Of Minds and Language (35 page)

Fig. 13.6. Tractograms for two brain regions: Broca's area (BA
_44/45
) and the frontal operculum (FOP) for 4 different subjects are displayed. Three-dimensional rendering of the distribution of the connectivity values of two start regions with all voxels in the brain volume. (
Left
) Tractograms from FOP: the individual activation maxima in FOP as a function of the Finite State Grammar (FSG) were taken as starting points for the tractography; from the FOP connections to the superior temporal gyrus (STG) via the fasciculus uncinatus were detected. (
Right
) Tractograms from BA
_44/45:
individual activation maxima in Broca's area as a function of the Phrase Structure Grammar (PSG) served as starting points for the tractography: from Broca's area connections to the posterior and middle portion of the superior temporal gyrus (STG) via the fasciculus longitudinalis superior were detected.

Source
: Adapted from Friederici et al. 2006a

The left part of the figure displays the fiber tracts in four subjects, with the fiber-tractography calculation starting from the frontal operculum which connects via the fasciculus uncinatus to the anterior portion of the superior temporal gyrus (STG). Interestingly enough, we usually do see the anterior STG active in the processing of local dependencies in studies on normal language processing. On the other hand, when starting the fiber-tractography calculation in Broca's area (right part of the figure), the connecting fibers go via the fasciculus longitudinalis superior to the posterior portion of the STG, and then along the entire STG.

With these data we now have evidence for a differentiation of the two areas in the inferior frontal gyrus, not only functionally but also structurally. Basically, we can describe two separate networks, one consisting of the frontal operculum and the anterior portion of the STG, and the other including Broca's area and the posterior portion of the STG extending to the entire STG. The first network, we hypothesize, is responsible for processing local phrase structure building, while the second network may be responsible for processing hierarchical structures.

What this means with respect to the evolutionary issue is the following. The human ability to process hierarchical structures could be based on the fully developed, phylogenetically younger cortex, that is Broca's area comprising BA 44/45, whereas the older cortex, that is the frontal operculum, may be sufficient to process local dependencies.

C
HOMSKY
: There were three languages. There was AB AB, A
_n
B
_n
, and then the third is the nested one, ABC CBA, with all the optional variations. Two questions. First, I didn't understand in the presentation whether you found a physical difference in the brain between the second type and the third type – the A
_n
B
_n
and the nested one. Was there any difference between those two?

F
RIEDERICI
: No, for both these types of artificial languages, that is the second and the third one, we saw Broca's area activated, and I think it would be hard to make a claim of more activation in the third grammar than the second grammar on the basis of the present data because here we are looking at different subject groups. I think the conclusion from this may be that even for the processing of the second language, the A
_n
B
_n
, you already use Broca's area, but you certainly need it for the third grammar. So the argument that you can process the second grammar only with a simple counting mechanism perhaps cannot be ruled out, but at least for the processing of the third grammar it can.

C
HOMSKY
: Yes, well, there is a possible experiment here. I mean, humans do have the third type, we're sure about that. We do not know if they have the middle type. So they may only have PSG and finite state options, but not counting mechanisms. That's one possibility. So therefore, when they're doing the counting system, they may be using the richer system, which doesn't require a phrase structure grammar. The other possibility is that they also have a counting system and that it's being obscured here. But if you looked at the famous starlings, that's what you'd find, because they do not have a PSG. So is there a way to test that?

F
RIEDERICI
: I think the data of the third grammar may be the most conclusive of all the experiments. With respect to the second grammar I can only for the moment argue only on the basis of the similarity between the brain activation for the two grammars, that at least our subjects are not using a counting mechanism, but are going for hierarchical structure processes.

C
HOMSKY
: But see that's possibly in fact plausible for a subject, a human, which has the third mechanism.

F
RIEDERICI
: Yes, you are right, the starling data (Gentner et al. 2006) of processing the A
_n
B
_n
grammar could be explained by a counting mechanism. But the prediction would be that starlings should not be able to learn the third grammar.

C
HOMSKY
: But you might expect that you're getting a masking effect in the humans where some might be using the counting mechanism and some might be using the richer mechanism, and get a muddled conclusion. But I'm just wondering if it's possible to tease it out? Have you done, for example, a pure counting study?

F
RIEDERICI
: No, we haven't done that.

C
HOMSKY
: That might be interesting to do, because then you could extract that out of the data for the two phrase structure types to see if they differ in that respect. The other question is just a kind of technical point. Finite state and local dependency are not the same thing. So you can have FSGs with arbitrarily long dependencies. I do not know if anybody has looked at this, but you can have a language which is AB
_n
A and CB
_n
C, and that's an FSG but it has indefinitely long dependencies.

F
RIEDERICI
: Yes, but from the data we have for the moment, I think we can only draw conclusions about the local dependencies. But you are right, maybe the same sort of network also deals with the non-local probabilistic dependencies.

C
HOMSKY
: Just take a guess. I mean, all this confusion about finite state grammar goes back fifty years, and the things that people call FSGs are almost always ones with local dependencies. But that's just a special case. So it's possible that they're not studying FSGs at all, they're just studying kind of associationist structures, which do have local dependencies. And yes, they are a subclass of FSGs, but they're not using its capacities.

F
RIEDERICI
: Yes, you are exactly right, so there are at least two more experiments, if not more, that we have to do.

