To establish, verify and prove connections between languages in historical linguistics, i.e. the affiliation to language families, the the primary and basic method is the comparative method. For (possible) distant etymologic relationship however additional specific methods and procedures are employed in order to carefully evaluate the data and avoid erroneous conclusions. Following is a summary of according methods, but also traps, based for a great part on an overview by Campbell (2003).
1. Lexical comparison
Words comparison is an obvious and frequently adopted method. For supposed far-reaching relationships just a handful of possible cognates is no convincing evidence, because similarities could be due to chance and other factors like phonological preferences or indirect contact, etc. In order to be convincing a critical mass and a certain level of quality — both in form and semantics — is required. If the number of potential cognates is high enough, accidental similarities become less likely. With regard to semantics, consistently good matches in meaning (especially with very specific or rather complex glosses) may counter weaknesses in isomorphism (cf. semantic constraints below). The reliability of lexical comparison may vary significantly depending on properties of the languages in question.
Worth discussing is the role of chance in lexical comparison: what is the probability of accidental similarities in both form and semantics? It certainly depends primarily on the length of the compared words (though obviously we can't simply apply statistical mathematics) and on the found number of similar pairs; i.e. with growing number of pairs the probability of accidental matches diminishes. Another aspect that should be mentioned is that for lexical comparison basic vocabulary be used (like kinship terms, animal names, etc.), such that mostly nouns are consulted and, to a lesser extent, verbs, cf. Swadesh lists. However, also words like 'toward, many, up' and adjectives can be rewarding, if a sufficient number can be found to exclude diffusion / borrowing.
Examples of false similarities which are baffling due to their semantic correspondence are Kaqchikel mes ‘mess, disorder, garbage’/English mess; Jaqaru aska ‘ask’/English ask, and there are many others which must be considered accidental. Phonetically similar and sometimes almost identical, but semantically dissimilar examples are poeta ‘drunk’ in Maipure, but ‘poet’ in Italian; putta in Otomaco 'head', while in Italian 'prostitute'.
Glottochronology is rejected by most linguists. It isn't suited to detect relationships, but it rather attempts to date base vocabulary splitting points in time for hypothetical or accepted relationships between languages. The fading of similarities in lexeme pairs over time at an assumed rate is applied for the dating: "comparable lexemes must inevitably diminish to near the vanishing point the deeper one goes in comparing remotely related languages" (Bengtson 1989: 30, by Campbell 2003). Though there is no doubt that in time similarities diminish, this is a very problematic assumption, because there is no fixed or reliably determinable rate for such bleaching or divergence, like a coefficient. There might be significant differences depending of the language pair, including phonological and morphological particularities, and historical circumstances like displacement and migration, number and intensity of contacts with other languages, size of the group/tribe/people, etc.
This method is summarized by Greenberg (1987: 23, by Campbell) as "looking at (...) many languages across a few words" instead of "a few languages across many words". The lexical similarity shared across many languages is considered evidence of genetic relationship. According to Campbell, the results of this method have proven mostly erroneous or at best controverted.
2. Sound correspondences
Sound correspondence is a very important — if not crucial — method to determine and account for genetic relationships. The sounds of word pairs don't necessarily need to be similar, more important is the resolution of phonological processes that lead to the resulting sound shapes. For example "Armenian hing/English five ((...) derived by straightforward changes from original
Indo-European (IE) *penkwe ‘five’), French boeuf/English cow (from PIE *gwou-), French /nu/ (spelled nous) ‘we, us’/English us (from PIE *nes-; French through Latin nōs, English from Germanic *uns [IE zero-grade *ns])" (Meillet 1948 : 92–3, by Campbell).
3. Grammatical evidence
Grammatical evidence is considered very conclusive to establish true relationship between languages.
"The more singular the facts are by which the agreement between two languages is established, the greater is the conclusive force of the agreement. Anomalous forms are thus those which are most suited to establish a 'common language'." (Meillet 1925, quoted from 1967: 41, by Campbell 2003: 269)
According to Campbell, Meillet's method is still considered standard practice, and Sapir's "submerged features" are considered close to Meillet's method.
When one passes from a language to another that is only remotely related to it, say from English to Irish or from Haida to Hupa or from Yana to Salinan, one is overwhelmed at first by the great and obvious differences of grammatical structure. As one probes more deeply, however, significant resemblances are discovered which weigh far more in a genetic sense than the discrepancies that lie on the surface and that so often prove to be merely secondary dialectic developments which yield no very remote historical perspective. In the upshot it may appear, and frequently does appear, that the most important grammatical features of a given language and perhaps the bulk of what is conventionally called its grammar are of little value for the remoter comparison, which may rest largely on submerged features that are of only minor interest to a descriptive analysis.
(Sapir 1925: 491–2; by Campbell)
Swadesh tested Sapir's method (and ability to distinguish loaned from true relationship) with English and French and was impressed by the result, where irregularities were unlikely to have overcome by borrowing (Swadesh 1951: 7, by Campbell). An example is English good/better/best vs. German gut/besser/best- (Greenberg 1957: 37–8, 1987: 30, by Campbell).
