University of Massachusetts Occasional Papers in Linguistics
Volume 8 Occasional Working Papers in Cognitive Science Article 4
1982
Computational Neurolinguistics
Helen M. Gigley
Follow this and additional works at: https://scholarworks.umass.edu/umop
Part of the Linguistics Commons
Recommended Citation Gigley, Helen M. (1982) "Computational Neurolinguistics," University of Massachusetts Occasional Papers in Linguistics: Vol. 8 , Article 4. Available at: https://scholarworks.umass.edu/umop/vol8/iss2/4
This Article is brought to you for free and open access by ScholarWorks@UMass Amherst. It has been accepted for inclusion in University of Massachusetts Occasional Papers in Linguistics by an authorized editor of ScholarWorks@UMass Amherst. For more information, please contact [email protected]. Gigley: Computational Neurolinguistics 77. COMPUTATIONAL NEUROLINGUISTICS
Helen M. Gigley Computer and Information Sciences Department
Computational neurolinguistics is an interdisciplinary approach to studying natural language processing that uses methodologies from
artificial intelligence (AI), and neural modelling to define, design,
and implement a simulation model of language. The initial design,
implementation, and development of HOPE, a neurolinguistically constrained processing model of natural language comprehension is the
first such simulation model developed (cf. Baron, 1974a; 1974b; 1975; Cunningham and Gray, 1974; Lavorel, 1982) for discussion of other
neuro-based models). The model is based on evidence from normal and
aphasic language behavior as reported in the literature of psycholinguistics, neurolinguistics, and linguistics.
We will provide a brief overview of the behavioral constraints
incorporated in the design, the neurologically plausible control
strategies of the implementation, and describe a II first effort" in
validation of an aspect of the model in its current stage. (A more
The author wishes to thank Dr. Michael Arbib, Dr. Joseph Duffy, Mrs. Esmat Ezzat, Dr. Pierre Lavorel, Dr. Victor Lesser, and Dr. Barbara Partee for their help in making this research possible.
Published by ScholarWorks@UMass Amherst, 1982 1 n. University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4
detailed presentation can be found in Gigley, 1982). The need for such simulation models as aids in our investigation of brain function, and in language has been discussed elsewhere (cf. Arbib and Caplan, 1979;
Lavorel, 1980; Lavorel, 1982) and will not be included.
1. Behavioral Constraints and HOPE
The most critical behavioral constraint on HOPE's design was the
fact that the model had to simulate processing in what could be assumed to be a "normal" way,, and that it could be artificially "lesioned" and
continue to function without redesign or re-implementation. Further
behavioral constraints were imposed by observation of performance of
aphasic individuals and by clinical and neurolinguistic evidence as
documented in the literature. These constraints dictate the knowledge
structures included in the design, PHONETIC, morphological relationships, case-control information for verbs, GRAMMAR, word-meaning
representation, and a contextually determined memory representation of
the meaning of an utterance distinct from world knowledge, called a PRAGMATIC representation.
https://scholarworks.umass.edu/umop/vol8/iss2/4 2 Gigley: Computational Neurolinguistics 79.
Hypothesized "lesion" conditions included in the design are:
1. degradation of knowledge representations 2. inability to access knowledge representation(s),
3. short-term memory capacity problems, and
4. that mistiming could be a factor underlying observed aphasic deficits if timing-coordination was considered a critical factor of normal processing.
The first three hypotheses are suggested from neurolinguistic and
clinical studies (cf. Brookshire, 1978; Duffy, 1979; Geschwind, 1965;
Goodglass and Kaplan, 1972; Jenkins, Jiminez-Pabon, Shaw, and Sefer,
1975; Lesser, 1978). The fourth hypothesis is based on the fact that we are simulating a neural, time-dependent, coordinated process, and supported by several studies of the effect of presentation time of
auditory input on aphasic patients' processing (Brookshire, 1971;
Laskey, 1976; Liles and Brookshire, 1975; Katz, Blumstein, Goodglass,
and Shrier, 1981).
