The evolutionary psychology of irregular morphology

« previous post | next post »

Yesterday, Mr. Verb asked some questions about morphology and politics:

On News Hour just now, I swear I heard Bush talk about the Tibe[tʃ]an people. I'm puzzled. This is a case of /t/, like the last sound in Tibet, affricating, that is, becoming a 'ch' sound. That is hardly in and or itself striking — actually is regularly pronounced a[ktʃ]ually. But this doesn't usually happen in this environment. Put an -an on Montserrat and see if you get a [t] or an affricate for the adjective form for that place. […]

Is there some pattern here I don't know about? Bush wasn't obviously reading, so that kind of reading-based pronunciation error is probably out. Is Bush treating this (by analogy?) like -tion suffixes? Was he extending the pattern of affrication noted above? Is he really and truly not a competent speaker of English? What's happening?

As it happens, this is a question that I can answer.

Back in January of 2004, I noted that President Bush is prone to "a culpable consistency in the derivation of ethnonyms". In particular, he sometimes adds -ian when some other ending, or even full-out suppletion, is standard: Kosovians, East Timorians, Grecians. And if you add an i-initial ending to a t-final stem, what do you get? Palatalization, hence [tʃ].

The funny thing is, these mistakes make a very messy system somewhat more regular and rational; and consistency and rationality are goals that prescriptive theorists often urge us to strive for, when they're not berating us for abandoning some instance of historically-sanctioned irregularity. Here's what I've written about this in my lecture notes on morphology for Linguistics 001:

[O]ne of the ways that morphology typically differs from syntax is its combinatoric irregularity. Words are mostly combined logically and systematically. So when you exchange money for something you can be said to "buy" it or to "purchase" it — and we'd be surprised if (say) groceries, telephones and timepieces could only be "purchased," while clothing, automobiles and pencils could only be "bought," and things denoted by words of one syllable could only be "acquired in exchange for money."

Yet irrational combinatoric nonsense of this type happens all the time in morphology. Consider the adjectival forms of the names of countries or regions in English. There are at least a half a dozen different endings, and also many variations in how much of the name of the country is retained before the ending is added:

-ese Bhutanese, Chinese, Guyanese, Japanese, Lebanese, Maltese, Portuguese, Taiwanese
-an African, Alaskan, American, Angolan, Cuban, Jamaican, Mexican, Nicaraguan
-ian Argentinian, Armenian, Australian, Brazilian, Canadian, Egyptian, Ethiopian, Iranian, Jordanian, Palestinian, Serbian
-ish Irish, British, Flemish, Polish, Scottish, Swedish
-i Afghani, Iraqi, Israeli, Kuwaiti, Pakistani
-? French, German, Greek

And you can't mix 'n match stems and endings here: *Taiwanian, *Egyptese, and so on just don't work.

To make it worse, the word for citizen of X and the general adjectival form meaning associated with locality X are usually but not always the same. Exceptions include Pole/Polish, Swede/Swedish, Scot/Scottish, Greenlandic/Greenlander. And there are some oddities about pluralization: we talk about "the French" and "the Chinese" but "the Greeks" and "the Canadians". The plural forms "the Frenches" and "the Chineses" are not even possible, and the singular forms "the Greek" and "the Canadian" mean something entirely different.

What a mess!

It's worse in some ways than having to memorize a completely different word in every case (like "The Netherlands" and "Dutch"), because there are just enough partial regularities to be confusing.

So why do we (like pretty much every other language) behave in such a logically disgraceful way?

I don't have an answer.

Or rather, I have too many answers, and no idea which of them if any are correct.

In "The theology of phonology", 1/2/2004, I gave two answers:

James McClelland, Mark Seidenberg and others think that quasi-regularity arises because the (partial) regularities are emergent properties of connectionist networks; Steven Pinker, Michael Ullman and others think that quasi-regularity arises because there are two distinct and competing brain mechanisms whose functions overlap, one a (temporal/parietal-lobe) semantic memory system for looking things up, and the other a (frontal-lobe and basal ganglion) procedural memory system for figuring things out.

(See "The curious case of quasiregularity", 1/15/2004, for further discussion.)

Another obvious idea is that irregularly complex morphology serves various evolutionary functions precisely because it's hard to learn. It helps to distinguish native members of a speech community from parvenus, for example. And mastery of such extravagantly wasteful systems might be useful in sexual selection, perhaps by virtue of the Handicap Principle.

On the other hand, there is a paradoxical effect that may apply to our current president. In some cultures, men signal high status by ostentatious lack of fluency: morphological over-regularization, inarticulateness, even speech defects. For some details, see "Status and fluency", 5/11/2004, and "Clarifying status in Wolof by fake disfluency", 5/20/2004.

[Update — Mark Seidenberg writes:

Our account of inflectional and derivational morphology has a number of parts. As applied to the question of ethnonyms, the basic idea would be that the form that people produce is that which best satisfies some number of probabilistic or "soft" constraints, which are picked up in the course of learning a language. There may be many of these and there may be occasional true exceptions. I can't specify what all the constraints are in the ethnonym case because I haven't studied it, and one certainly could not tell from that limited list of examples. However, as an exercise, a person might try to use a simple neural network to discover what the constraints are. With some understanding of the constraints, one might then have a basis for investigating how they came into the language, i.e., why these constraints and not some others?

One way would be to train a simple network to associate a country name with the ethnonym, using say a feedforward network with a layer of hidden units. Train the netwok on a suitably large number of country names and ethnonyms, possibly modifying the training to reflect relative frequencies of usage (chinese high, burmese lower). train the network until it learns this task pretty well (don't overtrain or it will just memorize the forms). then analyze the network to see what bases it uses for performing the task. This would be fun, if you have the right attitude.

One could analyze the underlying representation the model developed (over the hidden layer) or test the model on novel country names, seeing what forms it produces as generalizations, and comparing the model's performance to people's. Extra credit: why "Flatbushtuns" not "Flatbushistanis" (

Cassidy, et al. (Cassidy, K. W., Kelly, M. H., & Sharoni, L. J. (1999). Inferring gender from name phonology. Journal of Experimental Psychology: General, 128, 362–381), used a similar approach on a simpler problem, discovering phonological differences between male and female names in English.

Another possibility would be to use a set of candidate constraints as input (the country name represented at a reasonable level of phonological detail, geographical information, ethnic information, whatever), and training the network to produce the appropriate ethnonym as output.

Again there are pretty standard techniques for analyzing how the model learns to perform the task, i.e., which factors matter. (You experiment with the network, see how it generalizes, zero out inputs to see how the model's behavior changes, analyze the underlying representations using methods like cluster analysis, etc.) This requires have some a priori ideas about possible factors and so is not a fully automatic discovery procedure, but that's OK.

It would probably be important to capture some diachronic factors in the training, e.g., how the early establishment of one form affected forms that were created later.

Of course, a person would also want to account for individual differences by training different models, giving them knowledge of different potential constraints, different starting weights, etc. Bush's overgeneralizations suggest that he is working with far too few hidden units for the task. (He can pick up on simple generalizations but his network does not have the capacity to encode more complex interactions among factors. Of course he may also lack some relevant knowledge, e.g. of geography.) This account appeals to properties of Bush's brain, such as it is, rather than the behavior of male peacocks or assumptions about Mesolithic hunter-gatherer societies. This also has the advantage of being able to explain his tendency to exhibit the same behavioral limitations in other, very different domains, instead of invoking a different evolutionary principle for each one.

For a more recent paper on our approach to derivational morphology, see Gonnerman, L.A., Seidenberg, M.S., & Andersen, E. (2007). A distributed connectionist approach to morphology: Evidence from graded semantic and phonological effects in lexical priming. J. Exp. Psych: General, 136, 323-345.

This recent exchange:

Berent, I., & Pinker, S. (2007) The dislike of regular plurals in compounds: Phonological or morpholoigcal. The Mental Lexicon, 2, 129-181.
Seidenberg, M.S., MacDonald, M.C., & Haskell, T. (2007). Semantics and phonology constrain compound formation. The Mental Lexicon, 2, 287-312.

provides a good look at where the debate is at.

Both Mark Seidenberg and Steve Pinker are psychologists. Though their views on about morphological (ir)regularity are serious and interesting, they don't represent the views of most of the many linguists who have looked at such questions over the years. My somewhat unserious post also didn't try to cover linguists' ideas about the nature and sources of morphological patterns in general, or English ethnonyms in particular — some day I should take up that task. Meanwhile, there's a bit more discussion of one range of subcases in "Who let the 'n' in?", 1/22/2006. ]

Comments are closed.