Obscure vs. transparent cognates in linguistic reconstruction