What are the functional units in reading? Evidence for statistical variation influencing word processing

What are the functional units in reading?

Evidence for statistical

variation influencing

reading

Alastair Smith

Padraic Monaghan

The Debate Theoretical Models Contrasting Evidence Study Discussion

The Debate

What information is used to map orthography

onto phonology?

The Debate Theoretical Models Contrasting Evidence Study Discussion

Competing Models

Dual-Route Models– Dual-Route Cascade Model (Coltheart et al, 1993)

– Connectionist Dual Process Model (Zorzi et al, 1998)

– CDP+ (Perry, Zorzi & Ziegler, 2007)

Single Route Models– Parallel Distributed Processing Model

– Seidenberg & McClelland, 1989

– Plaut, McClelland, Seidenberg & Patterson, 1996

– Harm & Seidenberg, 1999

The Debate Theoretical Models Contrasting Evidence Study Discussion

Lexical Route

Sub-lexical Route

Serial Processing

Explicit level of representation for graphemes

Dual-Route Models

The dual-route cascaded model.

From The CDP+ Model of Reading Aloud.

By Perry, C., Ziegler, J.C. & Zorzi, M., 2007

Psychological Review, 114, p.275.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Single Route Models

Parallel Processing

Encodes statistical

relations between

patterns of letters and

their pronunciation

Single letters provide

input

The ‘triangle’ model.

From Computing the meanings of words in reading.

By Harm M.W., Seidenberg M.S., 2004

Psychological Review, 111, p.663.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Psycholinguistic Grain Size Theory(Ziegler & Goswami, 2005)

Type of processing that occurs in reading system

determined by statistical relations between orthography

and phonology.

Grain sizes:

– Language specific

– Allow for efficient mapping

Learning to read is learning to find shared grain sizes in

orthography and phonology.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Graphemes

Written representations of phonemes

Can be composed of multiple letters:

– Digraphs

– Trigraphs

From Exploring Grain-Size Effects in Reading.

By Pagliuca, G., Monaghan, P., 2008

Proc 30th Ann Conf Cog Sci Soc. Mahwah, NJ:

Lawrence Erlbaum.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Whammies and Double Whammies(Rastle & Coltheart, 1998)

Participants read non-words containing 3 phonemes (e.g. fooce) slower than control non-words containing 5 graphemes (e.g. fruls)

Behavioural study supported by simulation data from dual route model

– Non-lexical route processes non-word serially left to right, letter by letter

Conclusions:

– Reading system

is serial

– Functional unit is the

letter, not the digraph From Whammies and double whammies.

By Rastle, K., Coltheart, M., 1998

Psychon Bull Rev, 5, 277-282

The Debate Theoretical Models Contrasting Evidence Study Discussion

Grain-Size Effects in Reading(Pagliuca, Monaghan & McIntosh, 2008)

Findings seem to contradict those of Rastle & Coltheart, 1998

Indicates grain-size adapts according to statistics in the orthography - phonology mapping

Hypothesis:

– If graphemes are functional units within the reading system, then a word containing a multi-letter grapheme should be read more accurately than a word without given the same kind of perceptual noise to impair the orthographic input.

Modelling data from single route model supported by behavioural study

The Debate Theoretical Models Contrasting Evidence Study Discussion

Modelling: Pagliuca, Monaghan & McIntosh, 2008

Single route model based on Harm & Seidenberg, 1999

Orthographic input represented by 8 letter slots

Activation from input letter slots reduced along monotonic gradient from left to right

so that the lowest level of activation was in the left most slot

– two severities of impairment applied, severe and mild

Model tested on two sets of 62 words, all 5 letters in length and monosyllabic

– Set 1: Digraphs in initial position ch, sh, th

– Set 2: Control set (no digraphs) cr, st, tr

Words beginning with digraphs

were read more accurately

From Exploring Grain-

Size Effects in

Reading.

