Winner of Single Search Result Contest
My Googlewhack
Mentioned in the article on
Cambridge Spelling.
Note for the pedantic: OK, alright, this isn't officially a googlewhack
since it contains more than
two words.
Google Search: cambridge "order of letters" "first and last letter"
Dissatisfied with your search results? Help us improve.
©2003 Google
Sublexical
units and the split fovea
An anatomical perspective on sublexical units: The
influence of the split fovea
Richard Shillcock & Padraic Monaghan
University
of Edinburgh
January 10, 200311,522
Abstract
We discuss the problem of how to represent the internal structure of English words, and
consider the solutions adopted in a number of implemented models of visual word
recognition and naming. We describe two sets of simulations with the split-fovea model,
an implemented connectionist cognitive model of single-word reading, whose
architecture is based on the precise vertical splitting of the human fovea. We show that
the model can capture critical human data from two effects concerned with the parallel
activation of lexical competitors: (a) the transposed letters effect, in which pairs of words
like salt and slat, or clam and calm, interact during processing, and (b) the neighbourhood
effect, in which large lexical neighbourhoods facilitate naming. We discuss the results in
terms of the coarse coding generated by the architecture of the split-fovea model and the
naming task. Finally, we consider some of the implications for the processing of different
languages and different orthographies, and for language impairment.
An anatomical perspective on sublexical units: The influence of the split fovea
We have demonstrated elsewhere that the fact that the human fovea is precisely vertically
divided is a crucial point of departure for the computational modelling of visual word
recognition (see, e.g., Monaghan, Shillcock & McDonald, submitted; Shillcock, Ellison
& Monaghan, 2001; Shillcock & Monaghan, submitted). Researchers have proposed a
number of solutions to the question of how to represent the structure of words in reading.
In this paper we review some of these solutions and situate our own approach with
respect to them. We then present new simulations showing that the split-fovea model can
capture human data from experiments on the interaction of words related by letter
transposition, such as slat and salt, and on the recognition of words with large or small
lexical neighbourhoods.
One of the strengths of the computational modelling approach is that it requires
the researcher to be explicit about representational assumptions. In the computational
modelling of visual word recognition
and naming, the representation of the visually
presented word must be specified; the positions of the constituent letters must be coded.
How does information about letter location participate in lexical processing? Word
recognition is a task for which mammalian visual cognition did not specifically evolve.
The human visual system has evolved to remember and to distinguish such things as
conspecifics, objects and places in the natural world. Words, however, are unlike any of
these things. They are typically asymmetric, and one half of a word does not usually
determine what the other half contains. Every letter, at every point in the word, counts
towards the identity of the word. Finally, there are tens of thousands of words in an adult
lexicon, all different and unique. Given the nature of the problem, we are entitled to
return to first principles. We will first consider some of the cognitive requirements of
visual word recognition. We will then look at two observable effects, the transposed
letters effect and the neighbourhood facilitation effect, that have implications for how
items in the lexicon compete to be matched against a visual stimulus. We will discuss a
number of models of lexical processing from the perspective of these two effects, and
draw general conclusions about how best to specify letter position. Finally, we will
present simulations of the two effects, using an implemented version of the split-fovea
model. Our example language . and that of the simulations described . will be English,
but we will consider the issue of different orthographies when we explore one of the
critical dimensions in lexical processing, the coarseness of the coding required.
The requirements of visual word recognition
We will frame our discussion of word recognition in terms of two related questions: What
types of information are used, and how much information is necessary? In naturalistic
reading tasks, the participant needs to distinguish the word being read from all of the
other possibilities in the lexicon. Efficiently achieving this goal may not require the brain
to access all of the potential information about the word being read, or about the words
stored in the lexicon. (At the same time, it should be appreciated that the many different
relationships between words are often highly intercorrelated, and there is considerable
redundancy.) If the reading task is relatively demanding - for instance, proofreading,
reading degraded stimuli, or rejecting particularly word-like nonwords . we can expect
more of the potential information available to be accessed, and for strategic effects to
become more evident. In general, in this discussion, we will be concerned with the more
naturalistic reading tasks, particularly naming.
In order to appreciate the role of information about letter position and sublexical
units, it is useful to distinguish between the global and the local information available
about the word being recognized. Perhaps the most salient aspect of global information
about a word is its length: in the normal reading of text, it may be available parafoveally,
and it plays a pervasive role in determining fixation behaviour (e.g., just & Carpenter,
1980; Rayner, 1979). Indeed, word length might be a good candidate for the early, fast
uptake of low spatial frequency information by the magnocellular pathways. Word length
is a powerful predictor of lexical processing in certain tasks....
From index
