How do computational linguists abstractly represent a language?

Question

When building models of the evolution of languages or similar phenomena where many different languages are involved and change over time, how do computational linguists abstractly model a language?

It seems like the canonical computer science approach of just considering a subset S of all strings on 0 and 1 will ignore too much of the structure that linguists care about. Thus, if building a computer simulation, or trying to do mathematical/theoretical linguistics, is there a canonical way for researchers to represent languages?

The primary interest is in an abstract approach that is easy to work with computationally or mathematically but that captures some (or all) of the features of a language one might typically care about.

Answer 1

Speaking from personal experience:

We started with a model of what differentiates languages, in this case a set of 13 parameters which could either be on or off. Considering each possible combination of these parameters, we ended up with 3072 languages (less than the expected 2^13=8192 because many combinations were impossible). For each language, we generated a list of all possible sentences which could be constructed in the language without sentence-lengtheners like conjunctions and relative pronouns. Each sentence looked something like Noun Adverb Verb Adjective Noun.

NB. These models were used to test theories of language acquisition, not language change, but the methodology could be used there as well. Also, this computer model assumes that the linguistic model (i.e. parameter-setting) is accurate; that part unfortunately can't be tested.

Edit: The research I was directly involved in was never published. However, the same model was originally used by Janet Fodor and William Sakas of CUNY. A quick search finds this paper, which describes the "language domain" in more detail.

Answer 2

Sometimes in machine translation, a model called an "interlingua" is used to represent grammatical or semantic structures in a language agnostic way. The term does not specify the structure of the model in any way, but is a good search term for this topic. Wikipedia has an article on it. Also, here is a recent example from Google's research blog with some details on their model.

They use something called an 'interlingua'. This paper provides an overview.