In: Proceedings of CoNLL-2000 and LLL-2000, pages 13-18, Lisbon, Portugal, 2000. 
Pronunciation by Analogy in Normal and Impaired Readers 
R.I. Damper 
Image, Speech and Intelligent 
Systems Research Group, 
Department of Electronics and 
Computer Science, 
University of Southampton, 
Southampton SO17 1B J, UK 
Y. Marchand 
Cognitive/Clinical Neuroscience Unit, 
Department of Psychology, 
Dalhousie University, 
Halifax, Nova Scotia, 
Canada B3H 4J1 
Abstract 
The prevailing dual-route model of oral read- 
ing claims that a lexical route is used for the 
pronunciation of words and a non-lexical route 
processes nonwords. Neurological data from pa- 
tients with acquired dyslexias have been high- 
lighted to support this claim. Models using a 
lexicon alone are generally held to be incapable 
of explaining these data. However, by selec- 
tively impairing its component parts, it is easily 
possible to account for phonological and surface 
dyslexias using a single-route model based upon 
pronunciation by analogy. 
1 Introduction 
We have previously developed pronunciation by 
analogy (PbA) as a model of reading aloud and 
as a method for automatic phonemisation in 
text-to-speech synthesis (Sullivan and Damper, 
1993; Damper and Eastmond, 1997; Marchand 
and Damper, 2000). We have also demon- 
strated (Damper et al., 1999) that the perfor- 
mance of PbA in producing correct pronunci- 
ations is vastly superior to manually-written 
rules and significantly better than the competi- 
tor data-driven techniques of back-propagation 
(Sejnowski and Rosenberg, 1987; McCulloch et 
al., 1987) and the IBloIG method based on 
information gain weighting (Daelemans et al., 
1997). Although we cannot claim that PbA is 
absolutely the best method for pronunciation 
generation, it must be taken seriously. This 
view is clearly shared by other workers who 
are actively developing analogical methods for 
natural  processing tasks (Pirrelli and 
Federici, 1995; Jones, 1996; Yvon, 1996; 1997; 
Bagshaw, 1998; Pirrelli and Yvon, 1999). 
Explicit analogy (e.g., Dedina and Nusbaum, 
1991; Damper and Eastmond, 1997) retains the 
lexicon in its entirety, typically as a list of words 
and their spellings. PbA requires a dictionary in 
which text and phonemics have been aligned, so 
that pronunciations corresponding to matching 
orthographic substrings can be identified. How- 
ever, many of the necessary computational steps 
to assemble a pronunciation can be carried out 
in advance. Thus, in implicit analogy (e.g., Sul- 
livan and Damper, 1993), the lexical database is 
precompiled to yield a generalised phonological 
knowledge base which is consulted during pro- 
nunciation generation. This done, the (explicit) 
dictionary can be discarded. Implicit analogy 
may also attempt to compress the training data, 
so that some proportion is discarded. 
Here, we extend earlier work on modelling 
pronunciation by normal readers to impaired 
readers with acquired dyslexias. There are sev- 
eral forms of this: two of the most important 
are phonological and surface dyslexia. Cases 
of phonological dyslexia display good ability 
to read words (both regular and irregularly- 
spelled) aloud but poor nonword reading abil- 
ity (Beauvois and D~rouesn~, 1979). In surface 
dyslexia, however, patients misread irregularly 
spelled words, which tend to be regularised in 
their pronunciation (Coltheart et al., 1983). To 
simulate these dyslexias, we use explicit PbA 
without compression. The approach is to dam- 
age the model and then to observe its ability to 
replicate the neuropsychological data. 
2 Dual and Single Routes to Sound 
The nature of the cognitive processes underly- 
ing the act of reading aloud has spawned an im- 
portant and controversial debate in psychology 
(Humphreys and Evett, 1985; Seidenberg and 
McClelland, 1989; Coltheart et al., 1993; Plaut 
et al., 1996). One popular view is that there are 
13 
two routes from print to sound: a texical and a 
nonlexical route (Coltheart, 1978). The former 
involves access to lexical knowledge for familiar 
words. The second route concerns the pronunci- 
ation of unfamiliar words or pronounceable non- 
words and is thought to operate on the basis of 
a set of abstract spelling-to-sound rules. The 
strong version of this dual-route theory claims 
that nonwords are segmented at the level of the 
grapheme and that the pronunciation of non- 
words is not influenced by lexical information. 
