1
UALL 2004
PSYCHOLINGUISTICS
LECTURE 7
Topic 4: Word Selection and Recognition
2
Semantics Phonology
Selecting Words
Syntax
3
Stepping-stone Waterfall /
Model Cascade Model
Selecting Words
Interactive/Spreading
Activation Model
Stepping-stone Model 4
Stepping-stone Model 5
Two major stones: (1) meaning and word class, (2) sounds
First, a person picks a meaning, a word class, and then the sound
structure.
Example: Looking at a smallish wild animal and later identifying it as a
BEAVER.
However, there would be a signpost saying whereabouts to go in the
next component: a phonological ‘area-code’.
Stepping-stone Model 6
‘BEAVER’ phonological area-code: a two-syllable word beginning with
/b/ and ending in /ər/.
Yet, the semantic component activates other small wild animals
belonging to the same category: OTTER, BADGER, RABBIT
Phonologically, words like ‘beaker’, ‘badger’, and ‘bearer’ are activated
too.
This explains the error of assuming a BEAVER as an OTTER or calling
it a BEARER.
Stepping-stone Model 7
Waterfall/Cascade Model 8
Waterfall/Cascade Model 9
(McClelland, 1979)
There was a need for a model to indicate that a person still thinks
about the meaning as he/she selects a sound – The waterfall model
shows this.
In this model, all the information activated at the first stage is still
available at the next stage.
So, after selecting a word's meaning from the activated group of
words, all of them remain available as the person deals with the
sounds.
So, word selection is not just a case of following one word through
from beginning to end.
Waterfall/Cascade Model 10
Waterfall/Cascade Model 11
(McClelland, 1979)
Example:
A: Think up the names of some woodland animals.
B: Rabbit, squirrel…I can’t think of anymore.
A: Hmm….beginning with ‘b’.
B: Oh….badger, beaver…
The dialogue shows information can flow both ways: particular sounds
can enable a speaker to activate meanings, just as meanings activate
sounds.
Interactive/Spreading 12
Activation Model
In speech production, the current is normally initiated in the semantic
component, where a semantic field will be aroused and then narrowed
down.
The current flows to the phonological ‘area-code’ before the final choice
is made, where many words will be triggered, and those activated will
feed back into semantics (information flow backwards and forwards).
All the links between activated sections will metaphorically be lit up, with
electric current rushing backwards and forwards, exciting more and more
related words.
Interactive/Spreading 13
Activation Model
Interactive/Spreading 14
Activation Model
The related words are stimulated strongly while the unrelated ones fade.
Thus, the activated word groups gradually narrow to more likely
candidates.
Phonology activation follows the same pattern of activating any that fits
more than one animal and is later matched with semantically activated
meaning to narrow down the choice.
If a person doesn’t pay much attention, the wrong choice can be made.
Different words require different activation level.
Word Recognition 15
Basic Problems:
1. It isn’t easy to hear each
phoneme in normal speech.
2. Neighboring sounds alter sounds.
3. Sound segments cannot be
separated – they melt into each
other.
Spoken Word Recognition 16
Process
(1)Initial contact, (2)lexical decision (selection), and
(3)word recognition are the three stages involved in the
spoken word recognition process.
Initialcontact: The listener takes the speech as input and
generates the representations that contact the internally
stored form-based representations associated with each
lexical entry.
Spoken Word Recognition 17
Process
Lexical decision (selection): Accumulating sensory input
continues to map onto this subset until the intended lexical entry
is eventually selected.
Word recognition – the endpoint of the selection phase when a
listener has determined which lexical entry was heard.
Word Recognition 18
Logogen Model TRACE Model COHORT Model
First Generation Model Second Generation Models
WORD RECOGNITION
Third Generation Model
Simple Recurrent Network Model
Logogen Model 19
John Morton (1969)
A bottom-up driven system that takes spoken or visual input and uses it
to activate previously stored word form representations.
A set of processing units that would receive input from spoken or written
modalities and fire when their activated inputs exceeded some criterion
level (threshold).
A logogen represents each word in a person’s vocabulary, so words are
recognized when the activation levels of their corresponding logogen
exceed the threshold.
Logogen Model 20
John Morton (1969)
Second Generation Models 21
(Models of Speech Recognition – Interactive Models)
The Cohort Model and TRACE Model are concerned with
auditory word recognition.
For these models, the result is a meaningful utterance
rather than a meaningless syllable.
These models aim to describe the interaction between the
processes of phoneme recognition and word recognition.
Cohort Model
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
22
Marslen-Wilson & Tyler, 1980, 1981)
This model was
designed to account
only for auditory word
recognition.
Cohort Model 23
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
Marslen-Wilson & Tyler, 1980, 1981)
Involves three stages: STAGE 1: The acoustic-phonetic
information at the beginning of a target
word activated all words that resembles
Stage 1: Activation (or Contact) it, making up a “word-initial cohort”
John was trying to get some bottles
down from the top shelf. To reach
Stage 2: Selection them he headed to sta… stab, stack,
stagger, stagnate…
STAGE 2: All other sources of
Stage 3: Integration information (frequency, context) may
influence the selection of the target
word from the cohort
STAGE 3: Final word chosen when
only a single candidate remains in the
cohort.
