Toward a theory of reading acquisition as a synthesis of implicit and explicit learning

Material Information

Toward a theory of reading acquisition as a synthesis of implicit and explicit learning
Maranzana, Elisa
Publication Date:
Physical Description:
vi, 157 leaves : ; 29 cm.


Subjects / Keywords:
Learning ( jstor )
Linguistic morphology ( jstor )
Linguistics ( jstor )
Morphemes ( jstor )
Orthographies ( jstor )
Phonemics ( jstor )
Phonology ( jstor )
Syllables ( jstor )
Vowels ( jstor )
Words ( jstor )
Dissertations, Academic -- Linguistics -- UF ( lcsh )
Linguistics thesis, Ph. D ( lcsh )
bibliography ( marcgt )
non-fiction ( marcgt )


Thesis (Ph. D.)--University of Florida, 1997.
Includes bibliographical references (leaves 143-156).
General Note:
General Note:
Statement of Responsibility:
by Elisa Maranzana.

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright [name of dissertation author]. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Resource Identifier:
027600704 ( ALEPH )
37259565 ( OCLC )

Full Text








I am grateful to my committee chair, Gary Miller, whose interest and

support have made it possible for me to complete this project. I would also like to

thank my committee members: Caroline Wiltshire, Ann Wehmeyer, and Linda

Lombardino. This work has profited very much from their interest, comments and

suggestions. Thanks also go to Diana Boxer and Michel Achard for pinch-hitting at

the oral exams. A special thank you goes to Charles Wood for his ideas, advice

and editing skill, all of which have improved the quality of this work.

I would like to acknowledge the contribution of my experience at the

University of Texas at Austin which has galvanized my interest in the cognitive

sciences, and inspired the many of the ideas behind this work.

I wish to thank my friends and colleagues at the English Language Institute

and the Program in Linguistics for their moral support. I am also grateful for the

financial support that I have received from the PIL and ELI, and especially to Jean

Casagrande for making it happen. A sincere thanks goes to Cedar Key, the island

and the people, for hosting a portion of this work. Finally, a warm thank you goes

to my friends and family, especially Charles, whose love and support made it

possible for me to be here and to finish this project.


ACKNOWLEDGMENTS ....................................................... ii

ABSTRACT ..................................................................... v



Trends in ReadingTheory .................................................. 3
The Question of Representation: Evidence from Writing Systems and Language
Processing .............................................................. 5
The Principles of Implicit Learning ........................................... 7
Implicit Learning and Reading Acquisition: Processing Commonalities ......... 10
Practical and Theoretical Implications ............. .......................... 10
Concluding Remarks ........................................................ 11

2 TRENDS IN READING THEORY AND PRACTICE .......................... 13

Phonic Theories of Reading .................................................. 13
Whole-Word Theories ....................................................... 17
Dual-Route Theories ......................................................... 18
Interactive Activation Theories ............................................... 21
Reading and Morphology .................................................... 25
Teaching Philosophies: Code vs. Whole Language Approaches
to Reading Acquisition ................................................... 25
Strengths and Weaknesses of the Code Approach .............................. 27
Strengths and Weaknesses of Whole Language .... ............................ .... 30
Summary and Concluding Remarks ......................................... 31


A Survey of Script Types .................................................... 36
Alphabets ............................................................... 43
Orthographic regularities .......................... ................... 44
Linguistic structures .................................................. 45
Summary and Concluding Remarks ........................................ 51


Syllable Structure and Language Processing ................................ 53
Syllable Structure in Reading ............ ............................... 56
Some Evidence for Use of Onset and Rime .......... ..................... 56
Some Evidence for Use of Syllable Boundaries ................. .... ... 59
Morphological Structure and Language Processing ............................. 61
Morphological Structure in Reading ..... ........ ...... ........... ......... 63
Some Experimental Evidence of Morphological Processing ................. 64
Evidence of Morphological Processing from Children's Spelling ........... 69
Some Evidence from the Evaluation of the Reading Impaired ................ 70
Neurobiology of Multi-leveled Processing: Evidence from the Neuroanatomy of
Dyslexics ................................................................ 72
Summary and Concluding Remarks ........................................ 73

5 IMPLICIT LEARNING ..................................................... 76

Empirical Studies ............................................................ 77
Grammar Learning ..................................................... 78
Grammar learning and natural language learning........................ 85
Process Control .......................................................... 87
Sequence Pattern Acquisition .................................... ..... 88
Learning Mechanisms and Knowledge Representation ........................ 90
Implicit versus Explicit Learning Mechanisms ............................ 93
Implicit and Explicit Learning and Individual Differences ....................... 95
Developmental Issues ......................................................... 96
Some General Implications for Instruction .................................. 98
Concluding Remarks ....................................................... 90

PROCESSING COMMONALITIES ......................................... 101

The Implicit/Explicit Interface ................................................ 102
Orthographies and Regularity Salience .......................... ........ 102
Task Demands ........................................................ 110
Chunking ................................................................... 112
Learning is Gradual ......................................................... 113
On Consciousness ........................................................... 116
Implicit/Explicit Learning is Efficient ......................................... 117
Toward a Theory of Reading Acquisition: A Proposed Framework .............. 120

7 PRACTICAL AND THEORETICAL IMPLICATIONS ......................... 122

Implications for instruction ................................................. 122
When is Instruction Useful .............................................. 122
How Much Instruction is Optimal ....................................... 124
How Aware Must Learners be of Relevant Preexisting Knowledge .......... 124
Thinking Aloud .......................................................... 125
The Order of Implicit and Explicit Learning ............................... 126
The Role of Attention .................................................. 127
A Synthesis of Code and Whole-Language Insights ........................ 129
Theoretical Implications ..................................................... 130
The Optimality of Script Type .......................................... 130
Implicit /Explicit Learning and Individual Differences and IQ................ 134
On Connectionsim ....................................................... 136
Rule Based Vs. Association Based Models ................................ 138
Implications for Linguistics .............................................. 139
What does reading have to do with language? ........................... 139
Language acquisition and implicit learning ....................... ..... 139

REFERENCES ............................................................... 143

BIOGRAPHICAL SKETCH ..................................................... 157

Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the Requirements for the
Degree of Doctor of Philosophy




May 1997
Chair: Dr. D. Gary Miller
Major Department: Program in Linguistics

While theories of reading are presumably based on an understanding of

orthographic systems, the understanding of what in the written system represents

what in the spoken system is often not carefully considered. Studies of writing

systems and the typological classifications of the world's script types, however,

have provided insight into the various types and nature of orthographic

representations, and it appears that among seemingly diverse scripts, there are some

striking commonalities. Linguistically informed analyses of various scripts tell us

that orthographies represent different "levels" of linguistic knowledge. What is

often overlooked in reading research is that 1) orthographies are only partially

representative of any particular level of linguistic knowledge. For example,

alphabetic scripts only partially represent phonology. And 2), orthographies

usually represent aspects of more than one level simultaneously: phonological

(segmental and/or syllabic), morphophonemic, morphological, lexical. Hence,

orthographic systems are complex, incomplete, sometimes conflicting

representations of language.

The first component of this dissertation will bring reading acquisition
theories and empirical research relating to reading and visual language processing

together with the linguistically informed theoretical study of writing systems. My

intention is to demonstrate that orthographic systems (their commonalties and

differences) are themselves evidence as to how visual language is processed, and

that this information can and should inform reading theory. The second

component is a discussion of what I see as a major implication of the previous

argument. That is, the complexity of orthographic systems indicates that learning

the associations necessary for reading involves far more than what is generally

provided by way of instruction. Hence, reading acquisition involves what, in many

of the fields of cognitive science, is called implicit learning --a process whereby

structured information is induced from a complex stimulus environment, largely

without awareness of what exactly is learned or how. It is generally acknowledged

that learning to read and write does not proceed automatically (as does first

language acquisition) but instead requires explicit instruction. The notion that

reading acquisition also involves a large degree of implicit induction (a natural,

efficient cognitive process) has major theoretical and pedagogical implications.


Virtually all reading research' is based on the assumption that orthographies

-- how the visual symbols are systematically arranged -- are representative of some

aspect of the structure of language. The assumption that the representational

relationship between speech and writing is straightforward is a central premise in

reading research. For example, alphabets are thought to be representative of the

sound system of language. With English, it is commonly assumed that alphabets

represent phonological segments and that those segments combine to form words in

a fashion similar to the spoken language. Since reading research most often

involves alphabetic scripts, the bulk of experimental research has focused on how

phonological structure is used in reading. Although research with non-alphabetic

orthographies is comparatively limited, the premise of a transparent representational

relationship is the same. For instance, Chinese characters are generally assumed to

represent morphemes/words. Syllabaries used for Japanese and Korean represent

phonology at the level of syllables. The notion that the representational relationship

may be more complex than this is less often addressed, especially in the

experimental literature and in reading pedagogy.

While theories of reading are presumably based on an understanding of the

orthographic representation, the conception of what in the written system represents

what in the spoken system is often not carefully considered. Studies of writing

systems and the typological classifications of the world's script types, however,

I I refer to'reading research' throughout, as that which cuts across several disciplines--
psychology, education, linguistics, and others, all of which begin from the assumption that
writing is a representation of speech.

have provided insight into the various types and nature of orthographic

representations. Linguistically informed analyses of various scripts tell us that

orthographies predominantly represent different "levels" of linguistic knowledge

(see Daniels, P. and Bright, W, 1996 for a thorough, linguistically focused survey

of the world's writing systems). For example, alphabetic scripts are usually,

predominantly phonological. What is not so clear, indeed often overlooked in

reading research, is that 1) orthographies are only partially representative of any

particular level of linguistic knowledge. For example, alphabetic scripts only

partially represent phonology (Chomsky and Halle 1968; Chomsky 1970; Coulmas

1989; Miller 1994; Daniels 1992; 1996). And 2), orthographies usually represent

aspects of more than one level simultaneously: phonological (segmental and/or

syllabic), morphophonemic, morphological, lexical, phrasal (Gleitman and Rozin

1977; Henderson 1982; Sampson 1985; Coulmas 1989; Daniels 1992, 1996;

Miller 1994).

The first component of this study will bring together reading acquisition

theories and empirical research relating to reading and visual language processing,

with the linguistically informed theoretical study of writing systems. My intention

is to demonstrate that orthographic systems (their commonalties and differences) are

themselves evidence of how visual language is processed, and that this information

can and should inform reading theory. If reading does involve simultaneous

interpreting of various aspects of structure, then understanding the word

recognition process will require evaluating how and to what degree each level of

structure is encoded in an orthography. The second component is a discussion of

what I see as a major implication of the previous argument. That is, the complexity

of orthographic systems indicates that learning the associations necessary for

reading involves far more than what is generally provided by way of instruction.

Hence, reading acquisition involves what, in many of the fields of cognitive

science, is called implicit learning -- a process whereby structured information is

induced from a complex stimulus environment, largely without awareness of what

exactly is learned or how.

It is generally acknowledged that learning to read and write does not

proceed automatically (as does first language acquisition) but instead requires

explicit instruction. Further, the traditional stuff of education has been the

measurement of explicit learning processes, which involve the selective search for

information, the building of hypotheses and testing them, or the assimilation of

rules following explicit instruction. Since teaching intervention is known to be a

critical factor in learning to read, research has largely been devoted to developing

theories of explicit learning -- phonics, for example. The notion that reading

acquisition also involves a large degree of implicit induction (a natural, efficient

cognitive process) has major theoretical and pedagogical implications.

Trends in Reading Theory

Theories of reading generally proceed from the position that word

recognition involves the analysis of one level of linguistic structure. Research on

how words are best learned range from "phonic" theories, which suggest that the

graphic system is a code for the phonological system, to "whole word" theories,

which believe that words are learned as visual gestalts. Most current work in

reading suggests that both phonic and lexical-semantic strategies apply in learning

to read. These "dual route" theories suggest that readers can read words using one

of two routes. The lexical route leads directly into the lexicon and is accessed

semantically. The phonic route leads indirectly to the lexicon, with readers first

applying letter sound translation. Interestingly, most dual route theories posit that

these routes operate independently and are not interactive. So, while the

frameworks above differ with respect to the particular access routes) to word

recognition, all assume the reader need interpret only one level of linguistic

structure in order to recognize words. More recently, interactive activation models

of word recognition emphasize that word recognition is a single interactive process.

Instead of alternate recognition pathways, these models assume that access of

phonological information is an automatic consequence of word recognition.

Because theories which endorse phonological recoding have led to specific,

testable questions about reading, there has been a great deal of experimental

research in this area. The guiding premise of this view is that the relationship

between alphabetic scripts and what they represent is primarily phonological, based

on phoneme-grapheme correspondences. Until recently, it has been assumed a

priori that if the orthography is alphabetic, the representation must be phonemic.

There is a great deal of literature and experimental research which examines the

relationship between reading and phonemes wherein there is considerable evidence

that "phonological awareness" correlates with beginning reading success.

Phonological awareness is often defined as the ability to recognize and separate

words into their constituent phonological units. Though this notion may be subject

to different interpretations2, it generally refers to the segmenting and blending of

phonemic units -- phonemic awareness. While some researchers have explored the

role syllabic (as opposed to phonemic) awareness plays in reading acquisition

(Gleitman and Rozin 1973; Treiman 1992; Goswami and Bryant 1992), the greatest

emphasis is put on phonemic awareness since it is this skill that has been most often

correlated with beginning reading (Adams 1990; Byre and Feilding-Barnsley

1989, 1990,1993; Ehri and Wilce 1980; Ehri 1994; Liberman, Shankweiler

2 Empirical evidence of phonological awareness is taken from experiments involving phonological
awareness tasks. These tasks are of varying complexity and difficulty for both children and adults.
The easier tasks involve testing knowledge of more perceptually salient aspects of phonology:
syllables-- rhyme and alliteration tasks and syllable counting. Phonemic awareness or phoneme
discrimination tasks such as phoneme counting/tapping, phoneme deletion or addition are much
more difficult. Even very literate adults are not uniformly successful with phoneme deletion
tasks (Scholes and Willis 1990), making unclear the degree to which phonemic awareness results
from literacy instruction.

Liberman, Fowler and Fisher 1977; Morais, Bertelson, Cary and Algeria 1986;

Stanovich, Cunningham, and Cramer 1984). Consequently, it is widely believed

that words are recognized in accordance with their phonemic structure and that

beginning reading depends on awareness of that structure. Accordingly, beginning

readers should be taught to analyze strings of letter to sound correspondences and

blend them together to form words.

Chapter two gives an overview of the competing reading theories and

major empirical research associated with them. Two relevant points arise from this.

First, reading research is particularly focused on how readers use phonemic

structure3. And second, it is widely believed that use of phonemic structure in

reading does not develop automatically but must be explicitly taught Accordingly,

the more aware readers are of the phonemic correlation, the more reading

acquisition should be facilitated. I intend to show (in following chapters) that by

focusing primarily on phonemic recoding, reading research has overlooked some

obvious properties of alphabetic writing systems and has largely excluded from

investigation the processing of other aspects of linguistic structure. Additionally,

the phonemic perspective endorses the notion that reading acquisition is primarily

an explicit, conscious, goal directed learning process, a perspective that this

discussion calls into question.

The Question of Representation: Evidence from
Writing Systems and Language Processing

What I believe has been most problematic in theories of reading is the

ambiguity about what in the graphic system represents what in the spoken linguistic

system. The premise that alphabetic scripts represent phonological systems is

fundamental to theories of phonological recoding. Ideally, there would be a one-to-

one correspondence between graphemes and phonemes, mirroring the language's

3 This is due to the fact that most research has focused on reading alphabetic scripts.

phonemic system. However, owing to a complex of factors such as historical

change, dialect variation, intermittent spelling reform, and pressures for linguistic

preservation, most alphabetic scripts fall short of this ideal. Even scripts that are

quite phonemically regular (e.g. Dutch, Italian) deviate from this ideal in many

respects. With some alphabetic scripts (e.g., English, French), direct phoneme-

grapheme correspondence is minimal.

Yet, alphabetic orthographies do, albeit partially, represent phonology at the

level of segments. This in itself, however, is an inadequate basis for concluding

that the letter-phoneme relationships are the exclusive means of representation and

processing. Similarly, there is no reason to assume that lexical-semantic word

recognition occurs independently, wi thout associations to smaller linguistic units

present in the orthography. It is more likely that the letter-phoneme connection is

but one aspect of the representational system that must be decoded by the reader;

the lexical-semantic connection another. It is likely also that the representational

system includes linguistic associations between the phoneme and word levels.

Furthermore, there is no reason that word recognition should rely on any single

pathway at a time. In fact, neurobiological evidence suggests that processing any

complex visual or auditory stimulus relies on multiple pathways simultaneously

(Galaburda 1996).

I believe the research focus on one script type (alphabets) has resulted in

investigators missing what alphabets have in common with other scripts. Chapter

three discusses the variety among the worlds script types in order to show that

among diverse scripts there are some striking commonalities: All scripts represent

multiple aspects of linguistic structure, and they are all partial representations of

those structures. Having illustrated that multileveled representations are common

across various scripts and script types, I will argue that processing written language

(in general) is much more complex than phoneme-grapheme translation. Instead,

reading involves interpreting multiple levels of linguistic structure, and also,

extracting information that is not overtly encoded as linguistic units.

Because discussions of how readers utilize linguistic structure have focused

primarily on phonology, specifically phoneme-grapheme correspondences, there is

comparatively little focus on whether readers utilize morphology or syllable

structure. This oversight is striking considering the central role that the constructs

morpheme and syllable play in theoretical linguistic accounts. Chapter four will

discuss empirical evidence indicating that morphological and syllable structure is

used in the visual processing of various scripts. Differences in the way people

read morphologically complex words suggests that they do rely on morphological

cues. That phonological processing involves structures other than phonemic

suggests that readers extract a variety of phonological cues. Further, there is

considerable evidence that readers interpret orthographic regularities -- statistical

information about recurring symbol patterns even when they do not directly

correspond to linguistic units.

In examining the relationship between the various levels of linguistic

representation, orthographic regularities, and word recognition, the aim is to gain

insight into how these multiple processes affect reading and reading acquisition. I

suggest that commonalities across scripts are likely to be a result of human

preferences and constraints on processing visual language. Hence, orthographies

themselves are evidence of how people learn to read and such evidence can and

should inform reading theory.

