The Comparative Method

Chapter 2 showed that sound change is regular. This chapter uses that fact. If sounds change regularly, then two languages that descend from one ancestor will show regular matches between their words, and nothing else produces a pattern like that.

Cognates

Cognates are words in different languages that have been "inherited in direct descent from an etymological ancestor" in a shared parent language (Wikipedia, Cognate). Their sound and meaning can change so much that the relationship needs careful historical study to confirm.

All four come from Proto-Indo-European *nókʷts, according to the source. English father, French père and Armenian hayr also descend from a common ancestor, even though they look quite different.

Look-alikes that are not cognates

• False cognates: Latin habēre and German haben both mean "to have" but come from different roots. English much and Spanish mucho are similar but not cognates.
• Loanwords: English beef was borrowed from Old French boef.
• Chance: English day and Latin dies match in meaning and sound, but that is chance.
• Borrowing that mimics kinship: English taboo resembles Polynesian forms because it was borrowed from Tongan.
• Heavy borrowing is not kinship: Persian's heavy Arabic borrowing does not make it related to Arabic.

The Steps

Wikipedia says there is no fixed procedure. Its outline, based on Campbell and Crowley, has five steps:

Worked Example 1: Polynesian

The page compares the words for "one", "two" and "three":

Hawaiian k regularly corresponds to t elsewhere (kahi / tahi, kai / tai). That is a correspondence set: a regular sound match, not a coincidence.

Worked Example 2: Romance

The sets are k : k and k : [ʃ]. French [ʃ] appears only before a, so the sets are complementary (step 3) and reflect a single proto-phoneme, *k (spelled c in Latin). That is the logic of conditioned sound change from Chapter 2, run backwards.

Shared Innovations

Two languages can share a feature for two reasons, and only one of them shows a close relationship.

• Shared innovations indicate a closer subgroup. The source's example is Grimm's law in English and German.
• Shared retentions do not. German and Russian both keep the dative / accusative contrast; that does not group them together, because it simply shows the loss happened later in English.

Limits of the Method

• Time depth: the page gives no fixed age limit. It says reconstruction can be only partial when languages are scarcely attested, the time depth is too great, or internal evolution obscures sound laws. Relatedness is then "plausible, but uncertain".
• Contact: areal diffusion and Sprachbunds can create false classifications (Chapter 6).
• Irregular change: random mutations, analogy (Russian dʲevʲatʲ "nine", influenced by "ten") and gradual lexical diffusion can obscure regularity.
• Lost features: Latin declension was lost in Romance, so it cannot be fully reconstructed from comparison.
• Certainty: reconstructions of unknown proto-languages are "only likely, not certain".

Try It Yourself