Cohort Model 24
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
Marslen-Wilson & Tyler, 1980, 1981)
COHORT views contact as being influenced only by bottom-up auditory
information and not by contextual information.
So, activation in COHORT is an autonomous process – it is affected by
auditory stimulation but not other potentially relevant cognitive processes.
❑ Thus, stored representations of words that do not fit into the evolving
context are activated anyway as long as they match the acoustic
properties of the word stimulus.
Cohort Model 25
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
Marslen-Wilson & Tyler, 1980, 1981)
The selection depends on the bottom-up stimulus because bottom-up
information activates word candidates, but it also depends on context.
❑ Words that fit better into the context will have an advantage over words
that do not fit, especially in cases where the bottom-up input is
ambiguous between two or more stored word candidates.
During integration, properties of the selected word – its grammatical class
and meaning – are evaluated to how well they fit with the preceding
context.
Cohort Model 26
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
Marslen-Wilson & Tyler, 1980, 1981)
COHORT is called as such because the process of lexical access starts
with a contact phase in which all words that match the perceived acoustic
profile are activated.
So, within about 100-150ms of the onset of a word, a whole group of
matching candidate word forms become more available or accessible
than usual.
This group of activated word forms is called a cohort.
Cohort Model 27
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
Marslen-Wilson & Tyler, 1980, 1981)
One advantage of the model is that it makes very specific predictions
about when, exactly, a word can be recognised and its meaning
accessed.
Word recognition depends on reducing the set of activated words to the
one that matches the acoustic input. The point where the COHORT is
reduced to a sole survivor is called the ‘recognition point’. (Marslen-
Wilson, 1987).
The model allows for minor adjustments to the recognition point based on
semantic or syntactic requirements imposed by the context, which allows
words that are highly predictable in context to be recognized faster than
the less predictable ones.
Cohort Model 28
(Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978;
Marslen-Wilson & Tyler, 1980, 1981)
Word recognition in this model is contingent on two factors:
(1) There has to be positive evidence for the presence of the word (e.g.:
the input ‘tres’ provides clues that the word ‘trespass’ is the matching
word target).
(2) The input has to rule out the presence of other words (e.g.: the onset
‘tr’ rules out the possibility that the matching word target is ‘tap’, ‘top’
‘table’ or any other word that does not begin with ‘tr’)
TRACE Model 29
(McClelland & Rumelhart, 1981)
(a) indicates the basic organization of
processing units and information flow,
while (b) shows a detailed view of the
way the processing units are
connected to one another.
TRACE Model 30
(McClelland & Rumelhart, 1981)
A highly interactive model in which activation of processing units is
determined solely by stimulation provided by the input.
The activation of one processing unit in a bottom-up system does not
directly affect the activation of other processing units at the same system
level.
Activation at higher levels of a bottom-up processing system does not
affect activation at lower levels.
Phonemes affect the activation of word units, but word units do not affect
activation in the units that represent phonemes.
TRACE Model 31
(McClelland & Rumelhart, 1981)
Interactive processing systems have connections between processing
units that allow units within the same level to affect one another and that
allow processing units at higher system levels to affect units at lower
levels.
The model assumes that activation is ‘cascaded’, contrasting threshold
activation.
In cascaded activation, units receiving input begin to send output as soon
as any activation at all comes in from other units.
Visual features start to send activation forward as soon as they begin to
be identified, so letter-level processing units start to become active soon
after feature-level processing units start to become active.
TRACE Model 32
(McClelland & Rumelhart, 1981)
That means letter representations start to become activated as soon as
any visual feature has been identified, and you do not need to perceive
all of the features of a letter before you start to activate letter-level
processing units.
The model offers a good explanation of the “word superiority” effect that
indicates that we have an easier time recognising and processing letters
and phonemes when they appear in the context of a word than when they
occur by themselves or in the context of a string of letters that does not
make up a real word.
Cohort and TRACE Models 33
Distributed Feature Models 34
The parallel distributed processing idea continued to grow and develop
with the invention of newer and more advanced mathematical models of
lexical access.
Jeff Elman’s Simple Recurrent Network (SRN) model assumed that
words were represented as a pattern of neural activity across multi-
layered network.
Simple Recurrent Network Model
Jeff Elman (2004) 35
Simple Recurrent Network Model 36
Jeff Elman (2004)
This model took the TRACE model and added a set of context units.
The context units store a copy of the activations in the hidden units
between processing cycles.
The network would respond not just to the current state of the input units
but also to recent events, as reflected in the activity of the context units.
The explicit task that the network performed was to predict the upcoming
word in an utterance.
Simple Recurrent Network Model 37
Jeff Elman (2004)
In this system, word identities can be represented as an activation
pattern among the hidden units.
The patterns split neatly into two classes, corresponding to nouns and
verbs; within each class, the word representations are subdivided further
into subclasses, with similar representations assigned to words that we
would judge as being close in meaning.