By Pagliuca, G.,

Monaghan, P., 2008

Proc 30th Ann Conf

Cog Sci Soc. Mahwah,

NJ:

Lawrence Erlbaum.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Behavioural Study: Pagliuca, Monaghan & McIntosh, 2008

Same sets of words used in the behavioural study as used in the simulation. 84 additional filler words selected, each five letters long with different initial bigrams and initial letters to the experimental and control stimuli

Visual noise applied to stimuli from left to right,

similar to noise applied in simulation study

Participants completed naming task in which each word was presented for 250ms

15 university students participated

all native English speakers

Words with digraphs were reported more

accurately than words without, confirming

predictions made by the model

From Exploring Grain-Size

Effects in Reading.

By Pagliuca, G., Monaghan,

P., 2008

Proc 30th Ann Conf Cog Sci

Soc. Mahwah, NJ:

Lawrence Erlbaum.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Conclusions:Pagliuca, Monaghan & McIntosh, 2008

Modelling:

– For digraphs two letter positions contribute to the activation of a single phoneme, whereas for non-digraphs each letter only contributes to one phoneme’s activation

– Graphemes emerge in the course of a system learning the regularities between orthographic and phonological representations of words

Behavioural study:

– Indicates computational properties have a profound affect on reading, at least under conditions where visual input is impaired

Different computational properties of the mapping between letters and phonemes suggests psycholinguistic effects of words should vary according to the compositionality of the mapping

The Debate Theoretical Models Contrasting Evidence Study Discussion

Research Aims:

1. Using a computational model of reading based

on Harm & Seidenberg, 1999.

Can we extend the digraph effects found in

Pagliuca, Monaghan & McIntosh to non-

words?

2. Test predictions raised by model in

experimental studies.

The Debate Theoretical Models Contrasting Evidence Study Discussion

Modelling Study: Design (Model)

Computational Model:– Based on Harm & Seidenberg 1999

– Orthographic Input Layer:• 10 letter slots

• One of 26 units active in

each slot to represent letter

– Hidden Layer: 100 units

– Phonological Output Layer:• 8 phoneme slots

• Each phoneme represented in terms

of 25 phonological features

– 25 Clean-up units

From Phonology,

Reading Acquisition,

and Dyslexia.

By Harm, M. W.,

Seidenberg, M.S.,

1999,

Psychol Rev, 106,

491-528

The Debate Theoretical Models Contrasting Evidence Study Discussion

Modelling Study: Design (Training)

– Training corpus:

• 6229 monosyllabic words,

• Words 1 to 8 letters in length

– Training algorithm:

• backpropagation learning algorithm (Rumelhart, 1986)

– 5 million cycles of training, words submitted randomly

according to frequency

– 99.9% accuracy following training (tested on training corpus)

The Debate Theoretical Models Contrasting Evidence Study Discussion

Modelling Study: Design (Stimuli)

– Stimuli sets each containing 64 items:• Words with digraph in onset

• Non-words with digraph in onset

• Control Words

• Control Non-words

– All Words and Non-words 5 letters in length and Monosyllabic

– Onset pairings matched for same initial letter and similar bigram frequency

– Controls applied:• Word frequency

• Body Friends and Body Enemies

• Neighbours

• Unigram and Bigram frequency

• Partial View Predictability

– Non-words were formed by switching onsets and rimes within given word set

(Controls were performed on non-words following formation)

– Noise applied in three conditions:• No Noise

• Uniform 50% reduction in input activation

• Decreasing noise condition (replication of Pagliuca, Monaghan & McIntosh, 2008)

The Debate Theoretical Models Contrasting Evidence Study Discussion

Modelling Study: Results (Words)

Model performance on word sets:

– Both sets read with 100% accuracy before noise applied

– Digraph set read with greater accuracy when input uniformly impaired

(t(126) = 2.453, p < 0.01)

– Digraph set read with greater accuracy in decreasing noise condition

(t(126) = 4.396, p < 0.01)

**

**

** p<0.01

* p<0.05

The Debate Theoretical Models Contrasting Evidence Study Discussion

Modelling Study: Results (Non-words)

Model performance on non-word sets:

– Accuracy based on comparing output to target. Target formed by combining phonetic representation of onset and rhyme extracted from corpus

– Lower accuracy in reproduction of digraph set before noise applied

– Digraphs read more accurately in non-words when input uniformly impaired

(t(126) = 3.355, p < 0.01)

– Non-words containing digraphs read more accurately in decreasing noise condition

(t(126) = 2.495, p < 0.01)

**

++

** p<0.01, * p<0.05, ++ p<0.01 based on error in onset

The Debate Theoretical Models Contrasting Evidence Study Discussion

Model Predictions:

Both words and non-words containing digraphs

in the initial position will be identified with

greater accuracy than controls.

