Typology and the future of Cognitive Linguistics

2.1 Distributional variation

In order to make the connection between typology and Cognitive Linguistics, the cross-linguistic patterns of linguistic diversity – Greenbergian universals – must somehow emerge from the behavior of individual speakers in each speech community. Since Greenbergian universals are patterns of diversity, that is, of cross-linguistic variation, an obvious place to look for the connection is language-internal variation. But there are several different kinds of language-internal variation, and it is not obvious how to get from any of those kinds of language-internal variation to cross-linguistic variation.

The first kind of language-internal variation is the type illustrated by Greenberg’s Universal 18 above. Three language-internal constructions are referred to by the universal: adjective-noun order, demonstrative-noun order and numeral-noun order. Universal 18 constrains the variation in ordering among these three constructions in any language. If Greenbergian universals have anything to do with the grammatical knowledge of an individual speaker, then somehow the implicational universal is part of what a speaker knows about the grammar of their language, even if the implicational universal does not uniquely determine the ordering in the language (for instance, if the adjective follows the noun in the language).

Another classic typological universal, the NP Accessibility Hierarchy of Keenan and Comrie (1977), exhibits another type of language-internal variation. The NP Accessibility Hierarchy refers to the possibility of forming a relative clause with a head referent filling a particular grammatical role in the relative clause (subject, object, etc.). Keenan and Comrie propose the following hierarchy:

Their Primary Relativization Constraint is as follows:

1. A language must have a primary RC-forming strategy.

2. If a primary strategy in a given language can apply to a low position on the AH, then it can apply to all higher positions.

3. A primary strategy may cut off at any point on the AH.

For example, Basque’s primary relativization strategy applies to subjects, direct objects and indirect objects, while Toba Batak’s primary relativization strategy applies to subjects only (Keenan and Comrie 1977: 72, 68–69).

Keenan and Comrie, and many typologists after them, used distributional variation within a language, namely the distributional pattern of the grammatical role of the head referent in the primary relativization strategy. That is, the NP Accessibility Hierarchy constraints in (i)–(iii) are universals of distributional variation both within and across languages.

In fact, many morphosyntactic universals are universals of combined language-internal and cross-linguistic distributional variation. For example, the universals of parts of speech described in Croft (1991, 2001) are of this type. The universals are based on the generalization that object concept words in the referring propositional act function (Searle 1969), property words in the modification function, and action words in the predication function are typologically unmarked. What this means is that in languages with more than one construction for a particular function, the different constructions will define a language-internal distribution of object, property and action concept words; and the words will be distributed in the constructions such that for example, action concept words will occur in the predication construction with the least morphemes (e. g., no copula) and the most inflectional possibilities, and property and/or object concept words will occur in predication constructions with at least as many morphemes, e. g., with the addition of a copula, and the same or fewer inflectional possibilities.

In fact, English requires a copula for both object and property predication, and in addition an invariant article a for predication of countable object concepts:

The relative amount of coding of predication in these three constructions of English replicates the hierarchy of structural coding of the predication of the three different semantic classes of objects, properties and actions:

These examples demonstrate that distributional variation within a language and grammatical variation across languages are governed by the same universal principles (Croft 2001: 107). This is central to Radical Construction Grammar. Grammatical categories are defined by the constructions in which they occur. Grammatical categories are construction-specific as well as language-specific – that is language-internal distributional variation. But the principles, such as (3), governing language-internal distributional variation also govern cross-linguistic distributional variation. In this way, the analyses of typology and of (construction) grammars of individual languages can be unified.

These universal principles can be represented for many typological patterns using what has come to be called the semantic map model (Anderson 1982, 1986; Croft et al. 1987; Kemmer 1993; Haspelmath 1997a, 1997b, 2003; Stassen 1997; Kortmann 1997; van der Auwera and Plungian 1998). In the semantic map model, distributional facts for grammatical elements that occur in particular roles in grammatical constructions are represented as a grouping of the concepts encoded by those grammatical elements. For example, the Accessibility Hierarchy represents a linear arrangement of the participant role categories represented by Subject, Direct Object, etc., as in (4):

The linear relationship among those conceptual categories reflects the constraints on possible groupings of those categories in relative clause constructions within and across languages.

