distribution of the three types, but the quoted passages concern the assignment of languages to types/parameter values.
We were surprised by (…) the large size of the Type 3 area. (…) Several working factors, if re-examined, might change the map (…) [one] factor which would alter the shape of the map is our working assumption that if any “pronominal” morphemes occurred in a language at pre-stem and/or verb-finally, then we would classify that language as Type 3. But in fact there are languages with partial paradigm series, where most OMs occur pre-stem but the first person singular occurs verb-finally. We classified such a language as Type 3, but it might have been better classified as Type 1. On the other hand, certain of our sources assert that their languages have only pre-stem markers, without giving full examples or considering what happens outside the strict limits of the verb. (…) We suspect some of those languages are Type 3, not Type 1, which would again alter our map.
13Two highly recurrent issues come up here. The first is that reducing grammatical variation to discrete parameter values invariably involves an important amount of decision making on behalf of the linguist. In this case, numerous languages fail to comply to the straightforward case where the object indexes of all persons and genders behave the same: all of them prefixed (Type 1), all of them postverbal (Type 2) or all of them either way (Type 3). Logically speaking, such languages with split paradigms could be assigned to any of the three types. The authors rightly point out that binning them as Type 3 versus Type 1 is arbitrary, but meanwhile they also arbitrarily leave Type 2 out of consideration as a possible binning option for split paradigms. Therefore, the parameter values on which subsequent generalisations are based should not be mistaken for data. They are decisions. An additional practical problem with the decision making that underlies parametric “data” is that it is rarely fully documented, so that the binning process cannot be reproduced. The second typical problem that comes up in this short quote is another source of arbitrariness, namely that the influence of the absence of available data can be disproportionally strong on binning decisions that depend on them. Many linguists must be aware of these problems, because conclusions based on a reductionist approach commonly come with the caveat that more data or more research is needed. However, more data and more research can never change the inherently arbitrary nature of binning.
14The remainder of this section will critically discuss an example of a cross-linguistic generalisation that is based on a reductionist approach to language data, viz. Van der Wal’s AWSOM correlation. I will argue that the AWSOM correlation is an artefact of the way in which individual languages have been assigned to a value [Asymmetry]. AWSOM is short for Asymmetry Wants Single Objects. It states that “asymmetric languages greatly prefer single object markers” and “languages with multiple object markers are overwhelmingly symmetric” (Van der Wal 2020: 206). Thus, the AWSOM correlation links two of the domains of variation mentioned in (1), namely (1a) and (1b). The first is the maximal number of object markers that a verb form can take (1a). Many different values for this parameter have been proposed for the Bantu languages (Marlo 2015), which Van der Wal prefers to reduce to a three-way distinction between [Single], [1+] and [Multiple]. The [1+] value is assigned to languages where “object marking is generally restricted to one marker, but under certain circumstances allows ‘extra’ markers” (Van der Wal 2020: 203). The second parameter, (a)symmetry, concerns the question whether the objects of a ditransitive construction are the same [Symmetry] or different [Asymmetry] in terms of their behavioural potential. Van der Wal (2020: 202) uses the following, narrower definition: “languages are classified as symmetric as soon as the Theme can be object-marked in any ditransitive construction, even if not all constructions are symmetric.” These values can be illustrated with the examples from Swahili in (2) and Ha in (3).
| (2) |
Swahili (G41; Marten & Kula 2012: 244) |
|
a. |
ni‑li‑m‑p‑a |
|
|
1sg.sm‑pst‑1om‑give‑fv |
|
|
‘I gave him (it).’ |
|
b. |
*ni‑li‑i‑m‑p‑a |
|
|
1sg.sm‑pst‑9om‑1om‑give‑fv |
|
|
Intended: ‘I gave him it.’ |
|
c. |
*ni‑li‑m‑i‑p‑a |
|
|
1sg.sm‑pst‑1om‑9om‑give‑fv |
|
|
Intended: ‘I gave him it.’ |
15Swahili verbs have maximally one slot for an object marker. In (2a), it is filled by the class 1 object marker m‑, which indexes the primary object ‘him’. The verb forms in (2b‑c), where both the primary and the secondary object are indexed, are ungrammatical. Moreover, if an object is indexed in double object constructions, it has to be the primary object, i.e. the one expressing the Goal relation. The Theme can never be indexed, which is why Van der Wal classifies Swahili as [Asymmetric].
16In contrast, Ha allows the Theme and the Goal to be indexed on the verb in a ditransitive construction, as shown in (3b). The same example shows that the verb can have more than one object prefix. Van der Wal classifies Ha as [Symmetric], [Multiple].
- 1 The asterisk signals a pitch-accent in Harjula’s analysis. As will be argued in Section 4.2, follo (...)
| (3) |
Ha (JD66; Harjula 2004: 133) |
|
a. |
yamúhaaye umukáaté umwáana |
|
|
a‑á‑mu‑ha*‑ye |
umu‑kaaté |
umu‑áana1 |
|
|
1sm‑rpst‑1om‑give‑pfv |
3‑bread |
1‑child |
|
|
‘She gave the child bread.’ |
|
b. |
yawumúhaaye |
|
|
a‑á‑wu‑mu‑ha*‑ye |
|
|
1sm‑rpst‑3om‑1om‑give‑pfv |
|
|
‘She gave it to him.’ |
- 2 Note, however, that the Fisher’s exact probability value for the correlation is 0.167. The result (...)
17The AWSOM correlation is based on a sample of 51 Bantu languages, represented in Table 1. With one exception, the asymmetric languages in the sample have the values [Single] or [1+], leading Van der Wal to conclude that asymmetry correlates with single object marking.2
Table 1 — (A)symmetry and number of object markers in Van der Wal’s 51-language sample
|
Single |
1+ |
Multiple |
Total |
| Asymmetric |
13 |
3 |
1 |
17 |
| Symmetric |
19 |
5 |
10 |
34 |
| Total |
32 |
8 |
11 |
51 |
18The one exception of an asymmetric multiple object marker language in the sample is Sambaa. As pointed out by Riedel (2009), the Theme in Sambaa can be indexed on the verb in a double object construction if and only if the Goal is also indexed. This is shown in (4). The Goal ‘child’ is indexed on the verb in (4a). In (4b), both the Goal and the Theme are indexed on the verb. Example (4c), where only the Theme ‘book’ is indexed, is ungrammatical.
| (4) |
Sambaa (G23; Riedel 2009: 106; cited via Van der Wal 2020: 206) |
|
a. |
n‑za‑m‑nka |
ng’wana |
kitabu |
|
|
1sg.sm‑pfv.dj‑1om‑give |
1.child |
7.book |
|
|
‘I gave the child a book.’ |
|
b. |
n‑za‑chi‑m‑nka |
ng’wana |
kitabu |
|
|
1sg.sm‑pfv.dj‑7om‑1om‑give |
1.child |
7.book |
|
|
‘I gave the child a book.’ |
|
c. |
*n‑za‑chi‑nka |
ng’wana |
kitabu |
|
|
1sg.sm‑pfv.dj‑7om‑give |
1.child |
7.book |
|
|
Intended: ‘I gave the child a book.’ |
- 3 Another question is how significant a binary classification can really be if it potentially relies (...)