A line of evidence generally held to support the 
model comes from neuropsychological studies of 
acquired dyslexia. For instance, the patient WB 
studied by Funnell (1983) is considered a par- 
ticularly pure case of phonological dyslexia with 
good reading of words and poor reading of non- 
words. This case appears to conform to one 
of the main predictions of dual-route theory: 
namely, that neurological damage could selec- 
tively impair either processing route, so that 
a patient may have impaired processing in one 
system but intact processing in the other. 
Nonetheless, the dual-route model has been 
criticised by different authors (Marcel, 1980; 
Kay and Marcel, 1981; Glushko, 1981; Shal- 
lice et al., 1983; Humphreys and Evett, 1985; 
McCarthy and Warrington, 1986) who empha- 
sise that nonword pronunciation can be subject 
to lexical influences and/or argue for "multi- 
ple levels" of processing. Two main alterna- 
tives have been proposed to counter these ob- 
jections: a single-route framework and a modi- 
fied dual-route model. The first claims that all 
print-to-sound conversion is realised through a 
lexical route. That is, oral reading involves pro- 
cesses that all operate on a lexical database so 
that words and nonwords can be produced by 
the same mechanism. However, there has some- 
times been a lack of clarity in defining such a 
single-route mechanism. Often, some kind of 
analogy process is posited, but its precise form 
has rarely been specified. Hence, informed com- 
mentators have most often been inclined to re- 
form and repair the dual-route theory by re- 
laxing its strong assumptions, either to allow 
an interaction between routes (Reggia et al., 
1988) or to extend the notion of grapheme- 
phoneme correspondence (Patterson and Mor- 
ton, 1985) by introducing the notion of body-- 
the vowel-plus-terminal-consonant segment of 
monosyllabic words. 
The dual-route model has been more recently 
questioned by a plethora of single-route com- 
putational models based on connectionist prin- 
ciples (Sejnowski and Rosenberg, 1987; Seiden- 
berg and McClelland, 1989; Hinton and Shal- 
lice, 1991; Plaut et al., 1996; Bullinaria, 1997; 
Ans et al., 1998; Zorzi et al., 1998). Less often 
has analogy been used as the basis of a single- 
route model. The idea that pseudowords can be 
pronounced by analogy with lexical words that 
they resemble has a long history (Baron, 1977; 
Brooks, 1977; Glushko, 1979). In place of ab- 
stract letter-to-sound rules in dual-route models 
we have specific patterns of correspondence in 
single-route analogy models. 
3 Implementing PbA 
In PbA, an unknown word is pronounced by 
matching substrings of the input to substrings 
of known, lexical words, hypothesizing a partial 
pronunciation for each matched substring from 
the phonological knowledge, and assembling the 
partial pronunciations. Here, we use an ex- 
tended and improved version of the system de- 
scribed by Dedina and Nusbaum (1991), which 
consists of four components: the (uncompressed 
and previously aligned) lexical database, the 
matcher which compares the target input to 
all the words in the database, the pronuncia- 
tion lattice (a data structure representing pos- 
sible pronunciations), and the decision func- 
tion, which selects the 'best' pronunciation 
among the set of possible ones. The lexi- 
con used is Webster's Pocket Dictionary, con- 
taining 20,009 words manually aligned by Se- 
jnowski and Rosenberg (1987) for training their 
NETtalk neural network. 
Pattern Matching: An incoming word is 
matched in turn against all orthographic en- 
tries in the lexicon. For a given entry, assume 
the process starts with the input string and 
the dictionary entry left-aligned. Substrings 
sharing contiguous, common letters in match- 
ing positions are then found. Information about 
these matching letter substrings and their cor- 
responding, aligned phoneme substrings in the 
dictionary entry under consideration is entered 
into a pronunciation lattice--see below. One 
of the two strings is then shifted right by one 
letter and the matching process repeated, until 
14 
some termination condition is met. This process 
can be alternatively seen as a matching between 
substrings of the incoming word, segmented in 
all possible ways, and the dictionary entries. 
Pronunciation Lattice: A node of the lat- 
tice represents a matched letter, Li, at some 
position, i, in the input. The node is labelled 
with its position index i and with the phoneme 
which corresponds to Li in the matched sub- 
string, Rim say, for the mth matched substring. 
