The Romance languages again provide our best documented example. As of 500 BC, Latin was confined to a small area around Rome and was only one of many languages spoken in Italy. The expansion of Latin-speaking Romans eradicated all those other languages of Italy, then eradicated entire branches of the Indo-European family elsewhere in Europe, like the continental Celtic languages. These sister branches were so thoroughly replaced by Latin that we know each of them only by scattered words, names, and inscriptions. With the subsequent overseas expansion of Spanish and Portuguese after 1492, the language spoken initially by a few hundred thousand Romans trampled hundreds of other languages out of existence, as it gave rise to the Romance languages spoken by half a billion people today.

If the Indo-European language family as a whole constituted a similar steamroller, we might expect to find its trampled debris in the form of older non-Indo-European languages surviving here and there. The sole such vestige surviving in Western Europe today is the Basque language of Spain, without known relations to any other language in the world. (The remaining non-Indo-European languages of modern Europe—Hungarian, Finnish, Estonian, and possibly Lapp—are relatively recent invaders of Europe from the east.) However, there were other languages that were spoken in Europe until Roman times, and of which enough words or inscriptions have been preserved to identify them as non-Indo-European. The most extensively preserved of these vanished tongues is the mysterious Etruscan language of northwest Italy, for which we have a 281-line text written on a roll of linen that somehow ended up in Egypt as wrapping for a mummy. All such vanished non-Indo-European languages were part of the debris left from the Indo-European expansion. Still more linguistic debris was swept up into the surviving Indo-European languages themselves. To understand how linguists can recognize such debris, imagine that you, as a freshly arrived visitor from outer space, were given one book each, written in English by an Englishman, an American, and an Australian, about his or her country. The language and most of the words in all three books would be the same. But if you compared the American book with the one about England, the American book would contain many place names that were obviously foreign to the basic language of the books—names like Massachusetts, Winnepesaukee, and Mississippi. The Australian book would contain more place names equally foreign to the language but unlike the American names—such as Woonarra, Goondiwindi, and Murrumbidgee. You might guess that English immigrants coming to America and Australia encountered natives who spoke different languages, and from whom the immigrants picked up names for local places and things. You would even be able to infer something about the words and sounds of those unknown native languages. We actually know the native American and Australian languages from which those borrowings took place, and we can confirm that your indirect inferences from the borrowed words alone would have been correct.

Linguists studying several Indo-European languages have similarly detected words borrowed from vanished, apparently non-Indo-European languages. For example, about one-sixth of Greek words whose derivations can be traced appear to be non-Indo-European. These words are just the sort that one might expect to have been borrowed by invading Greeks from the natives they encountered: place names like Corinth and Olympus, words for Greek crops like olive and vine, and names of gods or heroes like Athene and Odysseus. These words may be the linguistic legacy of Greece's pre-Indo-European population to the Greek speakers who overran them. Thus, at least four types of evidence indicate that Indo-European languages are the products of an ancient steamroller. The evidence includes the family-tree relationship of surviving Indo-European languages; the much greater linguistic diversity of areas like New Guinea, that have not been recently overrun; the non-Indo-European languages that survived in Europe into Roman times or later; and the non-Indo-European legacy in several Indo-European languages. Given this evidence for an Indo-European mother tongue in the distant past, can one reconstruct something of this tongue? At first, the notion of learning how to write a vanished unwritten language seems absurd. In fact, linguists have been able to reconstruct much of the mother tongue by examining word roots shared among its daughter languages.

To take an example, if the word meaning 'sheep' were totally different in each modern Indo-European language branch, we could conclude nothing about the word for 'sheep' in the mother tongue. But if the word were similar in several branches, especially in ones as geographically distant as Indo-Iranian and Celtic, we might infer that the various branches had wherited the same root from the mother tongue. By knowing what sound shifts have taken place among the various daughter tongues, we could even reconstruct the form of the word root in the mother tongue.

In many modern Indo-European languages, as well as in some ancient ones that we know from preserved writings, the words meaning 'sheep' are quite similar. These words must have been derived from an ancestral form that is inferred to have been owis and that was used in proto-Indo-European (PIE), the unwritten mother tongue.

As the figure above shows, the words for 'sheep' in many Indo-European languages from India to Ireland really are very similar: avis, hawis, ovis, ois, oi, etc. The modern English 'sheep' is obviously from a different root, but English retains the original root in the word 'ewe'. Consideration of the sound shifts that the various Indo-European languages have undergone suggests that the original form was owis.

Naturally, the same word root shared among several daughter languages does not automatically prove shared inheritance from the mother tongue. The word could also have spread later from one daughter language to another. Archaeologists sceptical of linguists' attempts to reconstruct mother tongues love to cite words like 'Coca-Cola', shared among many modern European languages. The archaeologists claim that linguists would absurdly attribute 'Coca-Cola' to the mother tongue of thousands of years ago. In fact, 'Coca-Cola' illustrates how linguists weed out recent borrowings from old inheritances: the word is obviously foreign ('coca' is actually from a Peruvian Indian word, 'cola' West African), and it does not exhibit the same sound shifts among languages as do old Indo-European roots (in German it is still 'Coca-Cola', not Kocherkohler). By such methods, linguists have been able to reconstruct much of the grammar and nearly 2,000 word roots of the mother tongue, termed proto-Indo-European but usually abbreviated as PIE. That is not to say that all words in modern Indo-European languages are descended from "IE: most are not, because there have been so many new inventions or borrowings (like the root 'sheep' replacing the old PIE root owis in English). Our inherited PIE roots tend to be words for human universals that people surely were already naming thousands of years ago: words for the numbers and human relationships (as in the table on page 226); words for body parts and functions; and ubiquitous objects or concepts like 'sky', 'night', 'summer', and 'cold'. Among the human universals thus reconstructed are such homely acts as 'to break wind', with two distinct roots in PIE depending on whether one does it loudly or softly. The root for doing it loudly (PIE perd) gave rise to a series of similar words in modern Indo-European languages (perdet, pardate, etc.)—including English 'fart' (see figure on page 235, and sample text on pages 248—9). So far, we have seen how linguists have been able to extract, from written languages, evidence of a pre-literate mother tongue and steamroller. The obvious next questions are: when was PIE spoken, where was it spoken, and how was it able to overwhelm so many other languages? Let's begin with the matter of'when', another seemingly impossible question. It is bad enough that we have to infer the words of an unwritten language; how on earth do we determine when it was spoken?