19Like any other generalisation over parameter values, the AWSOM correlation is undermined by the arbitrariness and inevitable mistakes involved in binning decisions. The most obviously problematic of these is the assignment of Sambaa to the value [Asymmetric], because according to Van der Wal’s definition, languages should be “classified as symmetric as soon as the Theme can be object-marked in any ditransitive construction”, and (4b) is an example of a ditransitive construction in which the Theme is object-marked. Incidentally, the Sambaa case also shows the vulnerability of reductionist approaches to accidental gaps in the available data, already mentioned in the beginning of this subsection. In order to classify a language as [Asymmetric] with confidence, one would need to be certain of the inexistence of ditransitive constructions in which the Theme can be indexed, but constructions such as (4b) can easily escape the attention of grammarians, especially if additional conditions need to be met, e.g. if both the Theme and the Goal have to refer to humans, as in ‘They showed the thief to the chief’.3
20A reviewer and some discussants at a workshop where this paper was presented suggested that the values of the (a)symmetry parameter should have been better defined. I disagree with this criticism of Van der Wal (2020), as I cannot think of any alternative definition that would avoid the problems of reductionist classification. For further illustration, let us consider the Ha examples in (3) again, and the classification of Ha as [Symmetric]. According to Harjula’s (2004: 132) description, Ha is much more similar to Sambaa than their different classification may suggest, in that Ha too has strong restrictions on the indexation of the Theme in double object constructions. However, these restrictions differ from those found in Sambaa, so that both languages are non-canonical in different ways with respect to the parameter of (a)symmetry. In Ha, the Theme can only be indexed if no lexical Theme object is present. In terms of the parameter values used in (1c), the prefix indexing the Goal is [Doubling] and the one indexing the Theme [Non-doubling]. Clearly, then, any change in the definition of [Symmetry] made to fix the case of Sambaa is likely to run into trouble when confronted with Ha. It so happens that the classification of Ha as [Multiple] is not unproblematic either. Harjula states that her consultants never allow more than two object prefixes in one verb form. Since one of these two can only occur “under certain circumstances”, namely in the absence of a lexical Theme, the Ha facts are compatible with a possible interpretation of the [+1] value.
- 4 In fact, if languages were assigned to types for both parameters based on their double object cons (...)
21There is another problem with the AWSOM correlation that is due to the fact that it generalises over parameter values, versus linguistic data. The questionable classification of Sambaa as [Asymmetric] and [Multiple] hides the fact that the value [Asymmetric] logically entails the value [Single]. In other words, allowing maximally one object marker on the verb is a logical consequence of allowing only the Goal to be indexed in double object constructions. Therefore, the parameters of symmetry and maximal number of object markers are not logically independent, making the AWSOM correlation trivial. A reviewer disagreed with this observation, pointing out that languages could in theory allow a Beneficiary and a Recipient to be indexed, while disallowing the Theme to be indexed in a construction with three objects. This is not incorrect, but it points to another problem with AWSOM that is due to its reductionist approach, namely that the parameters symmetry and maximal number of object markers apply to different data sets. The first generalises over the double object constructions of the sampled languages, whereas the latter, in order for the AWSOM correlation to be non-trivial, should classify the languages of the sample into types based on a characteristic of their multiple object constructions (i.e. three objects or more).4 Therefore, in order to avoid triviality, the implicational universal of the AWSOM correlation should be reformulated as in (5), although I do not think that this can be done based on the evidence presented in Van der Wal (2020).
| (5) |
AWSOM bis |
|
If a language never allows the Theme to be indexed on the verb in double object constructions, then it can have no more than one object marker in any of its multiple object constructions. |
22The point here is not to criticise the AWSOM proposal, but to argue that generalisations made over binning decisions are strongly dependent on more or less arbitrary choices made by linguists, and therefore not reliable and hardly reproducible. In much of the literature, the notion of symmetry corresponds to a cluster of parameters, since it is also sometimes applied to aspects of the behavioural potential of objects other than indexation, such as word order, or accessibility to passivisation or relativisation (see, for instance, Bresnan & Moshi 1990). Van der Wal (2020) has the benefit of providing a clear, operational definition of the parameter. However, this definition is clearly much stricter than the more or less implicit definitions used by other authors, whose symmetry judgements appear to have been sometimes adopted by Van der Wal (2020), also when they are not compatible with her own definition. For instance, Bemba is classified as [Asymmetric] in Van der Wal (2020: 205), perhaps following Marten & Kula’s classification (2012: 250). Marten & Kula (2012) seem to use the same definition of asymmetry as Van der Wal (2020), so it is a bit of a mystery why they classify Bemba as [Asymmetric], while giving an example of a double object construction in which the Theme is indexed on the verb (6).
| (6) |
Bemba (M42; Marten & Kula 2012: 245) |
|
à‑chí‑m‑péél‑é |
|
1sm‑7om‑1sg.om‑give‑sbjv |
|
‘S/he should give it to me.’ |
23Thus, in practice reductionist approaches add another layer of unreliability when parameters and their values are compared across authors who define them differently. This is somehow related to a well-known issue in the field of linguistic typology, known as the problem of comparability (Evans 2020) or the problem of cross-linguistic identification (Stassen 2010), formulated by Stassen as “how can one be sure that the data which is selected from the languages in the sample form a coherent body of facts?” (Stassen 2010: 90). In this case, how can we know that parameters and their values mean the same thing across different authors and studies?
24The problem of arbitrary categorisation that is inherent to reductionist approaches is strongly compounded when languages are taken as the object of comparison, versus constructions. It is well known that morphosyntactic properties can vary across constructions within a language. Imagine we need to decide whether double object constructions are symmetric with respect to the mutual ordering of postverbal objects in the Gabonese Bantu language Orungu (Van de Velde & Ambouroue 2017). When both objects are lexical in that language, their mutual ordering is free, i.e. symmetric. When one is pronominal and the other lexical, the pronominal object has to precede the lexical object (a type of asymmetry). When both are pronominal, the Goal has to precede the Theme (another type of asymmetry). In such a highly typical situation, one has to choose arbitrarily which construction to consider as basic. Inevitably, different researchers will make different choices, unless some kind of consensus is reached, but that would at best hide the arbitrariness of categorisation.
25The problems of loss of information and arbitrary categorisation are noted in methodological contributions such as Marten et al. (2007: 261‑262, 268), and indeed in a great amount of publications that are based on reductionist approaches, as illustrated by the quotes at the beginning of this section. They are tolerated because apparently they are deemed inevitable in cross-linguistic comparison. The remainder of this paper argues against such defeatism. The next two sections will discuss alternative approaches that have been developed in the field of linguistic typology with the aim of solving the problem of comparability, viz. the use of comparative concepts and the use of points of reference such as canons. Section 5 argues for the usefulness of a scenario-based approach for the comparative study of genealogically related languages.
26In a series of publications, Haspelmath has argued for a clear distinction between language-specific descriptive categories defined by language-specific criteria and comparative concepts that can be used for cross-linguistic comparison (e.g. Haspelmath 2010). The distinction can be employed as a useful methodological tool for dealing with cross‑ and intra-linguistic variation, also if one disagrees with Haspelmath’s insistence on the fundamental ontological difference between comparative concepts and descriptive categories (as does Spike 2020, for instance). Following general practice, the names of descriptive categories are capitalised, versus those of comparative concepts. The correspondence between comparative concepts and descriptive categories is potentially a many-to-many relation.