– For digraphs in both words and non-words two letter

positions contribute to the activation of a single

phoneme

The Debate Theoretical Models Contrasting Evidence Study Discussion

Behavioural Study: Design (Stimuli)

– 4 stimuli sets taken from simulation: Control Non-words

Control Words

Words with digraphs in onset

Non-words with digraphs in onset

– 2-dimensional digital pixel noise applied across word in decreasing gradient from left to right

Example of control word with visual noise applied Example of control non-word with visual noise applied

The Debate Theoretical Models Contrasting Evidence Study Discussion

Behavioural Study: Design (Procedure)

Participants:– 15 university students

– All native English speakers

Lexical decision task:– departs from Pagliuca, Monaghan

& McIntosh, 2008

Procedure:– Short practice period

– Fixation cross presented before stimuli

– Stimuli selected at random without replacement

– Stimuli presented for 250ms

– Response recorded by key press

– 256 trials completed by participant

The Debate Theoretical Models Contrasting Evidence Study Discussion

Behavioural Study: Results (Accuracy)

Accuracy of Response:

Words containing digraphs were responded to more accurately than controls

(t(14) = 3.254, p<0.01)

Non-words containing digraphs were responded to less accurately than controls

(t(14) = 2.457, p<0.05)

*

**

** p<0.01, * p<0.05

The Debate Theoretical Models Contrasting Evidence Study Discussion

Behavioural Study: Results (Response Times)

Response Times:

Similar trends were found in participants reaction times although significance levels were not reached

The Debate Theoretical Models Contrasting Evidence Study Discussion

Summary

Modelling:– Greater accuracy reading both words and non-words containing

digraphs in the initial position in high level noise conditions.

– For digraphs two letter positions contributing to activation of single phoneme

Behavioural study:– Words containing digraphs identified with greater accuracy than

controls when visual noise applied in a decreasing gradient across word

– Non-words containing digraphs identified with less accuracy than controls when visual noise applied in a decreasing gradient across word

The Debate Theoretical Models Contrasting Evidence Study Discussion

Discussion (1)

Task differences:

– Word naming task:

• Pagliuca, Monaghan & McIntosh, 2008

• Rastle and Coltheart, 1998

• Modelling study

– Lexical decision task:

• Behavioural study

The Debate Theoretical Models Contrasting Evidence Study Discussion

Discussion (2)

Simulation and Behavioural data showed an advantage

for words containing digraphs:

– Replication of Pagliuca, Monaghan & McIntosh, 2008

– Indicates the grain size for reading in English is adaptable

according to statistics of the letter-sound mapping

– Challenges views on independence of letter recognition (Pelli,

Farrell and Moore, 2003) indicating word perception affected by

statistics in the language

The Debate Theoretical Models Contrasting Evidence Study Discussion

Discussion (3)

Combined findings:

– Single Route (Parallel Processing) Model:

• Provides explanation for increased accuracy in identifying

digraph words displayed by simulation and behavioural

data

(Pagliuca, Monaghan & McIntosh, 2008)

• Model predicted advantage for reading digraph non-words,

however behavioural data showed lower accuracy of

response and slower reaction times

The Debate Theoretical Models Contrasting Evidence Study Discussion

Discussion (4)

Combined findings:

– Dual Route (Serial Processing) Model:

• Provides explanation for reduced accuracy in digraph non-

word response

(Rastle & Coltheart, 1998)

• Digraph word advantage not predicted by models lexical

route

The Debate Theoretical Models Contrasting Evidence Study Discussion

Direction of Future Study

Non-word Naming Task

Digraphs in final position– If non-lexical route serial this should lead to slower response

times

(Rastle & Coltheart, 1998)

Use similar paradigm to investigate grain-size effects in languages with differing grain-size

Special Thanks & Acknowledgements

Experimental Psychology Society

Lancaster University

Questions: