Greg Detre
Wednesday, 05 June, 2002
Neonates as young
as two hours old show categorical perception of phonetic distinctions used in
all human languages (e.g. Emias et al., 1971), and are hence referred to as
�universal listeners�.
Towards the end
of the first year of life, they lose the ability to discriminate phonetic
distinctions in non-native languages (Werker & Tees, 1984).
Werker & Tees
(1984) found that at around 10 months old, English speaking infants cease to
become universal listeners
Best et al.
(1988) found that 10-12 month old infants are able to discriminate the
non-native phonetic contrast of Zulu clicks, even though they have no
linguistic experience of them.
Pegg & Werker
(1994) found that infants lose the ability to discriminate a phonetic contrast
that commonly occurs in English, but is not used to distinguish meaning, at
around 10-12 months of age
Pegg & Werker
(1994) found that most adults were able to discriminate the allophone variants
that the 10-12 month olds failed on. This suggests at least a partial recovery
of phonetic discrimination after the �loss� at around 10-12 months.
Werker & Tees
(1984) found that adults were able to discriminate a truncated portion of a
Nthlakampx contrast when told that they were being tested on sounds from water
dripping into different buckets
Werker (1995)
argues that these results can be accounted for by a model which centres around
the needs of the developing linguistic abilities of the infant, the FRT. Her
argument is summarised in the table below:
Adults can be trained to discriminate non-native contrasts, but
their abilities rarely reach those of native speakers (Polka, 1992).
Investigating the
ways that such a network develops over time may shed light on the functional
reorganisation of the linguistic system in the same way that Nakisa &
Plunkett�s (1998) network shed light on the way that genetically specified
structures could permit the rapid learning of the complete set of phonetic
categories from just two minutes of input from a single language.
Proponents of the
symbolic account of human language argue that morphological inflections are
added to stems by a dual-route process in all languages (Pinker 1991).
Both symbolic and
connectionist models (e.g. Plunkett & Marchman, 1993) are able to account
for the acquisition and end state of English inflectional morphology, and for
responses to novel forms
McCarthy &
Prince (1990) state that the Arabic �sound� plural acts as a default, despite
the fact that less than 20% of nouns in the language take this inflection
Plunkett &
Nakisa (1997) constructed a connectionist network that was able to produce the
correct plural inflection for Arabic nouns (disproving Prasada & Pinker�s
(1993) assertion that such a model would fail).
Plunkett &
Naskia used cross-validation to show that the single route network
out-performed the dual-route model (with the optimum t value) in correctly inflecting novel forms (real Arabic nouns
that they were not trained on) (64% vs. 57% for broken plurals, 63% vs. 20% for
sound plurals).
This clearly
challenges McCarthy & Prince�s (1990) assertion concerning the default
status of the sound plural in Arabic as the dual-route network should perform
far better on sound plurals
Marcus et al.
(1995) argue that inflections German are the �exception that prove the rule� �
the minority default status of the -s plural demonstrates that a symbolic rule
is being applied
However, Naskia
et al. (in press) constructed a network that out-performed a dual-route model
in generalisation to novel forms, though by smaller margin than for Arabic
plurals
Clifton &
Ferreira (1987) contend that there is a grammar processing system composed of
modular sub-components; a lexical-processing module, one or more distinct
syntactic sub-processing modules and a mechanism for interfacing between
grammatical representations and �general knowledge� of discourse and the world
Fodor (1983)
argues a similar position, and presents a definition of modularity that is
shared by Clifton & Ferreira � a module is distinct information-processing
element that is tightly constrained, domain specific, informationally
encapsulated (i.e. there is no non-syntactic influence on syntactic decision
making), fast and dumb
This essay will
review the evidence that such a module exists for syntactic processing before
arguing that the work of Altmann (1986, 1998) poses a serious challenge to this
positions, as it suggests that contextual information can have a �top-down�
influence on the on-line processing of syntactic information.
Rayner et al.
(1983) found that, of the locally ambiguous sentences below, the second had a
significantly longer reading time. � The burglar blew open the safe with the
dynamite � The burglar blew open the safe with the diamonds.