27I have used the distinction in an analysis of proper names in Rundi (Van de Velde 2009). Following Van Langendonck (2007: 87), proper names can be defined semantically and pragmatically as in (7).
| (7) |
A proper name is a noun that denotes a unique entity at the level of established linguistic convention to make it psychosocially salient within a given basic level category. |
|
The meaning of the name, if any, does not (or not any longer) determine its denotation. |
- 5 Other types of semantic agreement that are not available to underived common nouns can be used as (...)
28This definition of proper names is cross-linguistically applicable and can be used as a comparative concept. It is also possible to recognise a descriptive category of Proper names in Rundi, based on formal criteria that are specific to that language and therefore not cross-linguistically applicable. The main formal characteristic of Proper names in Rundi is that they systematically trigger a specific type of semantic agreement determined by the noun class of the common noun used to express the basic level category of the name. That is, names for persons agree according to the noun class of the noun for ‘person’, names for cows according to the class of the noun for ‘cow’ and names for rivers according to the noun for ‘river’.5 Another characteristic that distinguishes Proper names from Common nouns in Rundi is that the former lack an augment, a nominal prefix that can be found in many Bantu languages and that originates in a demonstrative (see de Blois 1970 for a detailed comparative study; and Van de Velde 2019: 247‑255; and Halpert forthcoming for more recent overviews). Both characteristics are illustrated in (8)‑(9). The forms in (a) are the Common nouns from which the Proper names are derived. The forms in (b) are the Common nouns for the basic level categories to which the proper names belong.
| (8) |
Rundi (JD62; Van de Velde 2009: 224) |
|
a. |
u‑ru‑kara ‘black(ness)’ (class 11) |
|
b. |
u‑muu‑ntu ‘person’ (class 1); i‑m‑bwá ‘dog’ (class 9) |
|
c. |
Rukara |
a‑rikó |
a‑rafuungura |
|
|
Rukara |
1sm‑is |
1‑eating |
|
|
‘Rukara (a person) is eating.’ |
|
d. |
Rukara |
i‑rikó |
i‑iraryá |
|
|
Rukara |
9sm‑is |
9‑eating |
|
|
‘Rukara (a dog) is eating.’ |
| (9) |
Rundi (JD62; Meeussen 1959: 189) |
|
a. |
etymology unknown to me |
|
b. |
u‑rú‑uzi ‘river’ (class 11) |
|
c. |
Maragaraazi |
ru‑gabanya |
u‑bu‑Ha |
n |
ú‑bu‑Ruúndi |
|
|
Maragaraazi |
11sm‑separates |
au‑14‑Buha |
from |
au‑14‑Burundi |
|
|
‘The Maragarazi separates Buha from Burundi.’ |
- 6 Alternatively, names for brooks can take a locative prefix and have locative agreement. In this co (...)
29It turns out that there is a two-way mismatch between proper names (the comparative concept) and Proper names (the language-specific grammatical category) in Rundi. That is, some words that correspond to Van Langendonck’s definition of proper names are not construed as Proper names in Rundi, and vice versa. For instance, contrary to river names, names for brooks do not have Proper name properties in Rundi (10). They have an augment and their agreement is determined by their noun class, which is marked by a prefix on the name itself.6
| (10) |
Rundi (JD62; Van de Velde 2009: 228) |
|
a. |
U‑ru‑bâanza |
rw‑áarakámye |
|
|
au‑13‑Rubaanza |
13sm‑dried_up |
|
|
‘The Rubaanza has dried up.’ |
|
b. |
A‑ka‑gaara |
ka‑ratwáara |
|
|
au‑12‑Kagaara |
12sm‑is_dangerous_for_drowning |
|
|
‘The Kagaara is dangerous for drowning.’ |
30Certain suppletive kinship terms such as nyina ‘her/his mother’ are an example of the opposite state of affairs: nouns that are grammatically construed as Proper names in Rundi, but that are not proper names in the sense that they have a meaning that at least partly determines their denotation. Evidence for their grammatical status of Proper name in Rundi can be seen in (11), where no augment is present and agreement is determined by the class of the noun for the basic level categories ‘person’ (class 1) and ‘cow’ (class 9), respectively.
| (11) |
Rundi (JD62; Van de Velde 2009: 233) |
|
a. |
nyina |
a‑raryamye |
|
|
3sg.mother |
1sm‑is.sleeping |
|
|
‘His/her mother is sleeping.’ (of a person) |
|
b. |
nyina |
i‑raryamye |
|
|
3sg.mother |
9sm‑is.sleeping |
|
|
‘His/her mother is sleeping.’ (cow) |
31The explicit distinction between descriptive categories and comparative concepts allows for easy navigation of categorical mismatches within and across languages. For instance, it allows to formulate such hypothetical generalisations as language names are never Proper names in the Bantu languages, regardless of the specific grammatical characteristics that Proper name categories may have in individual Bantu languages.
32In reductionist approaches, Van Langendonck’s definition of proper names could be used to delimit a morphosyntactic domain, whereas formal properties such as agreement type and presence/absence of the augment can be used as parameters with a finite set of values. The mismatches between proper names and Proper name categories mean that problems of categorisation will inevitably arise when languages need to be assigned to types. If, for instance, the augment parameter has the values [Absent], [Present] and [Both], one should either categorise Rundi (and most/all other Bantu languages) as [Both], or arbitrarily choose to ignore some types of proper names — such as those for brooks — and categorise the language as [Absent]. The first choice is unlikely to be made, as it would lead to a meaningless result. The second choice will run into difficulties in the highly likely case that different languages have different types of mismatch between proper names and their Proper name category/categories. The likely fix for that is to classify languages as [Absent] as soon as some proper names lack an augment (similarly to Van der Wal’s working definition of [Asymmetry]), but that would be an arbitrary decision that dismisses a lot of potentially interesting intra‑ and interlinguistic variation.
33In summary, comparative concepts allow for cross-linguistic comparison by mapping the grammatical categories of individual languages to a cross-linguistically applicable concept that can be defined according to the researcher’s needs. This method is especially useful where a functional domain is carved up differently in the grammar of individual languages.
34In a paper on the methodological underpinnings of comparative research, Round & Corbett (2020) point out that, as long as the dimensions of variation which we study have not yet been fully established, we need points of reference, which “enable all other objects in a domain to be compared back to the same origin”. They distinguish four different types of point of reference; namely standards, idealisations, sample summaries and canons. Standards, such as one hundred degrees Celsius, are real world properties of objects within a domain (Round & Corbett 2020: 498). I am not sure whether there are examples of standards that are or could be useful in comparative linguistic research. In what follows, I will give examples of how the three other point of reference types can be used in comparative linguistic research as alternatives to reductionist approaches.
- 7 Edelsten et al. (forthcoming) seem to claim that only parameters with binary values can be used fo (...)
