(The following conversation reflects some of the discussions that we had over the last few years, and particularly at a recent mini-workshop at WIKO Berlin.)
Martin Haspelmath: In the typological literature of the last decade, one finds more and more instances of people claiming that this or that typological generalization actually has a diachronic explanation. Quite a few authors appeal to “common pathways of change” (e.g. Anderson 2016), and Bybee (2006) has even claimed that “the strongest universals are universals of change”. But I have argued (in this recent post, point (A)) that if one wants to explain universal distributions, one needs to appeal to more than just “common pathways of change“ – one actually needs constraints on the kinds of change that are possible, what I call “mutational constraints”.
Sonia Cristofaro: I am not sure that any diachronically oriented typologist (e.g. Bybee) has actually ever invoked mutational constraints, implicitly or explicitly. Typological “universals” always have exceptions, so the idea is usually that there are communicative preferences, not mutational constraints. Exceptional patterns can always arise, except that the changes leading to those patterns are less frequent. Mutational constraints, on the other hand, are postulated by synchronically oriented linguists such as Hawkins or Kiparsky, who explicitly assume that implicational universals capture constraints on what types of changes can take place.
Martin Haspelmath: Yes, Kiparsky (2008) is a remarkable paper which argues that some or many “typological generalizations” are due to mutational preferences, and he claims that one can see this because they can have exceptions (unlike true universals, which are due to UG (universal grammar); or in my terms, exceptionless representational restrictions). I find this approach conceptually perfectly coherent, but I have problems with Bybee (2006). Maybe she does not want to distinguish between universal tendencies and frequent (vs. less frequent) patterns. But my problem with the notion of “cross-linguistically frequent patterns” is that it is not testable: Whether something is frequent or not is subjective (is 3 cases frequent? or 30? or 300? this depends on the context), whereas whether something is a universal tendency is objectively verifiable. So these changes must represent universal tendencies (i.e. they must be mutational constraints) to be valid explanations of universal distributions.
Sonia Cristofaro: I am not sure that a distinction between universal tendencies and frequent patterns would be theoretically useful. If you regard “universals” as a product of specific diachronic processes (which most Greenbergian typologists do, even if they don’t investigate the actual diachronic processes), then the frequency of individual patterns is a result of the frequency of the processes that give rise to those patterns. Different processes and the resulting patterns will be more or less frequent, but there will be no substantial difference between more frequent and less frequent processes/patterns (in the same way, for example, as there is a conceptual difference between patterns originating from inbuilt constraints in a speaker’s mental grammar and patterns originating from recurrent cross-linguistic diachronic processes). In any case, Kiparsky (2008) regards the animacy/referential hierarchy as an exceptionless pattern, but this hierarchy, has long been demonstrated to have many exceptions, and many of the patterns described by the hierarchy have diachronic explanations (see Cristofaro 2013 and references therein).
Martin Haspelmath: Some authors who invoke diachronic explanations of typological distributions note that there is a convergence of developments (point (B) in my earlier post). But if several unrelated sources converge on the same result, we have a remarkable coincidence, which could not really be due to accident. Thus, convergent developments point to “pull forces”, or what I would call functional-adaptive constraints that favour changes in the direction of a certain (desirable) result. On the other hand, if only three or four cases are involved (as in some of Anderson’s 2016 examples), the observed convergence could of course be accidental. But in that case it cannot explain a universal distributional tendency.
Sonia Cristofaro: I don’t think the coincidence is remarkable (and in some cases there may not even be an actual convergence, as noted later). Individual typological distributions usually arise from a limited number of sources and developmental mechanisms. For example, ergative alignment restricted to perfective contexts usually originates from a limited number of constructions typically associated with those contexts (when ergative alignment originates from constructions not associated with perfective contexts, on the other hand, it is not restricted to those contexts). Likewise, overt marking restricted to plural originates from just a handful of processes, usually the grammaticalization of an expression of multitude, and occasionally phonological erosion of a singular marker or metonymic transfer of plural meaning from a quantifier to a co-occurring element (e.g., in expressions such as ‘many of us’ a partitive case marker on ‘us’ can take on the plural meaning originally associated with the quantifier). It is not particularly surprising that a few distinct processes independently lead to patterns that can be described in the same terms synchronically, particularly if these terms are very general. For example, “overt marking for X (as opposed to Y)” is a very general description of a situation that could be the result of two distinct processes, the grammaticalization of a marker for X as opposed to Y or the loss of an existing marker for Y due to phonological erosion. These two processes would not usually be assumed to reflect the same motivation just because they lead to the same result.