Frazier (1979)
explains that the �Human Sentence Processing Mechanism� follows the �Principle
of Minimal Attachment� (PMA) when reading syntactically ambiguous sentences
Osterhout et al.
(1997) measured the �event related potentials� (ERPs) of subjects who were
exposed to various semantic and syntactic stimuli
Altmann (1986)
tested the generality of the results obtained by Rayner et al. (1983) by using
the same test sentences, but preceding them with different contextual sentences
However, Fodor
(1983) argues that the locus of such context effects lies outside of the module
itself, and instead acts on the results of the module�s analysis. Nevertheless,
the finding that reading times for non-minimally attached targets (2) were
faster than those for minimally attached targets (1), and that both were faster
than in a null context condition (a direct replication of Rayner et al.) is
incompatible with PMA, and indeed an contextually impenetrable syntactic
processing module.
Rather than
rejecting modularity completely, such an approach instead argues that there is
a greater degree of interaction between interaction between modules than Fodor
(1983) allows for (Crain and Steedman�s (1985) �weak interaction�).
Altmann (1998)
argues that these constraints are applied within the �interactive-activation�
model of language processing. Small segments of such an approach have been
implemented in connectionist models of language processing (e.g. Plaut et al.,
1996), but a fully implemented model of sentence processing has yet to be built
Piaget:
elaborated the role of sensori-motor skills in the development of a symbolic
capacity
2 sharp critics
of Whorf�s methodology and conclusions were Lenneberg and Feuer
Brown and
Lenneberg (1954) showed that differences in our ability to recognise and
remember colours were associated with availability of specific colour names
bring in: the
language tree may also be similar to the gene tree (Luigi Cavalli-Sforza)
Bloom�s
experiment intended to demonstrate that the the Chinese language�s lack of any
basic grammatical construction mirroring the subjunctive means that the Chinese
have great difficulty understanding counterfactual situations is another straw
man
Luria�s study on
the relative abilities of illiterate peasants, schoolchildren and students in
classifying and abstraction seemed to support this, showing that with education
came a marked increase in facility (indeed, spontaneous-seeming naturalness) at
the tasks.
Vygotsky took an
opposite perspective from Piaget, claiming that �inner speech� is a separate
form of internalised communication which develops after we learn to use
speech as a means of external communication to others
Bright (1984) �Animal language�
Payne (1995) recordings of humpbacks in Bermuda
Tyack (1986) used playbacks of the social calls, which
brought the singing whales to home straight in on the loudspeaker at 12km/hour
� considered that the song may be used as a spacing mechanism for courting
males
Norris (1991) studied the social structure and
behaviour of groups of Hawaiian spinner dolphins
Bastian (1964) conducted a famous experiment with two
dolphins, able to hear but not see each other
John Lilly probed dolphins� propensity to mimic by
getting them to follow a count up to ten and recognisably say simple English
words
Fisher & Ford (1983) sounds of killer whale pods are very
stable, comprised of about 12 distinct, stereotyped calls, that they use to
communicate when spread over a couple of miles
Hockett (1960, 1966) devised a set of 13 linguistic
universals:
semanticity and arbitrariness, productivity, interchangeability,
specialization, displacement, cultural transmission, duality of patterning; auditory-vocal
channel; broadcast transmission
and directional reception; discreteness;
rapid fading):
Gardner & Gardner (1989) Washoe
by the age of five, she had mastered 133 signs. She had been
spontaneously combining signs after learning only 8 or 10
Terrace (1979) Nim Chimpsky, less successful, rewarded
with approval rather than food, used symbols in the absence of their referents,
but not novel generation
Fouts et al. (1989) no sign language was used by humans in
Loulis� presence, yet he was still able to learn 51 signs just from the other
chimpanzees
Kanzi (Savage-Rumbaugh 1991)
Pepperberg (1991) Alex the grey parrot
our ethical
anthropocentricism may be regarded in the same light that slavery is now (Cavalieri & Singer 1993)
Marcus et al�s studies on over-regularisation
One influential
early connectionist model was a net trained by Rumelhart and McClelland (1986)
to predict the past tense of English verbs.
However, the
model is poor at generalising to some novel regular verbs, which Pinker &
Prince (1988) point to as a failing of connectionist models in general
Connectionist
modellers have defended themselves against the wider criticism that neural nets
are ill-suited to the generalising necessary to master cognitive tasks
involving rules (e.g. Niklasson and van Gelder, 1994).