Diffusion is an obvious source of similarities among languages that have come into contact geographically and implicit culturally, directly or indirectly (via intermediary peoples or even individuals like merchants or mercenaries importing words for which the language in question didn't have a notion). Distinguishing loanwords and other borrowed language features from cognates is a necessary procedure.
5. Semantic constraints
The closer word pairs are semantically (besides phonological similarity), the more reliable is the match. It is much easier to find phonetically matching words with dissimilar semantics, but of those, very few cases can be traced back to a common ancestor. Swadesh recommends to count only exact equivalences (Swadesh 1954: 314, by Campbell).
Sound-imitating words may develop independently from the same sound and are therefore to ignore in a comparative analysis for etymological relatedness.
7. Sound symbolism
Sound symbolism develops in languages primarily based on size and on shape. Therefore here too, caution is recommended.
8. Nursery forms (so-called Lallwörter)
Lallwörter like mama, nana, papa, dada, caca, pipi should be avoided because they show very high levels of similarity, evidently ( Jakobson 1960, quoted from 1962: 538, by Campbell) without common ancestry. Jakobson's argumentation and interpretation of the reasons for these striking similarities however, is not convincing.
9. Short forms and unmatched segments
The more segments / syllables of the compared words match, the better the likelihood for cognates."Monosyllabic CV or VC forms may be true cognates, but they are so short that their similarity to forms in other languages could also easily be due to chance" (Campbell). Very specific, matching semantics however (like grammaticalized markers) strengthens the hypothesis for genetic relatedness.
Similarly problematic is, if some word segments match, while others don't. But here too there are exceptions, namely with languages of more or less rich morphology lice agglutinative languages, which may have lexicalized the historical term together with other morphemes, like markers. Such cases can be found for example in what Quintero (2004) calls ki-words in Osage.
10. Similarities by chance
Avoidance of accidental matches is crucial.
Resemblances between languages do not demonstrate a linguistic relationship of any kind unless it can be shown that they are probably not the result of chance. Since the burden of proof is always on those who claim to have demonstrated a previously undemonstrated linguistic relationship, it is very surprising that those who have recently tried to demonstrate connections between far-flung language families have not even addressed the question of chance resemblances. This omission calls their entire enterprise into question. (Ringe 1992: 81)
Doerfer (1973: 69–72, by Campbell) discusses two types of similarity by chance.
- Statistical chance is dependent on how many and what types of words can be expected to be similar by accident, and he gives the example of the names of 79 Latin American Indian languages which all begin with na-: Nahuatl, Naolan, Nambicura, etc.
- Dynamic chance refers to similarity by convergence, where there is no original common ground, but where similarities develop due to sound change. An example is French feu 'fire' vs. German Feuer 'fire', where feu stems from Latin focus, while Feuer stems from Proto-Indoeuropean *pūr.
11. Sound/meaning isomorphism
Following Meillet, considering sound and meaning together is strongly recommended. Considering sound and intonation (in tonal languages), or meaning alone is entirely unreliable.
12. Non-linguistic evidence
It is a mostly accepted practice among historical linguists, not to accept non-linguistic evidence like folklore, mythology, technology and cultural traits, which must be avoided from being taken into consideration. Language is much more conservative than other cultural and technological exchange, which through diffusion may occur in significantly shorter periods of time. Therefore apparent affiliations based on domains like the aforementioned may be deceptive.
Here too there are exceptions, and those are known and documented historical facts, archaeological evidence (actually, archaeology and historical linguistics work effectively together), and human genetics: forensic genetics (graves) and palaeogenetics/human genetics (mitochondrial DNA).
13. Erroneous morphological analysis
For words morphologically decomposed to morphemic segments for etymological analysis, root, enclitics and affixes, there must exist (or be produced) evidence in the grammar, in order to ensure that morphological segmentation is correct. Incorrect or unmotivated segmentation discredits etymological analysis. Another frequent problem are undetected morpheme divisions, which leads to erroneous interpretation and may lead to resemblances which on proper segmentation fade away.
A completely invalid procedure is comparison of non-cognate forms within one language family with forms from another.
Forms of limited scope
It is recommended to include a certain minimum number of related languages into a comparative analysis in order to be able to achieve evidence status.
"When an initial 'proto language' is to be reconstructed, the number of witnesses which a word has should be taken into account. An agreement of two languages, if it is not total, risks being fortuitous. But, if the agreement extends to three, four or five very distinct languages, chance becomes less probable." (Meillet 1925: 38, quoted from Rankin’s 1992: 331 translation, by Campbell 2003: 279)
Neglect of known history
Neglecting known history — and thus omitting to verify plausibility — is so obviously a critically omission, that it doesn't need further discussion here.
15. Spurious forms
One must make sure the data is reliable (not erroneously transcribed, or interpreted in a grammar).
16. A single etymon as evidence for multiple cognates
Campbell states that "A common error in proposals of DGR [distant genetic relationship, SB] is that of presenting a single form as evidence for more than one proposed cognate set. A single form/etymon in one language cannot simultaneously be cognate with multiple forms in another language".