35For an example of the use of an idealisation in comparative linguistics, we turn to the field of areal typology. Studies that argue for the existence of linguistic areas tend to be based on a reductionist approach that favours binary values, in asking for each language in their sample whether it does or does not have a certain grammatical phenomenon.7 For instance, Clements & Rialland (2008) and Güldemann (2008) point out that African languages that have labial-velar stops (LVs) are concentrated in an area that they respectively call Sudanic zone and Macro-Sudan belt, an area that contrasts with the rest of the African continent, where LVs are virtually absent. However, we know that not all languages with LVs are equal. In the Bantu languages that have them, for instance, labial-velar stops are somewhat marginal phonemes, whereas in languages such as Yoruba they are as common as any other type of consonant. Therefore, a binary parameter +/‑ lv clearly does not do justice to the full extent of observed variation. The fix for this in reductionist approaches is “increasing granularity” (Edelsten et al. forthcoming), i.e. adding parameters or parameter values. On top of a Type 1 (no LVs), we could bin languages with labial-velar stops in a Type 2 (LVs are normal phonemes) versus a Type 3 (LVs are marginal phonemes). However, if we care to follow the best practices of, e.g. Van der Wal (2020), we need to unambiguously define the parameter values [Normal phoneme] and [Marginal phoneme], which is where arbitrariness will unavoidably invite itself, because of the gradient nature of phoneme frequency.
36In order to avoid such arbitrariness, Idiatov & Van de Velde (2021) made a cross-linguistically comparable estimation of the lexical frequency of labial-velar stops (FLV) for every language in a sample of 315 languages with labial-velar stops. We did this by counting all the LVs in our lexical source on that language (typically a dictionary) and dividing the result of that count by the expected amount of LVs in the ideal situation where each consonant has exactly the same lexical frequency in a source (the reference frequency). This idealisation, which we called the zero hypothesis, is obviously not expected to exist in any natural language. Instead, it serves as a stable point of comparison from which to measure the attested variation. Thus, the FLV of a language is expressed as a percentage, where 0% corresponds to the absence of LVs and 100% is the reference frequency. In this way, the entrenchment of labial-velar stops in the lexicon of the languages of our sample is characterised as exactly as possible with the currently available data, free of any kind of binning decisions on our behalf. By way of an example, (12) gives the estimated lexical frequencies of labial-velar stops in a number of Bantu (12a) and other (12b) languages.
| (12) |
a. |
Mekaa (A83) |
2,9% |
|
|
Mpiemo (A86c) |
6,4% |
|
|
Eton (A71) |
14,6% |
|
|
Mokpwe (A22) |
19,8% |
|
|
Likile (C55) |
28,6% |
|
b. |
Biafada |
4,1% |
|
|
Mangbetu |
53% |
|
|
Degema |
97% |
|
|
Yoruba |
110,8% |
|
|
Logba |
116,6% |
37We then visualised our data, using Generalised Additive Modeling. The resulting map clearly shows three hotbeds of high lexical frequency of labial-velar stops (Figure 1). The clarity of the pattern allowed us to formulate detailed hypotheses regarding its origin, and could not have been achieved if the map were based on a small set of parameter values.
Figure 1 — The lexical frequency of labial-velar stops
Heat map color scheme contour plot of the GAM regression surface of the FLV frequencies in percentages (including 0% for languages without LV stops) as a function of the combination of longitude and latitude using thin-plate regression splines (Idiatov & Van de Velde 2021: 84). The model summary: k = 13 (k-index = 0.99, p-value = 0.39, k′ = 195), family = Gaussian, edf = 70.76, deviance explained = 77.60%, AIC = 6048, intercept FLV = 19.95%, p < .001.
38Canons are useful for the comparison of grammatical constructions, because they contain multiple points of reference and can therefore compare constructions along more than one dimension of variation. In that sense, they can be compared to bundles of parameters in reductionist approaches. However, instead of carving up each dimension of variation that is relevant for the object of study into a discrete set of values, they provide a point of reference for each of them from which variation can be measured.8 Hyman (2012; 2006) has used a canonical approach to study cross-linguistic variation in the domain of word-level prosody in Bantu and beyond, and explicitly contrasts it with an approach in which languages are classified into discrete types. The problems he identifies with regard to the reductionist binning that has been customary in the literature on word-prosodic properties are reminiscent of those discussed in Section 2:
The question is whether there is any reality in classifying languages rather than analysing the properties of the relevant subsystems. This issue is particularly pronounced in the area of prosody, where the practice of distinguishing “stress languages” from “tone languages” is well established and has encouraged some to propose additional types: “pitch-accent language”, “restrictive tone language” and so forth. (...) A major aim of this study is to show that there is considerably more diversity in prosodic systems than such labels have thus far distinguished. (Hyman 2006: 226)
39Hyman points out that it is possible to imagine a logically perfect, canonical example of a tone system, and that the same can be done for the prosodic phenomenon of stress-accent, but that there is no such thing as a logically perfect pitch-accent system. Instead, the parameter value [Pitch-accent] has been applied to languages whose tone system departs in various ways from canonical tone, and which have little or nothing in common from a word-prosodic point of view. According to Hyman’s definition of tone, a language has tone if “an indication of pitch enters into the lexical realization of at least some morphemes” (Hyman 2012: 350). He reasons that since the main function of tones is to distinguish morphemes, a canonical tone system is one that fulfils this function in the most optimal way. Hyman proposes the characteristics in (13), illustrating canonical tone by means of a two-tone system, for the sake of simplicity.
| (13) |
Canonical tone properties (Hyman 2012: 357) |
|
a. |
Binarity: |
|
|
Both H and L are phonologically activated. |
|
b. |
Omniprosodicity: |
|
|
Every tone-bearing unit (TBU) has a H or L. |
|
c. |
Unrestrictedness: |
|
|
All combinations of H and L occur. |
|
d. |
Faithfulness: |
|
|
Every /H/ or /L/ is realized on its underlying morpheme and TBU. |
|
e. |
Lexical: |
|
|
/H/ and /L/ should contrast on lexical morphemes. |
|
f. |
Contours?: |
|
|
HL and LH contours should be possible on a single TBU. |
|
g. |
Floating tones?: |
|
|
H and L tonal morphemes and lexical floating tones should be possible. |
40Actual tone systems can and do depart from the canon with respect to each of these properties. Such departures are mostly gradient. For instance, regarding property (13c), Unrestrictedness, tone systems can depart from the canon in imposing tonotactic or positional restrictions on the distribution of tones over words: languages may disallow word-final H tones, allow maximally one H per word, require that each word has at least one H, and so on. Crucially, there are no non-arbitrary cut-off points along the lines of increasing departures from the canon. That being said, tone systems that are extremely far removed from the canon may become ambiguous, in the sense that they can alternatively be analysed as an instance of another grammatical property. If, for instance, a tone system allowed maximally one H tone per word and also required that every word has a H tone, that high tone can be analysed as an accent, and the language as having the property stress-accent. Stress-accent systems too can be more or less canonical (see Hyman 2012: 362 for a proposal).