Lenneberg (1967)
to postulate a critical window of language-learning, ending before the onset of
puberty, after which our ability to acquire language has markedly diminished
Eimas (1971)
observed signs of decline in our ability to recognise perceptual categories
with his voice onset time experiments on infants
�phonetic sixth
sense� (Pinker) feeds us
�fast-mapping�
(Carey) vocabulary spurt
Gall and Spurzheim
made premature steps towards parcelling the brain up geographically into
functionally distinct areas in the early 19th century under the spurious guise
of �phrenology�
Paul Broca first
argued for there being a language centre in the left hemisphere (1861)
Broca�s aphasia dialogue
(Gardner, 1975):
Karl Wernicke was
able to pinpoint a centre for language comprehension
(1874)
Oliver Sacks
mentions about patients in his casebook on the aphasic ward.
Hughlings Jackson
was the first to point to right hemisphere specialisation for sensory
processing, implying a language/visuospatial divide that we still retain to
some degree
Levy has since
argued for an analytic-holistic distinction between processing paths on the
left and right side
as Saussure
noticed, its symbols are arbitrary � there is nothing which connects the word,
�dog�, for instance, with dogs, other than our language community�s rote
memorisation of the term
Modularity. Do
children learn language using a "mental organ," some of whose
principles of organization are not shared with other cognitive systems such as
perception, motor control, and reasoning (Chomsky, 1975, 1991; Fodor, 1983)? Or
is language acquisition just another problem to be solved by general
intelligence, in this case, the problem of how to communicate with other humans
over the auditory channel (Putnam, 1971; Bates, 1989)?
Language and
Thought. Is language simply grafted on top of cognition as a way of sticking
communicable labels onto thoughts (Fodor, 1975; Piaget, 1926)? Or does learning
a language somehow mean learning to think in that language? A famous
hypothesis, outlined by Benjamin Whorf (1956), asserts that the categories and
relations that we use to understand the world come from our particular language,
so that speakers of different languages conceptualize the world in different
ways. Language acquisition, then, would be learning to think, not just learning
to talk.
The scientific
study of language acquisition began around the same time as the birth of
cognitive science, in the late 1950's. We can see now why that is not a
coincidence. The historical catalyst was Noam Chomsky's review of Skinner's
Verbal Behavior (Chomsky, 1959)
Chimpanzees
require massive regimented teaching sequences contrived by humans to acquire
quite rudimentary abilities, mostly limited to a small number of signs, strung
together in repetitive, quasi-random sequences, used with the intent of
requesting food or tickling (Terrace, Petitto, Sanders, & Bever, 1979;
Seidenberg & Petitto, 1979, 1987; Seidenberg, 1986; Wallman, 1992; Pinker,
1994a).
This allowed them
to benefit by sharing hard-won knowledge with their kin and exchanging it with
their neighbors (Pinker & Bloom, 1990).
Normal subjects Evetts
+ Humphreys (1981) mask-prime-target-mask
Interactive
Activation model � Rumelhart & McClelland (1981)
Word superiority
effect � Reicher 1969
Scarborough et al
(1977) � reading a word aloud
Oldfield &
Wingfield (1965) � name a picture
Bloomfield tried
to make linguistics more scientific. agreed with Watson that thought is just a
kind of concealed musculature
in his 1957 Syntactic Structures, he described three
increasingly powerful descriptions of language
Markman et al.