Martin Haspelmath: So are you saying that it only seems that we find universal tendencies? Or would you say that one can get real universal tendencies from the accidental convergence of a few unrelated types of change?
Sonia Cristofaro: We do find universal tendencies, but, as also argued in Bybee (2006), these are tendencies for particular source elements to evolve in particular ways, not tendencies for languages to develop particular patterns because of some general preference for that pattern in itself. As for convergence, one possibility is that a few unrelated types of change independently lead to the same result, as noted earlier. Another possibility is, however, that there is no actual convergence. Not all of the processes that can lead to a particular result are unidirectional, in the sense that they always or usually lead to that result. For example, the evolution of number markers from expressions involving some type of number meaning seems to be unidirectional in this sense, in that cases where an expression of multitude evolves into a plural marker appear to be more frequent than cases where expressions involving singular meanings evolve into singular markers. Other processes that shape number marking, however, lead to opposite results in different cases. This is the case, for example, when overt number markers develop through metonymic transfer of number meaning from a quantifier to a co-occurring element: expressions such as ‘many of us’ give rise to plural markers, which can lead to a situation where plural is overtly marked and singular is zero marked, but expressions such as `one of us’ give rise to singular markers, which can lead to a situation where singular is overtly marked and plural is zero marked. Similarly, in languages where both singular and plural are overtly marked, phonological erosion can target the singular marker, leading to zero marking for singular and overt marking for plural, or it can target the plural marker, leading to overt marking for singular and zero marking for plural. This suggests that the higher frequency of some results as opposed to others (for example, overt marking for plural and zero marking for singular, rather than the opposite pattern) may actually not be due to the fact that different processes all converge on that result. Rather, some process usually leading to that result is more frequent that other processes that may or may not lead to that result. In this case, there is no actual convergence, and universal preferences must be preferences for particular processes, rather than preferences for the result of those processes in itself. Then of course one needs to explain why certain processes are preferred. This, however, need not be related to the result of the process, so one cannot make assumptions by only looking at the result.
Martin Haspelmath: What are the best arguments for your “source-determined” approach to cross-linguistic grammatical regularities?
Sonia Cristofaro: There are two major arguments:
(i) Alternative triggers: The actual diachronic processes that lead to the patterns described by typological universals are not obviously triggered by the principles traditionally postulated based on the synchronic properties of the pattern. In classical explanation, these patterns arise because they comply with functional principles, e.g. economy or processing ease. Grammaticalization studies and historical linguistics in general, however, suggest that the constructions involved in individual patterns arise from pre-existing constructions, through processes related to the properties of these constructions and the contexts in which they are used, rather than properties of the resulting patterns. For example, overt markers for particular grammatical meanings (e.g. particular number values, particular argument roles) arise from processes of reinterpretation of pre-existing elements. These processes are plausibly triggered by (often context-induced) associations that speakers establish between the original meaning of the element and the new grammatical meaning, so there is no evidence that they are driven by the need to give overt expression to the grammatical meaning (as assumed in classical explanations for overt marking).
(ii) Differential development: Again and again we observe that whether or not particular patterns arise depends on specific source constructions and developmental mechanisms. For example, constructions not applying to first and second persons give rise to ergative markers restricted in the same way, but ergative markers originating from constructions also used with first and second person (e.g. possessor markers in nominalizations) can be used with these person as well. Likewise, grammaticalization gives rise to overt markers restricted to plural if the source construction does not apply to singulars. Yet, in cases where number marking originates from elements applying to both singular and plural (such as gender markers with distinct singular and plural forms), both singular and plural are overtly marked. Some mechanisms can also give rise to patterns prohibited by particular synchronic universals. For example, when ergative markers are eliminated, they could be eliminated for pronouns rather than nouns, which would lead to a configuration conformal to the animacy/referential hierarchy, but there actually are a number of cases where they are eliminated for nouns by not for pronouns, leading to exceptions to the hierarchy. Similarly, as noted earlier, particular processes can give rise to patterns where singular is marked overtly and plural is zero marked, leading to exceptions to the number hierarchy. This suggests that (a) particular patterns are motivated in terms of the properties of particular source constructions and developmental mechanisms, not properties of the pattern in itself, and (b) the frequency of particular patterns (e.g. ergative markers restricted to particular NP types, or overt marking restricted to plural) depends on the frequency of particular processes that lead to those patterns, as opposed to processes that could lead to different patterns.
Martin Haspelmath: I also argued (this is point (C) in my earlier post) that functional-adaptive constraints do not lead us to expect uniform mechanisms of change. One could have a set of diverse mechanisms that lead to relatively uniform outcomes.
Sonia Cristofaro: True, but, if there is evidence that different mechanisms of change are motivated in terms of different factors, then, in order to claim that these mechanisms all reflect some overarching functional constraint, one needs to show in what ways exactly the constraint plays a role in individual mechanisms. In principle, this could be in two ways:
(i) The functional constraint is actually an additional factor triggering particular mechanisms. For example, zero marking for particular meanings as opposed to others can be a result either of the phonological erosion of an existing marker or the fact that zero marking was originally used for all of the relevant meanings, but then got restricted to some of those meanings because the language developed overt markers for the other meanings (for example through grammaticalization). These are very different processes, motivated in terms of different factors, but in principle one could claim that they are also both favoured by the fact that the meanings that end up being zero marked are the ones less in need of disambiguation. The only reason to make this claim, however, is that the grammatical patterns that do not comply with some postulated functional constraint (in this case, the need for disambiguation) are unattested or rarer than those that comply with that constraint. This could be due to reasons other than the postulated functional constraint (for example, the rarity of particular source constructions or uses thereof that could give rise to particular patterns, see above), so it should be investigated independently.
Martin Haspelmath: Yes, I agree that the main reason to propose a functional explanation is that the patterns disfavoured by the functional principle do not occur (or hardly occur). And yes, alternative explanations in terms of source constraints should be considered, but if none are available, I find functional-adaptive explanations plausible.
Sonia Cristofaro: Let me continue: (ii) Alternatively, the functional constraint could lead to differential transmission rates for different grammatical patterns, independently of how those patterns arise. This would be the equivalent of biological evolution, where natural selection leads to particular distributions independently of the various mechanisms that give rise to the traits involved in the distribution. In biological evolution, however, this is because adaptation to the environment leads to better transmission rates for a trait, in terms of higher survival rates of the organisms carrying the trait. In language change, on the other hand, we don’t usually have a clear equivalent of natural selection, in terms of differential survival rates for constructions depending on whether they comply with particular functional principles. For example, once overt marking for less frequent meanings arises in a language, it is usually maintained even if it is not economically motivated.
Martin Haspelmath: I don’t think so. I think that there is both differential survival of better adapted patterns, and differential creation of better patterns, as I argued in some detail in an old paper (Haspelmath 1999).
Sonia Cristofaro: You would have to demonstrate that for individual patterns described by typological universals, and that’s not been done. In fact, there is a long tradition of linguistic thought in which the transmission of individual constructions within a speech community is entirely determined by social factors independent of particular functional properties of the construction.
Martin Haspelmath: One example of different paths leading to the same functionally adaptive outcome is the various types of word shortening, such as clipping, abbreviation, and shortening by phonetic erosion. These are all motivated by the high frequency of the word (or morpheme) that gets shortened, but the mechanisms of change are different.
