Elliot Murphy

It is a common, and very reasonable question to ask what the predictions of the Minimalist Program would be concerning what would happen if the core language system (in the brain) was disabled (see here). Theoretical linguists have not, on the whole, done the best job at clarifying how we might go about answering these kinds of questions, and below are some brief notes sketching out how this might be achieved.

Consider this new pre-print arguing that the brain’s core language system exports domain-specific information to various cognitive systems, with their own unique cortical topography. To me, this looks like a nice neural model for the Minimalist Program’s central claim that the core language system (compositional syntax-semantics) provides instructions to distinct cognitive systems, and that these systems have their own unique ‘interface conditions’ for readouts. The core language network establishes hierarchical constituency structure, proprietary representations, and grammatical relations between elements, but the full scope of what inferences we can make with linguistic instructions does not end there. According to Casto and colleagues, when we read sentences about people’s views on others, we activate the theory of mind network. When we read sentences about complex emotions, the emotional centers are involved. When we read sentences about balls rolling down hills, our intuitive physics centers are involved. And so on. This yields some important experimental predictions for real-time language processing, with some exciting possibilities for exploration. Relatedly, it is one thing to draw up cartographic, localizationist maps of different networks that the language system might interface with, but what are the experimental predictions for how this is achieved?

I would argue that the way the language network sends information to other systems is via PAC and travelling wave dynamics, as outlined in the experimental predictions discussed in this recent paper on the ROSE model for language. Many usage-based and construction grammar-aligned researchers regularly ask, rather sensibly, “what are your experimental predictions?” to generative grammar-aligned researchers, and ROSE offers one possible means to address these important challenges to linguists.

The type of PAC in question would depend on the endogenous rhythmic properties of these language-external networks (e.g., it wouldn’t all be through alpha-gamma coupling, but would presumably differ based on network). ROSE outlines which neural signatures are relevant to distinct scales of linguistic complexity. So if the language network sends basic Representations to one network it would be via gamma coupling (e.g., measured via gamma-derived partial directed coherence, or similar measures in intracranial EEG), but if it sends more complex Structures (categorized multi-unit linguistic objects) it would also involve the delta-theta range. Depending on the type of information that the language network exports to these varying non-language networks (polysemous lexical inferences on single morphemes, phrasal idiomatic inferences, or truth-evaluable complex nested relative clauses, etc.), the dominant mode and frequency band would be modulated.

What does all of this have to do with the Minimalist Program in linguistics? This area of theoretical linguistics doesn’t actually have strong experimental predictions for how this ‘exporting’ of linguistic information happens, since it’s an atemporal computational account of what “knowledge of language” constitutes. For example, there’s nothing in Marcolli, Berwick and Chomsky’s formalization of Merge as a free non-associative commutative magma operation that necessitates it be instantiated in one brain region over another, or in one frequency band over another!

This answer might be somewhat unsatisfying, but that’s also why models like ROSE exist, to explore how the brain might neurally enforce those algebraic properties. For example, if you ask “What are the predictions of the Minimalist Program for such and such an issue in neural dynamics or neuropsychology”, the answer would depend on how you cash out computational-level models at the algorithmic or implementational Marrian levels. Meaning, the way I personally do it in ROSE is different from how other generative grammar-compliant researchers have done it, such as Friederici and Matchin and Moro, amongst many others.

In this lecture, at around the 28-30 minute mark, I discuss some of the issues about misconceptions of the neuropsychology literature, and its supposed consequences for theories of linguistic knowledge, and address the question of how language contributes to cognition under one possible read of the generative grammar architecture.

On a slightly tangential note, the thesis from Casto and colleagues that deep understanding of language requires exporting information to other areas and networks is already a core proposal in ROSE, and discussed in my 2020 book, and I’m pretty sure it’s also assumed in the Matchin & Hickok model too (a very influential cortical model for ‘core’ language computations that is not discussed by Casto and colleagues).

Back to theoretical linguistics: The Minimalist Program’s general architecture about how narrow syntax interfaces with general conceptual knowledge can be very informative for theory-formation at the psychological and neural levels, but the point to be made here is that you still have to do the relevant homework – this doesn’t happen for free! There is no direct or necessary psycholinguistic predictions from a computational/level account of grammatical knowledge. It depends on how you decide to implement this theory in novel domains of psychology and neurobiology. As I mentioned already, the way I choose to do this is not the same as how Moro, Matchin, Friederici, Trettenbrein, Berwick and others do it, even though we are all loosely generative-aligned researchers.

So, my question in return here would be: surely there is also an onus on those who explore language in the brain to offer a clear formal description of the target behavior? What are the non-negotiable design features and formal properties and algebraic structures that any theory of language must have?

Why is this important? Consider how Marcolli and Berwick recently used a specific formalized model of linguistic knowledge via Hopf algebra and category theory to mathematically guide them towards certain neural mechanisms over others. They discuss ROSE as a candidate model for neurolinguistics, and show that they can use their mathematical model of syntactic knowledge to connect with the mathematics underlying phase synchronization and phase-amplitude coupling, thereby narrowing the candidate search space for neurobiologically plausible causal-mechanistic structures for syntax. Thus, the formal properties of syntax (Merge) can be shown to be, in principle, directly compatible with known neural mechanisms postulated as critical for language in ROSE. That kind of multi-disciplinary proposal is exactly what is needed in the field — using higher-level theories to guide the selection of candidate mechanisms at the implementational level.

Other interesting possibilities emerge here. For example, how does Adger’s mereological model of syntax (which has its own explanatory benefits at the computational level in terms of accounting for certain grammatical phenomena) differ in terms of how it might help point towards particular implementation mechanisms over others (and if it doesn’t help with this project, does that necessarily mean it’s an inferior theory to Marcolli et al’s?).

In summary, then: If someone says “Here is my mathematical set-theoretic or category-theoretic model of how grammatical knowledge structures semantic objects and precedes the organization of morpho-phonological assemblies”, and their friend replies to them “So what are your experimental predictions in the right hemisphere specifically for the neural signatures of impaired orthographic parsing of passivization in common language deficits in Alzheimer’s disease patients?”, a very reasonable answer to this question would be “I have no idea — it depends on how you migrate the computational level account over to an algorithmic or implementational space.” ROSE offers one set of candidate mechanisms and regions, but my point here is that if we haven’t done the work to establish what Poeppel and Embick would call ‘linking hypotheses’ between domains of cognitive (neuro)science, then we cannot expect to hear a fruitful answer.

A certain older, now retired linguist used to argue that science requires specifying the object of inquiry before explaining it. This remains the case for cognitive neuroscientists of language.