Generative and nongenerative diversity linguists do not often engage in debates on general issues, but the topic of word-classes (nouns, verbs and adjectives) seems to be an area where some kind of dialogue seems to be not impossible (cf. Baker 2003, Croft 2009, and the 2005 LSA Institute class taught jointly by Baker and Croft, presenting their contrasting approaches to the students). Now the open-peer-commentary journal Theoretical Linguistics has published a very nice target article on word-classes in Chamorro by Sandy Chung (Chung 2012a), with commentaries by various linguists, including three nongenerativists (Bill Croft, Eva van Lier, and myself). In a nutshell, Chung argues that a more detailed study of Chamorro shows that it has a noun-verb and a noun-adjective distinction after all, contrary to Topping’s (1973) analysis, where Chamorro has just two major word-classes, Class I (transitive verbs) and Class II (others). Chung thus sees her in-depth study as supporting the claim that the categories of noun, verb and adjective are universal (Baker 2003). The really nice thing about her article is that it provides a wealth of new data on Chamorro (with many naturally occurring examples), plus a general discussion that is accessible also to non-insiders.
The commentaries by Croft & van Lier (2012) and by myself (Haspelmath 2012) are very similar in the general thrust: We argue that if there is no principled limit on the kinds of constructions that can be used as “diagnostics”, many different groupings of words could be justified, and there is no particular reason to privilege the noun/verb/adjective classification (unless one has decided in advance that one wants to find evidence for it). In my own commentary, I sharpen the critique by noting that Chung’s generative (and thus aprioristic) approach is in danger of leading to consequences quite opposite to what she had intended: In the target article, she calls for “understudied languages to be investigated more intensively and in greater depth”, because “only then will we be able to come to grips with the full potential of universal grammar” (Chung 2012a: 50). But if one assumes a priori categories, then often the kind of am priori categories that dominate the discussion are the ones of English and other overstudied languages. So we fall into the ethnocentric trap: We look for evidence for our familiar categories, find some, ignore all other patterns, and conclude that all languages are like English or Latin. A non-aprioristic approach that assumes no pre-established categories thus has a greater chance of uncovering the full potential of linguistic diversity and unity.
In her reply to the commentaries, Chung (2012b: 141-142) writes:
“I take these remarks…to raise the following objection to the generative enterprise. First, generative grammar is a theory with deductive structure (“aprioristic”). Second, the universals proposed in generative grammar are often supported by evidence from what Haspelmath calls “the major languages”… Because of this, understudied languages cannot have a real impact on theoretical questions.”
She then objects that
“the conclusion does not follow. Theories with deductive structure lead to hypotheses that can be evaluated on the basis of empirical evidence, including evidence from understudied languages.”
Of course Chung is right that the conclusion does not follow in a strict sense. What I meant was that in practice, the influential hypotheses about universal grammar tend to come from the major languages, simply because these are known by most linguists. Someone claiming that all languages are like Chamorro in that they have two major word-classes (Class I and Class II) would not find many adherents, even though the claim is actually quite reasonable (transitive and intransitive words are strikingly different in that only transitive words take an object, and this applies to all languages).
The problem is not the “deductive structure” of generative linguistics. All scientific activity is partly deductive and partly inductive. It may be true that generative linguistics puts more emphasis on deduction that some other approaches, but in the area of word-class universals, this is surely not the case. Baker’s (2003) theory is not more deductive than Croft’s (2000) theory. What does distinguish the two is the crucial role of a priori categories in Baker’s theory and the emphasis on non-aprioristic language-particular analysis in Croft’s. Like all generative linguists, Baker and Chung assume that languages select from a universal, pre-established set of categories or features that is given a priori to its learners. But this is not the same as “deductive structure”. One may assume a priori features and not deduce them from anything else, or a priori features that can be deduced from something more general, or language-particular features that can be deduced, or purely inductive language-particular features.
Unfortunately, generative grammarians have so far not addressed the powerful arguments against universal, a priori features and categories, and Chung’s short reply to the commentaries is no exception to this general silence. So the real debate about the core issues separating generative and nongenerative linguists has still not taken place.
Chung, Sandra. 2012a. Are lexical categories universal? The view from Chamorro. Theoretical Linguistics 38(1-2). 1–56.
