Can computational techniques solve historical problems that couldn't otherwise be solved?

Question

Recently I've read that machine learning has been used to apply the Comparative Method (example with references here). Also, there are other mathematical approaches that have been applied to the problem of reconstructing proto-languages. For example, this is a very interesting work where the researchers employ a probabilistic model of sound change to reconstruct parts of the proto-austronesian language.

My question, aimed at professional linguistics working on these fields, is: How relevant do you think are the computational/mathematical techniques in the fields of historical linguistics/language reconstruction? Are there interesting unsolved problems that could be (potentially) solved with these methods?

I am asking because I've also read some critics say that these methods have not brought anything new, only already known stuff, and all I know about the subjetc comes from mathematicians, physicists and computer scientists, but I have not heard much about it from linguists.

Answer 1

You say "... some critics say that these methods have not brought anything new ..." From my recollection of some old results (well outside my areas of expertise), I would say the problem is rather that automatic methods have not brought anything old. To have confidence in such a method, linguists would need to compare the classifications it comes up with known results to decide whether the method is sound. Early results in this area were extremely disappointing. I'm sorry that I can't be very specific about some work that was done back in the 70s. I recall a persuasive paper by Eric Hamp, a renowned Indo-Europeanist, but now I can't find it on the Web.

I also recall an LSA paper claiming to have detected a close family relationship between English and Icelandic. That was amusing.

I had my own brief adventure in lexicostatistics trying to construct a genetic tree for the Je/Ge languages of NE Brasil, and in particular the two languages Cherente and Chavante (or Xerente/Xavante), whose peoples separated recently enough, around 1900, that the parent tongue was actually partially documented. So I could compare what a reconstruction method gave me starting from the modern pronunciations and reconstructing back, and the actual earlier pronunciations.

Glottochronology gave me a separation time for the two languages of over 600 years, which was wrong. It should have been less than 200. The reconstructions didn't look much like the earlier forms from the parent dialect. It was just nonsense. Maybe I did it wrong.

So, my suggestion is that you look first for any work done to validate the methods used in such studies as you referred us to.

Answer 2

One thing that computational technology might in principle be useful for is managing huge amounts of data. This could be useful if all of the sudden we discovered word lists for a thousand Austronesian languages and didn't know that there was an Austronesian phylum. But we've know about Austronesian for many years, and automated reconstruction is not an advance on already-existing manual reconstruction. Reconstruction is a problem that can be solved with just pencil and paper (and data).

Elementary lexicostatistics can in principle be done with pencil and paper, but there are modern methods that are sufficiently compute-intensive that they can't be done manually. As Greg Lee pointed out, though, there is a real question as to the validity of the method, as a technique for arriving at subgroupings. The two main problems with lexicon-based sub-grouping is that it does not distinguish shared retention from shared borrowing; and it fails to weight shared innovation above mere shared retention. These are not permanent problems of computational approaches, and in principle some good metric of relatedness could be developed using computational technology. But these results have to be calibrated using the gold standard for historical linguistics, shared grammatical innovations. If the two methods give different results, well, we already have an explanation for why vocabulary is historically very fluid.