USING PET TOWARD A NATURALIZED MODEL OF
HUMAN LANGUAGE PROCESSING
Robert S. Stufflebeam
Abstract. My purpose for this paper is to show that
there are 'facts' about brains revealed by Positron
Emission Tomography [PET] that preclude serial models
of human language processing [HLP] from being
implementable. For this end, I first explain how PET
works. I also identify some of its limitations. I then
apply the data from some recent language-related PET
studies to evaluate the realizability of the
Wernicke-Geschwind model [W-G model]. I focus on the
W-G model because it is the most pervasive neurological
model of HLP; it's also a serial model. I show that the
PET data is inconsistent with what ought to be the case
if the W-G model were accurate: (1) language is not
processed exclusively in a serial, feed-forward manner,
but also along parallel pathways; (2) visual language
stimuli are not transformed into an auditory code, and
neither are visual stimuli [always] fed-forward via
Wernicke's area; and (3) Wernicke's area cannot be the
only area where meanings are stored. Additional
research is required. Nevertheless, I show that HLP
depends on a much richer set of functional areas and
interconnections than the W-G model [or any serial
Few cognitive scientists refer to the brain or its organization in
their models of human language processing [HLP], this 'decade of
the brain' notwithstanding. A commitment to naturalism makes such
references essential: the model ought to be implementable, and its
realizability depends on whether it is consistent with a host of
facts, including those about brains. But given the disparate
emphasizes of the various disciplines engaged in modeling HLP, it
is unlikely that any single discipline will be informed about all
the relevant facts. Hence, an interdisciplinary effort is
required, both to constrain one's portion of the overall model and
to avoid the pitfalls discovered outside one's domain. Toward the
interdisciplinary end of a naturalized -- and thus implementable
-- model of HLP, the neurofunctional imaging of Positron Emission
Tomography [PET] has already made significant contributions. I aim
to reveal how.
My purpose for this paper is to show that there are 'facts' about
brains revealed by PET that preclude serial models of HLP from
being implementable. But for all its potential, I also intend to
show that PET isn't a panacea. Toward these rather modest ends, after briefly
showing how PET works, I appeal to data from recent language-related PET studies to evaluate the realizability of the Wernicke-Geschwind model [W-G model]. I focus on the W-G model
because: (1) it is clinically useful and (2) it is the paradigmatic serial, neurological model of HLP. I show that the W-G model isn't accurate, for among other failings, it makes predications that are
inconsistent with the PET data.
2. What is PET?
The performance of any cognitive task supervenes on discrete
portions of the brain, functional areas dedicated [though not
necessarily exclusively] to specific types of information
processing. For example, visual perception would not be possible
without an antecedent increase in the neuronal activity of the
primary and secondary visual processing areas of the occipital
lobe. And as with any brain-mediated task, increases in neuronal
activity alters local blood flow [and metabolism].
Enter PET. The PET scanner is designed to detect the distribution of gamma rays within its oval aperture. Because two gamma rays are emitted whenever a positron collides with an electron, if a
positron-emitting isotope tracer is administered into a subject's blood, the scanner can detect the emission of paired gamma rays as blood flows through the brain. Thus, the functional image computed by PET directly measure changes in the local blood flow [or metabolism] of discrete functional areas of the brain, and indirectly measure changes in neuronal activity [see Figure 1].
To image brain blood flow during a language-related task, the
subject's head is stabilized and then inserted into the scanner
aperture. A radioactive isotope [such as 11C, 13N, 18F, or 15O] is
then administered. Most activation studies in normal subjects use
15O, which is incorporated into a saline solution and then
injected into the blood stream. Given its short half-life of 122.3
seconds, an additional dose of 15O-labeled water must be
administered before each scan. Each scan lasts approximately 40
seconds; a total of 5 to 10 scans are performed during a session.
The resolution of current PET-produced images is between 2-5 mm.
