Beyond universals and particulars in language

In his target article “General linguistics must be based on universals (or nonconventional aspects of language)”, Martin Haspelmath invites us to reconsider the distinction between general linguistics and the linguistics of a particular language. The distinction is clearly useful to demarcate foci of interest and job descriptions. Also, it helps clarify what we expect our theories and models to achieve in terms of data fit: In general linguistics, we expect theories and models to capture something about the general nature of language. Now, since we can test the relevant hypotheses only in a relatively small sample of currently known languages, we want to avoid what statisticians call overfitting our particular sample. This means that we don’t expect models to perfectly fit every single language in this one sample; instead, we expect models to have good predictive performance in any sample we choose, allowing for random historical fluctuations and a healthy modicum of uncertainty. From a language-particular study, by contrast, we expect theories and models to capture one specific language, from its most general regularities to its most specific exceptions. We can’t overfit at that level. Where we do want to avoid overfitting is at the level of the idiosyncrasies that characterize individual speakers/signers or individual utterances (or, say, individual glyph tokens when we want to generalize not over an entire language but only over its inscriptional record). Beyond clarifying these differences, I don’t think the distinction between general and particular linguistics takes us much further. In important respects the two enterprises are heavily intertwined: Language-particular findings can generate hypotheses about the general nature of language (e.g. the discovery of free affix order in one language led to a learning-based hypothesis why most avoid it: Mansfield et al. 2020), and general linguistics obviously depends on samples of language-particular analyses. Haspelmath proposes that the distinction take us further nevertheless. The proposal rests on linking the distinction to two demands: general linguistics (1) “must study universals”, and (2) needs “comparative concepts” that are independent of the categories we need for describing particular languages. I am not convinced that these two demands are helpful