Transcript Phonology - University of Florida

Connectionist Models of
Language
Stephen E. Nadeau, M.D.
Malcom Randall DVA Medical Center
Department of Neurology, University of
Florida College of Medicine, Gainesville, FL
Other References
• Nadeau SE. Connectionist models and language. In
Nadeau SE, Gonzalez Rothi LJ, Crosson B, eds.
Aphasia and Language: Theory to Practice. New
York, Guilford Publications, 2000:299-347.
• Nadeau SE. Phonology: A review and proposals from
a connectionist perspective. Brain Lang 2001;79:511579.
• Nadeau SE, Gonzalez Rothi LJ. Rehabilitation of
language disorders. In Ponsford J (Ed). Cognitive
and behavioral rehabilitation: From neurobiology to
clinical practice. New York: Guilford, 2004:129-174.
• H.L. Roth, S.E. Nadeau, and K.M. Heilman, Naming
concepts: evidence of two routes, Neurocase
2006;12:61-70.
Plaut DC. Relearning after damage in
connectionist networks: toward a theory of
rehabilitation. Brain & Lang 1996;52:25-82
McClelland JL, Rumelhart DE. A distributed model of human learning and memory. In:
McClelland JL, Rumelhart DE, PDP Research Group, eds. Parallel Distributed Processing.
Cambridge: MIT Press, 1986:170-215. vol 2).
Rumelhart DE, Smolensky P, McClelland JL, Hinton GE. Schemata and sequential thought processes in PDP
models. In: McClelland JL, Rumelhart DE, the PDP Research Group, eds. Parallel distributed processing.
Cambridge, Massachusetts: MIT Press, 1986:7-57. vol 2).
Rumelhart DE, Smolensky P, McClelland JL, Hinton GE. Schemata and sequential thought processes in PDP
models. In: McClelland JL, Rumelhart DE, the PDP Research Group, eds. Parallel distributed processing.
Cambridge, Massachusetts: MIT Press, 1986:7-57. vol 2).
Rumelhart DE, Smolensky P, McClelland JL, Hinton GE. Schemata and sequential thought processes in PDP
models. In: McClelland JL, Rumelhart DE, the PDP Research Group, eds. Parallel distributed processing.
Cambridge, Massachusetts: MIT Press, 1986:7-57. vol 2).
Rumelhart DE, McClelland JL. On learning the past tenses of English verbs. In: McClelland JL, Rumelhart DE,
the PDP Research Group, eds. Parallel distributed processing. Cambridge, Massachusetts: MIT Press,
1986:216-271. vol 2).
Plaut DC, McClelland JL, Seidenberg MS, Patterson K. Understanding normal and impaired word reading:
computational principles in quasi-regular domains. Psychol Rev 1996;103:56-115.
Seidenberg MS, McClelland JL. A
distributed, developmental model of
word recognition and naming. Psych
Rev 1989;96:523-568
Nadeau SE. Phonology: A review and proposals from a connectionist perspective. Brain Lang 2001;79:511-579.
I’ve been retired since 1972 with
/cardimiapesun/ (cardiomyopathy)
Ten per cent [of] the people[with] the
/catraps/ (cataracts) has the [problem with
the] retina.
I was in the /bizzet/ (business) of
records…/fotegraph/ (phonograph)
records…for the/shusta (distribution?)…In
other words, I was a /eksiev/ (executive).
Look, I think it’s /porten/ (important).
I can’t [say]/tivelsha/, /diveltsher/
(television), uh TV.
Nadeau SE. Phonology: A review and proposals from a connectionist perspective. Brain Lang 2001;79:511-579.
Properties of PDP Networks
• The knowledge is in the connections
• A network supports processing, working and longterm memory
• Hidden units + nonlinearity support associations
between orthogonal representations
• Distributed representations
– Content-addressable memory
– Graceful degradation
– Inference, generalization, confabulation
• Bottom-up and top-down processing
• Learning of implicit rules, structure of knowledge
domain through experience
Roth HL, Nadeau SE,
Hollingsworth AL, CiminoKnight AM, Heilman KM.
Naming Concepts:
Evidence of Two Routes.
Neurocase 2006;12:61-70.
Grammar
• Syntax (knowledge of word order and
sentence structure)
– Sentence organization (verb argument
structure, embedded clauses, passive voice
constructions)
– Phrase structure rules
• Grammatic morphology (grammatic
modifications of words; words of
primarily grammatic function)
Genesis of Grammar
• Chomsky: a grammar generator
• A connectionist conceptualization:
– Sentence organization reflects the ways in which
the brain manipulates distributed concept
representations
– Phrase structure rules & grammatic morphology are
emergent products of pattern associator networks
responsible for articulatory and inscriptional output
that incorporate sequence knowledge
– Several aspects of grammar are dependent on
working memory systems (sustained selective
engagement)
Neural Network Representation
of Word Classes
• Concrete nouns
– Distributed concept representations
• Adjectives, intransitive verbs, auxiliar verbs
– Modification of distributed concept representations
• Main transitive verbs, locative prepositions
– Reciprocal modification and linkage of two or more
distributed concept representations
• Abstract words
– Juxtaposition of multiple distributed concept representations;
limbic distributed representations
Concrete Nouns
Adjectives, Intransitive Verbs:
Modification of Distributed Concept
Representations
• The obese, pockmarked, oily-haired,
slovenly, unctuous, check-shirted, plaidjacketed man with the striped pants,
food-stained paisley tie, goatee, wirerimmed glasses, bundle of pens in his
pocket, and clip board in his hand
Main Transitive Verbs, Locative
Prepositions
Reciprocal modification and linkage of
distributed concept representations
• Verb argument structure
– The old manAgent shot the burglarGoal.
– The manAgent gave flowersTheme to MaryGoal.
– HeAgent knew (Mary would arrive soon)Theme.
• Locative prepositions
– The book is on the table.
Abstract Words
• Juxtaposition of distributed concept
representations; limbic distributed
representations
– “Intellectual”
Broca’s Aphasia
Cinderella uh … scrubbing and uh … hard worker and wants to go a
ball. Step fa … mother uh I want to go. Can’t do it. Not … no well why
not? I don’t know. Because uh I uh uh … scrubbing uh uh watchacallit uh
uh working … object and so clean out bad.
Cinderella uh seems like animals love her. Because uh dress …
Horse help her. And stepmother uh uh … ruin dress. I don’t know why.
Probably because cute. Mad because uh uh … stepmother really ugly.
Dress broken and now can’t do it because what dress? Mother Teresa …
not exactly … uh uh magic godmother! That’s it. Godmother dress. Don’t
worry. I can fix it. And … beautiful.
Now carriage where? I don’t know how do you go here? Because
castle big. Probably uh mountain castle. How do you get uh here? Oh
don’t worry. I can uh … pumpkin and uh … servants and horse and
beautiful carriage and so magic. But, better midnight be here because uh
uh pumpkin carriage gone. Midnight be here ….
Thompson (2001)
Broca’s Aphasia
Broca’s Aphasia
• Agrammatism
• Impaired sentence organization
– Simplification of syntax
– Greater difficulty accessing nouns than verbs
– Greater impairment in use of verbs with complex
argument structure
– Limited lexical priming
– Resistance to concatenating adjectives
• Deficits in selective engagement of
distributed concepts representations and
manipulation of those representations
Phrase Structure Rules
• Word sequence knowledge — an
extrapolation of phonemic sequence
knowledge
Grammatic Morphology
• Free grammatical morphemes: articles,
auxiliary verbs, conjunctions, some
prepositions
• Bound grammatic morphemes: e.g.,
affixes specifying case, number or tense
Grammatic Morphology
• Articles
– Sequence knowledge underlying phrase
structure rules
– Lexical semantic knowledge: definite vs
indefinite (but also case, number, gender)
– Working memory of current linguistic
context
• Article production in Broca’s aphasias
Conduction Aphasia
I’ve been retired since 1972 with
/cardimiapesun/(cardiomyopathy).
Ten percent [of] the people [with] the
/catraps/(cataracts) has the [problem with the] retina.
I was in the /bizzet/(business) of
records…/fotegraphy/(phonograph) records…for
the/shusta/(distribution?)…In other words, I was a
/eksiev/(executive).
Look, I think it’s /porten/(important).
I can’t [say]/tivelsha/, /diveltsher/(television), uh TV.
Conduction
Aphasia
Wernicke’s Aphasia
Tell me what happened that brought you
here.
I been going all right caut so far. They got hold
of my lackid last night. Yard in the back. I don’t
know what all kapt it. Old himbone and boy I
hadn’t seen crossways…on time and best size.
You asked when they gave me tham garden
sizm made me so sore…just bugarden you
know. I didn’t go a slooten until way in the nighttime afterwards. Course there was a long lot of,
somebody’s bunyangin or something.
Wernicke’s
Aphasia
Transcortical
Sensory
Aphasia
Other Aphasias
• Anomic
• Transcortical motor
• Mixed transcortical
Summary:
Domains of Neural Network Knowledge
Underlying Language Function
• Lexical semantic function
–
–
–
–
Semantic
Sequence
Pattern associators linking
Semantic & sequence
knowledge (the basis for
the phonological lexicons)
• Grammar
– Elaboration of sequence knowledge (affixes, phrase structure rules)
– Lexical syntactic
• Selection of an articulatory sequence by configurations of reciprocally
related concept representations (transitive verbs, locative prepositions)
• Selection of an articulatory sequence by configurations of yet to be
spoken concept representations and working memory of already spoken
representations (articles, pronouns, many auxiliary verbs)
Summary of Aphasia Types
• Brocas
– Grammar, articulation
• Conduction
– Phonological sequence
• Wernicke’s
– Lexical semantic,
phonological sequence
• TCSA
– Semantic, lexical semantic
• Anomic
– Semantic, lexical semantic in
output
• TCMA