Principles of Implicit Learning

The larger objective of this study is to bring one major implication of the

preceding discussion to bear on reading theory. If, as I am arguing, the decoding

process involves decoding multiple levels of structure, then there is much

information readers learn in addition to what they are explicitly taught and indeed,

even before they are explicitly taught decoding strategies. Hence, in addition to

(sometimes in spite of) what readers are taught, there is much that they must

implicitly induce in order to learn to decode an orthographic system and learn to

read. If various levels of linguistic structure are demonstrably functional in

recognizing words, then learners must analyze much of the structure present in an

orthography without explicit instruction or explicit knowledge of the

representational relationship on which the decoding process is based. So, as

readers mature, they do not proceed on the assumption that the orthography is

phonemically valid, but use the various levels of structure at once to identify lexical

items. Hence, much of learning to read can be thought of as implicitlearning. In

short, most of us leam to read despite the considerable amount of information that

we are not taught.

The term implicit learning has been used to characterize the manner in which

people come to apprehend complex structured information from a stimulus

environment, where acquisition occurs largely independently from explicit

knowledge of how you are learning and what exactly you have learned. Implicit

lsirri.rg ilh_ h, run., .r lli r ..Ji.' information about the world in an

unconscious, non-rcflective way. This process is conceived as a by-product of the

application of attention to relevant rule-governed structures in the environment, be it

natural language, socially prescribed behaviors or of complex economic systems

(Winter and Reber 1994). Hence, implicit learning is a general inductive process

that derives information about patterned relationships in a stimulus environment,

and results in an abstract4 and tacit knowledge base (Reber 1967, 1976, 1993;

Berry and Dienes 1993; Berry and Broadbent 1984, 1988; Mathews et al. 1989;

Mathews 1991; Cleeremans 1993).

4 The issue of abstractness is a debated one which will be discussed further in chapter 5.

This is not to suggest that complex structures cannot be learned through

conscious analytical strategies, but rather that implicit operations can be engaged in

the analysis of structured patterns, even when these regularities have not been

consciously detected (Winter and Reber 1994). Furthermore, implicit mechanisms

may be 'better at' learning structures that are too complex for conscious analytical

strategies. That is, explicit processes are resource intensive, requiring explicit

memory, conscious attention, and overtly controlled hypothesis testing. On the

other hand, implicit mechanisms are automatic, unconscious and as such, less

resource consuming, thus freeing up explicit mechanisms for higher order


Various implicit learning paradigms have been designed to experimentally

explore the process by which people acquire complex knowledge about the world,

largely independent of conscious attempts to do so. Artificialgrammar learning is

the most widely replicated paradigm. In theses experiments, subjects learn

complex rule governed stimuli generated by a synthetic, semantic free, Markovian,

finite-state grammar, which are too complex to be learned in one afternoon in the

laboratory (Reber 1993). In the typical study, subjects memorize strings of letters

in the synthetic language and are later tested for their knowledge of the rules of its

grammar by being asked to make decisions concerning the well-formedness of

novel strings of letters, the associations of which are determined by arbitrary rules.

The empirical findings indicate that, when the underlying structure of a stimulus

network is highly abstract and complex, subjects become quite adept at judging the

grammaticality of letter strings that they have never seen before. The implication is

that subjects are learning abstract regularities of the stimulus network.

Although implicit learning experiments attempt to isolate implicit processes,

the literature suggests that both implicit and explicit mechanisms are involved. In

general, implicit and explicit processes should be viewed as cooperative processes

with interactive components. Furthermore, these processes may vary in their

accessibility to conscious awareness, degree of generality, and degree to which

these processes facilitate each other. Implicit and explicit mechanisms operate in

parallel, with the resultant knowledge being a synthesis of the two levels of

analysis. Chapter five will review theories and empirical research with implicit

learning phenomena, with the intention of making a parallel with reading


Implicit Learning and Reading Acquisition:
Processing Commonalties

It is widely believed that the prominent characteristic of reading acquisition

that distinguishes it from learning to speak is that reading requires conscious

analysis of how the written code maps on to spoken language. While this

observation is accurate to some degree, it does not adequately capture the

complexities of learning to read. I .,- L. I rh.lT what we know about written

language processing, the commonalities among writing systems, combined with the

insights of implicit learning research, permit a reconceptualization of the process of

reading acquisition. I believe that insight into the interaction between explicit and

implicit learning operations, provide a framework for this conceptualization.

Drawing on the specifics of prior chapters, in chapter six, I will argue that many

of the circumstances of learning to read parallel that of an implicit learning situation.

My goal is to suggest a framework whereby learning to read is characterized by an

interface of implicit and explicit learning processes, which operate concurrently in

the interest of efficient processing.

Practical and Theoretical Implications
Educators rely on the fact that learners bring crucial implicitly acquired

knowledge to the educational stage. Particularly young children acquire most of

what they know without intention to learn. Yet, at the same time, formal education

typically proceeds from the assumption that learning complex knowledge is an

explicit, goal-directed process where information is presented in more-or-less

factual form, and subsequently put to use. It is interesting that while formal

educational methods rely on implicitly acquired knowledge,just how implicit

learning plays a role in formal, explicit instruction is rarely examined. Instead,

instruction is focused on learning that is intentional and reasoned. Consequently,

formal education relies on the dynamic of implicit and explicit learning, yet tends to

restrict its scope to issues of explicit instruction. It is distinctively the interface of

implicit/explicit learning and knowledge, and the tendency of educators and

researchers to disregard it, that has resulted in many of reading's contentious

instructional issues. Moreover, the reconceptualization of reading presented here

has cross-disciplinary theoretical implications. Chapter seven discusses how and

why the proposed framework bears on both specific practical concerns and general

theoretical questions.

Concluding Remarks
This study is predicated on the assumption that the literatures in reading,

the linguistic study of writing systems, and implicit learning can be usefully

integrated in order to reconceptualize the fundamentals of learning to read. I believe

that there is considerable insight to be gained from this synthesis of literatures that

are not often joined. In turn, an analysis of reading acquisition from this

perspective expands the literature in each of these domains.

Much of the existing research on learning to read is dominated by the

following assumptions: alphabetic letters represent phonemes and therefore, the

most essential part of learning to read is mastery of letter-phoneme relationships.

Also, written language maps on to spoken language in ways that must be

consciously understood. I intend to show that both of these premises lead to an

inadequate characterization of reading, and so are in need of revision.

The literature in implicit learning can be effectively appropriated in order to

advance our understanding of reading acquisition. Within the disciplines of

psychology and computer science, there is abundant evidence that implicit

processes are efficient, principled cognitive functions that are very good at

processing certain types of information. On the basis of processing evidence and

commonalties across scripts, I suggest that decoding orthographic representation is

one of these types.

A natural proclivity for processing efficiency is the basis for the proposed

conceptualization of the learning strategies used in reading acquisition. The

capacity to interpret various aspects of linguistic structure simultaneously facilitates

decoding, making the reading process arguably more efficient than it would

otherwise be. Similarly, employing implicit and explicit processes to a learning

task is a matter of allocating resources where, and in a manner, that they can be

most effectively used. I believe that an understanding the dynamics of such

learning strategies can be used to advance discussions of reading in innovative



Theories of reading acquisition which attempt to specify how beginning

readers learn to recognize words are numerous and elaborate. Teaching practice

and educational policy are, to some degree, influenced by ongoing research in this

area. This chapter will review the dominant trends' in reading research and discuss

how they have influenced teaching practice and policy. This chapter will discuss

the strengths and weaknesses of various reading theories and instructional

philosophies in order to demonstrate that each captures some essential property of,

or insight into, the reading acquisition process. It is also my intention to point out

that despite the quantity of reading literature and the insights it provides, the scope

of reading research has been sufficiently narrow such that it has not given due

consideration to an understanding of the nature of human learning and language.

Bearing in mind that understanding the reading process must include an

understanding of language and the basic operating characteristics of the human

brain, this discussion is prelude to an integration of theoretical considerations which

are largely absent from the reading literature.

Phonic Theories of Reading
Phonic theories are premised on the idea that alphabetic orthographies

represent phonology systematically. Accordingly, word recognition is a matter of

translating the orthography's phonological code and phonological processing is

5 This is by no means meant to be a comprehensive review of the vast amount of existing
research. Rather, this is a general account of the dominant theoretical trends and their
contributions to understanding reading acquisition.

presumed to be the most important mechanism in learning to read with alphabetic

scripts. Basically, theories of phonologicalrecoding suggest that children learn to

read by converting letters into sounds, applying phoneme-grapheme

correspondence rules, and then recognizing the word from its pronunciation

(Liberman et al. 1977, Gough 1972). These relationships are encoded into the

lexicon of beginning readers.

Since alphabetic systems are thought to represent phonological systems, it

is believed that reading acquisition hinges on what has been widely called

phonological awareness -- generally, the ability to break words up into their

constituent phonological units. There is some confusion about this term in the

literature6. It is sometimes used to mean conscious analysis of speech at the level of

phonemic segments. Other times it is used to mean a general perceptiveness of the

units of speech, not necessarily phonemic. Empirical evidence of phonological

awareness is taken from experiments that involve tasks of varying complexity and

difficulty among children and adults. The easier tasks involve testing for

knowledge of more perceptually salient aspects of phonology such as syllables.

Evidence of syllable awareness comes from rhyme and alliteration tasks and

syllable counting. Preschool children (Maclean et al. 1987; Treiman and Zukowski

1991) as well as nonliterate adults (Morais et al. 1986) have been shown to be

successful with syllable awareness tasks. Phonemic awareness tasks try to assess a

persons knowledge of segments, which are much less perceptually salient speech

units. These tasks, such as phoneme counting or tapping, phoneme deletion or

addition, are much more difficult. Usually, children begin to acquire this skill as

6Stanovich (1992) suggests that the generic term phonological sensitivity be used'to cover the set
of processing constructs being tapped by the various tasks used in research. Phonological
sensitivity should be viewed as a continuum ranging from deep sensitivity to shallow sensitivity,
The deep sensitivity would be characterized both by the ability to give conscious, explicit reports
of speech units and also the ability to detect less perceptually salient units like phonemic
segments. Hence, phoneme segmentation would fall at the deep end of the continuum while
syllable counting would be at the shallow end.

they learn to read and spell. However, even very literate adults are not uniformly

successful with some phoneme awareness tasks (Scholes and Willis 1990).

Indeed, there is strong correlational evidence that some type of phonological

awareness is a key factor in learning to read with alphabetic scripts. However, it

is phonemic/segmental awareness that is most commonly associated with beginning

reading success (Adams 1990; Byrne and Feilding-Barnsley 1989, 1990,1993;

Ehri and Wilce 1980; Ehri 1994; Iverson and Tumner 1993; Morais et al. 1986;

Stanovich, Cunningham, and Cramer 1984). The significance of this correlation

has been widely cited to support phonics instruction as a pedagogical strategy, at

least among beginning readers. Yet, even this strong connection between reading

and phonological awareness appears to be only part of the story. Although children

with low phonemic awareness skills are more likely to lag behind in beginning

reading, not all children with good phonemic segmentation abilities become good

readers. Hence, it appears that phonemic awareness is not the only condition

necessary for efficient early reading (Juel, Griffith and Gough 1986).

While it is evident that beginning reading involves phonological processes,

phonic theories of reading leave many apparent components of reading acquisition

unclarified. Perhaps the most obvious oversight of phonological recoding theories

is that it is assumed that alphabets represent phonemic segments, and that the

visual-phonological code is based on letter-sound translations. This view has been

supported by the positive correlation between phonemic awareness and reading.

Yet, at the same time, this view has been very problematic for writing systems like

English, for which letter-sound translations involve a complex of rules and

numerous "exceptions." This fact has led some researchers to reconsider the

assumption that alphabetic representation is not strictly phonemic. For instance,

Goswami (1986, 1988), Treiman (1992), and others, have found that readers are

sensitive to syllable structure alphabetic scripts. In fact, if you consider the

phonological representation in English script to include word analysis into onset

and rime, the number of phonemic rules and exceptions are dramatically reduced

(Treiman 1992). Following this line of reasoning, we must consider other levels of

representation possible with alphabetic orthographies: morphophonemic,

morphemic, etc.

Another issue of some difficulty to phonological recoding theories is that the

extent of phonological awareness necessary to facilitate recoding is unclear.

Moreover, it is unclear whether or to what degree phonological awareness need be

explicitly taught. It is often argued that phonemic awareness only develops with

explicit instruction or acquisition of an alphabetic script (Adams 1990), and that the

"alphabetic principle" must be consciously understood to be used properly

(Liberman and Liberman 1992). This premise has provided a strong rationale for

phonics instruction. Accordingly, the more aware readers are of the phonemic

correlation, the more reading acquisition should be facilitated. Yet, even for those

who endorse this view, the degree of phonemic awareness necessary to facilitate

beginning reading, and the degree of instruction necessary to facilitate phonemic

awareness, are matters of debate. For instance, some researchers suggest that

phonemic awareness can develop spontaneously in learning to read and spell a large

number of words. In this view, readers implicitly internalize how sound maps on

to printed words. Hence, phonological awareness is consequent to reading,

regardless of instructional method (Gough and Hillinger 1980; Gough, Juel and

Griffith 1992). Also, many researchers regard phonemic awareness as much a

consequent as a facilitator to learning to read (Bradley and Bryant 1983; Ehri 1984,

1994; Mann 1987; Perfetti 1985; Perfetti et al. 1987; Morals et al. 1986; Prakash et

al. 1994). This bi-directional language awareness phenomena makes it difficult to

assess the degree to which phonemic awareness is necessary for beginning reading.

Finally, there is the issue that successful readers who are taught to recode

letters to phonemes use this strategy for only a brief period. Beginning letter-by-

letter readers decode words at the rate of 1 letter per second. Skilled readers

process approximately 100 letters per second This means that, with practice,

recoding becomes an automatic process a permanent set of associative

connections in long-term memory that develops with practice (Schneider and

Shiffrin 1977). What is problematic in phonological recoding theories is that it is

unclear whether these automatic connections operate at the phonemic level or at

some other level, such as the word. If automatic connections are to word-sized

units, then phonological processing may be may be used only in the very early

stages of reading acquisition. This issue, along with others discussed above have

led to competing theories of reading.

Whole- Word Theories
An alternative view of reading is that the representational system is

lexical/semantic. Readers learn associations between visual forms of whole words

and their semantic referents. In sight reading, or whole-word reading theories,

words are stored in memory as visual gestalts as a result of much practice reading

the words (Smith 1972). Readers recognize sight words without attending to the

individual letters or segmenting the words into their component sounds. This

theory presumes that even children who are explicitly taught phonemic recoding

rapidly abandon that strategy in favor of identifying whole words.

One source of evidence taken in support of whole-word theories comes

from research on eye movements in reading. Readers make a series of rapid eye

movements called saccades, separated by fixational pauses that last about 100-250

msec each. Virtually all new information is extracted from the text during these

fixational pauses. The function of the saccade is to move the eyes rapidly from one

position to another, with the average saccade length being about 7-9 characters, or a

bit over one word (Rayner and Pollatsek 1987: 239). Additionally, experiments

using a tachistope -- equipment that can present information to the eyes for a very

brief period -- have shown that readers can identify words as easily and quickly as

random letters (Kolers and Katzman 1966; Newman, 1966 in Smith 1988). It can

be argued that if words areju i J,- a. ., I..i ni ini .11 letters, then readers must

perceive words as units,just as they do letters. Also, if words can be identified just

as quickly as letters, then recognition need not involve letter by letter analysis.

Further evidence for the sight reading perspective comes from studies

which conclude that phonological recoding has limited application to learning the

graphic representations of words because it is a strategy employed only in short-

term memory. There is evidence that phonic recoding occurs only if the

orthographic material is to be temporarily remembered (Mattingly 1984), but that

long-term memory is semantically structured. Therefore, words must be

represented in memory as whole units.

Whole-word theories do reflect the intuitive notion that we read words

rather than strings of letters, however, they also have some apparent weaknesses.

Evidence collected to support this view does not explain how words are recognized

as wholes. Nor does it explain what words "look like" to the reader. That is, if

there is no reliance on letters, then how are word shapes are defined? Furthermore,

there is abundant empirical evidence that phonological processes of various types

are involved in reading for both children and adults.

Dual-Route Theories
Despite the inadequacies of whole-word theories, and despite the evidence

that reading involves some kind of phonological processing, everyone agrees that

skilled readers process written language much too fast to be performing serial letter-

to-sound translations. Dual-route theories have attempted to accommodate these

facts by integrating both phonological processing and sight-word analysis (Barron

and Baron 1977; Baron 1977; Coltheart et al. 1977; Bryant and Bradley 1980; Frith

1979). These theories suggest that readers can read words using one of two routes.

The lexical route leads directly into the lexicon and is accessed semantically. The

phonic route leads indirectly to the lexicon, with readers first applying letter sound

translation. Most dual-route theories posit that these routes operate independently.

Or, if both routes operate in parallel, the fastest process will yield recognition. In

either case, the pathways are separate, not connected. For words that are

previously learned and stored in memory, the visual-semantic route is faster

because it does not require phonological mediation. For unfamiliar or difficult

words, the phonological route would be engaged first.

When words are recognized via the visual-semantic route, their visual forms

are associated with the word's meaning which has been stored in memory. The

visual form itself may be recognized in terms of spatial cues that may take the form

of permissible orthographic sequences, sequence length, boundary letters, etc. The

associations are made through repeated contact with words until they are stored in

memory. Once stored in memory, word recognition is immediate, via direct access.

However, these visual sequences are word specific and they do not represent

linguistic structure below the word level. This means that any orthographic

systematicity within an individual word is not relevant to identifying other words.

The phonological route converts words into their pronunciations by

applying recoding rules. The pronunciations are then used to access the

semantically stored words. Although recoding improves with practice, it is never

as fast as whole-word access. Further, the letter-to-sound correspondences for

particular words are not part of the reader's memory representation for those

words. For those who have been taught to read by phoneme-letter analysis,

phonological recoding is primarily an initial strategy that is bypassed when they

have developed automaticity in word identification. Hence, when the words

readers have seen many times before become stored in memory, phonological

recoding is unnecessary and not used (Gough 1984; Barron and Baron 1977;

Singer 1984). In fact, the difference between good and poor readers is often a

question of moving beyond letter by letter analysis and developing automaticity in

word recognition. For skilled readers, phonological recoding is thought to be used

mainly for unfamiliar or difficult or nonsense words that have not been stored in


In effect, dual-route theories posit that readers will use both the semantic

and phonological routes to access words regardless of how they are taught to read.

