October 14, 2007

Fitchifying Language Change

This is the second of two posts today on the fitchification of aspects of historical linguistics. The first focused on a misunderstanding of the history of (modernish) linguistics in Tecumseh Fitch's recent Nature article. This one is on his odd views about the nature of language change. Some are just simple mistakes, or at best uncritical acceptance of highly controversial views on the topics he's addressing. Examples are his claim that Proto-Indo-European was spoken `some 10,000 years ago' (the prevalent view among Indo-Europeanists, who are the people who have the linguistic evidence, is more like 6,000 years); his belief that the family tree of related languages was `the crowning achievement' of 19th-century historical linguistics (see comments in my earlier fitchification post today); and his belief that people `don't generally invent words or grammatical forms'. On this last one, anyone who was ever a teenager, or who ever even met one, should know better -- in addition to the world of teenage slang and all those corporate inventors of terms for new products, consider novel constructions like the -f***in'- infix in words like abso-f***in'-lutely. (No, of course those asterisks aren't necessary on an enlightened forum like Language Log. I just think they're cute.) And community-wide examples of deliberate linguistic changes, at all levels of linguistic structure, are turning up quite frequently these days -- changes, that is, that alter an entire language, typically in a small speech community. But Fitch also reveals a more interesting misunderstanding in his conception of what language change, especially lexical change, is.

The problem is that Fitch seems to be conflating three very different change processes. First, there's lexical replacement, as when early English hound was replaced by dog as the generic term for the animal, or when Old English deer was replaced by the borrowed word animal as a generic term for a beast. Second, there's analogic regularization, as when the verb fly, without giving up its inherited past tense flew, acquired an alternative past tense form flied (with a slightly different meaning) in baseball expressions like he flied out to left field. There are also occasional analogic changes in the opposite direction, away from global regularization, as when wear changed from a weak past tense, which would have become weared if it had survived into Modern English, to a strong past tense wore on the analogy of rhyming verbs like tear, swear, and bear; it is not true that analogic change always favors the majority pattern, so Fitch is wrong when he says that irregular verbs `remain only as irregular residues'. And third, there is sound change, which (according to the regularity hypothesis, which dates back to the late-19th-century Neogrammarians) is blind to considerations of morphology, semantics, and -- crucially in Fitch's context -- frequency.

Fitch emphasizes frequency of occurrence of particular words because that is the focus of the two articles he is surveying in his essay (see Mark's post of a few days ago for the links). One set of authors argues that the most frequent words resist analogic regularization; the other set argues that the most frequent words resist change, at least in Indo-European languages. Fitch acknowledges that the `realization that frequency of use has a significant role in language change is nothing new', but claims that `the use of sophisticated methods...to quantify these relationships is an important step forward'. He does not say why it's an important step forward, but never mind. It's nice to have quantitative evidence to support what historical linguists have known for a hundred years or so, and it's not necessarily a trivial result.

Some of Fitch's examples are unfortunate choices as illustrations of his thesis. He says that `high-frequency English verbs retained their ancestral irregular state ("go/went" or "be/was")'. It depends on what he means by `ancestral', of course, but he might be interested to know that some of the irregularities in these two verbs are not all that old. Old English had a regular past tense for go; the earliest example of went given in the Oxford English Dictionary is from 1484, five or six hundred years after the earliest documented Old English. So this is another case like wear, where a regular verb has become irregular over time. The history of be is more complicated. According to the OED, the modern paradigm of this verb is an amalgam of three different Proto-Indo-European verbs -- PIE *es- `be' in the present tense (am, are, is, are, with analogic changes merging several forms); PIE *bhew- `become' in non-finite forms (be, being); and PIE *wes- `remain, stay, continue to be' in the past tense (was, were, again with several analogic mergers). The inherited verb for `become' was still a distinct verb in Old English, and only later merged with the other two combined verbs as the infinitive. So this most irregular of Modern English verbs, like go and wear, has in fact become more irregular in recorded history. Of course these quasi-counterexamples don't affect the statistical patterns that show that frequency affects the tendency for regular formations to spread analogically at the expense of irregular formations. They do suggest that it's a good idea to check one's facts.

Fitch's most serious mistake, though, is not his choice of examples that don't show what he thinks they show. His most hair-raising error is the flat assertion that `frequently used words are resistant to change'. This is clearly true if the type of change is analogic regularization. It may or may not be true of lexical replacement (and I admit that I haven't yet read the articles he refers to, to see if they make any distinctions according to the type of lexical change); it won't be surprising if it holds there too. But it most certainly is not true of lexical changes due to sound change. The Neogrammarians' hypothesis that sound change is inevitably regular runs into problems, primarily (but not entirely) having to do with the differential spread of sound changes through a speech community. However, the fact that the Comparative Method, which rests heavily on the regularity hypothesis, has proved to be effective in the establishment of language families and the reconstruction of proto-languages for all but a very small subset of the thousands of human languages is in itself evidence that regular sound change is the norm, not the exception. And regular sound change is indeed blind to frequency and all other nonphonetic contextual factors. So it is nonsense to say that frequent words resist change unless one qualifies the statement to exclude regular sound change. When Fitch speculates that `new phonological forms might arise less often for high-frequency words because errors of perception, recall or production are less common for frequently used words', his reference to perception and production clearly includes lexical changes due to sound change. He himself does not mention pronouns. A member of one of the two author sets for the quantitative Nature articles (I didn't catch the interviewee's name), interviewed the other day on a BBC radio program, did cite pronouns as examples of unlikely-to-change words, and certainly they are among the most frequent words in Indo-European languages. But pronouns aren't all that unlikely to change; see this earlier Language Log discussion of the `super-stable pronoun' hypothesis (it's a recurring theme).

I've been critical of Fitch's article, so I should emphasize one point on which we agree: he is quite correct (if unoriginal) to say that an `adequate explanation for [what he quaintly calls] glossogenetic phenomena must incorporate individual and collective levels of description, and show why they are necessarily related'. I'm skeptical about the likelihood that statistical analyses imported from other disciplines will solve this difficult problem. It's not that statistical explorations of relationships between frequency and lexical replacement and/or regularization, and in other areas of language change as well, should be ignored; they have produced very interesting results in a variety of domains. But errors of the sort that Fitch makes show clearly that new approaches to historical linguistic analysis will be successful only to the extent that they take into account the results of historical linguistic investigations over the past hundred and fifty years or so. Failing to learn something about a field one wishes to contribute to is all too likely to lead to reinvention of the wheel at best, and to a garbage in/garbage out problem at worst.

Posted by Sally Thomason at October 14, 2007 10:00 PM