An arc is placed from node i to node j if there 
is a matched substring starting with Li and 
ending with Lj. The arc is labelled with the 
phonemes intermediate between Pim and Pjm 
in the phoneme part of the matched substring. 
Additionally, arcs are labelled with a 'frequency' 
count which is incremented each time that sub- 
string (with that pronunciation) is matched 
during the pass through the lexicon. 
Decision Function: A possible pronuncia- 
tion for the input corresponds to a com- 
plete path through its lattice, from Start to 
End nodes, with the output string assembled 
by concatenating the phoneme labels on the 
nodes/arcs in the order that they are traversed. 
(Different paths can, of course, correspond to 
the same pronunciation.) Scoring of candidate 
pronunciation uses two heuristics. If there is 
a unique shortest path, then the correspond- 
ing pronunciation is taken as the output. If 
there are tied shortest paths, then the pronunci- 
ation corresponding to the best scoring of these 
is taken as the output. 
This also offers a way of simulating the 'word 
segmentation' test of Funnell (1983), in which 
patients have to find words 'hidden' in letter 
strings. First, there is an initial segmentation 
in which the input string is segmented in all 
possible ways, as in 'regular' PbA. Then, if 
any of these substrings produces a lattice with 
a length-1 arc, this identifies a lexical word. 
A single-route connectionist model or abstract 
rules (or, for that matter, implicit PbA) can not 
do this without some extension to maintain ex- 
plicit knowledge of lexical status. Of course, it 
is possible that a patient can perform the first 
of these steps, but not the second. This is the 
difference between our 'unconscious' and 'con- 
scious' segmentations (see below) so-called be- 
cause, in the latter, the patient is aware that 
he/she has to find a hidden word. 
This particular implementation of PbA does 
not guarantee an output pronunciation. A com- 
plete path through the lattice requires that all 
nodes on that path (except the first and last) 
are linked by at least one arc. Clearly, each arc 
must have a node at either end. Although an 
arc may have an empty label, a node cannot. 
Hence, the minimum matching segment length 
corresponds to a letter bigram. It may be that 
no matching bigram exists in some cases. So 
there with be no complete path through the lat- 
tice and no pronunciation can be inferred--the 
'silence problem'. 
Recent Improvements: The implementa- 
tion used here features several enhancements 
over the original Dedina and Nusbaum (D&N) 
system (Marchand and Damper, 2000). First, 
we use 'full' pattern matching between input 
letter string and dictionary entries, as opposed 
to the 'partial' matching of D&N. That is, 
rather than starting with the two strings left- 
aligned, we start with the initial letter of the 
input string Z aligned with the last letter of 
the dictionary entry YV. The matching process 
terminates not when the two strings are right- 
aligned, but when the last letter of Z aligns 
with initial letter of \]/Y. Second, multiple (five) 
heuristics are used to score the candidate pro- 
nunciations. Individual scores are then multi- 
plied together to produce a final overall score. 
The best-scoring pronunciation is then selected 
as output. Marchand and Damper show that 
this 'multi-strategy' approach gives statistically 
significant performance improvements over sim- 
pler versions of PbA. 
4 Modelling Phonological Dyslexia 
By selective impairment of component parts 
of the PbA model, we have simulated read- 
ing data from the two phonological dyslexic pa- 
tients (WB and FL) studied by Funnell (1983). 
(The reader is referred to this original source for 
specifications of the tests and materials.) While 
the first of these patients has often been cited 
as a key individual strongly supporting dual- 
route theory, we believe that FL (who has been 
largely ignored) is actually a counter-example. 
FL was unable to supply a sound for single 
letters (which argues that the abstract rule- 
based route is impaired) although she could 
read non-words normally (which contradicts the 
15 
Table 1: Reading performance of patient WB and versions of faulty and non-faulty PbA. 'Words 
(712)' refers to a random sampling of words from the dictionary. 
Patient Faulty PbA Non-faulty Tests 
WB Version 1 Version 2 PbA 
Lexicon Words (712) 85% 85% 79% 100% 
Nonwords 
Single letters 0/12 0/12 0/12 10/12 
Nonsense words 0/20 0/20 0/20 17/20 
Pseudo-homophones 1/10 0/10 0/10 7/10 
Isolated suffixes 1/10 1/10 1/10 7/10 
Parkin's test 0/10 0/10 0/10 10/10 
Segmentation 
Test 1 Parent words 
Segmented words 
Test 2 Parent words 
Segmented words 
Test 3: Hidden words 
15/15 12/15 7/15 13/15 
30/30 30/30 26/30 30/30 
14/15 10/15 6/15 15/15 
22/30 24/30 21/30 28/30 
15/15 15/15 14/15 15/15 
Table 2: Reading performances of patient FL and of faulty and non-faulty PbA. 