41In the domain of word-level prosody the canonical approach helps to identify the full range of logically possible variation in tone systems and stress-accent systems, to investigate which kinds of variation are attested (versus which are not) and to situate each attested system with respect to a cross-linguistically stable point of calibration, viz. the canonical types. It also allows to make simple generalisations. Disregarding differences in canonicity between systems, we can determine whether a language has tone, stress-accent, neither or both. In contrast, reductionist approaches to word-prosodic systems have tended to bin languages into either of three values: [Tone], [Stress-accent], or [Pitch-accent] (or the like). Languages have been classified as [Pitch-accent] when their tone system differs more or less significantly from canonical tone. Consequently, the value is impossible to define in a cross-linguistically applicable way and considerable disagreement exists as to what counts as pitch-accent. In this sense, the parameter value [Pitch-accent] is reminiscent of many type twos or threes of reductionist studies in the domain of morphosyntax, such as Beaudoin-Lietz et al.’s Type 3 mentioned in Section 2. In that study of the morphological nature and position of object indexes, Types 1 and 2 correspond to the “clean” cases where all object markers are in the same slot of the verbal template. All the rest, i.e. most of the actual variation, is assigned to a hopelessly incoherent third type.
- 9 Van de Velde (2013) used R1 and R2, instead of N1 and N2, where R is short for relatum. Although t (...)
42In the domain of comparative Bantu morphosyntax, I have adopted an approach inspired by canonical typology when I found myself struggling with the amount of variation in connective (aka associative) constructions while writing what initially seemed to be a straightforward overview chapter (Van de Velde 2013). The Bantu connective construction is a genitive construction, i.e. a construction used to modify a noun by another noun. Starting from this characterisation, a number of basic properties of the grammar of Bantu languages (such as noun-modifier order and generalised noun class agreement within the noun phrase), and the idea that one grammatical meaning (here the genitive relation) is ideally signalled by one invariant form, I arrived at the characterisation of the most perfect, canonical example of a Bantu connective construction in (14). N1 is short for the first (head) noun and N2 for the second (modifying) noun, and ag stands for an agreement marker.9
| (14) |
The canonical connective construction relates nominal constituents headed by a canonical Bantu noun by means of an overt relator. (…) Canonically there is a uniform element that relates N1 to N2 [called] (…) the connective stem (…). The canonical form of the connective relator is ag‑a. The connective relator is placed iconically in between N1 and N2. Since it marks N2 as being syntactically dependent on N1, it is formally more closely linked to N2 than to N1 (…). |
43Clearly, this is a Bantu-specific canonical situation, adapted to the study of variation within the Bantu family, not one that is cross-linguistically relevant. It so happens that actual examples of the canonical situation are relatively easy to find in the Bantu languages (15), which is not always the case for canons.
| (15) |
Kagulu (G12; Petzell 2008: 86) |
|
m‑eji |
g‑a=mu‑nyu |
|
6‑water |
6‑con=3‑salt |
|
‘salt water’ |
44Taking the properties of the canonical connective construction as a starting point, I then explored variation in terms of departures from the canon along five dimensions. This approach allowed me to obtain a detailed but elegant overview of the very rich range of attested morphosyntactic variation in Bantu connective constructions. For instance, one of the five dimensions of variation concerned departures from the canonical second relatum (R2). From the start, I had been aware of examples of connective constructions with an adjective as R2, rather than a noun, but the canonical approach helped me realise that the space between a canonically nominal R2 and a canonically adjectival R2 is occupied by a full range of intermediate constructions that can be organised in terms of degrees of departure from the two canonical word classes (canonical nouns and canonical adjectives). One of the strengths of the canonical approach is that it creates a clear picture of the range of variation that is in theory possible. It pushed me to look for other types of departures from the canonical R2 than those that I was already aware of and helped me discover that R2 can have increasingly verbal properties culminating in the canonical inflected verb that is found in relative clauses introduced by a connective relator. I then started actively looking for other word classes that can be found in R2 position and found a whole range of them, some of which are illustrated in (16). R2 is an ideophone in (16a) and a prepositional phrase in (16b).
| (16) |
Mongo (C61; Hulstaert 1966: 246, 270) |
|
a. |
ba‑úta |
b‑ǎ |
sέnjεlε |
|
|
6‑oil |
6‑con |
pure |
|
|
‘pure oil’ |
|
b. |
bo‑búngá |
w‑ǎ=la |
j‑ángo |
|
|
3‑error |
3‑con=with |
5‑plan |
|
|
‘an intentional error’ |
- 10 This is an illustration that does not reflect how I would approach the question of alignment in mu (...)
45The canonical approach is adaptable to different research goals and theoretical preferences. In the domain of Bantu object indexation, for instance, a canon could be established that takes into account the theoretical assumptions on which much of the modern literature on the subject appears to be based, namely that a notion of symmetry is relevant; that it is a property or parameter of entire languages (versus individual constructions); and that it subsumes a number of dependent parameters related to the behavioural potential of objects, such as word order, or accessibility to passivisation, relativisation and indexation. Given these assumptions, we can push each aspect of symmetry to its logical extreme and establish the properties of canonical symmetry as in (17).10
| (17) |
Canonical symmetry |
|
a. |
Accessibility to indexation: |
|
|
Every object can be indexed on the verb. |
|
b. |
Obligatory indexation: |
|
|
Every object has to be indexed on the verb. |
|
c. |
Accessibility to relativisation: |
|
|
Every object is accessible to relativisation by means of the same relativisation strategy. |
|
d. |
Accessibility to passivisation: |
|
|
Every object is accessible to passivisation using the same passive construction. |
|
e. |
Order of object indexes: |
|
|
All object affixes are in the same position in the verbal template and their mutual ordering is entirely free. |
|
f. |
Order of independent objects: |
|
|
All morphologically independent objects occupy the same linear position in the clause, and their mutual ordering is entirely free, irrespective of their morphosyntactic status (pronouns, versus nouns, versus phrases). |
|
g. |
Cross-constructional coherence: |
|
|
The properties in (a‑f) apply to each and every construction with more than one object. |
46Languages with such canonical symmetry are extremely unlikely to exist, but the canon provides a fixed point of comparison from which all attested patterns encountered in individual languages can be measured. Departures from the canon can then be studied for every axis of variation in (17) and axes of variation can be compared in table format to see which types of departures do and do not co-occur.
47The term canonical is sometimes used in the literature as more or less synonymous with (proto)typical, but canons are different from prototypes. Since canons are at the logically perfect extreme in every relevant aspect of their makeup, they are likely to be outliers in their grammatical domain, and possibly unattested. Prototypes, in contrast, are by definition frequently attested. They are central in their domain, in the sense that existing variation can be characterised in terms of departures from the prototypical core around which the domain is organised. Prototypes only exist in relation to a given data set. They are a type of sample summaries in Round and Corbett’s (2020) typology of points of reference. The schematic representations in Figure 2 try to visualise the difference. The lines represent dimensions of variation and the ellipses a possible range of attested variation. In the scheme on the left, the dot represents a canon. The heart in the centre of the scheme at the right hand side represents a prototype. See, for instance, Taylor (1995) for a discussion of prototypes in linguistic theory.
Figure 2 — Canons (left) versus prototypes (right)
- 11 For reasons outlined in Van de Velde (2020), the term external possessor construction is misleadin (...)