Because multiple scans may be made during any one
session, each subject for a given study will normally
perform a specific task for each of the scans. In so
doing, it is possible to identify changes in local blood
flow [and indirectly changes in neuronal activity] by
comparing images from two sets of task conditions, a
control task and an active task. That is, to isolate
changes in local blood flow, the image generated during
performance of a control task is subtracted from the
image generated during the performance of an active task.1 This subtracted image is called the difference image. To increase the signal-to-noise ratio, the difference images across subjects performing the same task can then be averaged, thus increasing the likelihood of identifying the functional significance of discrete brain regions.
Before presenting any actual PET studies, it is worth noting that
PET is not a panacea. Like any methodology, PET too has
limitations. For example, (1) the actual imaging site is once
removed from the site of positron formation, and blood flow
measurement is itself once removed from the neurons being
activated. As such, the resolution of PET images employing current
techniques and positron-emitting isotopes will [likely] never
approximate the anatomical imaging of magnetic resonance imaging
[MRI]. And (2) given a scanning time of even 40 seconds, much
useful information may not be imaged, due either to a transient
increase in neuronal activity, or a functional area being
maintained at or below threshold.2
No single discipline or methodology [including PET] will likely be
informed about all the relevant language-related facts. Hence,
given that no intertheoretically consistent model of HLP has yet
to be devised, an interdisciplinary approach is required. As such,
we can progress toward an intertheoretically consistent model of
HLP by evaluating the implementability of the most common
neurological model of HLP; viz., the W-G model.
3. The W-G Model
Much of our knowledge of the brain structures [or areas] involved
in HLP comes from the study of aphasia. The W-G model is an
abstraction from that data. As such, in addition to being the most
pervasive neurological model of HLP, the W-G model remains
diagnostically and conceptually useful (Kolb & Whishaw, 1990, p.
581; Mayeux & Kandel, 1991, p. 843).
According to the W-G model, initial visual and auditory language
stimuli are processed in their respective primary and secondary
cortical sensory areas of the left hemisphere: (l) Broadmann's
areas 17, 18, and 19 of the occipital lobe for vision; (2)
areas 41 and 42 of the temporal lobe for audition. The sensory code is then conveyed to (3) the angular gyrus, an area of association cortex specialized for the integration of visual, auditory, and tactile information into a phonetic or auditory code important for both
speech and writing. This integrated representation of the initial
stimulus is then fed forward to (4) Wernicke's area, where the
code is not merely registered as language [i.e., not mere noise],
but also associated with meaning. On this model, Wernicke's area
is the [only] site of the stored meanings of words. Should it be
damaged, the ability to comprehend language is lost, a disorder
called 'Wernicke's aphasia'. Once associated with meaning, the
modified neural code for the initial stimulus is relayed via
(5) the arcuate fasiculus to (6) Broca's area. The arcuate fasiculus
is a fiber tract connecting Wernicke's area in superior temporal
cortex with Broca's area in prefrontal cortex. A lesion here will
result in 'conduction aphasia' -- a language disorder
characterized by impaired repetition and naming, as well as
paraphasic errors. Broca's area, the site where the
memory of word articulation is stored, lies near the facial region
of (7) motor cortex. Hence, Broca's area sends the signals that
the motor cortex transforms into commands that cause the mouth and
tongue to form words. It is in Broca's area, therefore, where the
code [with attending meaning] is transformed from a sensory
representation to a grammatical motor representation that can be
used as spoken or written speech. If damaged, a language
production deficit called 'Broca's aphasia' will obtain. I
summarize this serial pathway in Figure 2.
If the W-G model accurately depicts how language processing
occurs, then each of the following predictions about HLP should be
consistent with the PET data:
- Because visual linguistic stimuli should be processed along the same pathway as auditory linguistic stimuli, reading should be processed in a serial, feed-forward manner;
- visual linguistic stimuli should be transformed into an auditory representation, which is then fed forward to Wernicke's area; and
- words not processed through Wernicke's area should have no meaning.
If the evidence weighs strongly against these
predictions, then the greater the unlikelihood that the
W-G model is an accurate model of HLP. Since time and
space considerations preclude a thorough analysis, I
assume that what follows is nevertheless sufficient.
