While I do not dwell here on the different ways in which terms such as “comparison” and “relationship” have been used in historical linguistics, it is worth briefly considering both how we define language relationships and the consequences of these definitions for historical study. The comparison of languages to reconstruct their common ancestors - and to draw family trees - has typically been based on a notion of “normal” or “regular” language transmission (Thomason and Kaufman 1988). Such transmission is assumed to proceed from parents to children who are acquiring language in largely or wholly monolingual communities. Under this model, changes accrue when children adduce grammars with slightly different properties from their parents’ grammars (compare Hale 1998; Kroch 1989). The different patterns could be due to spontaneous innovation, reanalysis, or differences in the frequency of the relevant features in the speech to which the child is exposed. We realize, of course, that this is an overly simplified picture of both acquisition and change, and relies on an idealized picture of what a language is. Children’s peers are just as important an influence on acquisition as their caregivers are (Stanford 2006; Aitchison 2003), and adults too are capable of innovations. Thus the parent-to- child transmission model is at best an idealization of how linguistic features are passed on; more accurate is a population-based model where learners deduce the features of their language based on input from their whole community (for more on agent-based models of this type, see Croft 2000). However, given that learners, on balance, come to almost identical conclusions about the properties of their language, generational models are a useful way of conceptualizing the most frequent type of linguistic transmission. This allows us to compare child-learner-centered transmission with other situations, such as creolization and mixed language formation, where both the transmission facts and the linguistic outcomes differ (Ng 2015).
A further point of idealization comes from how we define a “language”. The input to language comparison is typically taken to be uniform. Either we are working with features which do not usually vary across speakers (such as basic vocabulary) or we abstract away from variation for the purposes of comparison by treating one speech variety as representative. Clearly, such assumptions will matter more in some areas than others. Internal linguistic diversity clearly matters in models of language transmission, as the learner’s input is never uniform. How learners abstract away from variation (and also acquire the patterns of variation) is crucial to understanding the role of acquisition in change. The transmission model gives us a working definition of language relationship. Two languages are related to one another if they show systematic similarities (that is, ‘correspondences’) across grammar and lexicon.
Linguistic comparison has been conducted for a much longer time than formalized comparative methods. However, there is also much work which compares languages without direct reference to either their evolutionary history or transmission processes. It is probably the case that whenever two people speaking different languages come into contact with one another, they notice similarities and differences between their languages. Some cultures have well-developed theories of folk-linguistic comparison which ascribe causes to similarities between two languages (Schebeck 2001; Niedzielski and Preston 2003). Such theories are also found in the Roman world, where we find frequent comparisons between Latin and Greek, as well as etymologizing within word families.
However, such comparisons were unsystematic and as such cannot be used for reconstruction. That is, while such theories suggest relationships between individual words, because the comparisons are unsystematic, they cannot be used to infer which changes happened at which times and to which words. This is the breakthrough of the late 18th and early 19th centuries; the discovery that similarities between related languages are systematic, while those between unrelated (more precisely, not demonstrably related) languages are ad hoc and unsystematic. All languages show resemblances, but only those which descend from a common ancestor have regular resemblances.