2. Neurologically Plausible Control
The control strategies employed in HOPE are drawn from current
knowledge of neural networks and psychological theories of memory. HOPE
uses feedback and feedforward interconnections of active information, in
conjunction with automatic-decay and threshold-firing propagation to
converge to a representation of the meaning of a sentence over time.
Computations which affect the interactive processing occur both serially
Published by ScholarWorks@UMass Amherst, 1982 3 80. University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4
and in parallel. Processing of information is time synchronous.
Information becomes active in HOPE when new words are "heard" or
introduced during the processing of a sentence. Each word has distributed aspects of its meaning that are activated over time using a
fixed-time spreading activation (Collins and Loftus, 1975; Quillian,
1980).
Once information is active, i.e. has an associated activity
vector, its activity value will affect the contextual state of the system depending on its interconnections to other information. All information that is active, is updated for a Compute-Time-Interval.
Control is maintained internally to the knowledge structures and is
effected through activity value propagations and interactions over time
(cf. Amari and Arbib, 1977; Anderson, Silverstein, Ritz, and Jones,
1977; Reiss, 1962). As activity values propagate, they are summed with
activity values of appropriately linked active information until they
reach or surpass threshold and fire. The effect of threshold-firing can
be either excitatory or inhibitory. Its effect is determined by the
state of the grammar and other active information in the system. The
amount of activity propagated after firing is modifiable for
experimental purposes.
There are two memory states for active information. All new
information enters the short-term state on its initial activation. Only
https://scholarworks.umass.edu/umop/vol8/iss2/4 4 Gigley: Computational Neurolinguistics 81.
information which fires can reach the post-refractory state. 'Ibis state
is hypothesized to be necessary in processing garden path sentences although not tested in the model.
3. HOPE as a Modelling Tool
As a modelling tool, HOPE permits experimental design of a "normal"
model including setting "normal" processing parameters. 'Ibe
experimenter can define his own lexicon (phonetically), the syntactic category-meanings associated with each lexical item (multiple meanings
are permitted), morphological relationships, case-control constraints,
the grammar (represented as a Categorial Grammar, cf. Ajdukiewicz,
1935; Bar-Hillel, 196lt; Bach, 1980a; 1980b; 1981; Lewis, 1972) and
interpretation functions (encoded in LISP) to activate the meaning of a
word within the PRAGMATIC representation.
Modifiable parameters included in the design are a basic set that
enable us to demonstrate the feasibility of the model's simulation
results. The parameter values must be determined for the
model-specified syntax to produce what is considered to be a " normal"
performance for the comprehension process on a cover set of sentences for the grammar. (A cover set of sentences includes all well-defined
syntactic combinations of the model-defined categories. In addition,
each sentence of the set must be clinically testable.)
Published by ScholarWorks@UMass Amherst, 1982 5 82. University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4
In the currently defined version of HOPE. the parameters permit modification of
1. the rate of word input relative to the fixed rate of activity propagation through memory.
2. the rate of decay in either of two different memory states. 3. the length of the decay interval (defined in terms of Compute-Time-Interval updates of information. and relative to the effect of fixed-propagation through memory. and the rate of new word introduction).
Other parameters which may be modified directly and which affect
propagation include:
1. the amount of activity propagated after thresholding (reaching or surpassing a parameter defined threshold value). and 2. the amount of inhibitory or excitatory propagation among competing meanings.
The parameters. once tuned for the "normal" state. can be modified
to represent clinically suggested processing hypotheses that result in
observed aphasic behavior. An example of a parameter modification that
can be interpreted as a short-term memory deficit is to increase the
rate of decay in the short-term state leaving all other parameters as tuned for the "normal" condition.
q_ Validation and Evolution of HOPE
We have conducted clinical studies to validate the present
experimentally defined model. One study analyzes whether aphasic
https://scholarworks.umass.edu/umop/vol8/iss2/4 6 Gigley: Computational Neurolinguistics • 83.
patients have multiple meanings of words spoken in isolation available
(Gigley, in preparation). It uses a picture selection task. Briefly, the task is to point to all possible meanings for the spoken word and to indicate when there are no more meanings available.