4. Using PET Data to Evaluate the Predictions
Petersen et al. (1988, 1983) were among the first to use PET to
study lexical processing -- the processing of single words. Their
results do not bode well for serial models, especially the W-G
Subjects were asked to perform three tasks. First, common English
nouns were passively presented visually [for one set of scans] and auditorily [for another set of scans]. The sensory task was simply to process the word. In the second task,
an output task, subjects were asked to repeat aloud all the
presented nouns. For the third task, an association task, subjects
were asked to generate aloud a verb corresponding to the presented
noun; e.g., one could say 'bake' or 'eat' when presented with the noun 'cake'.
Difference images were obtained by subtracting the image generated
from the lower-level active task from the higher-level active
task; i.e., the initial control task was subtracted from the
sensory task; the sensory task was subtracted from the output
task; et cetera. They found that (1) visual and auditory stimuli
are processed in modality-specific processing areas during the
passive sensory task. (2) During the output task, there was
bilateral activation of the motor and sensory face areas, as well
as activations in the cerebellum. And (3) during the association
task, there was bilateral activation of the anterior cingulate,
left anterior inferior prefrontal cortex, and the right inferior
lateral cerebellum (Petersen & Fiez, 1993, 515-517).
These results [having been duplicated and expanded upon
in subsequent PET studies] falsify each of the
predictions. Let me explain.
If visual stimuli are fed-forward to be rendered into a phonetic or auditory code, then there should have been some overlap in visual and auditory processing reflected in the difference images. There was not [see Figure 3]. Instead, there wasn't any significant temporal lobe activation during any level of lexical processing upon visually presented words
(Petersen & Fiez, 1993, 522). This finding is significant. First, given that Wernicke's area was altogether bypassed during even the verb generation [association] task, contrary to the W-G model, some other area(s) must be involved in word comprehension. The surprising finding was that the frontal lobe [rather than the posterior ones] seems to play a key role in the associative aspects of lexical processing.
Second, because visually presented words can be processed without having been modified in Wernicke's area, visual linguistic stimuli need not be processed along the same pathway as auditory linguistic stimuli.
And third, even low-level linguistic processing need not be fed-forward in a serial manner.
Thus, HLP is not exclusively a serial process. As such, despite its utility, the W-G model isn't consistent with the 'facts', including: (1) the cerebellum and some subcortical
structures -- viz., the left thalamus and the left caudate nucleus
-- have also been implicated in language processing. And (2) PET
studies have revealed that because visual linguistic stimuli are
not transformed into an auditory representation, visual and
auditory linguistic stimuli are processed independently by
modality-specific pathways that have independent access to Broca's
area. Moreover, (3) because the linguistic processing of visual
stimuli can bypass Wernicke's area altogether, other brain regions
must be involved with storing the meaning of words (Mayeux &
Kandel, 1991,p. 845; also see Kolb & Whishaw, 1990, pp. 582-583).
Thus, not only do there seem to be separate -- parallel --
pathways for processing the phonological and semantic aspects of
language, language processing clearly involves a larger number of
areas and a more complex set of interconnections than just those
identified by the W-G model (Mayeux & Kandel, 1991, p. 845).
Indeed, the PET studies support the notion that language
production and comprehension involve processing along multiple
routes, not just one:
No one area of the brain is devoted to a very complex function, such as 'syntax' or 'semantics'. Rather, any task or function utilizes a set of brain areas that form an interconnected, parallel, and distributed hierarchy. Each area within the hierarchy makes a specific contribution to the performance of the task. (Fiez & Petersen, 1993, 287)
[Hence, modular models ought to go by the boards, for
there is no one area of the brain devoted to reading, or
any other language-related cognitive ability.] Simply
put, it is the parallel, distributed feature of HLP that
precludes any serial, feed-forward, hierarchical model
from being implementable.
My purpose for this paper was to show that there are 'facts' about
brains revealed by PET that preclude serial models of HLP from
being implementable. For this end, I first focused on the nature
of PET and its limitations. I then applied the data from recent
language-related PET studies to evaluate the realizability of the
W-G Model. I have shown how the PET data is inconsistent with what
ought to be the case if the W-G model [or if serial models in
general] were an accurate account of HLP. Among the failings of
such models include: (1) language is not processed exclusively in
a serial, feed-forward manner; rather, it is processed along
parallel pathways; (2) visual language stimuli are not [always]
transformed into an auditory code; (3) visual linguistic stimuli
are not [always] fed-forward via Wernicke's area; and (4)
Wernicke's area cannot be the only area where meanings are stored.