48In an exploratory comparative study of external possessor constructions (EPCs) in the Bantu languages,11 I tried to establish the circumstances in which such constructions can be used, and found that these cannot be formulated in terms of a set of necessary and sufficient conditions of use (Van de Velde 2020). Instead, EPCs are used to express a prototypical situation, namely one in which a familiar person is affected by a change of state brought about in one of their body parts. An example is provided in (18).
| (18) |
Lunda (L52; Kawasha 2003: 252) |
|
mukwenzi wakotoka kánsi mwendu |
|
mu‑kwenzi |
wu‑a‑kotok‑a |
ka‑ánsi |
mu‑endu |
|
1‑young.man |
1sm‑pst‑break‑fv |
12‑child |
3‑leg |
|
‘The young man broke the child’s leg.’ |
49In every Bantu language that has an EPC, this construction can be used to express the prototypical situation exemplified in (18). Depending on the language, the speaker and the context, a number of more or less important departures from the prototypical situation can also be expressed by means of an external possessor construction. Moreover, when used to express the prototypical situation, external possessor constructions are always unmarked (19a), whereas in other contexts an external possessor may require the presence of a valency-increasing applicative suffix on the verb (19b).
| (19) |
Tswana (S31; Creissels 2006: 108) |
|
a. |
ngw‑ana |
o |
tlaa |
go |
gat‑a |
le‑tsogo |
|
|
1‑child |
1sm |
fut |
2sg.om |
crush‑fv |
5-hand |
|
|
‘The child will crush your hand.’ |
|
b. |
ngwana |
o |
tlaa |
go |
j‑el‑a |
dinawa |
|
|
1‑child |
1sm |
fut |
2sg.om |
eat‑appl‑fv |
8/10.bean |
|
|
‘The child will eat your beans.’ |
50Thus, recognising the prototype structure of the category of external possessor constructions in individual Bantu languages allows us to make cross-linguistic generalisations, e.g. (i) the prototypical situation type can always be expressed by means of an EPC and (ii) when an EPC is used to express the prototypical situation, its verb is never applicativised. Second, it creates hypotheses that can be checked against a richer data set, such as (i) the use of an EPC is obligatory to express the prototypical situation in every Bantu language, (ii) EPCs that express the prototypical situation must be frequent in discourse, or (iii) the frequency of use of EPCs used to express a situation X should be inversely proportional to the number of ways in which situation X departs from the prototype. According to the last hypothesis, we expect examples such as (19b) to be less frequent than examples such as (19a), because they do not involve a body part. Further departures from the prototype are expected to have even lower text frequency, for instance when the external possessor (or Concernee) is potentially less familiar (e.g. ‘The child will eat his beans’) and/or when he or she is less clearly affected (e.g. ‘The child will like your beans.’).
51Since we know that the Bantu languages are genealogically closely related, we also know that similarities between constructions are due to inheritance or convergent evolutions and that variation is due to divergent evolutions. In other words, the morphosyntactic constructions of the Bantu languages are linked to each other by means of a diachronic scenario. It would be a pity not to take advantage of this knowledge in our comparative research. Therefore, I propose to add scenarios to our range of methodological approaches to the discovery and management of grammatical data in comparative Bantu research. Using diachronic scenarios as a framework in which to situate Bantu internal grammatical variation is methodologically useful also if one is exclusively interested in studying synchronic variation. I have become aware of the strength of scenarios, e.g. for finding and recognising patterns of variation, during recent comparative work on agreement in relative clause constructions, and I will use that topic as an illustration in this section.
- 12 Object agreement is not immediately relevant for the topic, and will here be ignored.
52It has been well known since at least Meeussen (1967) that relative verb forms in the Bantu languages vary with respect to the element in the sentence with which they agree.12 Comparative overviews such as Nsuka-Nkutsi (1982) and Henderson (2007) recognise three major types of agreement. In the first, here called Type sbj and illustrated in (20), the relative verb agrees with its subject, which is the default situation in the finite verb forms of the Bantu languages.
| (20) |
Haya (JE22; Duranti 1977: 121) |
|
è‑bì‑tòòk’ |
[éby’ |
é‑m‑bwá |
y‑á‑lyà] |
|
au‑8‑banana |
8.dem |
au‑9‑dog |
9sm‑pst‑eat |
|
‘the bananas that the dog has eaten’ |
53A second generally recognised type can be called Type nprel, where the relative verb agrees with the relativised noun phrase (21). Markers that are used to index the relativised NP are glossed as rel.
| (21) |
Orungu (B11; Van de Velde & Ambouroue 2017: 623) |
|
ábà [mágòlìn óꜜŋwánt ꜜógà] |
|
á‑bà |
má‑à‑gòl‑in‑ì |
óꜜŋw‑ántò |
ó‑gà |
|
6‑mango |
6rel‑rpst‑buy‑appl‑rel.rpst |
1‑woman |
1‑chief |
|
‘the mangoes that the woman bought for the chief’ |
54In the third generally recognised type, the relative verb agrees with both the relativised NP and the subject, in that order. This Type nprel‑sbj is illustrated in (22).
| (22) |
Kaonde (L41; Kawasha 2008: 46) |
|
mu‑nzubo |
ayo |
[mo‑ba‑sa‑kú‑lal‑á |
ba‑njimi] |
|
18‑9.house |
9.dem |
18rel‑2sm‑fut‑inf‑sleep |
2‑9.farmer |
|
‘the house where the farmers will sleep’ |
55The variation represented by these three agreement types can easily be accounted for by a mechanism of recurrent change called the Bantu Relative Agreement (BRA) cycle, which consists of a succession of relatively minor changes, schematised roughly in Figure 3 (Van de Velde 2021). The first change is the emergence of a relativiser. This relativiser can have a multitude of origins, including demonstrative forms and a connective relator, but it strongly tends to agree in noun class with the relativised NP. This relativiser can be subsequently morphologically integrated into the relative verb form, where it can undergo partial reduction and be ultimately reinterpreted as an agreement marker. The endpoint of this evolution is step 2. Finally, in step 3 the resulting immediate succession of two agreement markers is simplified, either through merger or deletion. This simplification can result in a single agreement marker that indexes the relativised NP. It is easy to see how the outcome of step 2 in Figure 3 corresponds to Type nprel‑sbj agreement, and the outcome of step 3 to agreement of Type nprel.
Figure 3 — Schematic representation of the BRA cycle
56A first great advantage of such diachronic scenarios is that they create very precise expectations about the types of morphosyntactic variation that may and may not exist, thereby helping the researcher to find and recognise new patterns of variation. For instance, the great variety of relativisers that can be found in the Bantu languages is one of several indications that BRA cycles are easily and recurrently started. This creates the theoretical possibility of relative verb forms that agree twice with the relativised NP (an agreement Type nprel‑nprel), resulting from subsequent BRA cycles. It turns out that such constructions are attested. An example is provided in (23).
| (23) |
Dzamba (C322; Bokamba 1976; cited via Demuth & Harford 1999: 52) |
|
i‑zi‑bata |
[i‑zi‑eza‑áki |
oPoso |
ba‑butu |
loome] |
|
au‑5‑duck |
5rel‑5rel‑give‑ipfv |
[1]Poso |
2‑guest |
today |
|
‘the duck that Poso gave the guests today’ |
57Additionally, the existence of alternative agreement types is typically claimed to be a property of non-subject relatives in the Bantuist literature. However, the BRA cycle should logically apply to subject relatives as well as to other types of relative clause constructions. Therefore, our scenario leads us to expect finding the same range of agreement options in subject relative clauses as elsewhere. Again, this turned out to be the case. The Yao sentence in (24) is an example of a subject relative clause whose verb has Type nprel‑sbj agreement. The relative verb agrees twice with the independent pronoun uwe ‘we’, once as the relativised NP with the third person plural (class 2) relative prefix u‑ and once as its subject, with the 1st person plural prefix tu‑.