Tests Patient FL Faulty PbA Non-faulty PbA 
Single letters 0/15 0/15 12/15 
'Easy' Nonwords 25/34 26/34 31/34 
'Difficult' Nonwords 4/6 1/6 3/6 
presumption of impaired rules). 
For patient WB, two different versions of im- 
paired PbA have been studied. Version 1 sup- 
poses that brain damage has induced a partial 
loss of words from his mental lexicon (the 15% 
that he can not read aloud) and a total break- 
down of his concatenation mechanism. Ver- 
sion 2 supposes that WB's impairment results 
from injury to one component only; namely, the 
process of segmentation into all possible sub- 
strings is partially damaged. In Version 2, we 
stress the distinction made earlier between this 
basic (unconscious) segmentation process and 
Funnell's (conscious) segmentation. The un- 
conscious segmentation is that embodied in the 
PbA pattern matching when WB is asked to 
read some string. For this specific patient, we 
postulate damage to the segmentation compo- 
nent such that it can only process substrings of 
length between 5 and 7. The conscious segmen- 
tation is that used when WB is asked to find 
words within strings and to read them aloud. 
This process is assumed to be fully operational. 
For patient FL, a single 'faulty' version of PbA 
has been developed which postulates a deft- 
ciency of (unconscious) segmentation such that 
substrings of length less than three cannot be 
used in pattern matching. 
Table 1 shows reading accuracy for pa- 
tient WB for the various tests performed by 
Funnell together with the corresponding results 
of simulations of impaired and non-faulty PbA. 
Table 2 shows the results for patient FL read- 
ing aloud and the corresponding simulation of 
faulty and non-faulty PbA. Evidently, it is 
possible to reproduce quite well both patients' 
symptoms. Indeed, with Version 1, we can in- 
terpret WB's condition very directly: The con- 
catenation process involved in nonword reading 
is completely destroyed but the mental lexicon 
is relatively spared. Because of the absence of 
some compound words (e.g., gentlelman ) from 
the dictionary, the simulations concerning "par- 
ent words" (e.g., father is the parent of.fat and 
her) for both Test 1 and Test 2 are not perfect. 
Version 2 is slightly poorer but still close to the 
neuropsychological data. For patient FL, the 
faulty version reproduces her impaired reading 
of single letters and 'easy' nonwords very well, 
but does so less well for 'difficult' nonwords. 
16 
The simulations also handle the fact that these 
patients were completely unable to read single 
letters: the silence problem (see above) can oc- 
cur for single letters by virtue of the form of 
the pronunciation lattice used, which requires 
matching bigrams (at least) at all positions to 
produce a pronunciation. 
5 Modelling Surface Dyslexia 
We have also modelled data from patient KT 
described by McCarthy and Warrington (1986). 
KT was able to pronounce regular words and 
nonwords very well but had serious difficulty in 
reading irregular words, tending to produce reg- 
ularisation errors. (Again, limitations of space 
mean we must refer the reader to the original 
source for details of the reading tests and mate- 
rials.) Together with WB, these patients have 
been taken as almost an existence proof of dual 
routes which can be differentially damaged. 
We suppose that KT's impairment re- 
sults from injury to two components of the 
PbA model. First, as in phonological dyslexia, 
we assume that the process of segmentation 
into all possible substrings is partially dam- 
aged. More specifically, we postulate a defi- 
ciency concerning the size of the window in- 
volved in the pattern matching. Second, it is 
assumed that one or several (of the total of five) 
multi-strategies may be degraded. 
The simulations were obtained for a model 
with damage in the third and fourth multi- 
strategies (see Marchand and Damper, 2000, 
for detailed specification) and only substrings 
of length between 2 and 4 can be segmented in 
pattern matching. Table 3 shows KT's mean 
reading accuracy over the various tests per- 
formed by McCarthy and Warrington together 
with our corresponding simulation results for 
impaired and non-faulty PbA. Clearly, it is pos- 
sible to reproduce quite well the patient's car- 
dinal symptoms: his ability to pronounce regu- 
lar words much better than irregular ones. The 
incorrect pronunciations show a clear regulari- 
sation effect (not detailed here). 