Sonia Cristofaro: I find that clipping and phonological erosion are one thing, and cases such as the contrast between (frequent and short) horse and (rare and long) hippopotamus are quite another. Clipping and phonological erosion lead to the shortening of longer words, and are plausibly due to entrenchment and automatization of phonetic gestures. In this case, shortening of a frequent form is adaptive, in the sense that it leads to less articulatory effort. In the case of hippopotamus however, speakers do not shorten an existing form, they rather create a new, longer form. The new form is longer because it involves additional meaning components, which are encoded by additional formal components (a hippopotamus is a particular type of horse, namely a river horse). Such cases don’t prove that having a short form is adaptive, all they show is that additional meaning components are encoded by additional formal components, which is a general organizational principle of linguistic structure (for example, verbal roots describe actions, but in order to encode the participants involved in the action additional elements must be used).
Martin Haspelmath: But as I noted in my (2008) paper on frequency vs. iconicity, words containing more meaning components are not necessarily longer. For example, cat contains more meaning components than animal, and star contains more meaning components than celestial body. As anthropologists have noted, “basic level” notions tend to be encoded by short forms, in contrast both to more specific items (longhair cat) and less specific items (animal). It seems that the reason for classifying something as basic level is that it occurs frequently.
Sonia Cristofaro: This doesn’t disprove the point. If a word (e.g. hippopotamus as opposed to horse) is longer because it involves more formal components, each encoding a specific meaning component, this doesn’t mean that words involving more meaning components should always be longer.
Martin Haspelmath: It seems to me that understanding the precise mechanisms of language change may not always be needed in order to achieve partial explanatory success. Scientific explanation can be successful even if the mechanisms of change are not known (for example in the case of biological evolution, where the correct explanation was proposed decades before the mechanism of genetic inheritance and mutation was discovered). So I would be happy to accept functional explanations of synchronic distributions even if little is known about the ways in which the changes happen.
Sonia Cristofaro: One may have correct intuitions about language change even if the mechanisms of change are not known, but in order for these intuitions to be proven correct (“explanatory success”) one has to demonstrate that some postulated principle really plays a role in the processes that one is trying to account for. While Darwin was not aware of the precise mechanisms of genetic inheritance and mutation, he produced compelling evidence for natural selection through adaptiveness. Likewise, while the mechanism for continental drift was not known when the theory was proposed, such a mechanism was discovered at a later stage. For language change, on the other hand, principles such as economy are postulated based on synchronic distributions, but then the evidence that we have about the actual origins of these distributions points to different principles, and there is no obvious evidence that the principles postulated on synchronic grounds play a role either in the development or in the transmission of the elements involved in the distribution (as noted earlier). This does not rule out that these principles might actually have a role in the relevant processes, but this has to be demonstrated, and cannot be assumed based on the synchronic distributions alone.
Martin Haspelmath: Many thanks, Sonia, for this discussion – let’s continue this debate at the SLE workshop in Naples.
Sonia Cristofaro: Here’s a last question to you: I have had problems with your notion of “pull forces”, which I find rather vague and very difficult to apply in practice in a scientifically coherent fashion. I must say it sounds very much early nineteenth century linguistics to me.
Martin Haspelmath: Yes, the term is indeed vague. But the idea is very simple: A “force” is just another word for ‘causal factor’, and a “pull factor” is one that can explain a variety of diverse phenomena leading to the same result. Human motivations are clearly a pull factor (we do many different things to achieve the same sort of results, e.g. to become rich), but of course languages don’t have motivations, and speakers don’t have goals for their languages (though earlier linguists sometimes sounded as if they did, as you note). So the “pull factors” in language structures must work differently, for instance via an adaptive invisible-hand process of the sort described by Keller (1994). There is still much we don’t understand well, so yes, this notion is difficult to apply. But sometimes I find it compelling.
Sonia Cristofaro: Thank you, I am looking forward to continuing the discussion in Naples or elsewhere.