Monday, July 7, 2008

Summer Institute for Cognitive Neuroscience: Trade-off between neuroscience and linguistics

I just returned from giving a talk in the language session at the Summer Institute for Cog Neuro, which was held in Lake Tahoe. Beautiful venue, great weather (until the wind shifted and smoke from a fire in the area blew into the valley), and an interesting session.

The program was chaired by Alfonso Caramazza and included talks by Peter Hagoort, Kevin Shapiro (Caramazza co-author), Franck Ramus, David Caplan, Laurent Cohen, and me. I'm not going to go into detail about any of the talks - you can read all about them when Gazzaniga's The Cognitive Neurosciences Volume IV hits the bookshelves. Instead, I want to comment the one thing that made a serious impression on me as I listened to the talks and fielded questions on my own talk, namely, that there remains a serious gap between psycholinguistics and biology.

I have to confess up front that I didn't hear all of the talks: I had to miss Caplan's and Cohen's talk, which were given on the day after mine. Based on the talks I did hear, the contrast was striking between those of us who approach the cognitive neuroscience of language, with an emphasis on neuroscience versus an emphasis on (psycho)linguistics. For example, Hagoort talked about semantic unification, spending a good deal of time dissecting the ERP patterns associated with a range of linguistic constructions; he had little to say about the functional anatomy underlying these unification processes other than reference to some brain regions that might be generating these effects. Shapiro and Caramazza's talk was mostly concerned with the patterns of association and dissociation in morphological processes that one finds in brain damage patterns; again not much in the way of detailed functional anatomic networks. Franck Ramus talked about the genetics of language, providing details on the foci of genetic abnormalities associated with various language-related conditions, but provided little psycholinguistic detail using terms like "morphosyntactic deficits" to describe the language phenomena. Likewise, my own talk highlighted the role of specific brain structures such as the posterior half of the STS (bilaterally), posterior planum temporale (Spt) in various language processes, but I had little to say about the exact psycholinguistic processes involved and, like Franck, resorted to terms like "some aspect of phonological processing" and "sensory-motor integration."

Neither approach is right or wrong, or more or less important. They just reflect a different emphasis.

In some sense, listening to a talk from someone with a different emphasis can be unsatisfying. For example, Peter Hagoort asked me how Levelt's notion of syllabification fits into our Dual Stream model. My response: I don't know. Caramazza and Caplan both pressed me on what kind of representations live in the STS. My response: Maybe a phonological lexicon of some sort. Good questions that I hope to be able to answer some day, but to date, not very satisfying responses. Although I didn't, I just as easily could have asked Peter, or example, how the notion of semantic unification fits with the connectivity pattern between the anterior temporal lobe and BA45 (or whatever).

In the end, it is often hard to find points of connection between these different approaches to brain and language research, and I think the tendency is either to ignore findings/hypotheses from the research program at the other end of the spectrum, or to reject it. I certainly feel this tendency in myself (semantic unification or dissociations in morphological abilities don't help me understand the response properties of Spt, so why pay attention), and I can see how people from a psycholing model approach might tend to ignore or reject what we're doing (it doesn't clarify the nature of syllabification or morphological affixation, so why pay attention).

But this is the wrong mindset, of course. We are all interested in understanding the relation between neural systems and language models. We need to pay attention to what's going on at both ends of the spectrum and actively seek connections. We also have to understand that neuroanatomical models are not intended (at this stage) to be psycholinguistic models, and are necessarily (at this stage) more coarse in terms of processing stages/representations. Despite the fact that both the Dual Stream model and Levelt's model of speech production have boxes and arrows, they are not aiming to characterize the same set of facts. Hopefully the two types of models will converge and constrain each other in the end, but there need not be a one-to-one relation between boxes and Brodmann areas. At this stage in our field, it helps to approach a given paper, talk, or grant proposal with an "emphasis adjusted" mindset.


Anonymous said...

I think Indefrey & Levelt's (2004) meta-analysis of neuroimaging studies of speech production is very helpful in aligning the two approaches. Obviously, I'm of the opinion that a marriage of connectionist models and neuroimaging will save the day. As for teasing apart the representation of syllabic information, there's hope, as Matt Goldrick's work shows. So perhaps we need to focus on more talk, less action for the future (and yes, that is a dig at mirror neurons...).

Greg Hickok said...

While I do like the I & L 2004 paper (it was in our special issue after all!), and while it is a step in a useful direction, on its own, it is darn near vacuous.

Ouch. That's a strong statement. Let me explain.

The paper by itself basically assumes a psycholinguistic theory and then looks for the neural correlates of that theory. Said neural correlates were identified. What have we learned? Some geography, and that's about it. Maybe this "where" information is useful for neurosurgeons (in practice, probably not), but if we stop there, it doesn't get us any closer to telling us *how* the brain processes language. For example, suppose the organization of language systems in the brain turned out to be different in every individual. In such a case, "where" is useless information, but there still must be a "how" that is worth understanding. So "how" is the endgame, not "where."

This is not intended as an attack on I&L or anyone else who is looking for "where." I spend a good deal of time doing neural cartography myself. The hope, in this effort, though, is that "where" will provide clues to "how." Assuming there is a relation between structure and function, it is worthwhile mapping "where" as a step toward our ultimate goal. It important, though, that we keep the goal in focus, and don't get hung up making brain maps of our favorite cognitive model.

As for Goldrick's work... maybe, Greig, you could guest edit a summary of it for Talking Brains????

Anonymous said...

I use I&L for the basic representational architecture that underlies most 2 step models of LP. Levelt's is a serial or feedforward processing model, whereas other interactive activation (IA) models (by Dell or Harley etc) include feedback between lexical-semantic and word form levels. As I hope we've demonstrated in some of our studies, neuroimaging supports the IA approach, at least the basic requirement of cascaded processing involved in these models. So, we hopefully provide some "how" information relative to the "where" in I&L.

As for a precis of Matt's work, I'll see what I can come up with once I've finished with our grant rejoinders! This blog is too tempting to read when I need time out from responding to comments like "neuroimaging has no national benefit".

Jerry said...

I have been observing the same gap for my whole "academic existence" (~5-6 years) always hoping that with more insight and knowledge the gap would become smaller and smaller...
But I think I was wrong.
The Poeppel & Embick paper 3 years ago gave some new hope, but the pronounced cleft between linguists and neuroscientists at two meetings that specifically aimed at bridging the cleft (one on the Origins of Language and another on a Neural Theory of Language) made me lose faith. Now that you guys start casting doubts on the unifiability and the potential "cross-fertilization" of the two major disciplines which have been conducting extensive research on the very same subject (namely the probably best-examined cognitive faculty) for the past 50 years, how can I still believe in it?
I'm seriously looking forward to the conclusion in the upcoming Poeppel & Embick paper that DP mentioned earlier. I hope it'll be pre-released on talkingbrains...

Greg Hickok said...

No need to lose faith, Jerry. It *is* possible, and there are examples of steps in the right direction (see Greig's comments above, and comments on the "Top 10" post. We just have to keep reminding ourselves what the ultimate goal is, and work hard to nudge closer to it with every study. Sometimes this just means doing better cartography (you have to map the geography before you can understand its geological basis) or doing better psycholinguistics -- but always with an eye towards connections.