Towards a computational(ist) neurobiology of language: correlational, integrated and explanatory neurolinguistics

Abstract

We outline what an integrated approach to language research that connects experimental, theoretical and neurobiological (NB) domains of inquiry would look like and ask to what extent unification is possible across the domains. At the centre of the programme is the idea that computational/representational (CR) theories of language must be used to investigate its NB foundations. We consider different ways in which CR and NB might be connected. These are (1) a correlational way, in which NB computation is correlated with the CR theory; (2) an integrated way, in which NB data provide crucial evidence for choosing among CR theories; and (3) an explanatory way, in which properties of NB explain why a CR theory is the way it is. We examine various questions concerning the prospects for explanatory connections in particular, including to what extent it makes sense to say that NB could be specialised for particular computations.

Keywords:

• computation
• linguistic theory
• unification
• explanation
• reductionism
• levels

Notes

1. A comprehensive treatment would also take into account the substantial body of work from neuropsychology, which has a much longer history.

2. For general perspectives that are congenial to the one advanced here, see Carandini (2012) and Mausfeld (2012), as well as Henson (2005) and Page (2006). Marantz (2005) also argues for a similar perspective for neurolinguistics.

3. The idea that language can be approached in these terms is stressed in some recent work under the heading of Biolinguistics (see e.g. Chomsky (2005)). While we are sympathetic to many of the (mostly programmatic) suggestions in Chomsky's work, in practice much of the work that falls under that particular heading differs markedly in focus from the programme that we advance here.

4. On this theme, a difficult question is whether or not the CR theory should limit itself to hypotheses that are ‘neurobiologically plausible’, e.g. as was argued on occasion for connectionist architectures. We do not believe that it is advisable to do so at present, given our limited understanding of how any information is represented and computed in the brain. See Gallistel and King (2009) for extensive debate as well as section 11.

5. For a related discussion of the role of imaging data and its relation to psychological theories see Coltheart (2006).

6. For an important discussion of part of the history of this dynamic, see Phillips (1996) and Marantz (2005) on the derivational theory of complexity and its relation to the development of psycholinguistics and theoretical linguistics.

7. See Embick and Marantz (2005), Stockall and Marantz (2006), and Marantz (2013) for some specific proposals concerning how theories of morphology connect with psycho- and neurolinguistic data.

8. Part of what is at issue with the cognitive vs. brain data concerns what Chomsky (2000) calls methodological dualism.

9. Anatomically informed work is of this type can be productive. Our point is that, particularly in the domain of language, the area-based reasoning (and its concomitant assumptions about linguistic representation that implicate the GMP above) is prominent.

10. A further argument is that some approaches (e.g. ‘connectionism’) are restricted to mechanisms that are not computationally strong enough to compute what is required by the CR theory. See Gallistel and King for an argument of this type from the domain of memory.

Additional information

Funding

David Embick's work is supported by [grant number NIH R01HD073258], David Poeppel's work is supported by [grant number NIH 2R01DC05660].