Chung, Sandra. 2012b. Reply to the commentaries. Theoretical Linguistics 38(1-2).
Croft, William. 2000. Parts of speech as language universals and as language-particular categories. In Petra M. Vogel & Bernard Comrie (eds.), Approaches to the typology of word classes, 65–102. Berlin: Mouton de Gruyter.
Croft, William. 2009. Methods for finding language universals in syntax. In Sergio Scalise, Elisabetta Magni, & Antonietta Bisetto (eds.), Universals of language today, 145–164. Berlin: Springer.
Croft, William & Eva van Lier. 2012. Language universals without universal categories. Theoretical Linguistics 38(1-2).
Haspelmath, Martin. 2012. Escaping ethnocentrism in the study of word-class universals. Theoretical Linguistics 38(1-2).
Topping, Donald. 1973. Chamorro reference grammar. Honolulu: University Press of Hawaii.
I’m currently wondering if statistical learning theory (especially `2 part MDL) might be able to help us out here, by a) providing a quantitative basis for assessing the relative ‘importance’ of word-class divisions, and their hierarchy b) likewise for not worrying too much about how much splitting to do. On the basis that the more important divisions are the ones that are the most effective in reducing the amount of info needed to specify the data given the grammar, or, for a more intuitive way of thinking about it, the ones whose mastery will result in fewer learner errors. E.g., in English, noun verb and adjective are much closer to being equally important than they are in Greek, where there is relatively less difference between nouns and adjectives, while in Chamorro and Indonesian, the differences between verbs and adjectives would seem to be relatively slight (here’s a nice video giving the learner error perspective from the point of view of Indonesians learning English: http://www.youtube.com/watch?v=7-Dq1b3SHEI).
So, I don’t think any sensible person could deny the significance of the four classes that Martin puts in table 5 of his discussion of Sandy’s article, but whether adjective is better regarded as a subclass of verbs or as an independent class on its own is a live issue, which a quantitative approach might settle (in favor of having it a subclass would be my estimation) and there’s also of course the transitive/intransitive issue, which I would consider to be an independent factor of ‘subcategorization’ however you want to implement that. Greek would give a different subclassing, while treating any of the three as a subclass in English would probably not be very useful (since all three have very different behavior both as predicates of clauses and components of NP). But calculating actual numbers would surely be better than the guessing I’m indulging in here, if we can figure out how to do it.
Charaterizing generativism as “a theory with deductive structure (‘aprioristic’)”, as Chung (2012) does, calls for a comment. In linguistics, there are no exclusively ‘deductive’ theories. Regardless of its internal structure, every theory is – and must be – inductive in the sense of having at least some data base. Only if the data base is inordinately restricted, e.g. to English (as is – really – the case with generativism; cf. E. Bach 2004), does the impression arise that the theory has no data base at all, i.e. that it is purely deductive. This comes close to what Martin Haspelmath said in his original posting.
Quibble: I’d question the universal relevance of transitive vs. intransitive, because, in English, this is just one of a huge number of subcategorizational classes:
directional PP required:
Fred sidled *(off the podium)
NP + directional PP:
John put *(the baby) *(in the carseat)
[but librarian jargon has a no PP use: Mary put the book]
‘Idiomatic’ PP required:
John depends *(on Mary)
[one can debate how idiomatic these really are, but the choice of
preposition is fixed, and the ‘on of dependence’ is limited to a handful
verbs (depend, rely, count are the only ones I can think of ATM)
John said *(that Bill was an idiot)
John got *(depressed)
(get meaning ‘become’ needs an AP or adjectivally interpreted
So it seems to me that there are interesting tendencies for certain cognitive/semantic distinctions to get aligned with the features that distinguish word-classes, but, ironically, ‘transitivity’ is not really defensible as an important feature in English (nor, as far as I can see, in Latin or Greek!), although Dixon makes a good case that it’s important in certain Australian languages (and in any language with incontrovertible inflection for both subjects and objects).
‘Happening’ vs ‘State’ otoh seems to be much more ‘important’ in some languages than it is in English, tho I think we (or at least I) need to develop a better idea of what ‘importance’ amounts to.