Further research is clearly required -- to not only increase the pool of tested subjects over which to generalize, but to corroborate and improve upon the PET findings. It is just as clear, however, that HLP depends on a much richer set of functional areas and interconnections than the W-G model identities. As such, although clinically useful, the W-G model isn't implementable. The task now ought to be evaluating the implementability of the host of parallel, multiple-route models (see Pollatsek & Rayner, 1989). Agreed. But that is another paper.
1. The labels 'control task' and 'active task' are relative, for
the image generated from an active task during one scan may be the
control task in another. For a thorough account of how PET works,
see Posner & Raichle (1994).
2. Some limitations' of PET are more apparent than
actual. For instance, PET can't answer all the
questions. But neither can any other methodology or
discipline. Indeed, this 'limitation' serves to promote
the need for an interdisciplinary approach to modeling
HLP, an appeal PET researchers themselves have made on
several occasions (see Petersen & Fiez, 1993). And while
it's true that no two brains are exactly alike, neural
variability across individuals is small, and the
averaging techniques are themselves very effective
(Petersen & Fiez, 1993, 514). So, to demand exacting
specificity is to set one's standard unattainably high.
Fiez, J. A., & Petersen, S. E. (1993). PET as part of an
interdisciplinary approach to under-standing processes involved in
reading. Psychological Science, 4, (5): 287-293.
Kolb, B. & Whishaw, I. Q. (1990). Fundamentals of human
neuropsychology. (3rd ed.). New York, NY: W. H. Freeman and
Martin, J. H., Brust, J. C. M. & Hilal, S. (1991). Imaging the
living brain. In Principles of neural science. (3rd ed.). E. R.
Kandel, J. H. Schwartz & T. M. Jessell (Eds.). New York, NY:
Elsevier Science Publishing Company, pp. 309-325.
Mayeux, R. & Kandel, E. R. (1991). Disorders of language: The
aphasias. In Principles of neural science. (3rd ed.). E. R.
Kandel, J. H. Schwartz & T. M. Jessell (Eds.). New York, NY:
Elsevier Science Publishing Company, pp. 839-851.
Mohr, J. P., Pessin, M. S., Finkelstein, S., Funkenstein, H. H.,
Duncan, G. W., et al. (1978) Broca aphasia: Pathologic and
clinical. Neurology, 28: 311-324.
Petersen, S. E., Fox, P. T., Posner, M. I., Mintun, M. & Raichle, M.
E. (1988). Positron emission tomographic studies of the cortical
anatomy of single-word processing. Nature, 331: 585-589.
Petersen, S. E., Fox, P. T., Snyder, A. Z. & Raichle, M. E. (1990).
Activation of extrastriate and frontal cortical areas by visual words
and word-like stimuli. Science, 249: 1041-1044.
Petersen, S. E. & Fiez, J. A. (1993). The processing of single words
studied with positron emission tomography. Annual Review of
Neuroscience, 16: 509-530.
Pollatsek, A. & Rayner, K. (1989). Reading. In Foundations of
cognitive science. M. I. Posner (Ed.). Cambridge, MA: MIT Press, pp.
Posner, M. I. and Raichle, M. E. (1994). Images of mind. New York,
NY: Scientific American Library.
Raichle, M. E., Fiez, J. A., Videen, T. O., MacLeod, A. K., Pardo, J.
V., Fox, P. T. & Petersen, S. E. (1994). Practice-related changes in
human brain functional anatomy during nonmotor learning. Cerebral
Cortex, 4, (1): 8-26.
Stufflebeam, R. S. (1995). The brain, language, and PET: Revealing
constraints on a naturalized model of language processing. 10th
International Congress of Logic, Methodology, and Philosophy of
Science, Florence, Italy (Aug. 19-25). Volume of abstracts: 556.
Stufflebeam, R. S. & Bechtel, W. (forthcoming). PET: Exploring the
myth and the method. Philosophy of Science.