86 TENNET X!

REFERENCES

Bartolomeo, P., & Chokron, S. (1999). Egocentric frame of reference, its role in spatial bias after right

hemisphere lesions. Neuropsychologia, 37, 881-894.

Bowers, D., & Heilman, K. M. (1980). Pseudoneglect: Effects of hemispace on a tactile line bisection

task. Neuropsychologia, 18, 491-498; Neuropsychologia, 25, 735-738.

Chokron, S., & Bartolomeo, P. (1997). Dissociation between left hemineglect and deviation of the ego-

centric reference. Neuropsychologia, 35, 1503-1508.

Chokron, S., & De Agostini, M. (1995). Reading habits and line bisection: A developmental approach.

Cognitive Brain Research, 3(l), 51-58.

Chokron, S., & Imbert, M. (1993). Influence of reading habits on line bisection. Cognitive Brain Re-

search, 1, 219-222.

Chokron, S., & Imbert, M. (1995). Variations of the egocentric reference among normal subjects and

a patient with unilateral neglect. Neuropsychologia, 33, 703-711.

Delatollas, G., De Agostini, M., Jallon, P., Poncet, M., Rey, M., & Lellouch, J. (1988). Mesure de la

prrf~rence manuelle par auto-questionnaire dans la population fran~aise adulte. Revue de Psycholo- gie Appliquie, 38, 117-136.

Heilman, K. M., Watson, R. T., & Valenstein, E. (1997). Neglect: Clinical and anatomic aspects. In

T. E. Feinberg & M. J. Farah (Eds.), Behavioural neurology and neuropsychology. New York: McGraw-Hill.

Jeannerod, M., & Biguer, B. (1987). The directional coding of reaching movements: A visuomotor

conception of visuospatial neglect. In M. Jeannerod (Ed.), Neurophysiological and neuropsychologi- cal aspects of spatial neglect. Amsterdam: Elsevier.

Jeannerod, M., & Biguer, B. (1989). R~f~rence ~gocentrique et espace repr~sentr. Revue Neurologique,

145, 365-369.

Karnath, H.-O. (1997). Neural encoding of space in egocentric coordinates? Evidence for and limits of

a hypothesis derived from patients with parietal lesions and neglect. In P. Thier & H.-O. Karnath

(Eds.), Parietal lobe contributions to orientation in 3D space. Heidelberg: Springer-Verlag.

Karnath, H.-O., Dick, H., & Konczak, J. (1997). Kinematics of goal-directed arm movements in neglect:

Control of hand in space. Neuropsychologia, 35(4), 435-444.

Mattingley, J. B., Bradshaw, J. L., & Bradshaw, J. A. (1994). Horizontal visual motion modulates focal

attention in left unilateral spatial neglect. Journal of Neurology, Neurosurgery and Psychiatry, 57, 1228-1235.

Schenkenberg, T., Bradford, D. C., & Ajax, E. T. (1980). Line bisection and unilateral visual neglect

in patients with neurologic impairment. Neurology, 30, 509-517.

Ventre, J., Flandrin, J. M., & Jeannerod, M. (1984). In search for the egocentric reference: A neuropsy-

chological hypothesis. Neuropsychologia, 22, 797-806.

Age vs AIzheimer's: A Computational Model of Changes in Representation

Patrick Conley and Curt Burgess

University of California, Riverside

and

Guila Glosser

Department of Neurology, University of Pennsylvania

Previous research has demonstrated that the language of older adults leads to denser repre-

sentations in a high dimensional model of memory than does the language of younger adults

(Conley & Burgess, in press), and thus that density in the model (HAL or the hyperspace

analogue to language) may constitute a useful metric in comparing memory for younger and

older adults. This paper extends the previous research by examining the role of density in