| (24) |
Yao (P21; Sanderson 1922: 73) |
|
uwe |
[u‑tu‑li |
ŵa‑yao] |
|
1pl |
2rel‑1pl.sm‑be |
2‑yao |
|
‘we who are Yao’ |
58In contrast, the BRA cycle cannot lead to the emergence of new layers of subject agreement on relative verbs. We therefore have a clear hypothesis about the agreement types that are unlikely to exist, such as sbj‑sbj, sbj‑nprel or sbj‑nprel‑sbj. If we were to find those, they must have emerged through another mechanism, such as a new cycle of univerbation of independent subject pronouns, with the clear expectation that such a mechanism should touch non-relative verb forms as well. No such agreement types are attested, as far as I know.
- 13 Fortunately, these difficulties are laid out in great detail, making it possible for the reader to (...)
59As we just saw, reductionist approaches to language comparison are not as good at predicting probable but previously unattested patterns as are scenario-based approaches. Additionally, by focusing on classifying constructions or languages into pre-existing types, they have a way of obscuring patterns of variation that are actually attested in the available data, but that do not conform to any predefined type. My work on the BRA cycle relied heavily on Nsuka-Nkutsi’s (1982) comparative study of Bantu relative clause constructions, which is based on a sample of over one hundred expertly analysed Bantu languages and founded on a thorough understanding of comparative-historical Bantu linguistics. The approach Nsuka-Nkutsi adopted is reductionist of the autotypologising type, in that additional types/values must have been gradually added as more variation was encountered and recognised in the data. However, Nsuka-Nkutsi failed to recognise a number of patterns of variation that are clearly represented in his data. I am convinced that this can be attributed to his reductionist approach and that a scenario-based approach would have led him to recognise the full extent of morphosyntactic variation that is present in his sample. One missed pattern of variation has just been mentioned, namely that subject relatives can have the same agreement types as object relatives. Another type of variation concerns the paradigm of agreement prefixes that is used in relative verb forms. Nsuka-Nkutsi assumes the existence of two parameter values from the outset — [Verbal prefix] and [Pronominal prefix] — and subsequently struggles to categorise unexpected prefix forms into one of these two values (Nsuka-Nkutsi 1982: 114‑143).13 In contrast, the BRA scenario implies that agreement prefixes in relative verb forms can be the result of a merger of two successive prefixes, and it therefore predicts the existence of dedicated paradigms of Relative prefixes, thereby pointing out a previously unnoticed (or at least unmentioned) dimension of variation. Thus, scenario-based approaches create clear and detailed hypotheses regarding the types of variation that may and may not exist. Despite the intrinsically historical dimension of a scenario-based approach, this is a great advantage for synchronic studies as well.
60A second advantage of scenarios is that they are a great organising tool, because they provide an easy to remember storyline in which morphosyntactic variants can be situated. This way, one can take into account even the most detailed aspects of formal variation without losing overview. For instance, Meeussen (1971) distinguishes between a Lega type and a Luba type of constructions. Constructions of the Lega type have agreement of Type nprel‑sbj in all circumstances. In contrast, those of the Luba type have agreement of Type nprel‑sbj if the subject of the relative verb is of the 1st or 2nd person and agreement of Type nprel elsewhere, as illustrated in (25).
| (25) |
Luba (L31; Meeussen 1946: 174) |
|
a. |
Kabundi, |
[u‑tu‑asεki] |
|
|
1.Kabundi |
1rel‑1pl.sm‑mocked |
|
|
‘Kabundi, whom we mocked’ |
|
b. |
ba‑lumiana |
[ba‑mbidi‑yε] |
|
|
2‑man |
2rel‑address‑3sg.sm |
|
|
‘the men he addressed’ |
- 14 See Van de Velde (2021: 1005‑1006) for an explanation why subject prefixes of the 1st and 2nd pers (...)
61In a reductionist approach, one would have to bin Luba type constructions arbitrarily as either Type nprel‑sbj or Type nprel. Alternatively, one could add a [Luba type] parameter value, but this would have at least two disadvantages. First, the amount of different values needed at this degree of granularity would virtually cancel the simplification offered by reductionist approaches. Second, [Luba type] would be one of a dozen or so values in a list, and nothing would link it to the [Type nprel] and [Type nprel‑sbj] values with which it has close affinities. In our scenario, in contrast, constructions of the Luba type are easily characterised as halfway between step 2 and step 3 in a BRA cycle: they have simplified the succession of agreement markers in a subset of contexts.14 There is no limit to the amount of states-of-affairs we can account for in this way, i.e. no finite set of values.
62By way of a conclusion to this section, I will comment on a number of questions raised by reviewers of an earlier version of this paper regarding the applicability of a scenario-based approach to a wide range of topics, without claiming that a scenario-based approach is ideal, or even possible for the study of every grammatical domain. First, it may be useful to repeat that a scenario is most useful when applied to a sample of relatively closely related languages, where the range of possible variation is restricted in ways that are partly predictable. It cannot be used in the field of linguistic typology on a sample of the languages of the world as an alternative to the methods discussed in Sections 3 and 4, although scenarios of language change can still be invoked in such studies to explain observed patterns of variation, but this is another topic.
63Second, a reviewer suggested that it is hard to know what the scenario is for the evolution in a grammatical domain in the absence of a good reconstruction of that domain in the proto-language, in our case Proto-Bantu. The example of the BRA cycle already shows that this is not necessarily the case, because the BRA-scenario does not have a reconstruction as its starting point. The cycle can apply to many different types of initial state of affairs. The initial aim of my cross-linguistic study of agreement in Bantu relative clause constructions was to find an explanation for the range of variation in agreement types that was known at that time. It was only later, thanks to the BRA-scenario, that I first became aware of the existence of a much greater range of variation in this domain than had been previously recognised; and subsequently of the fact that Meeussen’s (1967: 113) reconstruction of agreement patterns in Proto-Bantu relative clauses is problematic (Van de Velde 2022). In fact, grammatical reconstructions should pay due attention to possible scenarios of change, rather than only take into account the contemporary distribution of grammatical features, following some kind of “majority wins” principle (see also Van de Velde 2022: Section 2.2). Therefore, figuring out likely scenarios of change logically precedes rather than follows disposing of reliable reconstructions. Figuring out the most likely scenario of change for a grammatical domain is a gradual process that starts from a hypothesis and may have to be adjusted when it is challenged by patterns for which the scenario cannot account. Testing the validity of a scenario can be achieved by checking whether the attested variation in any random sample is compatible with the predictions of the scenario. Crucially, this does not require any choices of categorisation, nor adopting any set of theoretical preferences. The more credible the scenario, the more it will be able to account for seemingly random formal details of constructions, on top of the kinds of grammatical properties that we set out investigating. In the domain of object indexation in Bantu, for instance, this would include the question why in some Bantu languages first person singular prefixes are the only object markers whose presence does not change the Final of imperative verb forms from ‑a to ‑e. That being said, finding a convincing scenario may not always be possible, for instance when several crucial steps that have led to the currently attested state of affairs have been lost.
