2 ms·

Explaining 25% of the variance seems pretty respectable to me... certainly consistent with a "human languages may be shaped by X" headline. No one's claiming th

Explaining 25% of the variance seems pretty respectable to me... certainly consistent with a "human languages may be shaped by X" headline. No one's claiming this is the _only_ thing shaping human languages. Also, 10% of languages (about 600 languages) is basically as good as it gets for linguistic typology studies. [1] argues that around 500 independent languages is what you need to make strong claims about linguistic universals. [1] http://tedlab.mit.edu/tedlab_website/researchpapers/Piantadosi_and_Gibson_2014_CogSci.pdf http://tedlab.mit.edu/tedlab_website/researchpapers/Piantado... (pdf)

panglott 11y ago
That's an interesting paper, but what they conclude is that is effectively impossible to get a sufficiently large sample to make strong claims for absolute linguistic universals. "It is important to emphasize one aspect of this analysis. The sampling procedure we use assumed independent samples from the true distribution. This means that what is really required is 500 independent languages, not 500 languages overall. For instance, Spanish and Italian do not count as two separate languages in this analysis since they are genetically related. This means that the real number of languages necessary may be much larger than 500 when sampling uses non-independent languages. Correlated samples will probably increase the number of samples needed to stay below a given false positive rate. Note, though, that the languages need not be independent in all respects: They need only be independent with respect to the relevant feature, which may be possible in some cases. To the best of our knowledge, it will in general not be possible to find 500 independent languages. There are, for instance, 212 language families in WALS, yet language families already are not independent samples. More aggressive independence methods—based on, for instance, geography (e.g., Dryer, 1989)—will likely arrive at much more independent samples, but orders of magnitude fewer of them. This means that it is very unlikely that statistical analysis will provide sufficient evidence to justify absolute universals." This is an interesting result, and confirms the poor history of reasoning about the so-called "linguistic universals". They seem to be taken for granted in generativist-influenced circles (rather than among linguistic typologists), but they frequently get debunked soon after they are hypothesized.

panglott 11y ago

That's an interesting paper, but what they conclude is that is effectively impossible to get a sufficiently large sample to make strong claims for absolute linguistic universals. "It is important to emphasize one aspect of this analysis. The sampling procedure we use assumed independent samples from the true distribution. This means that what is really required is 500 independent languages, not 500 languages overall. For instance, Spanish and Italian do not count as two separate languages in this analysis since they are genetically related. This means that the real number of languages necessary may be much larger than 500 when sampling uses non-independent languages. Correlated samples will probably increase the number of samples needed to stay below a given false positive rate. Note, though, that the languages need not be independent in all respects: They need only be independent with respect to the relevant feature, which may be possible in some cases. To the best of our knowledge, it will in general not be possible to find 500 independent languages. There are, for instance, 212 language families in WALS, yet language families already are not independent samples. More aggressive independence methods—based on, for instance, geography (e.g., Dryer, 1989)—will likely arrive at much more independent samples, but orders of magnitude fewer of them. This means that it is very unlikely that statistical analysis will provide sufficient evidence to justify absolute universals." This is an interesting result, and confirms the poor history of reasoning about the so-called "linguistic universals". They seem to be taken for granted in generativist-influenced circles (rather than among linguistic typologists), but they frequently get debunked soon after they are hypothesized.