Conceptual spaces can have a more complex structure, such as the conceptual space for indefinite pronouns posited by Haspelmath (1997a). Here the conceptual categories – different semantic types of indefinite pronouns – are arranged in a more complex network based on whether a language-particular form groups the meanings together.

Figure 1:

Conceptual space for indefinite pronouns, with the semantic maps of the indefinite pronouns of Finnish (Haspelmath 1997a: 293).

The categories of indefinite pronoun meanings and the links between them form the conceptual space, while the semantic maps for the indefinite pronouns of Finnish that indicate their meanings are illustrated with dashed/dotted lines; the Finnish forms are given in italics.

The conceptual space in Figure 1 is modeled as a graph structure: the nodes are the meanings and the edges (links) are the conceptual relations between the meanings inferred from the patterns of grammatical distribution within and across languages. That is, the conceptual space is represented as one of discrete meanings (for a statistical algorithm to compute a discrete graph-structural conceptual space, see Regier et al. 2013). It is also possible to model conceptual space as a continuous space, and use statistical methods such as multidimensional scaling (Levinson et al. 2003; Croft and Poole 2008; Rogers 2016 ; García Macías 2016) to construct Euclidean models of the conceptual space.

Multidimensional scaling is particularly useful for large and complex data sets such as the adpositional semantic data of Levinson et al. (2003). This data was obtained by eliciting in nine languages descriptions of the 71 pictures of IN/ON-type spatial relations designed by Melissa Bowerman and Eric Pederson (published in Levinson and Wilkins 2006, Appendix 4). Croft and Poole reanalyzed the Levinson et al. data; the analysis is described in greater detail in Croft (2010a). The data demonstrate a high degree of variation in categorization across the languages for this extremely fine-grained set of semantic distinctions in this restricted semantic domain. However, it also revealed a coherent conceptual space, ranging from degree of envelopment of the figure by the ground at the IN end of the continuum to degree of a relatively small figure in contact with and vertically supported by a relatively flat ground at the ON end of the continuum, with varying types of support, attachment and containment relations in between. In other words, what is universal is not a set of conceptual categories, but salient semantic dimensions in conceptual space that constrain – or perhaps better, guide – the grammatical categories that are defined in particular languages.

The semantic map model allows us to distinguish what is language-universal from what is language-specific. The structure of the conceptual space is language-universal, while the grammatical categories defined by continuous or connected regions in the conceptual space are language-specific. (There are other universal constraints on grammatical categories than those implied by the structure of the conceptual space; Croft 2001: 159–161, 2003: 140–142.) Some typologists describe the semantic map model as simply a methodological convenience to describe crosslinguistic generalizations (Anderson 1982: 228; Croft et al. 1987: 186; Kemmer 1993: 201; van der Auwera and Plungian 1998: 86). This stance is subject to the same criticism of typology by generative grammarians presented in § 1. On the other hand, Anderson, who is the immediate source of the modern semantic map model, later called conceptual spaces “mental maps” (Anderson 1986); Kortmann (1997: 177, 210) calls them cognitive maps, and I refer to the conceptual space as a cognitive structure. The universal structure of the conceptual space must be grounded in human cognition. The conceptual space and other constraints on grammatical categories provide a link between human cognition and typological variation and universals thereof. Likewise, the language-specific knowledge captured by the semantic maps is also a cognitive structure of the speaker, albeit not a universal one.