Beginning readers who have learned to read words by sight start out memorizing

whole-words later add phonological skills to the reading task. Phonics trained

readers may have superior recoding skills at first but whole-word readers also learn

phonological relationships as their written vocabulary grows. Research has

indicated that readers who are not taught phonics rules eventually induce letter-

sound relationships implicitly as they practice reading, and as they learn to spell

(Gough and Hillinger 1980; Ehri and Wilce 1987; Gough, Juel and Griffith 1992;

Perfetti 1992; Treiman 1993).

The dual-route models capture the notion that letter-by letter analysis is not

always prerequisite to word recognition. It also captures the fact that recoding is a

more time consuming process than directly accessing whole words. Yet, one

fundamental difficulty with dual-route theories (also with whole-word theories) is

that sight-word recognition does not rely on systematic patterning across words.

Consistent phonological and orthographic patterns do not influence the direct access

of words because each word is stored as an individual unit. Since a basic fact of

human cognition is that we exploit the structure available to us, it is highly unlikely

that readers would or could overlook patterned structure in the stimuli.

Another basic problem is that the semantic and phonological pathways are

not interactive. Except for difficult or unfamiliar words, the phonological pathway

should be bypassed completely for skilled readers. Yet, skilled readers have been

found to be sensitive to the phonological constituents of words (Treiman and

Zukowski 1988; Feldman 1987; Treiman, Goswami and Bruck 1990). Ehri (1992)

addresses this problem by suggesting that there are connections between the visual

(whole word) route and the phonological route. In this theory there are connections

in memory between a word, its sequence of letters in spelling, and its phonemes in

pronunciation. Thus, learning new words involves making systematic

phonological connections and storing the phonological information in memory with

the semantic information.

A third difficulty has to do with specifying only two pathways: phonemic

and semantic. Phonological recoding is thought to rely on a system of rules

whereby a visual word is translated into its phonological code. However, it has

proved immensely difficult to specify the rules governing phoneme-grapheme

correspondences. This undoubtedly has to do with the fact that orthographic

systems are more than just a set of rules that map phonemes to letters. Yet,

linguistic associations above the phonemic level, and below the word level, are not

part of the word representation in dual-route theories. I will address at length the

idea that orthographies encode several types of structure beyond the phoneme in

chapters 3 and 4.

Interacti ve Activation Theories
Interactive parallel activation theories of reading, which have developed

from interactive theories of speech perception, try to adjust some of the

shortcomings of dual-route theories (McClelland and Rumelhart 1981; Seidenberg

1985, 1987). One innovation of this approach to reading is that it emphasizes that

word recognition is a single interactive process. Instead of alternate recognition

pathways, these models assume that access of phonological information is an

automatic consequence of word recognition. Recognition is initiated by extracting

visual information from the input where readers do rely on systematic patenting

across words. As salient orthographic units are recognized, they activate

phonological representations. Consequently, phonological access lags behind

visual analysis. Direct access occurs when there is enough orthographic

information to recognize words prior to phonological access. This is the case for a

large number of common words. However, phonology is still activated for words

retrieved via direct access, though it is after recognition, making phonology a

necessary consequence of recognition .

Interactive processing theories posit that readers do not rely on a set of

explicit rules to make these systematic associations (e.g., rules of phoneme-

grapheme translation). Alternatively, specifying the processing system and the

knowledge encoded therein is a matter of specifying this pattern of connectivity

among the processing units. There is no storage bin of lexical items, the "lexicon"

is in the connections themselves; a map(s) of associations. The structural

representations are phonological and semantic. The phonological representations

are orthographic patterns of associations. The semantic representations are the

meanings of units. The connections between the structural representations will

form "maps," which will have similar items located near each other. The maps will

have densely interconnected associative connections. Output representation is

obtained from the weight vector of the most highly active unit. Maps and the

connections between them are organized simultaneously, based on examples of co-

occurring symbols and meanings. (Seidenberg 1989, 1992, Seidenberg and

McCLelland 1989).

The way an interactive activation model applies to reading is as follows.

The template (visual image) broadcasts activation to words in the lexicon

simultaneously, with each feature of the template contributing a small amount of

activation to each word that has that same feature. Many words will receive some

amount of activation, but the word that has more than any other is eventually

accessed. High frequency items need less activation than low frequency ones to be

accessed. Each word unit stands for a particular word and has connections to each

letter in the word. While there are no separate units for regularly occurring letter

sequences, all systematically occurring sequences are represented as maps of

associations, and hence, are an intrinsic part of the lexicon. The structural

representations are commonly thought to be phonological and semantic.

Theoretically, however, since activation will form associative maps for all

systematic, reoccurring sequences, morphemic and morphophonemic patterns (if

present in the visual stimulus) would be associated via the same process.

Because interactive activation models rely on identifying letter sequences

that recur in a systematic way, what are sometimes called analogy models fall into

this theoretical grouping. That is, people read words by activating known words

that have letter sequences in common with the stimulus. For instance, phonological

representations are associated with chunks of letter strings. So, an unfamiliar letter

string like vaid will activate a set of familiar words likepaid, maid, said and

vain. Various analogy models have been suggested. Glushko (1979, 1981),

Goswami (1986,1988), Goswami and Bryant (1992), Treiman (1992) suggest that

letter chunks represent syllabic constituents. So, when analyzing an unknown

word (or nonsense word) vaid, there is a preference for readers to segment the

word into v aid, the syllabic constituents of onset and rime, rather than vai d. In

this view, letter chunks that form linguistic units are more important than those that

do not. This is true because humans are predisposed to segment the visual stimulus

into the natural constituents of speech. Correspondingly, I would add also that this

is true because scripts themselves represent linguistic units.

The interactive models of Seidenberg and McClelland (1989) and

Seidenberg (1987, 1989, 1992) are similar to the analogy models above, except for

one point. The formation of letter chunks is based simply on distributional letter

properties, not on any perceptual preference for organizing the stimulus into

linguistic units. Sequence chunks are based on the frequency with which strings

occur in the visual stimulus and the consistency of their pronunciation only.

Syllable and morphological structure is not directly represented but rather falls out

of distributional probabilities of letter sequences. Accordingly, the preference for

chunking the orthographic sequence vaid as v aid over vai d is simply because

the former occurs more frequently with consistent pronunciation. Hence, the

distributional properties of letter patterns will activate certain sequences more than

others. Letter sequences will correspond to linguistic units below the word level

but not necessarily because people are naturally disposed to segment written

language into speech components like onset and rime.

I believe that the Seidenberg and McClelland view is trying to emphasize

that people will extract whatever regularities are present in the script It is true, for

example, that readers will extract particular linguistic structures like onsets and

rimes only where they are consistently represented. Yet, there is a difficulty with

suggesting that choosing particular sequences over others is not a result of

perceptual preferences. Such a suggestion obviates the reason why distributional

letter properties reflect linguistic structures and cues in the first place. Following

this logic, there should be no reason that a syllabic organization into onset and rime

is preferable to chunking arbitrary letter patterns (e.g. there would be no reason for

v -aid to occur with more consistent pronunciation than vai- d, or va -id).

Clearly, however, certain sequences occur more frequently and with more

consistent pronunciation than others because they reflect some property of the

spoken language. Hence, readers are able make associations between orthographic

form and their linguistic knowledge. The fact that orthographies all represent

linguistic structures suggests that there is a processing advantage for readers to

chunk the visual stimulus into units that correspond to those of spoken language. If

there were no processing advantage for chunking the visual stimulus into linguistic

units, we should be just as likely to see orthographic systems, say an alphabet,

whose distributional letter patterns were largely arbitrary. However, this is never

the case.

Reading and Morphology
Because most reading theories have focused on the phonological and lexical

aspects of visual language processing, attention to the morphological properties of

reading has been lacking. This is due in large part to the fact that most research

deals with alphabetic systems, English in particular, which are presumed to involve

phonological and/or lexical processing. The one exception to this rule is interactive

activation theories, which (at least in principle) can include morphological

processing in reading, However, there have been empirical studies which have led

some researchers to conclude that morphological structure is indeed used to

recognize words (Laudanna et al. 1989; Feldman 1987; Benton and Frost 1995;

Chialant and Caramazza 1995; Stoltz and Feldman 1995). I will review this

research in chapter 4.

Teaching Philosophies:
Code versus Whole Language Approaches to Reading Acquisition

In the above review of reading theories, I have mainly addressed issues of

how visual language is processed by the learner/reader. Reading research, to some

degree, has had an impact on educational policy. In this section, I will discuss the

dominant schools of thought which underlie teaching methods, and how those

methods are informed by reading research. I do not here presume to delineate all

the types of early reading teaching strategies, which undoubtedly are many and

eclectic. However, there are distinct schools of thought which make essentially

different assumptions about the learning process. As a matter of comparison, I will

present the poles of these views.

The first is what I will call the code 7 approach. In this view, learning to

read involves interpreting a fairly complex visual system and analyzing how that

system maps onto natural language. The code approach to reading has been directly

influenced by theories of phonological recoding and research which points to a

correlation between phonological awareness and reading success. Hence, the

principal goal is teaching beginning readers the "alphabetic principle" -- that letters

signify sounds and sounds combine to form words. Phonics, the study of the

relationship between the sounds in spoken words and the letters in written words,

is a standard tool for the code approach to reading. A distinguishing characteristic

of the code approach is that it assumes that the code must be explicitly taught to new

readers. Accordingly, learning the code involves conscious, directed analysis of

letter-sound relations and conscious understanding of the rules that govern them.

The other teaching philosophy which enjoys wide support is called whole-

languages. The whole-language approach to teaching reading is predicated on the

belief that comprehension and communication are pivotal to the learning process.

Hence, children should work with meaningful materials, and that they read and

write about topics that interest them. In this view (Goodman 1986), children are

7 This term, popularized by Chall 1967, has been widely used as a general label for a type of
reading instruction that emphasizes that written language is a code that children must learn in order
to read.
8 This is not to be confused with the whole-word theory of word recognition, which is a theory of
how people process written language. Whole-language is a philosophy of teaching, whose scope
is much broader than simply beginning reading. I will however, discuss it as a teaching method
only in regard to early reading acquisition.

good at understanding and producing language when they need to express

themselves. Hence, given an environment where the input is meaningful and

purposeful, children should be just as good at learning written language as they are

at learning spoken language. Meaningful materials include children's literature or

stories generated by the children themselves. A corollary is that children should

never be forced to read parts of text (words) in isolation or out of context since this

would distort natural language patterns and inhibit the communication of meaning.

Advocates of whole-language do not endorse teaching the relation between sounds

and letters or memorization of spelling rules. In fact, in this view, the teaching of

phonics could be harmful because it distorts language into small, abstract pieces,

which can interfere with the goal of communication. Instead, children should

figure out sound-letter relationships on their own from their experience with printed

and spoken words. Analysis of sounds in relation to print should only be

encouraged when the teacher observes children have already made the connections

on their own. Hence, each child is an "autonomous" learner, while teachers and

parents serve as "guides" and "facilitators," making sure the children are presented

with meaningful text.

Strengths and Weaknesses of the Code Approach

The code emphasis is bolstered by a great deal of experimental data which

point to a strong and consistent relationship between phonological awareness and

reading. Yet, even though there is little doubt that phonological awareness plays a

role in reading, it is far less clear how children learn the relationships between

sound and print. Whether children must be taught such relationships or figure them

out on their own is far from resolved. While it has been shown that children can

benefit from explicit phonemic awareness training and instruction in letter-sound

relationships (Adams 1990; Bradley and Bryant 1983; Byrne and Feilding-

Barnsley 1990, Byrne 1992), it is also true that normal readers become

phonologically aware as they learn to read and spell, regardless of the type of

instruction they receive (Gough and Hillinger 1980; Gough, Juel and Griffith 1990;

Perfetti et al 1987; Treiman 1993). Moreover, since reading and phonological

awareness seem to reciprocally facilitate each other, this is a difficult relationship to

sort out.

One incontrovertible fact in support of the code emphasis is that written

language, across scripts and script types, does indeed systematically represent

spoken language. It is a code for speech. Of course, successfully teaching "the

code" to beginning readers relies on a clear understanding of what the code is. It is

this understanding, or rather, lack of it, which I will argue often presents

difficulties. Presently, the code approach is primarily influenced by theories of

phonemic recoding. Consider English, where letter-to-phoneme correspondences

do not account for many words, or account for them only partially. The rules of

phonics are many and complex and have many exceptions. A careful consideration

of the facts of English orthography suggests that letter-to-phoneme translations, and

the rules that govern them, are only part of the code, only a fraction of the system

that the reader must internalize. I suggest that interpreting the orthographic code

involves deciphering linguistic patterns beyond the phoneme: syllable constituents,

morphophonemic patterns, lexical and functional morphemes, and so on, and that

this is the rule across scripts and script types.

An additional and very important point is that, if we accept that the code is

only partly a phonemic one, then the many phonics rules which try to force the code

to be strictly phonemic, are not likely to be the system that the reader actually

learns. Consider the "silent e" rule which makes vowels "long". By applying this

rule the reader knows that code is pronounced /cod/ as opposed to cod which is

pronounced /cad/. Yet it is possible, instead, that English orthographic

interpretation relies on some kind of analogy system whereby readers internalize the

pronunciations of onsets and rimes: c as /k/, ode as /od/, and od as /ad/. If this is

true, teachers who faithfully instruct their students in the rules of phonics may be

teaching a system that is different from the one the reader actually uses.

A crucial assumption of the code philosophy is that literacy acquisition,

unlike first language acquisition, does not proceed automatically but requires

explicit instruction. In this view, first language acquisition and literacy acquisition

follow basically different developmental routes. Spoken language has evolved as a

natural process of the human brain. Evidence of this is clear: grammatical

competence in spoken language is naturally acquired. It begins in infancy and is

largely completed by school age. Despite the fact that first language input is

incomplete and often irregular and indirect, the pattern and rate of acquisition

appears to be universal for normal children. This holds despite differences in

language form, and a fabric of social and cultural factors. Furthermore, acquisition

occurs without conscious effort or awareness, resulting in an impici or tacit

knowledge base.

Written language, on the other hand, is only a representation of natural

language according to the code view. It maps onto the spoken system in a way that

must be consciously understood if it is to be used properly (Liberman and Liberman

1992). Since teaching intervention is thought to be the critical factor in learning to

read, advocates of the code view endorse explicit learning/teaching methods like

phonics. Again this view assumes that such instruction will make clear all the

associations represented in an orthographic code. Yet, if it is true that the written

language code is not strictly phonemic, then even code-trained readers learn

considerably more than they are taught. It would seem then, that learning the

orthographic code does not entirely oblige explicit teaching and conscious analysis.

It is possible instead that learners implicitly induce many of the regularities of the

orthographic system.

Strengths and Weakness of Whole Language

The whole-language philosophy is intuitively appealing and widely

supported by teachers. Some of its basic tenets, such as the importance of print

exposure for preschool children, have influenced educational policy. Indeed, there

has been shown to be a positive correlation between exposure to print and reading

success (Cunningham and Stanovich 1990). The greater children's experience with

books and stories, the more they are aware of the print-language connection.

Whole-language captures the intuition that readers learn much more than they are

taught. Further, most readers are not wholly conscious of what they know about

the system. For example, it is difficult for most people to articulate how the written

system operates.

Although whole-language enjoys wide support, it also has received some

strong criticism. The essential (and I believe well-founded) criticism is that whole-

language does not identify the principles or mechanisms of learning upon which it

is based. Because this view assumes that readers make orthographic connections

spontaneously, it is unconcerned with the specifics of what the code is and how it

maps onto natural language. It does not address the similarities or differences

between spoken and written language since it presumes they are acquired the same

way. It does not address how children are able to learn so much in the absence of

direct instruction. What are the mechanisms that cause children to make the

connections between print and speech for themselves? This view suggests that they

are the same mechanisms as those that govern first language acquisition, though it

does not discuss what those are specifically. Among theories of language

acquisition, the most influential and widely received in linguistics and related fields

is that, as a result of evolutionary adaptation, humans are biologically predisposed

to learn language. Further, learning is constrained by basic principles of language

structure that are also biologically fixed. Whole-language advocates do not even try

to explain how the acquisition of written language could possibly parallel this type

of development. Certainly, nobody has claimed that written language structures are

also an innate component of the language faculty. Nor does whole-language posit

an alternative theory of language acquisition in which it would be possible that the

spoken and written forms could take the same developmental path. Additionally,

the whole-language philosophy does not employ reading theories or empirical

studies of word recognition and linguistic awareness to support its claims. Neither

have advocates of whole-language initiated research which would help identify

(within this framework) what it is that some children learn that make them good

readers, and what other children miss that become poor readers.

Summary and Concluding Remarks
Each individual theory of reading acquisition captures something important

and distinctive about the reading process. Phonic theories are correctly based on

the notion that reading involves phonological processing and that some kind of

emergent linguistic awareness is part of the learning process. Whole-word theories

rightly address the issue that normal readers identify words much too fast for letter-

by-letter analysis. Dual-route theories present the idea that there may be more than

one recognition pathway. This captures the idea that word identification via

phonological recoding is slow and there must be an alternative, more direct

(semantic) route to lexical access. Interactive activation theories reconceptualize the

notion that both semantic and phonological information is used to identify words.

They posit not that there are separate routes to word recognition, but that words

have both phonological and semantic representations which can be simultaneously

activated. Further, the phonological component does not rely on conscious rules of

phoneme-letter translation. Rather, phonological representations are an intrinsic

result of letter pattern recognition. Moreover, this theory can accommodate the

possibility of other levels of representation inherent in the orthographic code.

Likewise, both the code and whole-language teaching philosophies reflect

important generalizations about reading acquisition. The code approach, which is

founded not only on practical observation but a good deal of research, rightly

asserts that learning to read hinges on deciphering and internalizing the orthographic

system. Whole-language captures the idea, which is based on much practical

observation, that readers can naturally induce the principles of the orthographic

system (whatever they may be), without a lot of explicit instruction or conscious


It would seem that the code and whole-language views make opposing

assumptions about learning. Yet, I believe that basic insights of both views can be

usefully appropriated towards a theory of reading acquisition. The fact that written

language is a code for spoken language is well supported by research and widely

accepted (whole-language does not dispute this). Researchers continue to grapple

with the issues of what orthographic systems represent about spoken language and

how humans process this information. I believe that exploring a basic insight of

whole-language -- that readers can and do implicitly induce orthographic properties

-- can further the aim of discovering the what and the how of orthographic


The whole-language view endorses a kind of implicit learning process for

reading without bothering to specify what it is or how it works. This is due partly

to the fact that the implicit learning mechanisms involved are not well understood

and are outside the scope of traditional education. Yet, within other disciplines

such as psychology and computer science, there is abundant evidence that implicit


learning processes are efficient, principled cognitive functions that are very good at

processing certain kinds of information. I suggest that decoding orthographic

representation is one of these types. Connecting the principles of implicit learning

with the facts that are known about orthographic systems and processing, can only

further the discussion.


Understanding the process of reading acquisition relies on an understanding

of what an orthographic system is. Many theories of reading acquisition are

premised on the idea that processing visual language is influenced by what is

represented in the orthography. Reading research has relied on the assumption that

the representational relationship between speech and writing is straightforward.

For example, the interest in alphabetic scripts, which has dominated reading

research, has long presumed that alphabets represent phonemic systems. Attesting

to the currency of this assumption is an enormous body of literature which explores

the relationship between letters and words and their connection to phonemes and

words. Despite the fact that English script is known to be "irregular" in its

representation, and some have pointed to orthographic representation of

morphology (Chomsky and Halle 1968; Klima 1972; Read 1983), analyses of

English orthography have been principally concerned with describing letter-to-

phoneme mappings and the rules that govern them.

The idea that the alphabetic systems may encode other than phonemic

linguistic information is relevant to improving our understanding of how visual

language is processed. Yet, the continuing interest in how the phonemic aspects of

the system are learned has overshadowed such concerns and, with a few notable

exceptions, this issue has remained in the background of the literature on reading

acquisition. I believe that the study of writing systems, which specifically concerns

itself with what scripts represent, can inform reading research on this issue. This is

the case not only because reading research should look more closely at

representational relationships within particular scripts, but also because analyses of

the world's scripts reveal striking commonalities among seemingly diverse

systems. The idea that various scripts are very much alike in regard to the manner

in which linguistic information is presented has important implications for language

processing and can significantly advance reading research.

Two relevant observations emerge from a survey of the world scripts. The

first is that orthographies are usually only partially representative of any particular

level of linguistic knowledge. (Chomsky and Halle 1968; Chomsky 1970;

Coulmas 1989; Miller 1994; Daniels 1992, 1996). Even scripts that are thought to

be very consistent omit considerable information, making them incomplete

renderings of a linguistic system. For example, phonological scripts do not depict

all the phonological information present in the spoken language. Segmental scripts

may only partly represent phonemic systems. Korean, a segmental-syllabic script,

can unambiguously represent all the segments of the language but in some cases it

surrenders segmental uniformity in favor of morphological consistency. Many

phonological scripts omit phonemic stress (English, Russian). Some phonological

scripts represent phonemic tone (Vietnamese, Burmese, Thai) but others do not

(Serbo-Croation, Slovene, Lithuanian). Abjads leave out significant segmental

information (vowels) based on morphological predictability. No phonological

scripts give narrow phonetic transcriptions of their languages. Likewise,

logographic scripts supply incomplete morphological information. This is the case

with Chinese, where the morphological component of the characters sometimes

give substantial semantic information and sometimes only a vague hint (Coulmas

1989). Similarly, Sumerian cuneiform, in its early stages represented lexical but

not grammatical morphemes. (Cooper 1996).

The second observation is that orthographies usually represent aspects of

more than one "level" simultaneously. That is, scripts are likely to (partially)

represent some combination of phonemes, syllables, morphemes, and words all at

once. (Gleitman and Rozin 1977; Henderson 1982; Sampson 1985; Coulmas

1989; Daniels 1992, 1996; Miller 1994). Evidence of this is widespread across

scripts. In this chapter, I will discuss how various levels of linguistic information

(segmental, syllabic, morphological, lexical) are included in various scripts, giving

particular attention to how alphabetic scripts can represent information beyond the

phoneme. What emerges from this assessment is that apparently dissimilar scripts

are very much alike with respect to the ways in which they depict linguistic

information. That is, orthographies are likely to partially represent several aspects

of linguistic structure at the same time. Moreover, this discussion prefaces the next

chapter which reviews existing research that indicates that people do use the various

linguistic structures encoded in the orthography in order to process visual language.

A Survey of Script Types

The level of linguistic structure most widely represented among scripts is

the syllable. In addition to syllabaries which overtly represent syllables, both

alphabets and logographies include syllabic information to varying degrees9. As

Daniels (1992) argues, the preference for syllabic representation is likely a result of

the perceptual salience of the syllabic units of speech.

The distinguishing characteristic of a syllabary is that characters denote

particular syllables without reference to the phonemes that comprise them. Usually,

syllabaries represent CV syllables, having been adapted for languages with fairly

simple canonical structures. Syllabary characters almost never (with the exception

of Akkadian (Daniels 1992)) represent VC syllables'0.

9Typological terminology for various scripts is taken from Daniels 1996.
10 Interestingly, syllabaries represent CV's even in languages that have codas and long vowels.
For languages that allow coda's and long vowels, a CV syllabary usually some convention for
dealing with them--separate characters or diacritics. Hence, syllabaries may be moraic in the sense
that they distinguish the peak with post peak elements. However they are still syllabic in the
sense that the elements of the syllable are combined to form syllable units and syllable units are
spatially separated from each other. In addition, there may be perceptual reasons that syllabaries
always represent CV's. Generally the auditory cues present in CV's are more robust than VCs

Examples of a syllabary are employed in Japanese writing: hiragana and

katakana. The kana are complete scripts in that they represent all the syllables in

the language (Faber 1992) and in principle, anything in the language could be

expressed solely via the syllabaries. However, Japanese is typically written in a

mixture of kana and kanji, a logographic script, mainly of Chinese origin. The

kana syllabaries represent CV and CyV combinations. The hiragana, which

consists of 46 characters supplemented by diacritics, is used for grammatical

elements: particles, auxiliary verbs, and the inflectional affixes. The katakana is

used for foreign names, loan words, onomatopoeic and mimetic words,

exclamations and some scientific terminology. The kanji are used mainly to encode

primary lexical categories. Additionally, the kanji characters independently

represent various types of information. Each kanji character may have different

pronunciations. The on-readings are based on the original Chinese pronunciations;

the kun-readings correspond to a Japanese morpheme that corresponds to the

meaning of the character. The same character may stand for several different

morphemes, each with its specific meaning and reading. Also, when the characters

stand for a Chinese morpheme (on-reading), the phonetic component of the

character sometimes give clues to the pronunciation The mixture of these three

scripts in Japanese writing is an obvious example of a mixed phonological and

morphological system. (Shibamoto Smith 1996; Morton and Sasanuma 1984).

While "pure" syllabaries do not encode phonemes, some implicitly encode

some segmental information. The Mycenean (Linear B) script is an example of a

syllabary that represents syllables without explicitly indicating the segments that

comprise them. However, some segmental information is implicitly indicated.

Characters in the Mycenean script represent V, CV and sometime CCV. Except

(Ohala and Kawasaki 1984). There is a more reliable set of place cues in the CV transition than in
the VC transition. Therefore, listeners weight the former more heavily in deciding what they have
heard (Ohala 1990).

vowel initial syllables, Mycenean graphs do not indicate individual segments and

coda consonants are generally ignored. However, onset clusters are indicated by

the graphs for the appropriate consonants plus the nuclear vowel of the syllable.

Hence, there is a dummy copy of the peak vowel. For example, the onset cluster in

the syllable [tri] is indicated by the graphs for the syllables ti + ri. Further, coda

clusters that are possible onsets (according to the sonority scale) are indicated by

writing the sign for the first segment of the cluster. In these cases, according to

graphic conventions, the very final coda consonant is not represented. Miller

(1994) argues that these facts indicate that the Mycenaean system recognizes

segments as isolable units, and also that using it requires implicit knowledge of the

sonority hierarchy.

Vai script, an indigenous writing system of the Vai language of west Africa,

is a syllable/moraic representation. Mora is a prosodic unit, a formalization of

phonological weight, intervening between the phonemic string and the syllable. In

Vai script, the weight of the syllable determines the number of characters that

represent it. The characters represent CV's. Light syllables are indicated by one

CV character. Heavy (two mora) syllables are indicated with two characters. The

only closed syllables are those with velar nasals. This nasal coda is represented

with a separate character. The same nasal character may be syllabic in onset

position. Also, long vowels and diphthongs (two mora peaks) are written with two

characters. The spelling of syllables containing long vowels and diphthongs are

variable. For example [taa] is usually written with the graphs for ta.ha, not ta.a.

Similarly, [lakoa] is written la.ko.wa. The graph for the syllable [a] is reserved for

words that begin with [a]. However, the graph for the syllable [i] can occur both

word initially and elsewhere". Syllables with onset clusters are written with a

single character that represent that particular cluster plus vowel. This is consistent

1 These conventions could contribute to word identification since Vai does not explicitly signal
word boundaries -- there are no spaces between words or other graphic indicators.

with the convention that segmental constituents of the syllable are individually

indicated only when they contribute to syllable weight. Vai has lexical tone which

is not represented in the script, and there is no division between words, though

distributional properties among the characters make these omissions less confusing

then they might be. There is a character to signal the end of a sentence (Singler

1996). Hence, Vai script represents a variety of phonological information:

syllables, syllable weight, coda segments. Yet, it omits considerable information as

well, for example, it does not indicate individual segments within an onset cluster,

phonemic tone or word boundaries.

A neosyllabary12 represents syllables and their constituent segments. In a

neosyllabary, each character indicates a consonant plus a specific vowel. Other

vowels are denoted by a consistent modification of the consonant symbols. The

Indic scripts (Devangari, Punjabi, Bengali and others) and several Southeast Asian

scripts (Burmese, Lao, Tai, Khmer) are examples of this. Brahmi, the ancestor to

many modern scripts of this type, is based on open syllables: V, CV, CCV, etc. V

syllables are indicated by independent signs. CV syllables are represented by the

consonant plus a diacritic sign for the vowel which is attached to the consonant.

Consonant characters without any diacritic vowel are understood to represent the

consonant plus "inherent" default vowel /a/. A syllable consisting of a consonant

cluster is denoted by consonants that are joined in a conjunct character, indicating

the cancellation of the inherent vowel (Salomon 1996). Note that this is different

from a regular syllabary, which has no formal mechanism for denoting the

segmental constituents of the syllable. Neosyllabaries may have conventions for

distinguishing syllabic constituents. For example, coda consonants may be

distinguished from onset consonants, as in Tibetan (van der Kuijp 1996) and

Burmese (Wheatley 1996), and many neosyllabaries distinguish long and short

12 This term which I have taken from Daniels 1992, is taken from Fevrier (1948). The Ethiopic
name for this, alugida, is also used.

vowels (e.g., Brahmi). Further, a neosyllabary may have conventions for

representing tone. Tone may be marked by additional signs as in Thai (Diller

1996), or, as with the Punjabi script Gurmukhi, tone is indicated by the consonants

and vowel signs (Gill 1996). Also, a neosyllabary may indicate word boundaries

or phrase boundaries (e.g., Malayalam (Mohanan 1996)), or it may not (e.g., Lao

and Thai (Diller 1996)).

Korean script, the Hankul, (sometimes called an alpha-syllabary) encodes

syllabic, segmental and featural information. This orthography has a sign for each

phoneme in the language but it is written in syllable blocks. Any syllable must

begin with a consonant sign. To form the peak of the syllable block, the vowel

sign much attach to the side or beneath the consonant, following the principles of

stroke order. When the syllable begins with a vowel, the written syllable begins

with the zero sign, representing a zero consonant. The zero symbol also represents

syllable final /n/, which does not occur syllable initially. Consonant clusters are

denoted by the joining of consonant symbols. The hankul is called a featural

system because the shape of the basic consonants is a graphic representation of the

places of articulation. For example, the sign for /k/ is 7, which denotes the tongue

position for velar articulation. By doubling or adding strokes to the five basic

consonant shapes, one derives the shapes for the aspirates, tense unaspirates,

affricates and so on. Yet, despite its segmental accuracy, modern Hankul does not

consistently represent phonemes in the context of most morphophonemic

alternations in favor of keeping the morphemes consistently represented. This is

true for nouns and verbs but is particularly the case for nouns (King 1996).

Consequently, the Hankul represents a variety of linguistic information, from

features to morphemes.

Another type of syllabic script is the Cree syllabary, a syllabic "shorthand"

created for two Algonquian languages of Canada, Cree and Ojibwe and adapted for

Athabaskan and Inuit (Eskimo) languages. It is a shorthand-based script employing

geometric characters, some representing syllables, some segments. There is no

standardized spelling for any of the dialects using it and there are local conventions

for character inventory, shape, and writing conventions. However, some of its

characteristic features are the following: vowel initial syllables are written with a

triangle syllabic, rotated through four positions to show vowel quality. Most

consonant-initial syllables are written with syllabics in which the shape shows the

consonant and the orientation shows the vowel. Syllable-final consonants are

written with small alphabetic characters calledfinals, usually subscripted. The

initial C of a consonant cluster can be written with a final. Vowel length may be

marked with a dot over a syllabic (Nichols 1996). Interestingly, the amount and

specificity of information presented in this script will vary by individual.

Following the shorthand origin of the script, writers may use plain syllabics,

indicating basic syllable structure orpointed syllabics, adding diacritics to include

segmental information. With enough diacritics, a writer may give an exact

phonemic transcription, though this is rare.

The Pahawh Hmong script of Southeast Asia (Laos) is a phonological

system which exhaustively represents the syllabic constituents, onsets and rimes, in

the language. There are many possible consonantal onsets but only one possible

coda /n/. Each graph represents an onset (C or C cluster), or a rime (vowel, coda

/In, and tone, which is represented by diacritics). The graphs are written rime-

onset, opposite of the way they are pronounced, indicating the perceived primacy of

the rime (V plus tone) over onset consonants. The one coda consonant is always

part of the rime graph, the symbol of onset In/ is never used in coda position.

Hmong is an isolating language, consisting of mainly monosyllabic morphemes

which are usually represented by graph pairs. Each pair of rime-onset/morpheme

graphs are set apart by spaces and are written left to right across the page (Ratliff

1996). Hence, Hmong orthography represents syllables, syllabic constituents and

morphemes at once.

In logosyllabic scripts, characters represent words or morphemes as well as

specific syllables. This type of script is likely to represent a language where

morphemes comprise a single syllable. The Pahawh Hmong Script described

above is one kind of logosyllabic orthography. Sumerian cuneiform is a another

variant of such a script. In Sumerian, unbound morphemes are represented by

logograms, bound morphemes are represented by rebus-derived syllabic signs,

usually V, CV or VC (Cooper 1996). Likewise, in Maya writing, which has only

partially been deciphered, logograms appear in combination with CV syllable signs

(Macri 1996).

Chinese, which has been traditionally classified as a logographic script, is

another type of logosyllabic writing. The most numerous type of character in

modern Chinese script is that which contains a phonological and morphemic

component (Coulmas 1989). The phonetic element in the character represents a

particular syllable, some features of the syllable, or the syllable and accompanying

tone. The classifier or radical gives semantic clues for that particular morpheme

Both the phonetic and the classifier offer partial cues for word identification

although the relative importance of either element contributing to word identification

varies. For instance, some phonetics occur in many characters, some in only a

few. There are many homophonous forms but more than one phonetic element to

represent the various forms. For example, the character for /ting/ uses the same

phonetic component in 4 out of 6 homophonous forms; the other 2 use a different

phonetic. Similarly, the amount of information conveyed by the classifier varies

widely. In some characters the classifier is clear while in others it is vague,

providing only a clue. In still others, it merely differentiates similar characters. As

a result, classifiers and phonetics serve mutually diacritic functions. That is, in

some cases the phonetic is clear and the classifier is vague so, the major recognition

comes from the phonetic element and the classifier serves only to disambiguate

forms. In other cases most of information conveyed by character is semantic while

the phonetic serves distinguish similar forms (Coulmas 1989).

In an Abjad, characters represent consonant segments only. This type of

script was originally adapted for Semitic language structure (e.g., Arabic, Hebrew).

The Semitic scripts convey (partially) both phonological and conceptual-semantic

information. Word "roots," which are represented with a skeleton of consonants,

convey general semantic information. A specific meaning is established by

mounting a particular "word pattern" onto the root. The word patterns are vowel-

melody morphemes, which are not usually represented but are predictable by

context. For example, the root, represented by the letters for/k s r/, refers to the

concept of connection. in Hebrew. By inserting the word pattern [-e -e] on this

root, the resulting word is /keser/ means knot or connection or conspiracy. The

vowel-morphemes may be indicated by diacritics but usually are not Words are set

apart by spaces, providing some lexical cues. (Bentin and Frost 1995; Bauer 1996).

Hence, pronunciation of written Arabic and Hebrew relies on knowing the

languages' morphological patterns. Conceptual-semantic information along with

some phonological and lexical information are represented by the script.

Morphological information is usually supplied by the reader so that the meaning of

words can be inferred.


Most alphabetic scripts predominantly represent segmental information.

Yet, in addition to segmental information, an alphabetic script may represent

syllabic, morphophonemic and morphological and lexical information with varying

degrees of explicitness. Segments are most explicitly represented by individual

letters. Much information, however, is indicated by letter patterns.

Orthographic regularities. A shared feature of alphabetic systems is that

orthographicregularities -- the predictability of letter patterns within a word -- will

indicate recurring structures larger than the phoneme. For example, the distribution

of letter patterns within words can signal syllable boundaries in multisyllabic

words. Patterns of C's and V's will be reflected in written words, creating

inhomogeneous frequencies of different letter sequences. The letter distribution

across syllable boundaries is often distinct from letter patterns that occur within

syllables. Because there are more constraints on phonemes that can occur within

syllables than across them (Seidenberg 1987,1989), the distributional properties of

letter patterns insure that syllable boundaries are often marked by particular letter

tokens (Adams 1981 in Seidenberg 1987). For example, syllable boundaries are

often flanked by letter patterns with relatively low frequencies-- an vil or vodka.

The syllable boundary bisects the lowest frequency bigram in a word (Seidenberg

1987). By exploiting distributional facts, information about syllable boundaries can

be extracted.

In principle, the above is true for all sublexical units identified by recurring

letter sequences. Hence, an alphabetic orthography will also tend to reflect the

morphological composition of words. Like letter patterns across syllables, the

distribution of letter patterns across morpheme boundaries will differ from those

within morphemes. For instance, root plus affix morpheme boundaries will be

indicated by the frequency of particular letter patterns. That is, the letter pattern in

the word pressing reflects its composite morphemes as press + ing instead of pre

+ ssing since there is no morpheme -ssing.. Also, the sequence -ing occurs

frequently at morpheme boundaries.13

Linguistic Structures. It is true that letter pattern frequencies can identify

linguistic information other than phonemic that is present in alphabetic scripts.