C
HOMSKY
: Notice that these are the same two mistakes. It goes way back. Technically, A
_n
B
_n
is above an FSG, so in a particular hierarchy it's a context-free grammar, but it may not be using any of the capacities of a context-free grammar. Similarly, AB AB
is
a special case of an FSG, but it doesn't tell you that when you're studying it that you're studying FSGs, in fact you're studying a special case of local FSGs, which means maybe it's just local associationist nets. I mean, that hierarchy existed for a reason, but what people have been doing for fifty years is taking sub-cases of the hierarchy and studying them and thinking they're studying the hierarchy. But they're not, because the hierarchy has different properties. So the fundamental property of context-free is your third case, nesting, and the fundamental property of finite state I do not think anybody's studied, because it does include indefinitely long dependencies. So while that hierarchy sort of made mathematical sense and so on and so forth, the psychological experiments have not been investigated. They've been investigating sub-cases of it which have different properties. And it might be worth putting all this together and studying the real properties – which you did, in fact, in the third case there.

L
AKA
: In the original proposal about FLN there is the suggestion that the recursion mechanism could have originated from navigation, and, as you mentioned later on, music and math perhaps use these same mechanisms. My question is whether you have run experiments or whether you are aware of studies that have looked into navigation, music, or math that might show the circuits? Secondly, do you think there might be a connection, or do you have anything to say as to electrophysiological signatures and these two circuits?

F
RIEDERICI
: With respect to the first question, we have done experiments on music processing, and not surprisingly, it is the Broca's area that is active. However here I must say that it is very difficult to manipulate recursion without having memory involved. So I think we have to be very careful here. There are always memory issues involved because processing stretches over a certain time. Right now we are doing mathematics and I don't have data on that, but I think it is much more easily done, because with bracketing you can easily have embed-dings, and I am looking forward to those data. With respect to the electro-physiological signature, we find for the local dependencies – that is, within phrase dependencies – we do find very early negativity, which is maximal in the anterior portion of the left hemisphere. Dipole modeling of this effect using MEG shows us that we have two dipoles, one close to the frontal operculum and a second one in the anterior portion of the STG – so, exactly matching the first network I was proposing. The second network indeed involves Broca's
area.
⁴
The involvement of the posterior portion of the STG is a bit more complicated because in the posterior STG what we usually find is activation for semantic and syntactic integration. So this may be more an integration area of semantic and syntactic information.

R
IZZI
: If I remember correctly, there is this literature on the activation of Broca's area in pure memory tasks, in memory tasks that are allegedly independent from language, and the question is to see if they really are. Examples would be canonical tasks, such as card identification (one, two, three, etc.). So I guess one possible interpretation of your data could be that the processing of context-free dependencies really is whatever computational capacity is in the frontal operculum plus memory. But of course there is also the opposite interpretation, which is maybe more interesting, which is that for so-called pure memory tasks, we're really using grammatical knowledge which is crucially expressed in Broca's area, so that the effect observable in card-selection type tasks is derivative, in a sense, and uses some structure that is dedicated to language but then applied in a kind of instance to other types of more abstract tasks.

F
RIEDERICI
: Well, happily enough, these days we can be more specific than just talking about Broca's area. I mean, there is BA
₄₄
and BA
₄₅
. You're absolutely right, that for phonological memory issues, you get activation in Broca's area. This is the superior portion of BA
₄₄
. For our syntactic processes, we find the inferior portion of BA
₄₄
activated, and now the question is, can you really make a secondary argument of why there should be differentiation between the inferior and the superior portions? Given that the cytoarchitectonics of this area is the same, you may not have a good argument. However, recently we have information about the receptor architechtonics of the different areas and not surprisingly to me, but surprisingly to those who look at cytoarchitec-tonics only, we find a clear separation between the inferior and the superior portions. So what we certainly need to do is an experiment within subjects where we bury phonological memory aspects and syntax.
⁵

Cedric Boeckx, Janet Dean Fodor, Lila Glertman,
Luigi Rizzi

What I will be talking about is how I think generative grammar approaches syntactic universals, and I would like to start by saying that I think the topic of linguistic or syntactic universals is actually fairly odd. A legitimate reaction upon mention of this topic could be, what else? That is, basically what we are really interested in is explanation, and not so much in statements like
there is something or other
, but rather
for all X…, such and such happens
. That is, laws, or universals.

I think that it is useful to start with an article by a psychologist in the 1930s called Kurt Lewin, who was concerned with scientific explanations in particular and tried to distinguish between two ways of going about thinking about the laws in physics, biology, and other sciences (Lewin 1935). I think that his reflections carry over to cognitive science. In particular, Lewin distinguished between Aristotelian and Galilean explanations. Aristotelian laws or explanations have the following characteristics: they are
recurrent
, that is statistically significant; they specifically (though not always) target functions, that is they have a
functionalist
flavor to them; they also
allow for exceptions
, organized exceptions or not, but at least they allow for exceptions; and finally they have to do with
observables
of various kinds. Lewin contrasts these sorts of laws or universals with what he calls Galilean laws, which are very different in all respects from Aristotelian laws. In particular, they are typically
formal
in character, and they are very abstract mathematically. They
allow for no exceptions
and they are
hidden
. That is, if you fail to find overtly the manifestation of
a particular law that you happen to study, this does not mean that it is not universal. It just means that it is hidden and that we have to look at it more closely and we will eventually see that the law actually applies.