2.2 Variation in verbalization

We nevertheless still have not described how one gets from human cognition to typological diversity. The starting point is yet another type of language-internal variation, one that has not attracted much attention until recently. That is variation in verbalization: variation that occurs at the level of individual usage events. Most linguistic analysis starts from form and looks at meaning, for example, the many analyses of polysemy in Cognitive Linguistics and, for that matter, in the vast majority of reference grammars of indigenous languages. But one can also start from meaning and look at what grammatical forms are used to verbalize it.

The study of variation in verbalization is done most easily in experimental situations in which one or more speakers verbalize a set of the same stimuli. In fact, cross-linguistic elicitation tasks such as the Bowerman-Pederson IN/ON picture set referred to above, the World Color Survey (Kay et al. 2009) and the cutting/breaking video clips (Majid and Bowerman 2007; Majid et al. 2008) also document language-internal variation in verbalization (e. g., Kay et al. 2009: 55; Levinson et al. 2003: 503; Majid et al. 2008: 239–240).

Wallace Chafe published two groundbreaking papers on verbalization at the beginning of the cognitive linguistics enterprise (Chafe 1977a, 1977b). Chafe produced one of the classic narrative verbalization stimuli, the Pear Film (Chafe 1980); the Pear Film was later followed by the Frog Story, another narrative verbalization stimulus (Berman and Slobin 1994).

Examining the Pear Stories produced by the twenty English speakers at Berkeley that Chafe presents in his 1980 volume immediately reveals how pervasive variation is. Every verbalization of every scene in the film is unique in the entire corpus. Even when the verbalizations are broken down into their component parts (lexical categories, argument structure, etc.), variation is pervasive. This variation is due to the fundamental indeterminacy of language use. In each utterance, the utterance meaning in context is unique and hence at least slightly different from prior uses of the constructions replicated in the utterance. Although a speaker’s choice is determined in part by how she construes the experience to be verbalized, there are many ways to construe an experience and the hearer does not know which one the speaker would choose. Finally, the hearer cannot read the speaker’s mind; this is of course why one person’s experience must be verbalized in order to share it with another speaker. Yet speaker and hearer succeed because most of the time, perfect precision – impossible to reach anyway – is not necessary for the interlocutors to successfully carry out their joint actions that are coordinated by their utterances (Clark 1996).

So variation is pervasive in language use. But the variation is constrained in ways familiar to typologists (Croft 2010b). Two examples are given here. Second mention of referents pertains to how referents are verbalized after they are introduced in discourse. There are two common types of verbalizations in the Pear Stories: as possessive pronoun or as a definite article (numbers preceding each example indicate the speaker number and the number of the intonation unit of the speaker’s verbalization):

In second mentions of seven different referents from the Pear Film, there was systematic variation such that more animate referents that were less likely to be possessed were more likely to be verbalized with a definite article, and referents that were less animate and more likely to be possessed were verbalized with the possessive. Subtle semantic differences between the referents in the scenes governed the relative frequency of the two variants (Croft 2010b: 19–21). For example, the ladder is less likely to be owned by the pearpicker who is using it, hence it is less likely to take the possessive pronoun; but the bicycle is more likely to be owned by the boy riding it, hence it is more likely to take the possessive pronoun. These subtle semantic differences are reflected in different likelihoods of the use of the definite vs. the possessive pronoun.

These observations have far-reaching consequences for understanding the nature of grammar. Grammar is not a mapping of discrete forms to discrete meanings, or even sets of meanings. The mapping between form and meaning is a probability distribution of forms across very specific points (very specific meanings) in the conceptual space. Speakers make choices based on this probability distribution, giving rise to the most fundamental type of intralinguistic variation: variation generated simply in the process of communication. This is an exemplar-based model of grammatical knowledge, along the likes of exemplar models of phonological knowledge advocated by Bybee (1985, 2007, 2010) and Pierrehumbert (2001, 2003). In some cases, the probability distribution comes to be more categorical, and the result is the language-internal distributional variation that was discussed above.