Though all alphabetic scripts share this basic property, particular alphabetic scripts

will have various conventions for more explicitly representing different types of

linguistic information. Further, alphabetic scripts will vary with respect to the

degree of segmental, syllabic or morphological information they include. For

example, some alphabets include much straightforward phonemic information,

some very little. English is a good example of a script that has various conventions

(using letter patterns) for indicating several different types of linguistic information.

It has often been noted that English orthography is not straightfowardly

phonemic. Some English consonants are represented by a single letter. For

instance, /p/ is usually indicated by p, /b/ by b, and so on. Yet even just for

consonant- segmental level, the orthography quickly becomes less phonemically

regular. English script often uses different letters to represent the same sound: /k/

is represented by k, c, ck, x, and qu, /f/ is indicated byf and ph, /j/ byj, g, dg, /z/

by z, s, and x /r/ by d and t. etc. Conversely, many different sounds are
represented by the same letter or letter sequence: c represents /k, /s/, /s/, s

represents /z/, Is/, // and /s/ and so on. Pronunciation of letters that represent

several sounds are often predictable by context. That is, surrounding letter patterns

3 The fact that morphological and syllabic parsing will often segment different sublexical units
has been of issue to psycholinguists for some time (Coltheart 1978; Taft 1979) For example, pre-
ssing is syllabic parsing while, press-ing is morphological. The difficulty has resided with
determining which linguistic structures the reader is likely to parse. As it is clear that there are
various types of linguistic information encoded in scripts, my suggestion is that readers extract the
linguistic cues present in the visual signal simultaneously, and further this does not present a
conflict for processing, but an advantage.

cue the reader as to the correct pronunciation. For instance, word final s represents

/z/ after voiced segments and /s/ after voiceless.

Yet it is really English vowels which make clear that the orthographic

representation cannot be strictly phonemic. Since there are only six letters to

represent fourteen vowels, a system of rules which could account for vowel-to-

letter correspondences would be very complex with many exceptions.

Alternatively, as Treiman (1992, 1993) points out, while one can argue that vowels

are represented very inconsistently, if you consider the sequences of graphemes

permitted to compose an entire rime, the orthographic system suddenly becomes

exceedingly systematic. She suggests that English orthography represents syllabic

constituents: onsets and rimes. That is, onsets and rimes are indicated in the

orthography by a high probability of sharing certain groups letters. By way of

illustration, consider the representation of vowels in English. Instead of a single

letter indicating the vowel, in many cases it is the letter sequence composing the

entire rime that informs the reader of the vowel. Conversely, the consonant(s)

representing the onset never give information about the vowel. For example, for

any particular CVC syllable (e.g., /bask/, spelled back or /bek/, spelled bake), the

initial consonant-letter does not give any information about the pronunciation of the

vowel. However, the letters that denote the rest of the rime do inform the reader

about the vowel. Consider the letter a in monosyllabic words. In a phoneme-to-

letter mapping system, the letter a is thought to denote several different vowels as

in: blare, blank, blast, blah, blame, ball. In all these examples, the consonant(s)

that compose the onset do not give the reader any clue which vowel the letter a

represents. However, the letters that compose the entire rime (e.g., -ame, -all) are

crucial for recognition of the vowel/rime. This pattern of entire rimes being

consistently represented by particular grapheme sequences is the norm in

monosyllabic English words14. Consequently, English orthography marks onsets

and rimes as distinct units. So, instead of sound-spelling rules based on individual

phonemes, many phonological-graphic correspondences are based on larger

syllable constituents, onsets and rimes.

The representational system for vowels in multisyllabic words is often

vague. For instance, consider the following pronunciations of the letter a:

syllable /sllabal/, syllabic, /slaeblk/, syllabification, /salaebafakesIn/. Here, a

represents /a/, /l/ and le/. Particularly for multisyllabic, morphologically complex

words, vowel pronunciation is largely indicated by lexical stress which is not

represented in English script. This means word recognition depends largely on the

reader recognizing the morphemes or words in question. Since English

orthography tends to keep morphemes consistent, (e.g., syllable, syllabic,

syllabification), vowel pronunciation in multisyllabic words is determined largely

via consistent morphological representation, and less by phonological cues.

Another feature of English orthography that contributes to vowel

identification is the contrast between single and double consonants. In many cases,

double consonants do not contribute consonantal information, however, they do

give information about the preceding vowel. This is true for many bisyllabic

words. The double consonants in: apple, babble distinguish the preceding vowel

/a/ as opposed to /el in: staple, table. The same is true with

litter /I/ vs. litre i/
toddle /a/ vs. total lo/
setter/E/ vs. cedar/i/
syllable /I/ vs. stylish/ai/
muffle/A/ vs. scruple/u/

Similarly, English orthography uses the double vs. single consonant contrast to

distinguish vowels in some words that are made multisyllabic by affixation. When

14 I will discuss processing evidence for this structural division in the next chapter.

the one syllable, V initial suffixes ( -ing, -ed, -er, -y) are added to monosyllabic

words and some bisyllabic (with syllable final stress) words with the vowels /a/,

/1/, /EI, /a/, /U/ and /A/, the final consonant is doubled in order to indicate that the

preceding vowel remains the same. For instance,

/s/ in hatter vs. /e/ in hater
/I/ in pinned vs. /ai/ in pined
/ I/ in admitting vs./ ai/ in admiring.
/alin hopping vs. /o/in hoping
/U/ in inputting vs. /u/ in computing
/A/N in drumming vs /u/ in consuming

Though this fact about the orthography is usually explained in terms of a spelling

rule, it is neither arbitrary, nor applicable only to root-plus- affix environments.

The double consonants are a feature the orthography uses for distinguishing among

vowels15, which applies both word internally and across morpheme boundaries.

Hence, English script uses letter patterns to indicate segmental information.

A strictly phonological orthography would represent the words as they are

spoken, regardless of phonological alternations that are a result of morphological

processes. Yet some alphabetic scripts consistently indicate morphological

structures in favor of phonological ones. For instance, English, Czech (Henderson

1984), Albanian (Comrie 1996), 01 Cemet (Zide 1996), Korean16 (King 1996) are

partly morphophonemic. That is, in morphophonemic environments, they often

ignore phonemic contrasts in favor of consistently representing morphemes.

As previously noted, with English script there is a tendency toward

morphological consistency. This is true for both inflectional and derivational

15 In some cases double consonants can indicate consonantal rather than vowel information. In
cases where affixation results in geminate consonants, double consonants indicate consonant
length. For example, in openness loplnnes/ and unnatral /Annatr I/, double consonants are
not giving vowel information but indicating consonant length. Also, these spellings keep the
respective morphemes (e.g., open and -ness ) consistently represented.
16 an alphabetic-syllabic script

processes, regardless of phonological alternations. For instance, inflectional

affixes may be variably pronounced while being graphically consistent. The past

morpheme represented -ed indicates /t/ in tripped, /d/ in scarred, and /Id/ in parted

The plural -s indicates /z/ in dogs and /s/ in cats. The plural -es indicates /Iz/.

Where different inflectional markers are phonologically similar, they are

distinguished graphically. For instance, the -s plural and the -'s possessive

exhibit the same phonological alternations. Also -ing is frequently reduced in

speech, but is graphically consistent.

Likewise, English spelling consistently represents many derivational

morphemes regardless of predictable phonological alternations.


medic-ine /medic-in-al/medic -ate/medic -d

anx i-ous/anx i-ety

Vowel shifting with affixation is usually not represented in spelling., keeping the

base morphemes consistent.


However, in some cases, morphophonemic alternations may be represented in the


decide/decis ion
analyst is/analyz e

Alternatively, some alphabetic scripts such as Italian, Serbo-Croatian

(Henderson 1984), Turkish (Comrie 1996) favor phonological consistency. Such

scripts are quite phonemically consistent and do represent morphophonemic

changes in pronunciation in the spelling. As Henderson (1984) points out,

languages in which morphological processes lead to a lot of variation in the surface

forms tend to favor phonological regularity over morphological. Turkish, for

example is agglutinative and uses extensive vowel harmony. Hence, words tend to

be very long and have a reduced number of vowel distinctions, making it necessary

to attend to consonantal invariances.

Due to a number of factors such as historical change and linguistic

preservation pressures, almost all alphabetic scripts have some phoneme-grapheme

mapping inconsistencies. When orthographic systems remain unchanged for a long

period while pronunciations have changed greatly, some scripts (English, French,

Icelandic) become quite lexicalized. That is, when the phonological cues for

pronunciation of the word do not reliably predict the spelling, they must be

memorized logographically. English has many common words like this:


love, move

Also, many homophonous words are distinguished by their logographic

representation as in

Maine, mane, main.
plain, plane

Summary and Concluding Remarks

Orthographic systems are most often partial representations of multiple

levels of linguistic structure. They are alike in their simultaneous use of multiple

cues for word identification. Some combination of phonological, morphological,

and lexical cues appear in all scripts with varying degrees of explicitness.

Moreover, scripts seem to share the characteristics of being only partially

representative of speech and of our underlying linguistic knowledge.

Many of the world's scripts have evolved from one another and others have

arisen independently. There is little doubt that particular orthographies have been

adapted to serve their particular languages. As a result, scripts are as many and

varied as are languages. Yet, like the languages they represent, scripts are

constructed from and constrained by linguistic structures and our ability to process

them. I believe that the fact that diverse orthographic systems share such

fundamental characteristics is itself evidence of human preferences and constraints

on processing visual language.

The notion that visual language processing is influenced by the linguistic

information presented in scripts is not novel to the psychology of reading. What

has been problematic is that, to a great extent, reading research has assumed a prior

that alphabetic scripts represent phonemic systems. As a result, most theoretical

and experimental work has focused on delineating the essential phonemic features

of the orthographic system and how they are best learned. In doing so, many of the

regularities present in the system can be overlooked or misinterpreted. An appraisal

of a variety of writing systems obliges us to consider the idea that alphabets, like

various other scripts, include different types of linguistic information.

Furthermore, contrary to the notion that a complete and consistent phonemic script

is necessarily optimal, the fact that all orthographies are multileveled and incomplete

indicates that there is some processing advantage for information to be structured in

this fashion. The following chapter will discuss psycholinguistic evidence that

readers of various scripts do indeed extract cues form the various levels of linguistic

structure represented in orthographies.

The most important implication of these facts to the overall discussion is that

learning the varied, often incomplete features of an orthographic system is likely to

involve far more than what is provided by instruction. Further, many of scripts'

regularities are not obviously indicated, or complex enough so as to not lend

themselves to conscious analysis particulary by young children). Consequently,

orthographies are the kinds of systems that are likely to be learned implicitly.


Linguistic taxonomies of writing systems have some interesting implications

for theories of visual processing, word recognition and reading. Based on the

assumption that humans will attend to patterned stimuli when faced with a learning

task, it is highly likely that readers utilize the structure that orthographies present.

Furthermore, the fact that multileveled representation is the rule across scripts and

script types suggests that there is some processing advantage to extracting various

linguistic cues at once. Since spoken language is comprised of several structural

levels, it follows that humans would be predisposed to interpret the written code by

making connections to language representations already in place. Moreover, I

suggest that it is precisely because of human processing predilections that scripts

tend to show such remarkable commonalties. Hence, this chapter is a corollary to

the previous one. If reading does involve interpreting the various aspects of

structure represented in scripts, there should be evidence of this in the way people

process written language. Since there has been a great deal of study in regard to

phonemic structure in reading, and in an effort to emphasize the multi-

dimensionality of visual processing, I will discuss empirical studies which point to

reader's use of morphological and syllable structure.

Syllable Structure and Language Processing

The syllable is a constituent of phonological representation. Though it has

no conclusive phonetic definition, it generally corresponds to a single bump in

sonority (perceived loudness), and to the open-close cycle of the jaw. The syllable

is the smallest separately pronounceable unit of speech and perhaps the most
perceptually salient. Both adults and young children agree as to how many
syllables there are in a given form and agree to a large extent where the boundaries
are. Furthermore, spontaneous speech errors affirm the reality of syllables as a
speech unit. Spoonerisms, where an intended word is replaced by another, usually
have the same number of syllables as the intended word. Also, with the "tip of the
tongue phenomena," when searching for a word in long-term memory, the words
tried usually have the same number of syllables as the sought-for item (Brown and
McNeill 1966).
One widely accepted proposal for the internal structure of the syllable is that
it is divided into onset and rime. The onset is elements) before the sonority peak.
The rime consists of the nucleus (which is or includes the sonority peak) and coda,
the elements after the nucleus.
Fig. 3.1


onset rime

nucleus coda
The division of the syllable into onset and rime is justified on linguistic
grounds. The elements of the rime (nucleus and coda) may attract stress. For
instance, in quantity sensitive stress systems, a heavy syllable has something after
the sonority peak. Thus CVV and CVC are heavy and attract stress but CV and
CCV are not and so do not attract stress. Elements of the rime may be a tone
bearing unit. In some languages, a tone can appear on the peak or the element after
the peak (the second segment of a long vowel or a coda). Vowel length may be
affected by elements after the peak. For example, in some languages, a long vowel

can only occur in an open syllable. Conversely, onsets are not distinguished by

stress systems, they do not bear tone, nor do they affect vowel length.

There is behavioral evidence for this view of syllable structure from

spontaneous speech errors. Transpositions occur when an intended sound is

replaced by a sound belonging elsewhere in the utterance. In these cases, people

replace onsets with onsets and rimes with rimes ( Fromkin 1973; MacKay 1972,

Stemberger 1983). Or they may transpose onsets with onsets, peaks with peaks

and codas with codas (MacNeilage 1985). Additional evidence comes from

children's language games, such as "pig latin," in which segments, onsets, rimes,

or syllables are manipulated.

There is also experimental evidence which affirms the perceptual reality of

the onset-rime division of the syllable. Treiman (1986) conducted experiments

where college students heard pairs of CVCC words (e.g., /pakt/ and /nAts /) .

They were asked to blend the two words into one new word by taking parts of

each. Over 90% of all responses were blended by taking the onset on one word

and the rime of the other. For example, blending /pakt/ and /nAts/ resulted in a

C/VCC division, /pAts/ as opposed to CV/CC /pats/ or CVC/C /paks/

Even young children generally perform well at phonological awareness
tasks which manipulate syllables, onsets and rimes. For instance, in rhyming and

alliteration production tasks, some very young children (preschoolers) demonstrate

knowledge of spoken syllables at onset/rime boundaries (Maclean, Bryant and

Bradly 1987). Alliteration is defined as knowing when 2 words share an entire

onset (e.g., bread and brick). Kindergartners perform well on rhyming and

alliteration production tasks (Stanovich, Cunningham and Cramer 1984) and forced

choice tasks (e.g., what rhymes with bed ? sled or ring). In another series of

studies done by Treiman and Zukowski (1991), preschoolers and kindergartners

performed well on word pair comparison tasks which demonstrate knowledge of

onsets such as plank and plea vs. twist and brain. They also demonstrated

knowledge of rime, spit and wit vs. rai and snap. In phoneme deletion and

substitution tasks, first and second grade children have a very high rate of success

when asked to manipulate an entire onset or rime. However the error rate increased

greatly when asked to manipulate only part of the onset or rime (Bruck and Treiman

1990; Treiman 1985).

Syllable Structure in Reading
The syllable and its constituent structure are highly salient phonological

units and are crucial to theoretical linguistic accounts. This perceptual salience has

clearly marked the development of writing systems. Syllabic scripts are the most

common type of orthographies and often have developed from alphabets and

logographies (Daniels 1992). Given the perceptual salience of syllables, it is likely

that readers of alphabetic scripts may use syllable structure to recognize words. In

fact, there is evidence from English that this is the case.

Some Evidence for Use of Onset and Rime

As stated in the previous chapter, English orthography tends to consistently
represent onsets and rimes-- that is the pronunciations of onsets and times are more

consistent than phoneme pronunciations. This is particularly true for vowel

representation in English where every vowel symbol has multiple pronunciations.

However, if one considers the letters which compose the entire rime, the

pronunciation of vowel symbols are quite consistent. For instance, -ast is always

pronounced /ast/, where bla- can be pronounced variably as in blast, blame, blare,

blah. Also, consider the vowel symbol ea. It may be pronounced /e/ as in steak,
/i as in beat, or // as in head. For words that end in -ear and -each, ea is most

likely pronounced /i/. Yet for words that end in -ead, ea is most likely pronounced
/l/ (Treiman 1992). Consequently, the group of letters that form the rime are much

more likely to be consistently identified and pronounced than are the individual


It has been suggested that this consistent property of English orthography is

extracted during reading. Treiman (1992) and Goswami and Bryant (1992)

propose that children favor using orthographic units that correspond to onsets and

rime because they are more perceptually salient than are segments. Many young

children, before they learn to read are aware that the spoken syllable /blast/ consists

of the units Ibl/ and /est/. Accordingly, it has been suggested that children will tend

to use orthographic units that represent salient phonological units. As a result, it is

likely that young children will favor using onset and rime constituents in

recognizing and pronouncing written words: bl ast, as opposed to b last, bla st,

bias t, or even b I a s t. In this view, the correspondence between bl and /bl/ is

stored as a separately memorized unit and can be used whenever bl occurs at the

beginning of a word. Similarly, -ast is recognized as a unit.

In fact, there is some convincing empirical evidence which suggests that

onsets and rimes are used in learning to read English words. A series of studies by

Goswami (1986, 1988) indicated that pre-reading and beginning reading children

learned monosyllabic words by making "analogy" to the onsets and rimes of

previously learned words. Goswami performed a series of experiments with

children from kindergarten (prereaders) to second grade. In these tests, pairs of

"clue words" which shared the same orthographic sequence were given. The clue

words, which were pronounced for the children, shared either a rime sequence like

peak-beak, or an initial CV(V) sequence like beak- bean. Then the children were

asked to read a series of words and nonsense words that possibly shared either of

these orthographic sequences (e.g., leak or beam ). The results showed that while

children made some analogies from the orthographic sequences of the initial CV(V),

all the children were significantly more likely to make analogies from the rimes of

words. Even the prereading children were capable of using the orthographic rime

sequences to identify some words (Goswami 1986, 1988).

Also, studies have indicated that both children and adults are better at

recognizing nonwords whose rimes are orthographically similar to real words. In a

experiment by Treiman, Goswami and Bruck (1990), subjects were asked to

pronounce words like tain, goach, and others like goan, taich. The first two

nonwords share their rime with a number of real words like rain, main, coach,

roach. But the last two share their orthographic rime with few or no real words. If

just phoneme/grapheme correspondences were being used, subjects should respond

equally well on both types--they all contain the same graphemes. However, all

subjects performed better with nonwords with familiar real word orthographic

rimes, e.g., tain shares its rime with many real words such as main, rain, pain,

drain, etc. The more real words that shared the orthographic rime of the nonsense

word, the better the subjects' performance. The number of real words that

contained the nonword's initial CV sequence had no effect.

Children's spelling errors offer additional evidence of the onset-rime
division of the syllable. Children have difficulty analyzing a complex onset into its

constituent phonemes. Hence, they will often spell an onset cluster with the initial

consonant only (e.g., blow is spelled bow and stay is spelled say). Treiman

(1993) suggests that this reflects the isolable nature of the syllable onset. These

omissions do not reflect children's mispronunciations since children at this age can

pronounce the words correctly. Neither do these omissions reflect the sonority of

clusters since the second element of a cluster was likely to be deleted whether it be

an obstruent or sonorant. Treiman concludes that it is likely that this pattern of

spelling errors reflects children's tendency to analyze onsets as a single unit.

Some Evidence for use of Syllable Boundaries

Orthographic regularity 7 is the predictability of letter patterns within a

word. Readers have been shown to extract general information about the frequency

and position letter sequences, even when they do not refer to discrete linguistic

units. For example, ng is a regular bigraph that occurs syllable finally and across

syllable boundaries while gn never occurs word initially or finally, and only across

syllable boundaries. Taft (1991) has demonstrated that lexical decision times to

nonwords likeflink, with legal orthographic strings are slower than to random

strings like Ifkni, indicating that readers are sensitive to statistical regularities within

words. Young children (prereading) who have had print exposure will judge

sequences like AAAAA as unreadable, indicating that they have already abstracted

some letter frequency information (Sinclair and Papandropoulou 1984). Also,

Treiman (1993) found that young children, kindergarten through second grade,

already have some knowledge of orthographic regularities and that this knowledge

increases with age and print exposure.

There is also evidence that such orthographic regularities mark syllable

boundaries, and readers use this information for recognizing words. For instance,

when syllable boundaries are marked in the orthography by bigrams that only to

appear in different syllables (e.g., ABHOR and AN VIL), readers are likely to

utilize this information to recognize words (Prinzmetal, Treiman and Rho 1986;

Seidenberg 1987). In these studies, a word (ANVIL) is printed in two colors

(anVILi8) and displayed tachistoscopically for a duration that produces about 10%

errors over trials. Subjects' task was to report the color of a target letter, e.g. V.

The reasoning behind this method was that if subjects use syllabic units during

recognition, they should not respond to the target letterV incorrectly with the color

17 Some times called orthographic redundancy.
18 Different colors are represented here by upper-lower case

of N because the syllable boundary should act as a barrier to feature integration

errors. Conversely, if the display were ANVil, subjects may tend to report that V

was the color of il because it was part of the same syllable. Errors of the first type,

"violation errors," crossed the syllable boundary; while errors of the second type,

"preservation errors," preserved the syllable boundaries. It is thought that if

syllables influence the pattern of feature errors there should be more preservation

than violation errors.

Prinzmetal et al. (1986) who found that four of five experiments yielded

more preservation than violation errors, concluded that syllables do influence word

identification. However, the stimuli in the four experiments that show syllable

effects were words in which the syllable boundary bigrams only appear in different

syllables (e.g., AN VIL), or where syllable and morphological boundaries coincide

(e.g., TODAY ). The experiment that did not produce the syllable effect used

words where the syllable boundary was not marked orthographically. For

example, the letters that mark CA MEL and SA LAD do not uniquely occur at

syllable boundaries. These results were replicated by Seidenberg (1987). He

conducted similar experiments where he also found that readers are aware of

syllable boundaries where they are marked by orthographic patterns19.

19 Seidenberg suggests that this is true not just in words like ANVIL where the syllable boundary
bigrams only appear in different syllables, but also words like LAPEL and SONIC, where if you
take into account the bigram frequencies preceding, straddling and following the syllable boundary,
syllable boundaries are indicated. That is, syllable information can be induced from the overall
distribution of letter patterns, not just the syllable boundaries. Hence, Seidenberg concluded that
syllabic structure in orthographies falls out of letter frequency patterns in general, and that readers
will attend to all recurring patterns However, I believe there are some problems with drawing such
conclusions from this experiment. First, the stimuli include 5-letter words where Seidenberg
assumes that the syllable boundary is after the second (LA/PEL) or third (SON/IC) letter. This
syllabic division of words like SONIC is arguable--it can be considered ambisyllabic, where there
is no real boundary between the syllables Also, the entire stimuli list is not given so it is unclear
whether stress is potentially playing a role. That is, all the CV/CVC words may be stress final
(like LAPEL),which may give the reader additional cues about the syllable boundaries. Hence,
subjects awareness of syllable boundaries in stress-final words like LAPEL may be influenced by
phonological factors (e.g., stress), not simply letter pattern frequencies.

Morphological Structure and Language Processing

Though discussions about the nature of lexical representations is ongoing,

morphological structures have been central to linguistic accounts. Despite debates

about whether morphological structures are represented in a "lexicon" and

assembled via rules or are governed strictly by syntactic principles, it is generally

agreed that morphemes are the elements of word formation and interact with

syntactic and phonological processes. There is visible evidence of speaker's

knowledge of morphology from speech production and perception. Native

speakers know the system, (e.g., derivation, inflection, compounding) for

combining morphemes into words. Children and adults coin and understand new

words using productive morphology. Spontaneous speech errors involve

transposition of morphemes (Stemberger 1983).

In considering morphology from a processing perspective, Bybee (1995)

points out that morphological form and meaning are a consequence of human

processing capacities. For instance, in the context of word formation and language

change, in the gradual progression from a lexical morpheme to grammatical one,

changes occur in that morpheme's phonological shape, meaning, and grammatical

behavior. It is because bound morphemes are processed differently that they

develop properties that distinguish them from independent words. In this sense,

morphological structures are themselves evidence of morphological processing. I

believe that there is an analogous point for visual processing. Orthographies

indicate morphological structures in various ways because morphology is a

component of visual language processing.

Yet, theoretical accounts of word recognition and reading have tended to

minimize the role of the morpheme. Reading theories such as phonemic recoding,

whole word, and dual route theories generally maintain that words are identified

either by reference to individual phonemes or whole words. Similarly,

psycholinguistic word recognition models that have been proposed often discount

morphological processes. For instance, in Search Models such as Forster's

(1976), initial letters of a word are used to identify a 'bin' of similarly spelled

words and the bin is searched serially for an element which is sufficiently similar to

the stimulus word until a match is found. Since such models predict the first few

letters of a word will be effective to access the lexicon, they assume whatever

phonological or morphological information is encoded at the ends of words, is not

relevant to word recognition. A somewhat similar treatment of word recognition in

speech perception is Marslen-Wilson's (1980) Cohort Model. Cohort theory

asserts that stimulus information is used from the beginning of a word onwards to

reduce uncertainty until word identification is made. That is, starting from the

beginning of a word, when the word becomes uniquely distinguishable from all

other words in the language, identification is made. Again, this model specify

different treatment for morphologically complex forms, nor does not define how

word-final morphological information plays into rapid identification.

Still, there have been some models that have attempted to specify a level of

morphological processes in word recognition. In Decomposition models (Taft and

Forster 1975, Chialant and Caramazza 1995) lexical representations consist of

morpheme sized units, or both morpheme and word sized units. The challenge for

this type of model is to accommodate the different effects that are found in the

processing of regular and irregular complex words (e.g., walked and went). Also

problematic for decomposition models is that morphologically complex forms that

are very frequent fail to show significant morphology effects.

Interactive (spreading) activation models, such as Seidenberg's (1987) and

Seidenberg and McClelland (1989), address the issue of words' frequency effecting

recognition regardless of their complexity. In this model, the more frequently

connections are activated, the stronger the connections. Hence, high frequency

items need less activation than low frequency ones to be accessed and are

recognized more quickly. However, in these models, morphological structures are

not directly represented but are merely (possibly) a byproduct of the transitional

probabilities of letter sequences.

A somewhat different attempt to account for morphological processes in

word recognition is Taft's (1979, 1987) BOSS--Basic Orthographic and Syllabic

Structure-- model. He proposes that readers use access units called BOSSes,

which synthesize syllabic and morphological information without necessarily

referring to a syllable or a morpheme per se. A BOSS is the first part of the stem

morpheme of a word, up to and including all consonants following the first vowel,

but without creating an illegal consonant cluster in its final position, e.g.,

FAST/ER Empirical studies which have tried to assess the perceptual reality of

BOSSes have had conflicting interpretations (Taft 1979,1987; Seidenberg 1987;

Lima and Pollatsek 1983).

Morphological Structure in Reading
Perhaps the most important (and testable) prediction the various reading

theories and word recognition models make is that there should be (or not) some

differences that show up in the processing of morphologically simple versus

morphologically complex forms. For instance, for models that do not specify

morphological processing, there either should be no difference between processing

of simple and morphologically complex forms or such differences should be

accountable for by other means. Conversely, if morphology is functional in word

recognition we would expect to see differences in the way people process

morphologically simple versus complex words. Some empirical studies have

attempted to discover whether morphology does influence the processing of written

words. The following is a brief survey of studies that have been taken as evidence

for visual morphological processing.

Some Experimental Evidence of Morphological Processing

Priming effects have been taken as indicators of the role of morphology in

processing written language. Feldman (1987) has presented evidence that Serbo-

Croatian nouns, which can be inflected by affixation to six cases (singular and

plural) other than the nominative and form derived forms by adding a diminutive

affix, show priming effects. For example, diminutives can be formed from base

forms such as stan "apartment" and kora "crust" by adding one of the suffixes -

cc, -ica, -ence, -ak (e.g., stannic korica), where choice of suffix is constrained

by gender of the noun. This was a repetition priming experiment where base

forms are targets, they are preceded (7-13 items earlier) by a prime which was

either the identical word or a morphologically complex form. In this lexical

decision task,20 primes were either the base forms, the diminutive form, or a

"psuedodiminuative "--an unrelated monomorphemic word whose construction

inappropriately suggests that it contains the same base form plus a diminutive affix

(stanica, korak). The results indicated that decision latency to the target was

function of which type of prime preceded. Both base word and diminutive primes

significantly reduced target decision latencies while pseudodiminutives has no such

effect. That is, subjects recognized the targets as words significantly faster when

they were primed by morphologically related forms. Hence, these results show

significant facilitation for morphological relatives and no facilitation for

phonologically and orthographically similar but unrelated words.

In another priming experiment, Laudanna et al. (1989) presented evidence

from Italian that suggested that morphologically related words will prime each other

while merely orthographically similar ones do not. In a condition where the degree

20 Lexical decision tasks involve the subjects deciding whether the target is a word or not.

of orthographic similarity between words are controlled, only prior presentation of

a morphologically related word such asportare "to bring" and not merely a

orthographically similar word such asporte "doors", affects lexical decision

performance for a target word porta "bring." Lexical decision for such pairs is

faster when prime and target are morphologically related. That is, the presentation

ofportare, "to bring" facilitates recognition of porta (you) bring" but notporte

"doors," even though they are all visually similar. This is taken to indicate that

lexical access and storage is not determined simply by orthography or phonology,

but also morphology.

Evidence suggests (Caramazza etal. 1988 in Chialant and Caramazza 1995)

that lexical decision is significantly effected by words' parsability into morphemes.

In an experiment where orthographic similarity was controlled and word frequency

was matched, it was found that subjects are slower and produce more errors when

rejecting non words that can be parsed into actual root and affixes. For example,

Italian cant-evi and English walk-est are non words that are composed of

morphemes that do occur in the respective languages. In a lexical decision task,

subjects are much slower rejecting such forms as non words, than when rejecting a

non word that contains a pseudomorpheme (e.g., Italian cant-ovi and English

walk-ost). This is interpreted to mean that subjects are attempting to use whatever

morphological information that is present to identify the word. As a result, it is

easier to identify nonwords constructed with pseudomorphemes. Accordingly,

subjects were faster still when nonwords contain no morpheme-like sequences at all

(e.g. canz-ovi, wilk-ost), indicating that when there are no recognizable morphemes

in a pseudoword, subjects easily reject them as word options.

In a segment shifting task developed by Feldman, Stoltz and Feldman

(1995) compare subjects' ability to segment words along morphemic boundaries

with their ability to break up unitary morphemes. This method was an attempt to

simulate spontaneous speech errors (where morphemes are transposed) which are

taken to be evidence of morphological processing in speech production. The task is

to take part of the source word and attach it to the target, forming a morphologically

complex word and to name the resulting word as rapidly as possible. For example,

the source word is either harden or garden while the target word is bright. The

dependent variable is the time necessary to shift the designated segment -en from

the source to the target, bright-->brighten. If it takes longer to attach the

morpheme-en to the target, bright, after being primed with the single morpheme

garden than it does after the morphologically complex harden, then this argues for

morphological decomposition in processing. Results indicated that the

morphologically complex source words (e.g., harden, darken, soften) produced

faster shifting latencies than simple source words (e.g., garden) even though the

complex words had a lower average frequency. Furthermore, source words were

selected so that several letters close to the shifted portion, along with the shifted

portion, represented the same phonemes. Therefore, it is unlikely the outcome

reflects phonological factors, but instead that readers are sensitive to morphological

structure. A similar study of Serbo-Croatian was done by Feldman (1994) with

similar results.

So far evidence of morphological processing has been with languages

(Italian, English, Serbo -Croatian) in which complex words are constructed by

appending affixes to a stem, which is often a word itself. A study by Bentin and

Frost (1995) replicates Feldman's segment shifting experiment with Hebrew, a

language where words are composed of "roots" + "word patterns" which are

productive but are nonconcatenative. Roots are composed of a series of

consonants. Word patterns are vowel-melodies which "break up" the consonantal

root Words are formed by "mounting" word pattern morphemes onto the root.21

21 The authors are considering the word pattern as well as the root, which usually represents a
general concept, morphemes.

They hypothesize that if morphological effects generalize across different types of

morphologies, and over segments that are not final syllables, then morphology

effects would show up with Hebrew as well as English and Serbo-Croatian. That

is, if shifting infixed word patterns show the same effects as shifting concatenative

suffixes, this would indicate an effect of morphological organization on parsing

words that cannot be attributed to simple visual analysis of the letter string22. In

this experiment both C's and V's are represented, V's either by diacritics or letters.

For example, a source word might be (transliterated) moked The task is to extract

the vowel-morpheme [-o-e-] from the root [m-k-d], and to combine it with a target

string of consonants23 as quickly as possible in order to form a new word. There

are 2 types of roots: 1) Productive roots that can attach with several different word

pattern morphemes and 2) non-productive roots which can only have one word

pattern morpheme (indicating it is not decomposed morphologically). The authors

reason that if reading in Hebrew employs morphological analysis, it should be

easier to detach word patterns from productive roots than from non-productive

roots. The results indicated that subjects were much faster with productive root

primes than with non-productive. The authors concluded that in Hebrew word

identification, roots and word patterns are psychologically distinct, and also that

there is a stage where they are separated during word identification. This is

interesting since generally in connected text word patterns are not represented at all

and must be supplied by the reader. Hence, it would seem that Hebrew

orthography would oblige morphological decomposition and word pattern

morphemes would be inherently separable from their root. However, these data

22 Some researchers (e.g., Seidenberg's 1987 position on this has been discussed) assert that
extracting morphological patterns are simply a byproduct of visual analysis of the phonemic letter
string, NOT that morphology has an effect on parsing words.
23 The target string of consonants was not a meaningful root Therefore, mounting the word
pattern on it resulted informing a pseudoword. The authors argue that this is procedure insured
that the results could not be accounted for by other factors such as similarity or dissimilarity in
word class between the source and the target (Bentin and Frost 1995).

indicated that non-productive words, though written as though roots and word

patterns are distinct, are accessed as whole lexical items. In general, this seems to

indicate that there is some kind of processing distinction for bound versus free


Logographic scripts, where morphemes are explicitly represented, would

seem to necessitate morphological analysis by the reader. Taft and Zhu (1995)

investigate the status of morphemes vs. whole word representations in Chinese

writing with a series of experiments using "binding words." Chinese characters are

usually said to be equivalent of single morphemes. Each word is made up of one or

more (usually two) characters/morphemes, where each character is usually

monosyllabic. For example, the character for mil means "cleanse" and yi means

"bathe." The two characters together, mid yi mean "bath". However, not all

characters are free morphemes. Some characters can only occur with other

characters to form compounds or "binding words." For example: Hdo means

"vast" but is never used on its own, to use it must be combined with something like

hdoda where dd means "big." All characters are spatially distinct so there is no

way to tell orthographically if they are biding words or not. Certain characters in

binding words only occur in first position, certain others only in second position.

Further, some binding word morphemes are unique (non-productive) in that they

occur only in a single word, while others can occur in several words.

In one experiment which used a naming task, subjects were faster at naming

an isolated first morpheme of a binding word than the second. The authors initially

conclude from this that the morphemes of binding words do not have their own

lexical representation and the whole word must be accessed. In another experiment

the authors test the idea that a morpheme's productivity determines whether or not it

is independently represented. They found that non-unique binding morphemes

(those that occur in several words) were more easily accessed than those that

occurred in just one word. They conclude from this that characters develop

independent representations when their status as a morpheme has been fully

established--which involves each morpheme/character developing of a semantic

function. Hence, characters that are semantically independent are taken to be

independent morphemes and are recognized as such. Other characters that are

semantically bound to another character can only be recognized by accessing the

whole word. It appears that in Chinese, like Hebrew, words that are ostensibly

polymorphemic but non productive seem to be represented as wholes. Hence, a

conservative conclusion would be that in general, bound morphemes seem to have a

different psychological status2 than free morphemes and that difference shows up

in visual processing.

Evidence of Morphological Processing from Children's Spelling

In a study that followed two classes of first grade children over the school

year, Treiman (1993) found that children are significantly more likely to omit the

final consonant from a word when it is a regular inflectional morpheme than they

are if the word is monomorphemic or is an irregularly inflected word. For

example, killed and build are similar phonologically but children are more likely to

omit the -ed from killed than the d from build. Similarly, children were less

likely to omit the final consonant from irregularly inflected words likefound or

feet, than they are from killed. Omissions are most common when the suffix can be

predicted from the stem, subject to regular inflection. This error pattern is

potentially a result of children's recognition of morphological boundaries. This is

particularly likely because no phonological reason was found for this. For

example, it was not the case that they omitted final C's to avoid consonant clusters

because children did not, as a rule, do this anyway. Nor does the duration of the

4 The psychological status of bound versus free morphemes should depend on a number of factors
such as productivity, syntactic application, morphological typology of the particular language,
e.g., isolating, agglutinative, fusional, and so on.

final C's explain this omission. Since in adult speech final consonants are actually

longer when they are inflectional suffixes, one would expect them to be more likely

to be represented, which they are not. Treiman suggests that a possible explanation

for this morpheme omission is that children may forget they are writing the past

tense since they write so slowly. However, she notes that when reading back what

the child has written about yesterday, the child will say /Avd / where she has

written love. This tendency seems to indicate at least that it is clear to the children

that the word is recognizable without the inflection, that inflection is a separable

something extra, the necessity of which is not yet entirely clear.

Some Evidence from the Evaluation of the Reading Impaired

Reading disabled subjects (classified dyslexic) very often show significantly

poorer oral, as well as written, language processing than do normal readers. This

fact has often been hypothesized to mean that dyslexia is some kind of general

language disorder. Therefore, it is an impairment that effects both the auditory and

visual modalities. It is further hypothesized that this indicates a link between

phonological and visual processing and is taken as evidence that reading involves

recoding of phonological structure. However, the scope of oral language problems

that are associated with dyslexics suggests that linguistic deficits beyond

phonological are present. In some current work by Lombardino, Riccio, and

Hynd (in press), reading disabled (high IQ) subjects performed significantly more

poorly on an evaluation of language fundamentals in all areas of language

processing than did normal readers. The evaluations were presented orally. The

areas tested were subjects' knowledge of word class, semantic relationships,

sentence formulation, sentence recall, sentence assembly and word structure

(simple derivation and inflection). If the hypothesis that dyslexia is a general

language disorder and the auditory and visual modalities are both functional in

reading is correct, then this seems to be evidence that interpreting multiple

linguistic levels is part of the reading process. That is, this disorder is effecting all

(tested) areas of language processing and its effects are showing up in both reading

and speech perception and production.

The notion that dyslexia is a whole language learning disorder points away

from the hypothesis that dyslexia is particularly a result of some phonological

weakness. Scarborough (1991) suggests that oral language processing deficits

often precede reading disabilities and these oral deficiencies show up in syntactic,

morphological as well as phonological abilities of preschool children. Scarborough

followed the development of a group of children from two to eight years in order to

compare the development of children who became reading disabled with those who

became normal readers. The children in these groups were evaluated at ages 30,

36,42,48 and 60 months, at which times various tests and naturalistic

observations were recorded. These evaluations involved measuring sentence

comprehension, grammatical complexity, and measuring the length of utterances

by counting morphemes and productive compounds. The results indicated that

"syntactic abilities"2 at age 30 months were predictive of Grade 2 reading ability.

Dyslexic children showed deficiencies in sentence comprehension and syntactic and

morphological complexity throughout much of the preschool period that normal

reading children did not. However, by age 5, oral language difficulties were no

longer evident by the tests given. This is the potential result of a number of factors.

Scarborough notes that one of which may be simply that these tests were

insufficiently sensitive to the linguistic proficiency of older (5 and over) children.

The point relevant to the present discussion is that this study indicates that dyslexia

is not simply a phonological disorder or a reading disorder. Instead, it seems to be

a developmental language disorder that affects multiple levels of linguistic

processing. The fact that this disorder significantly influences reading ability is

25 Syntactic abilities defined and measured by the Northwestern Syntax Screening Test (sentence
comprehension), the Index of Productive Syntax and the Mean Length of Utterance evaluation.

evidence that reading involves multi-leved linguistic interpretation. That this

disorder is more problematic for reading than for speaking is a complex issue that

has not been resolved.

Neurobiologv of Multileveled Processing
Evidence from the Neuroanatomy of Dyslexics

I have argued that there must be a functional reason that orthographies are

organized to include several different types of information at once, and also that it

appears that readers interpret scripts' many different properties simultaneously.

Neurobiological research indicates that there is no single mechanism in the human

brain that is able to extract all the information from an auditory or visual stimulus.

Hence, the stimulus is broken up and little bits of the information are processed by

different subsytems and along different pathways Galaburda (1996) contends that

this general property of the brain applies to processing linguistic stimuli. That is,

we are at all times splitting up the stimulus in order to analyze its particularly

different properties. We do this at various stages, and at various levels, and along

different pathways, which extract different kinds of information from the stimulus.

Information comes into the system in a bottom up fashion but top-down

mechanisms adjust the attentional systems in order to regulate the incoming

information and make expectations about what may or may not be coming in.

Galaburda postulates that reading involves mental functions of various types,

which function on many levels and pathways at the same time. Hence, to have

clinically evident dyslexia multiple systems would have to be impaired, at different

stages and along different pathways. This is widely agreed by researchers in the


In fact, there is evidence that there are anomalies in both perceptual and

cognitive systems in dyslexics. Dyslexics have fewer margnocells, which process

rapid information and handle location and orientation in space. There is a delay of

this type of low level, physical information into the brain of the dyslexic, hence,

dyslexics are slower at this type of perception. That is, they show a slowing down

of transmission of the stimulus before it ever begins to be analyzed linguistically or

otherwise. Dyslexics also show neurophysiological anomalies in their cognitive

systems. There is evidence of phonological and syntactic problems, problems with

verbal memory and attention.

Galaburda concludes that multiple processing systems are involved in the

neurobiology of dyslexia. He suggests this is the result of the connectivity between

different systems and pathways in the brain. At the neurophysiological level, this

can be seen with anomalies that begin in one place in the brain and propagate to

other areas. Due to the connections between the frontal and temporal cortex

(cognitive processors) and the visual and auditory areas of the thalamus (perceptual

processors), damage in one area will cause problems in the other. Hence,

anomalies occurring within both perceptual and cognitive systems result in effecting

the other.

Summary and Concluding Remarks
The empirical evidence reviewed in this chapter points to the fact that

readers do make use of the various linguistic structures encoded in scripts. In

particular, studies indicate that morphemes and syllables as well as segments play a

role in visual processing. Moreover, this processing evidence is simply counterpart

to the fact that scripts represent multiple aspects of linguistic systems. In spoken

language, phonemic segments, syllables, onsets, rimes, morphemes and words are

all elements of linguistic representations that are active simultaneously. It is

unsurprising then that orthographies, whose function is to represent spoken

language, reflect these components and further, that readers make use of them for


The facts of multileveled representation and processing have several
implications that are relevant to reading acquisition. The most important implication

to this discussion is that the diversity and number of cues readers extract from the

signal must effect the learning process. Orthographic cues are several and of

several types. As discussed, scripts typically cue several aspects of the spoken

language simultaneously. Also important is that while scripts represent many

things at once they do so only partially. That is, they do not, probably can not,

represent all the information in the spoken signal. Additionally, information about

the linguistic structures presented in scripts is not uniformly obvious. For example,

linguistic cues may sometimes be in the form of discrete units, e.g., a

segment/letter, a syllable/syllabogram, or a morpheme/logogram. Yet, very often,

units such as segments, syllables and so on, as well as larger units like words, are

indicated by grapheme patterns. Sometimes, information about a particular unit is

indicated by cues outside the boundary of the unit itself. For example, double

consonants in English script give information about a preceding vowel and often

across a syllable boundary, e.g., a-pple. Information about morpheme and

syllable boundaries are reflected by letter sequence frequencies. Cues are often

overlapping and sometimes conflicting. For example, morpheme and syllable

boundaries may overlap as in pre-view or they may conflict as in press-ing and


Scripts have adapted to serve a communicative purpose. Hence, the

organization of information in scripts is unlikely to have arisen by accident. Instead

it is likely that orthographic organization is functional and there is some processing

advantage for information to be structured in this fashion. Indeed, it stands to

reason that it is because of processing constraints that scripts have the formal

properties that they do. The final question is then, what are these constraints and

how do they influence learning? That is, what type of learning system can


accommodate the facts (as we know them so far) of multileveled representation and

processing. Learning an orthographic system involves attending to a set of

complex and often diverse cues. Furthermore, reading instruction most often

provides learners with only a subset of the relationships they must internalize.

Hence, I suggest that reading involves learning mechanisms -- implicit learning --

that are particularly "good at" processing information that is complex, incomplete,

and at least partly unavailable to explicit, conscious learning strategies.


Traditional methods in education presume that learning is an intentional

process that involves a conscious awareness of the nature of the stimulus

environment and the application of explicit attention to problem solving tasks.

Likewise, traditional models of cognition (Newell and Simon 1972) posit an

explicit information processing model of human learning, where knowledge is

sequentially constructed by goal-directed processes. Explicit cognitive processes

include reasoning, problem solving, overt recall from long and short term memory,

awareness of self, and so on. However, information that is complex, large in

quantity, and for which stimulus cues are incomplete or conflicting is not easily

apprehended by explicit processes. Learning such information typically proceeds

automatically and without conscious control. Hence, in many (perhaps most)

learning situations it appears that a good proportion of knowledge is acquired in an

unintentional, non-reflective manner. This type of knowledge acquisition has come

to be called implicit learning.

Implicit learning is a process whereby complex knowledge of richly

structured stimulus domains is, for the most part, acquired independently of

awareness of both the acquisition process and the knowledge base that is acquired

(Winter and Reber 1994, Reber 1967, 1976, 1989, 1993, Berry and Dienes 1993,

Berry and Broadbent 1984, 1988, Mathews et al. 1989, Mathews 1991,

Cleeremans 1993). This learning is not assumed to take place in a wholly

incidental way, however. It is conceived as a by-product of attention to relevant

rule-governed structures, be it natural language, socially prescribed behaviors or of

complex economic systems (Winter and Reber 1994). That is, while the learner

explicitly attends to certain salient stimuli, underlying structured information is

implicitly induced and represented in a tacit knowledge base. Hence, implicit

learning is a general inductive process that extracts patterned relationships from the

environment, creating a tacit knowledge base.

This is not to suggest that complex structures cannot be learned through

conscious analytical strategies. The point rather is that implicit operations can be

engaged in the analysis of structured patterns, even when these regularities have not

been consciously detected (Winter and Reber 1994). Furthermore, explicit

processes are resource intensive, requiring explicit memory, conscious attention,

and overtly controlled hypothesis testing. On the other hand, implicit mechanisms

are automatic, unconscious and as such, less resource consuming, thus freeing up

explicit mechanisms for higher order functions. Implicit mechanisms may therefore

be 'better at' learning structures that are too complex for conscious analytical


Empirical Studies
Implicit learning experiments have explored the process by which people

acquire complex knowledge about the world, largely without conscious attempts to

do so. Three different implicit learning paradigms have yielded consistent and

robust results: artificialgrammarlearning (see Reber 1993 for a comprehensive

review), process control (see Berry 1993 for a comprehensive review), and

sequence learning (see Dienes and Berry 1993 for a comprehensive review). The

following review does not attempt to address all the complex issues that have arisen

through this research. My goal is primarily to illustrate the kind of systems and

tasks that have been shown to consistently engage implicit operations, and to

highlight some of the issues that I believe relevant to language and reading


Grammar Learning

Artificialgrammarlearning is the most prominent and widely replicated of

the implicit learning paradigms. Grammar learning has been used particularly to

examine the relationship between implicit learning and conscious awareness. These

experiments, first explored by Reber (1965,1969), have subjects attend to complex

rule-governed stimuli (usually strings of letters or geometric shapes) generated by a

synthetic, semantic free, Markovian, finite-state grammar. There is a well-

developed branch of finite mathematics that underlies these grammars. Further, the

stimulus domains are novel and arbitrary so as to rule out the effect of preexisting

knowledge. These systems are too complex to be learned by simple testing of

consciously held hypotheses.

1.TXS 5. PVV

Fig. 5.1 The first and simplest of the Markovian artificial grammars used by Reber
(1965) with several grammatical strings that it generates (Reber 1993).

The typical study proceeds in two phases. Subjects first memorize strings
of letters generated by the synthetic language. Later they are tested for their

knowledge of the grammar's rules by being asked to make decisions concerning the

well-formedness of novel strings of letters. Subjects are not informed that they are

working with rule-governed stimuli but are asked only to memorize the strings of

letters printed on cards. Another group of subjects perform the same task but with

random strings of letters. Twenty learning stimuli are used in the acquisition phase,

usually three to eight characters in length.

In experiments of this type, there were some interesting differences between

the subjects presented with the grammatical strings and those presented with

random strings. With practice, both groups of subjects became increasingly good

at this task. However, subjects who worked with grammatical strings made an

average of eighteen errors before being able to reproduce all four strings of the first

set correctly. By the seventh set they made an average of three. Subjects who

worked with the random strings showed a different pattern. They too began with

eighteen errors with the first set but after several sets make an average of eight

errors. The improvement that occurred among both groups was attributed to a

"learning to learn" effect that occurs regardless of structure. The difference in

performance between the two groups was attributed to the first group's ability to

exploit the structure in the stimuli. Furthermore, the subjects that confronted the

rule-governed strings were then able to use what they had learned about the rules of

the grammar in order to discriminate new strings that conformed to the grammar.

These findings are very robust and have been replicated many times under diverse

conditions (Reber 1965, 1969,1976; Brooks 1978; Fried and Holyoak 1984;

Mathews, Buss, Stanley, Blanchar-Feilds, Cho and Druhan 1989; Morgan and

Newport 1981; and Servan-Schreiber and Anderson 1990).

Equally interesting is the observation that subjects transfer implicitly

inquired knowledge from one environment to another. Reber (1969) asked subjects

to memorize a set of strings generated from a finite state grammar. He later

presented them with the task of memorizing a set of strings generated from the same

grammar, but with a different set of letters, or a different grammar. He found that

subjects' performance improved on the second set of strings when the set was

generated from the same grammar, even when the strings were made up of different

letters. Mathews et al. (1989) replicated this basic finding. The researchers

concluded that subjects learned abstract regularities of the stimulus network.

Another interesting issue is the interaction of explicit leaning strategies with

implicit learning. Reber (1976) found that encouraging subjects to "look for rules"

during the memorization phase resulted in a decrease in performance on the

grammaticality test, indicating some kind of interference effect. Reber concluded

that subjects were at a disadvantage when they explicitly searched for patterns

because the properties of the stimulus array (Markovian grammar) were not easily

detected by conscious problem solving strategies. That is, the grammatical rules

were too complex to be deciphered by consciously held hypotheses. However,

when the rules of the stimulus array were made more salient by manipulating the

properties of the material, Reber, Kassin, Lewis and Cantor (1980) found that the

interference caused by explicit rule seeking disappeared. They concluded that

looking for rules works only if you can find them.

Further experiments found that giving the subjects specific information

about the stimulus array effected learning. Reber et al. (1980) addresses this issue

with another grammar learning procedure that gives subjects both an explicit

training period and a passive observation period. Three groups of subjects were

given a 7-minute course using a schematic diagram of the grammar which illustrated

the rule system and were instructed how the grammar generated strings. They also

simply observed a set of correct exemplars. This explicit training was combined

with the implicit learning procedure (observation only) in various ways. All three

groups of subjects received the explicit training session, but received the training at

different times. One group received training at the outset; another at midcourse,

when some of the exemplars had been seen. The last group received training at the

end, when all of the exemplars had been seen. As usual, learning was tested by the

subjects' grammaticality judgments.

The authors found that the earlier the explicit instructions are given, the

more effective they were. Reber suggests that providing explicit instructions at the

outset, provided they are correct, serves to focus and direct the subjects' attention.

The instructions given did not teach the grammar in any complete way, but simply

oriented the subjects toward the relevant invariances in the stimuli, so that the

subjects were able to teach themselves. On the other hand, when subject were

given explicit training during orafter the implicit observation period, the training

was less likely to facilitate learning. The authors suggested that giving explicit

information about the system after subjects have passively observed it imposed a

formalization of structure not in concord with the system that subjects were

inducing. Moreover, individual learners may build representations that vary quite

widely from each other. Several researchers have observed that there seems to be

wide variation among individuals in regard to implicitly acquired knowledge of

artificial grammars (Dulany et al. 1986; Reber 1989, Mathews et al. 1989).

Holland et al.'s (1986) competitive rule induction model predicts wide individual

variability in the initial features induced from artificial grammars because many

different cues (exemplar fragments) could be used to select valid strings. Once a set

of cues become strengthened by a sufficient number of successful predictions, there

is no reason for the learner to further modify the knowledge base. Consequently,

the learners' representations are likely to vary across individuals. Likewise, learner

representations are likely to differ from the formalization represented by the

schematic of the Markovian grammar. Hence, if information about the grammar's

formalization is given after learners have developed their own representations of the

system, it causes interference and learning delay.

Mathews et al. (1989) further investigated the interaction of

implicit/explicit processes on learning. They compared subjects performance with

two types of artificial grammars: one Markovian finite-state grammar and one based

on simple logical rules. The authors hypothesized that, where structure is opaque

as in the case of the finite state grammar, explicit learning strategies should not

facilitate learning beyond what was implicitly induced. The grammar based on

simple logical rules however, should be accessible to explicit processes. In order to

examine the interaction between implicit an explicit learning with both types of

grammars, two types of tasks were devised. The first was an implicit "match"

short-term memory task. Subjects were presented with a single valid string from

the grammar and told to hold it in memory for a few seconds until five choices

appeared on the screen. They were then asked to select the identical string from

these choices, unaware that the strings were rule generated. The second task was

an explicit "edit" task. Subjects were exposed to the same set of items, generated

by the same grammar as in the match task. Yet, one of five letters in each string

was incorrect. Subjects were told that they would see items that were flawed

strings generated by a grammar. Their task was to figure out the rules of the

grammar so that they could identify the incorrect letter in each string.

There were five experimental groups. One group only performed the
implicit match task, another only the explicit edit task. The other three groups

performed both tasks in equal amounts. Of these groups, one (match/edit) did the

implicit match task first on half the trials, then the explicit edit task on the other half.

Another, (edit/match) did the edit task first then the match. The last group

alternated match and edit trials.

The results using the finite-state grammar were that there were no significant

differences among groups, though the groups that did the match task only and the

match task first performed slightly better than the others. The authors concluded

from this that implicit induction is capable of extracting complex patterns without

conscious attempts to discern those patterns, and that explicit learning strategies did

not aid knowledge acquisition of a finite state grammar. These findings support

earlier conclusions (Reber et al. 1980) that explicit learning strategies do not aid

learning unless the structure is sufficiently obvious to find.