(1980s): the child is shown unseen/unnamed objects, and then has to pick them
out (usually) for a demanding puppet
Naigles found
that when two-year olds are shown a rabbit feeding a duck, and either hear: �the
rabbit is zorking the duck� or �the duck is zorking�
Brown: if
children are corrected for producing untruths, and not corrected for producing
bad grammar, how is it that the result of this is an adult adept at telling
untruths but whose sentences are perfectly grammatical
Chomsky and
Pinker: the only viable explanation is that we�re endowed innately with crude
knowledge about types of words and their role in language, as a basis on which
to determine which other types of words there are and where they can be found
relative to one another
in some ways,
Pinker�s account is not that different from Gleitman�s, about mapping onto the
structure of the world but Gleitman�s is simply about covariation in the world
(in this case with language), whereas Pinker�s is specifically linguistic
Morgan: proposed
that the input sequences that babies here have cues (clues) to their internal structure, e.g. prosody � infants as
young as 4� months are sensitive to the prosodic patterns that accompany some
of the boundaries between the major
constituents of a sentence
Morgan and colleagues created a small artificial language composed
of sequences of meaningless syllables ordered according to a set of rules they
made up which specified which orders of syllables were grammatical
although in
Samoan and Javanese apparently the speech to children is not markedly different
studied by
Bickerton: noticed that the children of pidgin speakers did not speak the same
pidgin as their parents
comprehension
vocabulary of between 60,000 and 75,000 words, during speaking selects from
about 30,000
Mehler et al.:
original motivation was to discover whether the syllable functions as a perceptual unit (= the chunk used to
access + organise the mental lexicon), using a syllable-monitoring task:
Marslen-Wilson
& Warren lexical decision task with phonemes spliced together: people were
slower at ruling out non-words whose starting phoneme could be a real word, i.e.
we do use the smallest detail
possible to distinguish between alternative words in the mental lexicon
Marslen-Wilson
(1980s) � showed that we can recognise a word even while it�s still being heard
(before the speaker has finished saying it) i.e. we access the lexical entry
before its acoustic offset) when shadowing (= repeating aloud what they
heard on headphones), they�d frequently start to vocalise a word before it had
finished being said
Marslen-Wilson
suggested that entries in the mental lexicon are not simply accessed, they are activated
Zwitserlood &
Marslen-Wilson: cross-modal priming =
(e.g.) auditory priming word, visual target word, played them the first half of
the priming word (e.g. �captain�), then flashed up the target words, if the
target word was related (e.g. �ship�), subjects responded faster
Gaskell: showed
with priming experiments that while the �thim� in �thim book� would activate
the lexical entry for �thin�, the �thim� in �thim slice� would not
Swinney (late
1970s) first used cross-modal priming: demonstrated that the alternative
meanings of ambiguous words are activated, people listened to a sentence, then
he immediately flashed up a word � people�s lexical decisions were faster for
related words, i.e. a sentence with �bank� in primed both �money� and �river�, however,
he found that if he presented the target word two or three syllables after
(downstream from) the ambiguous word, only the target related to the
contextually appropriate sense of the word was primed
however, some later studies failed to show the same results: general
consensus is that Swinney�s results were right � we do activate all meanings of
an ambiguous word
Tanenhaus et al.: can we eliminate from the lexical search all
words from inappropriate syntactic categories?
apparently not � used words ambiguous between a noun and a verb
(e.g. �watch�), and found that both senses were activated when listening to
sequences such as �John began to ��
Shillcock (late
1980s): wanted to see whether listeners activate the lexical entry for (e.g.)
�bone� in the sentence �He carefully placed the trombone on the table� � used
the cross-modal priming: played the sentence, and flashed up �rib� (related to
�bone�) at the offset of �trombone� � he found that �trombone� primed �rib� as
much as �bone� did, i.e. the lexical entry corresponding to �bone� is activated
even when� bone� is heard simply as part of the word �trombone�
Crain &
Hamburger � realised that relative clauses are usually used to show which of
many elephants
children take a
long time to learn what pronouns can/can't refer to � errors on both kinds can
be found up till about 10, though they�re usually pretty good by 4-6
Frisian does
allow �him� to be used like �himself�, but only in situations where there�d be
no ambiguity
Tanenhaus et al.