An assumption in HOPE is that all meanings of words are
simultaneously accessed (cf. Marslen-Wilson and Tyler, 1980;
Seidenberg and Tannenhaus, 1980a; Seidenberg, Tannenhaus, and Leiman,
1980b; and Swinney, to appear), and that during the on-line sentence
comprehension process, disambiguation of the contextually correct meaning is determined over time. We tested a weaker hypothesis: are multiple meanings available at all, i.e. in the off-line condition.
The use of the results with respect to the model is to claim that for
those patients having multiple-meaning lexical access within normal
range, the parameter-controlled, hypothesized lesions could be factors
underlying their clinically determined performance deficit during
comprehension.
A second study on the same set of patients uses a picture selection
task to determine the meaning of isolated spoken sentences. Tite task
uses simple declarative sentences which include some sentences whose
meaning is dependent on correct processing of closed class words, such
as determiners.
Evidence, such as that from the first study, is to be used to
determine "lesionability" hypotheses between the model and patient
Published by ScholarWorks@UMass Amherst, 1982 7 University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4 .. 84.
population. The second study is intended to determine the validity of the model's performance, or to aid in determining how to modify the model to make it a better approximation to aphasic performance. The
model and clinical studies need to be designed hand-in-hand to effectively continue the evolution of the model and hopefully, to
enhance our understanding of brain processing.
5. Lesion Experimentation on HOPE
Lesion experiments consist of two parts. First, the model must be
run in its intact "normal" state, tuning parameters to produce an
accepted "normal" pattern of response from the model on a cover set of sentences. Then, lesions are encoded by modifying parameters such as:
1. the rate of short-term memory decay (clinically interpreted as a short-term memory deficit), 2. the rate of new word introduction relative to propagation effect (interpreted as a slow-down of propagation), 3. the propagation constant (clinically interpreted as defective firing),
LI. the elimination of specific knowledge within a representation, a complete knowledge representation, or access to a representation (clinically interpreted as a specific knowledge deficit), or
5. the usefulness of a representation (such as eliminating the interpretation of determiners as opposed to their grammatical level of representation).
Evidence from the "lesion" experiments on HOPE to date has
demonstrated that current testing procedures have omitted evidence that
https://scholarworks.umass.edu/umop/vol8/iss2/4 8 Gigley: Computational Neurolinguistics
85.
could refute or support the model. Often, model demonstrated
interpretations of a sentence have not been included in the possible
patient response selections of existing neurolinguistic studies. The
reason thi~ has happened is that the studies, based on linguistic
theory, have not been able to consider language as a process. HOPE is a
process model, and because of the complexity of interaction during
processing, demonstrates possible interpretations of an utterance that
are not obvious from a theoretical analysis. Hence, the valid~tion of
the model requires future testing using studies designed to include a
sufficient level of detail to relate to the evidence from the model.
6. HOPE as a Process Model
HOPE is suggested as a new methodology for the study of language as
a process. It demonstrates the differences between linguistic theory,
linguistic performance, and processing of language. It is not claimed
that the model is the model of language processing in the brain, but
rather, that it is a first attempt to develop such a model including
evidence from lesion and stimulation studies as constraints. As such it
will require extensive testing for validation and for evolution to a
better approximation of a model which subserves language behavior. "How
to validate?", "How to obtain adequate data?", "How to evolve?", are
current questions being address~d as planning for future work continues.
Published by ScholarWorks@UMass Amherst, 1982 9 University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4 86.