TENNET XI 87

semantic representations that emerged from the language generated by adults with Alzheimer's

and comparing the results with age-matched normal controls. We found that, just as older

adults have denser representations in semantic space than do younger adults, adults with Alz-

heimer's have still denser representations than normal older adults. These results support the

hypothesis that greater density, normally associated in the model with good semantic depth,

may in fact reach a "saturation point" and affect retrieval in older adults and especially adults

with Alzheimer's. 2001 Academic Press

The effects of aging and dementia on semantic memory have been the subject

of a large body of literature in recent years. It is well established that the normal

aging process causes some deterioration in the memory process. Age has been shown

to cause deficits in recall and recognition memory (Burke & Light, 1981). How-

ever, the nature and ultimate cause of these deficits is not clear, as demonstrated by

cases in which aging does not affect memory processes. Implicit memory, for in-

stance, has been shown in several studies to be resistant to aging (Light & Singh,

1987; Hashtroudi, Chrosniak, & Schwartz, 1991; Schacter, Cooper, & Valdiserri,

1992).

Similarly, the effect of Alzheimer's disease (AD) on semantic memory is difficult

to determine; clearly, deficits are present and detectable, but are these deficits due

to actual semantic degradation or are processing requirements overwhelming the

capabilities of the AD patients (Nebes, 1989)? This sets up the traditional "processing

vs representation" argument also prevalent in other domains of cognition. Most ex-

isting models of AD, as well as normal aging, are processing based. However, our

memory model, called HAL (hyperspace analogue to language; Burgess &Lund,

1997; Lund & Burgess, 1996) is a representational model. This allows us to examine

this problem from another direction, and in this paper we attempt to determine to

potential role of representation in memory deficits associated with AD and normal

aging.

The HAL Model

The HAL model of memory has been used to investigate semantic priming and

many other processes in memory and cognition (Burgess & Lund, 1999). The HAL

model of memory produces semantic representations for words from a 320 million

word corpus text, derived from the Usenet. HAL computes weighted cooccurrence

information by parsing a 10-word window over the corpus of text and stores this

cooccurrence information in a 70,0002 dimensional matrix. By combining row and

column information (preceding and following cooccurrence values), a vector for each

of the 70,000 most frequent words in the text corpus is generated. This vector thus

contains the coordinates for plotting the word's location in high-dimensional seman-

tic space. The contextual similarity between words is inversely related to their dis-

tance in this semantic space; closer words are more semantically related.

Previous research (Conley & Burgess, in press) has demonstrated that the language

produced by older adults leads to denser representations within the semantic space

generated by the HAL model than is the case for the language of younger adults.

Density, in these studies, is operationalized as the mean distance between any word

and the closest 10 neighbors to that word in semantic space. Normally, increasing

the size of the text corpus available to the model will result in an increase in density

in semantic space, but in the cases previously mentioned, text size was held constant

across the corpora being compared. These differences in density have been replicated

across both Usenet text and interview transcript data. Other factors--such as the

type/token word differences between the older and younger corpora, topic breadth,

88 TENNET XI

and possible oversensitivity of the density metric to any changes between different

text corpora--have been controlled for or experimentally examined. Therefore, we

believe that the HAL model is detecting actual representational differences between

the language produced by younger and older adults and that the variation in mean

density between matrices generated from each group is not merely an artifact of word

frequency differences or topic breadth.

The next question we address is whether differences in representation will be de-

tected between matrices when factors other than age are considered. Specifically, we

wanted to determine whether the HAL model, and specifically, the density metric,

would differentiate between adults with AD and normal older adult controls.

Method and Results

Two corpora were generated, both based on interview text provided by Glosser

(from Glosser & Deser, 1990, 1992). Each experimental group was asked the same

interview questions. The first text corpus (AD corpus) consisted of 19,000 words and

was generated from the interview texts of 23 patients (mean age 64 years) meeting

research diagnostic criteria for probable Alzheimer's disease (McKhann et al., 1984).

The second text corpus (normal corpus) also contained 19,000 words and was gener-

ated by 41 normal older adult controls (mean age 55 years). As the second text corpus

was truncated to match the length of the first, only 15 of the older adults' interview

texts were included in the second text corpus. The two corpora shared 877 words.

Each corpus was then used to generate a high-dimensional matrix. To obtain a

measure of density in semantic space for each matrix, the mean distance of the 10

closest neighbors to each word in the matrix was computed. Calculating a mean

distance for every word in the matrix ensures that any density changes detected are

systematic (i.e., not due to sampling error). The obtained mean distances, therefore,

are a measure of density in semantic space, and these distances were then compared

across the matrices.

