Reading 3 Flashcards
computational models of visual word recognition
- dual route model
- interactive activation model
- PDP model
central issues of computational models of visual word recognition
- how is print converted to sound and meaning
- symbolic knowledge vs statistical learning
- how is correct word selected from similar alternatives
- how is lexical knowledge learnt
dual route model (colheart, 1978)
computational implementation: dual route cascade model (DRC)
two routes from print to sound:
- orthographic/lexical route
- grapheme-phoneme/non-lexical route
- -> pronunciation most frequently associated with each grapheme, used to predict regular pronunciation of nonwords
NEED two routes to explain:
(1) Exception words: eg colonel, pint, head
- -> this should use orthographic route
(2) Nonwords: eg slint, pead
- -> this should use grapheme-phoneme route
both routes operate in parallel, response determined by their combined influence on phoneme units in response buffer
lexical route operates more quickly esp. for high frequency words, low frequency word can complete slower and cause competition at response buffer
critical evidence for DRC model
- regularity x frequency interaction
high frequency word: no difference in RT for retrieving regular vs irregular words
low frequency word: slower in retrieving irregular word
- cognitive neuropsychology: acquired dyslexia
-phonological dyslexia:
words>nonwords
lexical route intact, GPC route damaged
-surface dyslexia:
nonwords>words
GPC route damaged, lexical route intact
- ->double dissociation
- ->independent systems
Interactive Activation (and Competition) Model (McClelland & Rumelhart, predecessor of TRACE)
• Hierarchical layers of interconnected nodes
(feature letter word)
- Parallel, interactive activation
- activated nodes send positive and negative activation to nodes at high and lower levels
- Identification occurs when activation in a node exceeds ‘threshold’
- Threshold depends on frequency: less evidence needed for common words
- Lateral inhibition within levels
- Competition between activated nodes to select best matching word
benchmark phenomena
- word frequency effects:
- identification threshold lower for common words - word superiority effect
- present a stimulus briefly, present two alternatives for last letter
- letters in words receive top-down support from word nodes, rather then when the stimulus is a non-sense word - pseudoword effect
- wordlike nonwords activate nodes for similar words - semantic priming: active
- word nodes activate their semantic features at concept level
- top-down effects from concept level
can IA model simulate regularity effects
Full model assumes that
spoken and written words
activate same word nodes
Positive and negative
connections between letters
and phonemes
➔phonological influences
on visual word recognition
➔‘regularity effects’ due to
consistency of pronunciation
of letters/graphemes NOT
non-lexical rules
rules vs statistics
regularity vs consistency
regularity: regular according to grapheme-phoneme rules e.g. ea in bean, a in came
consistency: statistically consistent, same letters combination in many similar words
e. g. all in fall
DRC model compared regular consistent words with irregular inconsistent words, confounds regularity and consistency
andrews (1982): no significant effects of regularity
significant effects of frequency, consistency, AND frequency x consistency interaction, consistency only affects low frequency words
DRC cannot explain why LF regular consistent (bean) and regular inconsistent (bead) should be different: both follow GPC rules
interactive activation vs PDP
both connectionist models
interactive activation: •‘Symbolic’ nodes for letters, words •Computational implementation: ‘hardwired’ lexical knowledge •Does not explain how the ‘nodes’ are learned
PDP:
•No hardwired knowledge
•Learn “distributed” representations
Parallel distributed processing (PDP) models
orthography - phonology - semantics
Connectionist (“neural”) networks
• set of interconnected processing nodes
• a “propagation rule” for spreading activation through the network
Learning in PDP networks
• learning algorithms (e.g., delta rule, back propagation)*
• Multi-level architecture: ‘hidden [internal] units’ facilitate learning of complex [non-linear] associations
• Extract statistical regularities between Orthography, Phonology, semantics
Memory structure is an emergent property
of the distributed dynamic processing
assumptions of PDP models
Spreading activation
• Based on analogy with neural firing.
• Each input to a unit has a level of activation and a weight on its connection – positive or negative
• Net input to a unit is sum of (activations) x (weights)
– neti = Σiwijaj
– Output is transformed by a (non-linear) transfer function
back propagation learning rules
Error correction learning
• connection weights initially random
• compute output for particular input pattern
• compare with desired output
–>adjust weights (a little bit) to reduce error
REPEAT
PDP model of lexical access
- PDP models provide an explanation of how knowledge is acquired (error correction learning eg back propagation) that is lacking from ‘symbolic’ models like dual route and IA model
- PDP model consistent with knowledge of neural mechanisms
- “Regularity effects” due to inconsistent O-P associations, as in IA model, NOT separate lexical and rule systems
➔ all knowledge represented in associative
networks NOT discrete rules
evidence for PDP model
- words with inconsistent O-P mappings are more difficult to learn
- -> yield weaker O-P connections
- -> PDP network shows graded (LF vs HF) effects of consistency of pronunciation
e. g. buoy > bear > bead > bean - semantic information contributes to resolving pronunciation of low frequency inconsistent words
- high imageability LF inconsistent words suffer less in human
- also simulated in PDP model
how does PDP explain surface dyslexia
PDP model trained without semantic network has specific difficulty learning low frequency exception words
➔ Semantic information helps to resolve inconsistencies eg bead, tread, pint
(how do you remember pint is pint if you don’t know what if means?)
➔semantic impairment selectively disrupts identification of inconsistent words
➔ Semantic dementia