There were significant differences among experimental groups however,

when the same experiment was done using the simple, logical rule grammar. The

match/edit group performed significantly better than all the other groups. The

match only group performed the worst, nearly at chance levels. These results

suggest that implicit learning is operative both in environments where structure is

quite opaque (finite-state grammar), and where structure is more salient (logical

grammar). Also, it appears that when stimuli are obvious enough for explicit

learning strategies, the interaction of implicit and explicit learning produced a

significantly better understanding of the system, indicating a "synergistic effect."

That is, when implicit and explicit functions are allied, learning is more effective

than it would otherwise be. Moreover, the sequential order of implicit and explicit

learning is relevant. The match/edit group's greatly superior performance suggests

that a period of implicit observation benefits learning a more salient system. That

is, for systems equivalent to the simple logical grammar, the optimal procedure for

learning is to develop an implicit knowledge base before beginning to generate an

explicit model of the system.

Another relevant issue that has been explored with the grammar learning

paradigm is the availability to consciousness of the implicit knowledge base.

Though early research with implicit learning claimed a clear distinction between

implicit and explicit processes, most researchers now agree that implicitly acquired

knowledge is not wholly unconscious, and that various types of tacit knowledge

bases will differ in degree in its availability to consciousness. Reber and Lewis

(1977) suggest that a tacit knowledge base may become available to consciousness,

but the ability to explicitly state knowledge acquired from implicit learning

processes always lags behind what is known unconsciously. In a four-day study

subjects solved anagram puzzles generated by an artificial grammar. Subjects were

then asked to introspect and write the rules of the grammar, or the rules and/or

strategies they were using to solve the puzzles. Over the four days, there was a

general increase in the subjects' ability to communicate their knowledge of the

system. Yet, there was also an increase in ability to solve the anagrams. As

subjects increased their ability to verbalize the system's rules, they also acquired a

richer knowledge of the system. Consequently, task performance was always

ahead of explication.

The Mathews et al. (1989) study also addressed the consciousness issue.

The authors used a "teach aloud" technique whereby subjects were asked during the

task to stop and to explain to another subject26 what rules or strategies they were

using to learn the grammar. This information was given to second control group of

subjects who were tested in the same well-formedness task. With this information,

the second group performed about half as well as the first group. As the

experiment progressed and the proficiency of the first group increased, so did that

of the control group, yet those in the control group never caught up with the first

26 Subjects receiving instructions about the grammar were not present. Instead the subjects
giving instructions were tape recorded. They were told to record instructions for their partner so
that the partner could perform the task "just like you did."

group. The authors concluded that implicit knowledge of the artificial grammar is

available to consciousness, though incompletely. Moreover, the ability to examine

and report the knowledge acquired improves with practice, but more slowly than

the acquisition process.

A final issue involving an important aspect of the teach aloud procedure was

the type of information subjects communicated. Virtually all subjects expressed

their knowledge in terms of instructions to select or avoid certain letter groups.

For example, "select strings that begin with SCP" or "select strings that end in

VV." Successive verbal reports over trials included elaborations and exceptions to

"rules" provided on earlier trial blocks. This information included both item

specific and more general features of the grammar. Mathews et al. (1989)

suggested that subjects select patterns of common features (groups of letters) shared

by many exemplars. This is the type of acquired knowledge predicted by Holland

et al.'s (1986) rule induction model and is consistent with the idea that learners

build a body of knowledge by encoding small subsequences of strings (Reber and

Lewis 1977 and Servan-Schriber and Anderson 1990). For instance, it is

hypothesized that, given a string TTXVPXVS to be committed to memory, subjects

may first create string "fragments" or "chunks" TTX VP XVS. These chunks both

facilitate an constrain further learning. For example, the following string:

VXVPXXXVS is likely to be perceived in terms of previously analyzed chunks VX

(VP) XX (XVS). Hence, the new string seems less complex and constrains further


Grammar learning and natural language learning. Nations and McLaughlin
(1986) use an artificial grammar learning paradigm in order to suggest an

information processing approach to language learning. They proceed from the

assumption that mastery of complex tasks involve the integration of controlled

processing -- which is resource intensive, requiring a large amount of processing

capacity, and automaticprocessing -- which is low demand and requires little

processing energy27. They contrast three groups' performance in a grammar

learning task: monolinguals, bilinguals and multilinguals. All subjects participated

in two phases of the experiment. The first phase consisted of an implicit learning

task, which required subjects to attend to the stimuli without instruction as to what

they should learn, while the second phase involved an explicit task whereby

subjects were informed that the stimuli followed rules and were instructed to try to

discover them. After each phase subjects were asked make grammaticality

judgments of exemplar strings. The results were that multilinguals performed

significantly better on the implicit learning task than either of the other groups and

that there were no differences between groups on the explicit task.

The authors hypothesize that strategies used by "expert" learners

(multilinguals) differ sharply from "novice" (monolinguals and bilinguals) learners.

Basically, the contention is that multilinguals are successful with the implicit task

because they have automated the basic strategies of pattern recognition, while the

novice learner groups have not. Relevant to the greater discussion is that

multilinguals superior performance suggests that pattern recognition is a significant

part of the language learning task. Moreover, expert learners tend to use implicit

learning strategies in the proper environments, which allow them to allocate their

cognitive resources more efficiently than novices.

27 The automatic/controlled distinction is generally used to account for the progression from a
slower, more cognitively demanding stage to an automated stage of performance. Automatic
processing occurs when a particular response has been built up by processing the same input over
many trials. Controlled processing is deliberate, temporary activation of memory though
attentional control (Schneider and Shiffrin 1977; Anderson 1983; McLaughlin 1990). While the
authors use an implicit learning paradigm they do no specify how exactly the automatic/controlled
distinction correlates with the implicit/explicit distinction or specify the differences between them.
The relationship between implicit/ explicit and automatic/ controlled is not necessarily equivalent
since explicitly learned material can become automated. Implicit learning, also, becomes more
automated with practice. However, I believe that the results of this experiment contributes to the
empirical data on implicit learning and bears on the overall discussion.

Process Control

Other evidence that demonstrates implicit learning comes from process

control studies which have found that people are able to control complex systems

even when they lack the ability to verbalize the knowledge required to perform such

tasks. Berry and Broadbent (1984) looked at the effects of the learning task, verbal

instruction and subject's ability to verbalize their knowledge. Subjects were

required to control the output of a simulated system, a sugar production plant for

example, by setting the value of one or more input variables. In the role of manager

of a sugar production plant, subjects were asked to modify production to achieve

some preassigned level. The control was implemented by adjusting production

variables such as wages, labor peace, and number of workers hired. The computer

was programmed with rules that related production levels to these factors. The

subjects' task was to make shifts in the value of the relevant variable in an attempt

to achieve the target production levels. The computer interacted with the subjects

by displaying production levels in response to their adjustments. Following the

experience subjects were required to complete a written questionnaire that asked

about the relationships within the system.

The results were that task performance improved significantly with practice

but performance on post-task questions did not improve. In fact, subjects' post-

task responses were no better than compared to subjects with no experience at all.

Conversely, explicit verbal instruction on how to reach and maintain the target value

significantly improved ability to answer questions, but had no effect on task

performance. However, explicit instructions did have a beneficial effect on task

performance when subjects were also required to verbalize a reason for each control

value they chose. Overall, however, there was no evidence of a positive

association between task performance and question answering. Also, across three

experiments with similar types of tasks, subjects who were better at controlling the

task were significantly worse at answering questions. Berry and Broadbent

concluded that it was likely that these type of tasks were learned implicitly.

Broadbent, Fitzgerald and Broadbent (1986) replicated these findings

using a city transport system control task and an economic model task. Again

results showed that task performance improved with practice but the same did not

hold with respect to answering questions about the task. Other studies (Hayes and

Broadbent 1988) indicated that inducing subjects to use explicit control strategies

(like Reber's "looking for rules" in an artificial grammar) resulted in a decrease in

task performance. This suggests that, like Markovian grammar learning, when the

system regularities are sufficiently opaque, explicit learning strategies do not help.

Likewise, Berry and Broadbent (1987,1988) found that when they manipulated the

stimuli such that it became more salient, subjects were able to answer explicit

questions about the material and performance was correlated with explicit


These results are strikingly similar to the results of the grammar learning

task. It appears that looking for structure only helps if structure is easy to find .

Similarly, it appears that whatever type knowledge base is acquired, it is difficult to

verbalize. Stanley, Mathews, Buss and Kotler-Cope (1989) showed that, in a task

which asked one group of subjects to explain the system to other subjects, the first

group of subjects attained high levels of performance long before they were able to

communicate that knowledge. Moreover, typically in process control studies,

subjects report that they made certain decisions because they "feel right" or may

simply believe that they are guessing (Berry and Dienes 1993).

Sequence Pattern Acquisition

Sequence pattern acquisition involves three different types of tasks: choice

reaction time tasks, prediction tasks and probability learning tasks. The rationale

underlying the sequence learning paradigm is that responses reflect sensitivity to

sequential structure present in the material (Cleeremans 1993). In a choice reaction

task (Lewicki, Hill and Bizot 1988) subjects were presented with a complex matrix

of numbers arranged in four quadrants on a computer screen. The task was to note

the location of the target number by quickly pressing a button indicating the proper

quadrant. A continuous sequence of target numbers appeared in one of the four

quadrants. The sequential structure of the material was manipulated to generate

sequences of five elements according to a set of simple rules. Each rule defined

where the next stimulus appear (e.g., which of four quadrants of the computer

screen). Locations were a function of the position of the two previous stimuli. The

first two elements of each sequence was unpredictable, while the last three were

determined by their predecessors. The dependent measure was reaction time.

Subjects were unaware that the sequence of successive trials contained regularities.

With practice subjects whose trial sequence was structured showed a decrease in

reaction time that was significantly greater than subjects who observed material with

no sequential structure. Also, subjects were unable to articulate the rules they

appeared to be using.

Prediction tasks use essentially the same design, but instead of reacting to

the current stimulus, subjects are asked to predict the next stimulus or where it will

appear. In this case the dependent measure is the percentage of correct predictions

over trials. The usual result is that subject's ability to predict the location of the

next trial improves with practice, yet they are unable to articulate how they came to

their decision (Cleeremans 1993).

Experiments using the probability learning task, introduced by Millward and

Reber (1968, 1972), asked subjects to observe a sequence of events and predict

what comes next in the sequence. Subjects anticipated the changing probabilities of

events even when the anticipatory response required and integration of information

across 50 preceding events (Reber 1993). The experiment consisted basically of

subject's watching two lights rapidly flashing on and off in particular patterns.

Subjects were given a thousand trials at the rate of two per second and were asked

to predict the successive events before they occurred. The probability of any event

was systematically increased and decreases as they moved through the trials.

Subjects learned to anticipate shifts in the likelihood of events but their performance

was random when asked to recall crucial trials by which they formed their

decisions. The findings from sequence pattern acquisition experiments are similar

to that of the other learning paradigms in the following respects. Subjects appear

to be exploiting the structure of the stimuli in order to perform the task. Learning is

gradual and improves over many trials. Subjects are not easily able to identify how

they made their decisions or articulate the "rules" they seem to be using.

Learning Mechanisms and
Knowledge Representation
Empirical research has led to much speculation about the mechanisms by

which implicit learning occurs. In fact, there is much debate in the literature as to

the nature of implicitly acquired knowledge. Reber's early definition of implicit

processes gave them the general characteristics of producing an abstract, tacit

knowledge base that is representative of a structured environment. He further

suggested that, implicit learning optimally occurs independently of conscious

attempts to learn. Ongoing research has found that this view to be an

oversimplification. The idea that implicit learning produces an abstract and

unconscious knowledge base has been particularly disputed. This is largely

because of the difficulty of demonstrating (or defining) the notions of

"abstractness" and "consciousness". In general, implicitly acquired knowledge has

been consistently shown to be resistant to conscious inspection. Yet, empirical

research has indicated that depending on the system being learned and the type of

task employed in learning, the knowledge base will vary in its availability to

conscious awareness. That is, conscious awareness is a gradient property of

implicit knowledge.

Abstractness also appears to be a matter of degree. Early research

characterized implicit knowledge as abstract rule-based knowledge. What was

generally meant by abstraction was that the implicit knowledge is not about specific

exemplars or based on relevant features of these exemplars, but instead that the

knowledge is rule-like. However, there is abundant evidence that what subjects

acquire is not an abstract rule-governed schematic (like the actual schematic of the

Markovian grammar), but fragmentary knowledge about local constraints. Hence,

the acquired knowledge is not abstract or rule-governed in this early sense. Later

rule-based theories specify relationships between exemplar fragments, however.

This network of local relationships are thought to be abstract in nature. For

example, Reber and Lewis (1977) suggested that bigram or trigram invariance rules

are used to select valid strings. Holland et al. (1986) posit a competitive rule

induction system in which local (fragmentary) features are abstracted from

exemplars. The cues identified become strengthened with each successful

prediction. Implicit learning processes continually modify the strength of

competing rules.

Some researchers adopt the view that implicit learning of exemplar

fragments involves a memory association-based system where the resulting

knowledge base need not be abstract at all. As described above, learning involves

chunkingg" short groups of elements to form a network of local relationships,

however, grammatical knowledge is tied to a specific letter set (Servan-Schreiber

and Anderson 1990). In fact, most researchers now acknowledge that encoding

small substrings is central to the grammar learning task (Dulany et al. 1984;

Perruchet and Pacteau 1990; Reber and Lewis 1977; Holland et al. 1986; Mathews

et al. 1989; Reber 1993; Cleeremans 1993; Winter and Reber 1994). This is true

for rule induction models as well as memory based models. The debate now turns

on the issue of whether the resultant network of chunks becomes an abstract

knowledge base (Reber 1989,1993; Reber and Lewis 1977; Mathews et al. 1989),

or whether grammatical knowledge is tied to a specific letter set (Servan-Schriber

and Anderson 1990; Perruchet and Pacteau 1990). The most convincing evidence

that tacit knowledge is abstract is subjects' ability to transfer their grammar

knowledge to letter strings they have never seen before. This effect has been

replicated many times. Mathews et al. (1989) suggest that while grammar learning

relies on association-based chunking, the resultant knowledge, a network of

chunks, is abstract. Moreover, this combination of item specific and general-

abstract information is the type of acquired knowledge the Holland et al. (1986) rule

induction model predicts.

The issue of rule-based versus association-based implicit learning

mechanisms has been very difficult to sort out empirically. Indeed, the mental

representation of structure may fall between storage of specific instances in

memory, to more general abstract descriptions. In fact, systems based on memory

for specific exemplars and systems based on abstraction of rules may be

functionally equivalent (Cleeremans 1993). That is, although these respective

models make different predictions about knowledge representation, there is

considerable overlap in terms of their simulating human performance. While this

continues to be a theoretical point of contention, most researchers agree that implicit

learning tends toward the following characteristics:

* Implicitly acquired knowledge is fragmented, consisting of many units (e.g.,

"chunks" or "fragments") that encode local information (Dulany et al. 1984;

Perruchet and Pacteau 1990; Reber and Lewis 1977; Holland et al. 1986;

Mathews et al. 1989; Servan-Schriber and Anderson 1990; Reber 1993;

Cleeremans 1993; Winter and Reber 1994).

Implicit learning and the implicit knowledge base are strongly constrained by the

learning task and subjects' (cognitive) ability to perform that task. For

example, learning the finite-state grammar will be constrained by whether the

task is to memorize the strings or to figure out the system of rules that generate

the strings. The various experimental paradigms reviewed differ greatly in the

nature of the task and hence will point to different constraints on processing

(Reber etal. 1980; Berry and Broadbent 1989; Hayes and Broadbent 1988;

Mathewsetal. 1989).

Processing is unselective (Berry and Broadbent 1988). Implicit learning

strategies allow the system to remain sensitive to all dimensions of the stimulus,

at the cost of slow, gradual learning. By contrast, explicit learning selects

specific dimensions of the stimulus which (if correct) results in faster learning.

Implicit learning proceeds automatically, unintentionally and is resistant to

conscious inspection. However it can be used to guide subsequent task related

judgments and decisions (Berry and Broadbent 1988; Reber 1989; Berry and

Dienes 1993; Winter and Reber 1994)

For most learning situations, both implicit and explicit mechanisms operate

simultaneously. I will elaborate on this point in the next section.

Implicit versus Explicit Learning Mechanisms

Although implicit learning experiments attempt to isolate implicit processes,

the literature suggests that both implicit and explicit mechanisms are involved in

learning complex systems. Indeed, many of the tasks (in this review) used to elicit

implicit processes involve both explicit and implicit learning. Furthermore, implicit

and explicit learning involve processes which may vary in their accessibility to

conscious awareness, degree of generality, and degree to which these processes

facilitate each other. The relevant question then becomes, what factors are likely to

influence the implicit/explicit learning interface for a particular task and a particular

system to be learned?

Cleeremans (1993), following Holland et al. (1986) suggests that particular

systems and particular tasks will elicit learning strategies that range from purely

implicit to explicit depending on "regularity salience" and "task demands." The

regularity salience of a particular system depends on a mixture of the following: 1)

The number of cues and features present in the stimuli. Implicit learning is active

typically when the number of cues and features present in the system exceeds the

learner's ability to attend to them. Where a system's features are restricted enough,

learners are likely to engage explicit problem solving techniques. 2) The amount of

external feedback about the specific relevance of features and cues. In implicit

learning situations, external feedback is usually minimal. Explicit learning is

typically guided by feedback. 3) The validity of cues. Implicit learning situations

are typically noisy environments, where cues are only partially valid and have many

exceptions. Systems where cues are valid, complete and obvious are likely to

engage explicit reasoning strategies.

A learning situation that has a very wide or diverse number of cues, little

external feedback, and presents learners with noisy environments would have low

regularity salience. Indeed, research has indicated that the lower the regularity

salience, the more likely learners will engage in "pure" implicit learning, and the

less likely they are to explicitly seek regularities. On the other hand, the higher the

regularity salience, the easier regularities are to find, the more likely learners will

engage in explicit problem-solving. The demands of a particular task will interact

with regularity salience and influence learning. The higher the task demands, the

more likely learners will attempt to use explicit strategies. For example, if learners

Full Text
xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID EZF57ZS0U_6GMB82 INGEST_TIME 2013-03-25T13:00:55Z PACKAGE AA00013598_00001