(1980s/90s) � used two almost identical sentences:
�which woman did Bertie
present a wedding ring to ___?�
�which horse did Bertie
present a wedding ring to ___?�
they found that the implausibility (of presenting a wedding ring
to a horse) is discovered when you get to �wedding ring�, before getting to the
gap
alternatively, use an EEG � Kutas et al. (early 1980s) found a
wave in response to contextually implausible words
and languages
like German (or Japanese or Turkish), where the verb indicating the roles may
be left till the end of the sentences, tend to be case-marking
between 18 months
and 6 years, children learn upwards of 9 words a day
Johnson-Laird
(late 1970s/80s): what happens when we hear language (i.e. its meaning) has
much in common with what would happen if we directly observed the situation
that the language described
Garnham: trying
to distinguish between meaning as the mental equivalent of the language used to
describe something or as the mental equivalent of what happens when one
observes that something directly � they were unable to remember which they�d
read, suggesting that the meaning of the text had been stored in terms of the
mental equivalent of something like a film of what had happened, rather than in
terms of a script which would have specified things in terms of the �man by the
window� or �the man with the martini�
(isomorphism
(Hofstadter)) � perhaps the neural activity that happens when we think the of
the arch reflects what is common to all the different patterns of neural
activity evoked when we have experienced arches
a normal adult
speaker:
has passive knowledge of about 50,000 words
can recognise + produce 3 words/second without any difficulty
semantic network (Collins & Loftus, 1975) = words as
conceptual nodes
Warrington &
McCarthy (1983, 1987):
problems localised to specific semantic categories, e.g. animals
vs objects
PET reveals how
dissociations in neurological patients can be identified in normal brains:
|
naming: |
activates: |
|
pictures of animals or tools |
ventral temporal lobe bilaterally |
|
animals |
also the left medial
occipital lobe (associated with the early stages of visual processing) |
|
tools |
left premotor area (activated by imagining hand movements) |
H. Damasio et
al� (1996)
naming task (large population of patients with
lesions), under three conditions: naming famous faces, naming animals, naming
tools
30 patients � showed impairment
29 of these had a left hemisphere lesion
7 patients � deficit naming faces
5 patients � deficit naming animals
7 patients � deficit naming tools
11 patients remaining � combination of problems in word retrieval
for faces/animals/tools, faces/animals, animals/tools, but never for
faces/tools
could correlate naming deficits with specific regions:
|
word-retrieval problem(s): |
lesion area: |
|
persons |
left temporal pole (TP) |
|
animals |
anterior part of the left inferior temporal (IT) lobe |
|
tools |
posterolateral part of the left IT lobe plus lateral
temporo-occipito-parietal junction (IT+) |
indicates that the brain has three levels of representation for
word knowledge:
1.
top level �
conceptual preverbal level, containing the semantic information about the word
(e.g. that �bird = beak + feathers + wings�)
2.
lexical
level � the word form that matches the concept is represented (e.g. �bird�)
3.
phonological
level � the sound information that corresponds to the word
how do we know
how to pronounce correctly when we read aloud � two ways (dual routes):
1.
grapheme-to-sound conversion
2.
direct lexical route = directly from reading to pronunciation,
i.e. from whole-word orthographic input to representations in the mental
lexicons
i.
deep/phonological dyslexia = cannot read aloud words that do not
have a representation in the mental lexicon
ii.
surface dyslexia = rely only on regularity rules
Petersen et al.
(1990) PET study:
areas in the human extra-striate visual cortex light up when processing
visual word forms
cohort model (Marslen-Wilson & Tyler, 1980)
components of word/lexical processing:
word initial cohort = activates all potentially-correct word
forms (e.g. �cap�, �capital�, �caption� etc.)
1.
lexical selection = selecting from the activated word
representations for the one that best matches the total sensory input
a word is selected at its uniqueness
point = when it is uniquely distinguishable
2.
lexical integration =
modular (autonomous)
models = normal language
comprehension is executed within separate + independent modules
interactive models = all types of information participate in
word recognition
hybrid models = lexical access is autonomous (not
influenced by higher-level information), but lexical selection can be
influenced by sensory + higher-level context information
growing evidence that at least lexical selection is influenced by
higher-level context information
baseline
condition:� 300 msec to press a button from the onset of the target word
�kitchen�
semantically
absurd but grammatically normal:� response to target word slowed by 60 msec
syntax also
disrupted:� response further slowed by 45 msec
garden-path model (Frazier et al., 1987) = theory of
syntactic structure based on sentences� preferred interpretation
1.
minimal attachment = tries to produce a structure where the
minimum number of aditional syntactic nodes must be computed
2.
late closure = tries to assign incoming words to the
syntactic phrase or clause currently being processed
Levelt (1989,
1993) model for language production:
1.