This essential, fundamental process generating linguistic variation is the ultimate source of grammatical change, as argued in Croft (2010b). For example, variation in verbalization of so-called “light verbs” such as put correlates closely with the more contentful verbs that are the historical sources of the light verbs. The verb put was used by at least one speaker in Chafe’s Pear Film narratives in five scenes, and in four of the five scenes, there was variation in verbalization, with other speakers using other verbs; see Table 1.

As with second mentions, there are subtle semantic differences between the scenes which are reflected in the frequency distribution of forms, even with the low overall number of instances in the data. What might be called prototypical PUT, with a singular object and the agent’s hand as the instrument, is verbalized with put by all the speakers. Moving away from the prototype, to scenes with distributive and plural objects and an instrument other than the agent’s hand, leads to increasing frequency of verbalization with other verbs.

When the different verbs used in the scenes in the Pear Stories are compared to the etymological sources of the verbs in Indo-European (Buck 1949), we see that basically the same verbs are found; see Table 2.

These and many other examples described in Croft (2010b) demonstrate that grammatical change is largely gradual change in the probability distributions of the form-meaning mapping. As some variants increase in frequency and others decrease to zero, they come to look like the “discrete” language changes that are presented in historical linguistics textbooks.

2.3 Socially-governed variation, language change and typology

This “discrete” language change is misleading. Language is always variable; variation is one of the fundamental properties of language. But there is still a missing step (actually, two missing steps) between the fundamental variation just described and the typological diversity we observe in the languages of the world. The (first) missing step is socially-governed variation: the situation when two or more variants acquire social valuation. This is the realm of sociolinguistics, a theoretical approach to language that emerged around the same time as generative grammar (Weinreich et al. 1968) but is unfortunately often ignored by functionalists and cognitive linguists. Sociolinguistics itself is somewhat isolated from other theoretical approaches. Variationist sociolinguistics focuses on the social factors determining the propagation of variants through a speech community (Labov 2001; Milroy 1992; Trudgill 1983, 2002). Sociolinguistics generally takes for granted the existence of language-internal variation but does not offer a mechanism by which the variation is generated in the first place.

The usage-driven variation described above can be integrated with socially-driven variation in an evolutionary model of language use, variation and change (Croft 2000). Language use is the replication of sounds, words and constructions. As argued above, language use generates variation. Evolutionary models are models of change by replication. Evolutionary models are also two-step models: generation of variation, and selection of variants. I hypothesized that the first step, generation of variation, is driven by functional factors, namely meaning and discourse, as in the subtle semantically-driven variation in frequency described above; and selection is driven by social factors, most importantly the differential social weighting of linguistic variants (Blythe and Croft 2012).

We can now describe how typological diversity is connected to a speaker’s knowledge about her language. Speakers generate variation in language use; we can call this first-order variation (Croft 2010b: 3). A speaker’s knowledge about her language is usage-based: it is a probability distribution of forms over meanings in the conceptual space, inferred from past usage events and constantly changing as forms are replicated to verbalize new experiences in new usage events. These probabilities can change, both as a speaker’s experience changes and also because speakers also observe frequencies of variants used by particular social groups in her speech community, and come to associate certain variants with their social valuation of those groups. This gives rise to second-order variation. Second-order variation is socially-governed variation; it assigns first-order variants differential social values. At this point, frequencies of variants are now also governed by social factors. This last association can lead to propagation of more highly valued variants at the expense of less highly valued variants. Finally, as speech communities diverge and ultimately split, patterns of propagation of the original speech community will become independent and possibly lead to different outcomes. As the descendant speech communities become socially autonomous units, different variants end up being propagated. The result is the typological diversity of languages, that is, third-order variation.

This typological diversity has its ultimate origins in first-order variation. The functional factors that drive first-order variation survive in third-order variation because the social factors that drive second-order variation, and ultimately selection of a linguistic variant, are independent of the verbalization process. The output of selection in a large enough language sample will therefore reflect the frequency distribution of the verbalization process. Hence, patterns of typological diversity will reflect the factors that determine verbalization.