7. HOPE and AI
HOPE demonstrates the application of a new control methodology
within AI. Toe control strategy requires a different problem decomposition than is found in current AI systems. Production system control and ATN control both require an explicit enumeration of each
precondition set or state. The cooperative-control strategy instead
emphasizes a generalized interaction which implicitly includes all precondition sets or states. It frees the modeller from defining the
enumeration and instead employs the interactive computations to enumerate the preconditions or states for the modeller. The potential
strength of this methodology will be demonstrated in planned expansions
of the research.
8. Summary
We have developed a new approach to studying natural language as a
process that integrates our current level of knowledge of neural
function with cognitive modelling techniques from AI and simultaneously
employs constraints reported in studies of neurolinguistics, psychology,
and linguistics. HOPE provides a research facility to analyze language
as an on-line interactive process by including time as a critical factor
in the process.
The contribution of the research to the study of aphasia is in
focusing the analysis of the degraded performance of the clinical
https://scholarworks.umass.edu/umop/vol8/iss2/4 10 Gigley: Computational Neurolinguistics 87.
population on modes of processing abilities in conjunction with
quantitatively characterizable performance abilities, rather than relying solely on the latter.
The contribution of the research to computer science will come from its relation to many aspects of the problems of man-machine interface,
as well as in applications of the control strategy to other problem
domains. HOPE demonstrates that one can adequately analyze human
processing behavior in a computationally definable way. This is an
important result for Intelligent Computer Aided Instruction (!CAI) studies. Furthermore, the development of evolutionary methodologies for the HOPE model are applicable to the human studies which underlie the
design of !CAI models, and to the design and development of natural
language interfaces to data bases and computer systems.
Published by ScholarWorks@UMass Amherst, 1982 11 88. University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4
References
Amari, S. and Arbib, M.A., "Competition and Cooperation in Neural Nets," Systems Neuroscience, J. Metzler (ed.), Academic Press, 1977, pp. 119-165. Anderson, J., Silverstein, J., Ritz, S., and Jones, R., "Distinctive Features, Categorical Perception, and Probability Learning: Some Applications of a Neural Model, 11 Psychological Review, Vol. 84, No. 5, 1977, pp. 413-451.
Arbib, M.A. and Caplan, D., "Neurolinguistics must be computational," Behavioral and Brain Sciences, 2, 1979, pp. 449-483.
Bach, E. , "In Defensive of Passive, 11 Linguistics and Philosophy, 3, 1980a, pp. 297-341. Bach, E., "Discontinuous constituents in generalized categorial grammar, 11 paper presented at NELS XI, 1980b, to appear in the Proceedings. Bach, E., Generalized Categorial Grammars and the English Auxiliary, University of Massachusetts/Amherst, Feb. 1981.
Bar-Hillel, Y., Language and Information, Addison-Wesley Publishing Company, Inc., Reading-:-liassachusetts, 1964.
Baron, R. J., "A Theory for the Neural Basis of Language. Part 1: A Neural Network Model, 11 International Journal of Man-Machine Studies, 6, 1974a, pp. 13-48.
Baron, R. J., 11 A Theory for the Neural Basis of Language. Part 2: Simulation Studies of the Model," International Journal of Man-Machine Studies, 6, 1974b, pp. 155-204.
Baron, R. J., 11 Brain architecture and mechanisms which underlie language: an information processing analysis, 11 in Proceedings of the New York Academy of Sciences Conference on the Origins of Language, 1975.
Brookshire, R. H., "Effects of Trial Time and Inter-Trial Interval on Naming by Aphasic Subjects," Journal of Communication Disorders, 3, 1971, pp. 289-301. Brookshire, R. H., An Introduction to Aphasia, B R K PUblishers, Minneapolis, Minnesot.a, 1978.
Cunningham, M. A., and Gray, H. J., "Design and Test of a Cognitive Model," International Journal of Man-Machine Studies, 6, 1974, pp. 49-104.