prepare the
message � two crucial aspects:
a)
macro-planning = determining what you want to express
b)
micro-planning = how the information is expressed, which
means taking perspective (�the house is next to the park� vs �the park is next
to the house�)
conceptual message = constitutes the input to the formulator
2.
formulator = puts the message in a grammatically +
phonologically correct form
three levels in Levelt�s mental lexicon model (see Damasio�s
model):
1.
conceptual
level
2.
lemma level
3.
lexeme or
sound level
anomic patients = deficit in naming
semantic paraphasias (often accompanies Wernicke�s aphasia) =
produce related words
dysarthria (often accompanies Broca�s aphasia) = hinders
articulation and results in effortful speech
A simple model of
language organisation
emphasised
word-level analysis, leading to the classic
localisationalist view/connectionist model of language
re-popularised by Geschwind (1967), first described by Lichtheim
(1885)
predicts seven possible aphasias � lesions in the production or
comprehension area, or any of the tracts to/from the three areas
|
auditory |
�/span> |
phonological lexicon |
�/span> |
conceptual area |
|
|
|
�/span>�/span> |
|
|
|
motor |
�/span> |
speech planning + programming |
�/span> |
Broca�s area
however, Dronkers (1996) reported that only 10/22 patients with
lesions in Broca�s area had Broca�s aphasia
Wernicke�s area
includes the posterior third of the superior temporal gyrus.
However, language comprehension deficits also arise from damage to the junction
between the parietal and temporal lobes, including the supramarginal and
angular gyri
in one study of 70 patients with Wernicke�s aphasia, about 10% had
damage confined to regions outside of Wernicke�s area
Dronkers suggests that white matter underlying Wernicke�s area may
be key
H.W. = an
intelligent businessman, studied by Baynes
stroke in his left hemisphere (large lesion, including large
regions of the posterior language areas), left with anomia and almost no other
deficits of note (except a slight right-side hemiparesis and slight deficit in
face recognition)
Lateralisation
right hemisphere lesions � Hagoort et al. (1996): normal priming
effects for words that are associatively related (e.g. �cottage cheese�)
Chiarello (1991): found a left visual field/right hemisphere
advantage for processing words that come fthe same semantic category but have
no associative relation
ERPs to index
aspects of semantic and syntactic processing during language comprehension
named N400
because it is a negative polarity voltage peak in brain waves for words that
usually reach maximum amplitude around 400 msec after onset of the word
stimulus
discovered by Kutas & Hillyard (1980): the amplitude of the
N400 to the anomalous words ending the sentence was increased when compared to
that of the N400 to congruent words (= the N400
effect) � in contrast, the semantically congruent but physically deviant
words elicited a positive potential (P560) rather than an N400
Hagoort, Brown
and Groothusen (1993): one ERP component that has shown up is the syntactic positive shift (SPS = a large
positive component elicited by words after a syntactic violation
a positive shift emerges in the ERP about 600 msec after the
syntactic violation (an SPS)
M�1993) found a negative wave over the left frontal areas of
the brain = syntactic anterior negativity
(SAN)
the N400 and SAN have different scalp topographies, implying that
they are generated in different neural structures in the brain
Ojemann et al.
(1989):
electrodes are used to pass a small electrical current through the
cortex (direct cortical stimulation) of an awake patient, momentarily
disrupting activity, thus probing where a language process is localised (since
it varies among patients) in order to leave critical language areas intact when
removing epileptic areas
when the patient makes an error in naming, or is unable to name
the object, the deficit is correlated with the region being stimulated during
that tiral
aspects of language representation in the brain are organised in
mosaic-like areas of 1-2cm2
these mosaic usually include regions in the frontal + posterior temporal
areas
the correlation tewen these effects in either Broca�s or
Wernicke�s areas was weak � some patients had naming disruption in the classic
areas and others did not
aphasia:
processing or representation losses?
Swaab, Brown and Hagoort (1997):
non-aphasic, right hemisphere-damaged controls and aphasics with
light comprehension deficit had an N400 like that of normals
aphasics with moderate to severe comprehension deficits had a
reduced + delayed N400 effect
Sacks (1985) President�s speech
Found in the RH processing of undamaged
fully Rhanded people.