The matrix generated from the AD corpus had denser representations (M = 437

RCUs or Riverside context units, a distance measure scaled to resemble human RTs)

than the normal corpus (M = 452 RCUs), t(876) = 3.657, p < .001. Thus, we see

that the language generated by the adults with AD is leading to denser representations

within the model than is the case with language produced by normal older adults,

even though the factors that normally influence density in the semantic space (amount

of text, topic breadth, number of unique words) have all been controlled for. This

finding extends previous research (Conley & Burgess, in press), in which we detected

a similar density difference between younger and older adults, with older adults'

language producing denser representations in the HAL model.

Discussion

In this study, we detected a density difference between the memory matrices gener-

ated from language spoken by adults with AD and language spoken by normal older

adults. This finding is similar to an earlier study in which the language of older adults

led to denser representations in high-dimensional space than the same amount of text

from younger adults. Thus, density in the HAL model appears to change as a function

of population, with younger adults producing the sparsest representations, older adults

generating denser representations, and adults with AD leading to yet denser represen-

tations in semantic space. We previously postulated density (Burgess & Conley,

1998) as being a source of the difficulty commonly experienced in the retrieval of

proper names. The results of the current study support a cautious extension of this

TENNET XI 89

hypothesis: perhaps "density saturation" is one source of the memory difficulties

commonly found in aging and exacerbated still further in Alzheimer's disease.

As an example, disproportionate difficulty with proper names is frequently reported

by older adults (Cohen & Burke, 1993). In the HAL model, names occupy their own

area of semantic space, and name neighborhoods have been found to be denser than

the neighborhoods of frequency-matched common nouns. Perhaps names offer dis-

proportionate difficulty because they are already denser than common nouns and thus

more susceptible to this saturation effect. Errors of retrieval in common nouns would

usually result in a word that is semantically related to the desired word, which the

retriever could then use to recover to the desired word. While this process would

result in the desired word, it would also require more time and processing power

than just accessing the correct word during the original retrieval attempt. Thus, den-

sity, which usually provides the model with better representations and greater seman-

tic depth, could, in time, cause the system to decrease in efficiency if the density

was increased beyond a level of diminishing returns or if the individual's ability to

retrieve information efficiently was compromised in some fashion, as might be the

case in normal aging and especially in Alzheimer's.

HAL is more of a learning and representational model than a cognitive processing

model, yet when differences in mental processing are operationalized as the language

produced by groups with different memory characteristics, we hypothesize that the

model is "borrowing" the processing of information provided by each of the groups

studied--when they are simply trying to answer the interview questions in an appro-

priate manner--and that it is this difference in quality of processing between different

groups of adults that is leading to the variation in representations between matrices.

Most theories of AD and normal aging necessarily focus on processing issues (see

Nebes, 1989, for a review), yet we argue that the memory differences brought on by

both normal aging and abnormal dementia might be reflected in representation as

well as process. These results demonstrate how representational issues might affect

processing: density in HAL would normally suggest greater semantic depth, and thus

a richer representation; words cluster closely in the semantic space, signifying that

the model has a very good semantic "understanding" of words located in that space.

However, if the processing (or retrieval) has been adversely affected by other factors,

such as AD or normal age-related deficits, then the very factors that lead to richer

representation may lead to increased difficulties in processing, as the large number

of close semantic competitors will require processing resources that might not be

available to resolve the confusion and retrieve the desired word.

REFERENCES

Burgess, C., & Conley, P. (1998). Developing semantic representations for proper names. Proceedings

o/" the cognitive science society (pp. 185-190). Hillsdale, N J: Erlbaum.

Burgess, C., &Lund, K. (1997). Modeling parsing constraints with high-dimensional context space.

Language and Cognitive Processes, 12, 177-210.

Burgess, C., & Lund, K. (1999). The dynamics of meaning in memory. In Dietrich & Markman (Eds.),

Cognitive dynamics: Conceptual change in humans and machines.

Burke, D. M., & Light, L. L. (1981). Memory and aging: The role of retrieval processes. Psychological

Bulletin, 90~ 513-546.

Cohen, G., & Burke, D. M. (1993). Memory for proper names: A review. Memoo', 1,249-263.

Conley, P., & Burgess, C. (in press). Age effects in a computational model of memory. Brain and

Cognition.

Glosser, G., & Deser, T. (1990). Patterns of discourse production among neurological patients with fluent

language disorders. Brain and Language, 40~ 67-88.

90 TENNET XI

Glosser, G., & Deser, T. (1992). A comparison of changes in macrolinguistic and microlinguistic aspects

of discourse production in normal aging. Journal of Gerontology, 47, 266-272.

Hashtroudi, S., Chrosniak, L. D., & Schwartz, B. L. (1991). Effects of aging on priming and skill learning.

Psychology & Aging, 6, 605-615.

Light, L. L., & Singh, A. (1987). Implicit and explicit memory in young and older adults. Journal of

Experimental Psychology: Learning, Memory, & Cognition, 13, 531-541.

Lund, K., & Burgess, C. (1996). Producing high-dimensional semantic spaces from lexical co-occur-

rences. Behavior Research Methods, Instruments and Computers, 28, 203-208.

McKhann, G., Drachman, D., Folstein, M., Katzman, R., Price, D., & Stadlen, E. M. (1984). Clinical

diagnosis of Alzheimer's disease. Neurology, 34, 939-944.

Nebes, R. D. (1989). Semantic memory in Alzheimer's disease. Psychological Bulletin, 106, 377-394.

Schacter, D. L., Cooper, L. A., & Valdiserri, M. (1992). Implicit and explicit memory for novel visual

objects in older and younger adults. Psychology & Aging, 7, 299-308.

Visuospatial Working Memory in Turner's Syndrome

Cesare Cornoldi, Filippo Marconi, and Tomaso Vecchi

Dipartimento di Psicologia Generale, Universitgt di Padova, Italy

Turner's syndrome is a genetic disorder, specific to women, in which one of the X chromo-

somes is partially or completely deleted. This syndrome is associated with physical features

such as short stature or failure in primary and secondary sexual development, together with

a specific pattern of cognitive functions. It has been suggested that women affected by Turner's

syndrome perform poorly in tasks measuring visuospatial abilities and have a verbal IQ sig-

nificantly higher than performance IQ. Although this result has received strong empirical

support, the nature of the visuospatial deficit is still unclear. Recent studies on visuospatial

processes have highlighted that the underlying cognitive structure is more complex than previ-

ously suggested and several dissociations have been reported (e.g., visual vs spatial, sequential

vs simultaneous, or passive vs active processes). In the present study we analyze in detail the

characteristics of the visuospatial deficit associated with Turner's syndrome by presenting four

young women with a comprehensive battery of tasks designed to tap all aspects of visuospatial

working memory. Results confirm that Turner's syndrome is associated with a general visuo-

spatial working memory deficit, but the pattern of performance of different cases can be differ-

ent, with a greater emphasis on active visuospatial processes, and on either sequential or simul-

taneous spatial processes. 2001 Academic Press

Introduction

Turner's syndrome (TS) is the most common sex-chromosome disorder in females.

It consists of an abnormality of one of the X chromosomes, which is completely or

partially deleted (see Saenger, 1996). Approximately 50% of all cases of TS are of

the X0 type, consisting in the deletion of the second X chromosome (Magenis, Breg,

Clark, Hook, Palmer, Pasztor, Summitt, & Vandyke, 1980).

This syndrome was first identified by Turner (1938) and its main characteristics

are short stature, gonadal dysgenesis or agenesis, and congenital malformations such

as webbed neck or deformity of the elbows. A review of neurological studies (Rovet,

1995) indicates variable dysfunctions involving different regions of both hemi-

spheres, but more frequently concerning the right hemisphere.

From a psychological point of view, this syndrome presents a typical cognitive

profile: normal IQ and verbal abilities are associated with a weakness in visuospatial

tasks. More specifically a discrepancy between verbal and performance IQ is often

reported (see Rovet, 1990). From a neuropsychological point of view, most studies

have found a specific visuospatial deficit in TS individuals, concerning tasks such