Yes, exactly! The transitive/intransitive distinction isn’t very “important” in English, just as the adjective/verb distinction isn’t very “important” in Chamorro, in some intuitive sense. Topping was trying to express this intuition, but this is of course not very rigorous. But neither is Chung’s solution, namely to look at “unimportant” aspects of Chamorro and take them as sufficient evidence for an adjective/verb distinction. As long as we don’t know what “importance” amounts to, asking questions about category universality is not very helpful. Instead, I find it more interesting to ask about the universality of the ingredients of categories, e.g. “Do all languages have arguments with “idiomatic” PPs?”, or “Do all languages with no special relative-clause marking of attributive property-words require a copula when the property-word is predicated?”, and so on.
The way I want to think about word-class universality is in terms of where part-of-speech splits do and don’t occur, for example words for ‘kinds of things’ almost always (maybe absolutely always, if Evan’s ‘kinship verbs’ can be explained away by claiming that kinship relations aren’t kinds) all fall into one open class part of speech, which we happily call ‘noun’ if it contrasts with a part of speech that has some kind of association with events (in Wierzbicka’s NSM, ‘says that something happens’). I haven’t yet managed to figure out if people such as yourself and Bill Croft are already thinking about it that way. Being more precise about this requires pinning down what a ‘kind’ is, of course.
Exactly why generativists have usually assumed that parts of speech had to be universal is something I’ve never really understood.
Roughly speaking, it’s of course correct that ‘thing’ words tend to fall into one open word-class in all languages, but the problem is that multiple criteria could be used for determining word classes. Often these criteria cross-cut each other, so the resulting word-classes are not mutually exclusive. There is no unique way of determining word-classes within a language – it depends on the researcher’s tastes. And cross-linguistically, one needs to apply the ‘same’ criteria to get comparable classes. I cannot say that both Latin and Mandarin have an adjective class, if I define adjectives in Latin as words that have a comparative degree form, and adjectives in Mandarin as words that can occur prenominally without -de.
The assumption of cross-linguistic (universal, or universally available) categories or features is deeply ingrained in generative grammar. I don’t think it’s absolutely crucial – UG could consist just of innate architectures, with no innate features. But Chomsky has always assumed innate features (maybe as part of his naive 17th century rationalism that simply ignores diversity), and almost everyone else has followed him in this.
Rather fussily, I think one should be careful to refer to ‘kind of thing’ word, since nouns don’t really refer to ‘things’ (some noun phrases do). Especially when nouns are incorporated. And I’ve never heard of a subdivision of ‘kind of thing’ words that had an effect on word order, as opposed to the concord and agreement phenomena. E.g. a language where ‘masculine’ nouns took their modifiers in front, ‘feminine’ after. So I think more could be said than has been about what conceptual categories ‘fracture’ into different kinds of syntactic categories. ‘Evaluative’ concepts are for example highly fracturable, since, in English verb adjective, and noun formats are all available (that script is a dog; first two illustrated in my post above).
One thing I notice missing from the article and discussion of it is the possible role of Aktionsart, as well as intransitive event verbs with an external argument vs those with only an internal one (run vs fall(accidentally)). The differences between ‘noun’ and ‘intransiverb+adjective’ seem to be more substantial (that is, harder to explain away in terms other than a syntactic category difference) than those between ‘adjective’ and ‘intransiverb’, and the latter seem to me to be plausibly, on the basis of what appears in the article and discussion, just a consequence of something like an eventive/stative contrast. If the two behavioral differences Chung adduces can be triggered by a semantic/conceptual difference, there is no clear need for a part of speech distinction. We can’t do this for verbs vs adjectives in English because there are essentially synonymous constructions of both types: that movie is (very) bad/*that movie sucks.
It would seem that apriorists haven’t learnt the lessons which Franz Boas laid out for the analysis of languages in Handbook of American Indian Languages just over a century ago. By all means have classes such as ‘nouns’ and ‘verbs’ as working hypotheses, but be prepared to jettison them if they don’t correspond to what one finds in a particular language. This is even truer of languages such as Tetun Dili (and in large measure English) than it is of Chamorro. I’m not yet able to access Sandy’s article (devoted to Chamorro though I am) and wonder what she makes of Topping’s Class III words.
There is a PDF of this article on her webpage: