Linguistic archeology The tree of knowledge

The Economist, Aug 25th 2012, 13:46 by J.P.

http://www.economist.com/blogs/johnson/2012/08/linguistic-archeology

TREES are a gift to students of the past. An entire discipline, known as dendrochronology, is devoted to using tree rings to date ancient wooden objects and buildings. Linguistic archaeologists, it seems, share these arboreal inclinations, though the trees they examine are of an altogether different species.

In 2003 a team led by Quentin Atkinson, of the University of Auckland, in New Zealand, employed a computer to generate a genealogical tree of Indo-European languages. Their model put the birth of the family, which includes languages as seemingly diverse as Icelandic and Iranian, between 9,800 and 7,800 years ago. This was consistent with the idea that it stemmed from Anatolia, in modern-day Turkey, whence it spread with the expansion of farming. A rival proposal, that their origin amid the semi-nomadic, pastoralist tribes in the steppes north of the Caspian Sea, supposes their progenitor to be several thousand years younger. 

Some proponents of the steppe hypothesis remained unconvinced. They pointed out that the computer-generated phylogeny, to give the tree its technical name, showed only how Indo-European tongues evolved over time. It said nothing about how they spread across space. As Dr Atkinson and his colleagues report in Science, this issue has now been addressed. The results lend further credence to the Anatolian theory.

Linguistic archaeologists have even less to go on than their peers in other past-oriented disciplines, who can at least pore over the odd trinket for clues to mankind's prehistoric ways. The earliest written records date back less than 6,000 years, long after "proto-Indo-European" is believed to have emerged. Researchers do, however, enjoy an abundance of data about contemporary languages. Because tongues change less chaotically than other aspects of culture, this is more useful to someone studying linguistic prehistory than it might appear.

Dr Atkinson began by collecting basic vocabulary terms—words for body parts, kinship, simple verbs and the like—for 83 modern languages as well as 20 ancient ones for which records are available. For each family, Dr Atkinson and his team identified sets of cognates. These are etymologically related words that pop up in different languages. One set, for example, contains words like “mother”, “Mutter” and “mere”. Another includes “milk” and “Milch”, but not “lait”. (Here is the whole list; known borrowings, such as "mountain" and "montagne" were excluded, as they do not stem from a common ancestor.) Then, for each language in their sample, they added information about where it is spoken—or is thought to have been, based on where ancient texts were discovered—and in what period. The result is a multidimensional Venn diagram that records the overlaps between languages.

Each of the 103 languages, with its cognate sets, temporal and geographical range, constituted one leaf of the Indo-European family tree. The tricky part was filling in the branches. Here, Dr Atkinson resorted to rolling of the dice, using a method called Markov-chain Monte Carlo. This generates a random set of boughs (each assigned its own randomly generated cognate sets, time and place) that fits the known foliage. Next, an algorithm calculates how likely it is that this tree would sprout the modern leaves given the way languages evolve and travel. For instance, it is assumed that a cognate can only be gained once, by an ancestral language, but lost many times, whenever it disappears from any of the descendants. And languages, or at least their speakers, might migrate in any direction, though less readily across water or mountain ranges, say, than through plains and valleys.

The first rolls of the dice are unlikely to offer a good fit. They might, for example, have Icelandic and Iranian as siblings, as opposed to distant cousins. So the algorithm tweaks the tree, again at random, and decides whether the new branches are any better. If so, they are kept; if not, the algorithm reverts to the previous tree in the series. Repeat this process long enough, typically millions of times, and a point is reached where no further improvement is possible. Let a forest of such equally likely trees grow, then look at the number of those with roots in Anatolia and the steppes. The proportions reflect the relative likelihood that either of the hypothesis is correct. 

Dr Atkinson's findings leave much less room for doubt. The Anatolia-rooted trees are orders of magnitude more numerous than those growing out of the steppes (see picture; an animated version of Indo-European peregrinations is available here). The researchers verified the method's validity by getting it to retrace the evolution of modern romance languages from its Roman roots. The model returned an accurate reconstruction, closely in keeping with historical records. In linguistics, then, cultivating trees pays. So does a bit of gambling.

(Picture credit: Quentin Atkinson)