Duffy, J. R., "The Boston Diagnostic Examination," Evaluation of Appraisal Techniques in Speech and Language Pathology, F. L. Darley (ed.), Cambridge, Mass.:--"Iddison Wesley, 1979.
https://scholarworks.umass.edu/umop/vol8/iss2/4 12 Gigley: Computational Neurolinguistics 89.
Geschwind, N., "Disconnection Syndromes in Animals and Man, 11 Brain, 88, 1965, pp.237-294.
Gigley, H. , 11 A Computational Neurolingui stic Approach to Processing Models of Sentence Comprehension", COINS Technical Report 82-9, University of Massachusetts, Amherst, 1982. Gigley, H., "Processing Word Ambiguities: Availability of Multiple Meanings of Ambiguous Words in Aphasic Patients", COINS Technical Report, University of Massachusetts, Amherst, (in preparation). Goodglass, H. and Kaplan, E., The Assessment of Aphasia and Related Disorders, Lea and Febige~Philadelphia, 1972. Jenkins, J., Jimenez-Pabon, E., Shaw, R., and Sefer, J., Schuell's Aphasia.!!!_ Adults, Harper and Row, Publishers, Hagerstown, 1975.
Katz, B., Blumstein, S., Goodglass, H., and Shrier, R., "The Effects of Slowed Speech on Auditory Comprehension in Aphasia, 11 Paper presented at the 19th Annual Meeting of the Academy of Aphasia, London, Ontario, October, 1981.
Laskey, E. Z., Weidner, W. E., and Johnson, J. P., "Influence of linguistic complexity, rate of presentation, and interphrase pause time on auditory verbal comprehension of adult aphasic patients, 11 Brain and Language, 3, 1976, pp. 386-396. Lavorel, P. M., Aspects de la performance linguistique, These pour le Doctorat d'Etat, Universite Lyon II, Presses de l'Universite de Lille, 1980.
Lavorel, P. M., "Production Strategies: A Systems Approach to Wernicke's Aphasia," in Neural Models of Language Processes, Arbib, M.A., Caplan, D., and Marshall, J. C.~(eds.), Academic Press, 1982, pp. 135-164.
Lewis, D. , "General Semantics, 11 Semantics of Natural Language, Davidson and Harman (eds.), 1972, pp. 169-218.~
Lesser, R., Linguistic Investigations of Aphasia, Elsevier, North-Holland, 1978.
Liles, B. Z., and Brookshire, R. H., "The effects of pause time on auditory comprehension of aphasic subjects, 11 Journal of Communication Disorders, 8, 1975, pp. 221-235.
Marslen-Wilson, W., and Tyler ,L., "The temporal structure of spoken language understanding," Cognition, 8 (1980) pp. 1-71.
Quillian, M. R., 11 Semantic Memory, 11 in Semantic Information Processing," M. Minsky (ed.), MIT Press, 1980, pp. 227-270.
Published by ScholarWorks@UMass Amherst, 1982 13 90. University of Massachusetts Occasional Papers in Linguistics, Vol. 8 [1982], Art. 4
Reiss, R., 11 A Theory and Simulation of Rhythmic Behavior Due to Reciprocal Inhibition in Small Nerve Nets, 11 Proceedings AFIPS, Spring Joint Computer Conference, pp.171-194, 1962.
Seidenberg, M., and Tannenhaus, M., 11 Chronometric Studies of Lexical Ambiguity Resolution, 11 Proceedings of the Conference of the ACL, 1980a, pp.155-157. Seidenberg, M., Tannenhaus, M., and Leiman, J., "The Time Course of Lexical Ambiguity Resolution in Context," University of Illinois, Champaign, Illinois, 1980b.
Swinney, D.A., "Lexical Processing During Sentence Comprehension, Effects of Higher Order Constraints and Inplications for Representation," The Cognitive Representation of Speech, T. F. Myers, J. Laver-,--and J. Anderson (eds.), Amsterdam/New York: North-Holland Publishing Co. (to appear).
https://scholarworks.umass.edu/umop/vol8/iss2/4 14