64Third, the scenario-based approach by no means requires that all constructions in the investigated grammatical domain be cognate, in the sense of being reflexes of a single proto-construction. A scenario can consist of separate story lines that can diverge, converge or evolve in parallel. Again, the BRA cycle can partially illustrate this, because the relativiser that emerges in step 1 can have a multitude of origins, including different types of demonstratives and connective relators. The advantage of a scenario-based approach, even in the case of the parallel evolution of constructions with different sources within a functional domain, is that it traces the evolution of individual formal properties in the context of the constructions in which they evolve. This creates a frame of expectations about correlations between logically independent grammatical properties. For instance, according to the BRA scenario, the emergence of agreement with the relativised NP is due to the integration of an erstwhile independent relativiser into the relative verb form (step 2 of the cycle). This integration is potentially hampered by the presence of linguistic material in between the relativiser and the relative verb, which points to a potential correlation between the pre- versus postverbal position of lexical subjects in relative clause constructions and the type of agreement on their verbs, a statistical correlation that actually exists and had been pointed out previously in the literature (Givón 1972). An example of a lexical subject intervening between the relativiser and the relative verb form is provided in (26b). Examples (21) and (23) illustrate relative clause constructions with a postverbal lexical subject.
| (26) |
Haya (JE22; Duranti 1977: 121) |
|
a. |
è‑m‑bwá |
y‑à‑ly’ |
é‑bì‑tòòkè |
|
|
au‑9‑dog |
9sm‑pst‑eat |
au‑8‑banana |
|
|
‘The dog has eaten the bananas.’ |
|
b. |
è‑bì‑tòòk’ |
[éby’ |
é‑m‑bwá |
y‑á‑lyà] |
|
|
au‑8‑banana |
8.dem |
au‑9‑dog |
9sm‑pst‑eat |
|
|
‘the bananas that the dog has eaten’ |
65To summarise, when studying variation within a grammatical domain across a large family of closely related languages such as Bantu, trying to situate that variation in a scenario of change has a number of major advantages. First, constructions and their properties can be organised without any loss of information. Second, scenarios can be a powerful tool for creating hypotheses about the extent of possible and probable variation. Third, since scenarios trace the evolution of grammatical properties in the context of specific constructions, they also have the potential of detecting correlations between logically independent grammatical properties, such as the position of lexical subjects and the type of agreement in our relative clause example.
66One methodological question that has not been discussed so far is the issue of sampling. Some recent comparative studies, including Van der Wal (2020) and Marten et al. (2007), are based on a carefully selected predefined language sample. Creating a balanced sample of Bantu languages is complicated by the fact that the near totality of genealogical diversity is concentrated in a small north-western area, so that a sample can never be areally and genealogically balanced at the same time. In contrast, the comparative studies carried out in the tradition of the Lolemi programme in Tervuren, such as Nsuka-Nkutsi (1982), have taken into account all the information on the Bantu languages that was available at the time. Although theoretically ideal, this is obviously not always practically possible. The best approach is probably a combination of both, i.e. start out with a sample that is maximally balanced and then increase the granularity of the sample in the areas or subgroups that prove to be the most interesting for the topic at hand, if any. One practical restriction regarding the choice of languages in a scenario-driven approach is that publications dedicated to the relevant grammatical feature are of little use if they are not complemented by a full grammatical description. For instance, determining where exactly a construction is situated in the BRA scenario often requires a wide range of information on the language, including the segmental and tonal shape of paradigms of agreement prefixes or the paradigms of demonstratives. It is also important to have examples of relative clause constructions where the relativised NP is a first or second person pronoun. Such examples are hard to find even in reference grammars, of which relatively few have been produced for the Bantu languages. A scenario-based approach therefore provides a clear picture of the known unknowns, something that tends to be hidden by binning decisions in reductionist approaches to comparative data.
67Comparative studies of the Bantu languages are and have been carried out with a variety of goals and theoretical approaches. Such is the diversity in the latter, that we hardly agree on what counts as evidence or explanation. However, we cannot afford to disagree on what counts as data. The main aim of this paper is to argue that the parameter values that populate the majority of the databases on which comparative Bantu studies are based are not data. They are decisions made by linguists, decisions that inevitably involve a fair degree of opacity and arbitrariness. Generalisations made over such decisions are unreliable and hard to replicate, because they accumulate the inevitable errors and arbitrariness involved in every binning decision.
68Another disadvantage of starting a comparative study by reducing linguistic data to a small set of discrete categories, is that such a method is not conducive to discovering grammatical variation in the domain of study. In fact, it may actually be in the way of recognising variation that is clearly attested in the available primary data. This was illustrated, for instance, by Nsuka-Nkutsi’s failure to recognise dedicated paradigms of relative prefixes; subject relatives in which the verb agrees with the relativised NP; and relative constructions whose verb agrees twice with the relativised NP, in what is otherwise an impressive comparative study. With considerable difficulty and exemplary transparency, these phenomena were squeezed into the finite set of discrete parameter values with which Nsuka-Nkutsi (1982) set out.
69A second aim of this paper was to survey a number of alternative, non-reductionist approaches to data discovery and management in comparative linguistic research. The approaches illustrated in Sections 3 and 4 are applicable to sets of genealogically unrelated languages. When studying genealogically related languages, and especially the numerically and morphologically rich Bantu languages, a useful method is to try and situate the observed variation in a scenario of diachronic change, also when the aims of the study are strictly synchronic. The advantages of such an approach, which I have called scenario-based, were discussed and illustrated in Section 5.
70Irrespective of their goals, theoretical preferences or research traditions, most existing studies in comparative Bantu morphosyntax have been carried out using a reductionist approach, which involves digitising a complex analogue reality by classifying properties, constructions or languages into a finite number of types. Although the shortcomings of such an approach are well known and amply discussed in the literature, as briefly summarised in Section 2, they appear to be accepted as an inevitable evil in the quest for cross-linguistic generalisations. I hope to have shown that there exist several non-reductionist methods that are compatible with every research goal and theoretical approach. As for the scenario-based approach, it may not be possible to figure out the most likely scenario of change for every domain of morphosyntactic variation, but I have become convinced that it is always worth trying.
| 1, 2, 3, etc. |
numbered Bantu noun classes |
| appl |
applicative |
| au |
augment |
| con |
connective |
| dem |
demonstrative |
| dj |
disjoint |
| fut |
future |
| fv |
Final vowel |
| inf |
infinitive |
| ipfv |
imperfective |
| om |
object marker |
| pst |
past |
| pfv |
perfective |
| pl |
plural |
| rel |
relative |
| rpst |
remote pas |
| sbjv |
subjunctive |
| sg |
singular |
| sm |
subject marker |
71I wish to thank Dmitry Idiatov, Larry Hyman, three reviewers and the organizers and participants of the workshop Bantu Universals and Variation at the 10th World Congress of African Linguistics (WOCAL10) for their comments.