6 Conclusion 
Contrary to the claims of dual-route theorists, 
a single-route PbA model of reading is indeed 
able to explain both phonological and surface 
dyslexia, on the basis of selective impairment of 
its component parts. 

References 
B. Ans, S. Carbonnel, and S. Valdois. 1998. A 
connectionist multiple-trace memory model for 
polysyllabic word reading. Psychological Review, 
105(4):678-723. 
P. C. Bagshaw. 1998. Phonemic transcription by 
analogy in text-to-speech synthesis: Novel word 
pronunciation and lexicon compression. Com- 
puter Speech and Language, 12:119-142. 
J. Baron. 1977. Mechanisms for. pronounc- 
ing printed words: Use and acquisition. In 
D. LaBerge and S. Samuels, editors, Basic Pro- 
cesses in Reading: Perception and Comprehen- 
sion, pages 175-216. Lawrence Erlbaum Asso- 
ciates, Hillsdale, NJ. 
M. F. Beauvois and J. D~rouesn~. 1979. Phonologi- 
cal alexia: Three dissociations. Journal o\]Neurol- 
ogy, Neurosurgery and Psychiatry, 42:1115-1124. 
L. Brooks. 1977. Non-analytic correspondences and 
pattern in word pronunciation. In J. Renquin, ed- 
itor, Attention and Per/ormance VII, pages 163- 
177. Lawrence Erlbaum Associates, Hillsdale, NJ. 
J. A. Bullinaria. 1997. Modeling reading, spelling, 
and past tense learning with artificial neural net- 
works. Brain and Language, 59:236-266. 
M. Coltheart, J. Masterson, S. Byng, M. Pryor, and 
J. Riddoch. 1983. Surface dyslexia. Quarterly 
Journal o/ Experimental Psychology, 35A:469- 
495. 
M. Coltheart, B. Curtis, P. Atkins, and M. Haller. 
1993. Models of reading aloud: Dual-route and 
parallel-distributed-processing approaches. Psy- 
chological Review, 100(4):589-608. 
M. Coltheart. 1978. Lexical access in simple read- 
ing tasks. In G. Underwood, editor, Strategies o/ 
In\]ormation Processing, pages 151-216. Academic 
Press, New York. 
W. Daelemans, A. van den Bosch, and T. Weijters. 
1997. IGTree: Using trees for compression and 
classification in lazy learning algorithms. Artifi- 
cial Intelligence Review, 11(1-5):407-423. 
R. I. Damper and J. F. G. Eastmond. 1997. Pronun- 
ciation by analogy: Impact of implementational 
choices on performance. Language and Speech, 
40(1):1-23. 
R. I. Damper, Y. Marchand, M. J. Adamson, and 
K. Gustafson. 1999. Evaluating the pronunci- 
ation component of text-to-speech systems for 
English: A performance comparison of differ- 
ent approaches. Computer Speech and Language, 
13(2):155-176. 
M. J. Dedina and H. C. Nusbaum. 1991. PRO- 
NOUNCE: A program for pronunciation by anal- 
ogy. Computer Speech and Language, 5:55-64. 
E. Funnell. 1983. Phonological processes in read- 
ing: New evidence from acquired dyslexia. British 
Journal of Psychology, 74:159-180. 
R. J. Glushko. 1979. The organization and activa- 
tion of orthographic knowledge in reading aloud. 
Journal of Experimental Psychology: Human Per- 
ception and Performance, 5:674-691. 
R. J. Glushko. 1981. Principles for pronouncing 
print: The psychology of phonography. In A. M. 
Lesgold and C. A. Perfetti, editors, Interactive 
Processes in Reading, pages 61-84. Lawrence Erl- 
baum Associates, Hillsdale, NJ. 
G. E. Hinton and T. Shallice. 1991. Lesioning 
an attractor network: Investigations of acquired 
dyslexia. Psychological Review, 98:74-95. 
G. W. Humphreys and L. J. Evett. 1985. Are there 
independent lexical and non-lexical routes in word 
processing? An evaluation of the dual route the- 
ory of reading. Behavioral and Brain Sciences, 
8:689-739. 
D. Jones. 1996. Analogical Natural Language Pro- 
cessing. UCL Press, London, UK. 
J. Kay and A. Marcel. 1981. One process, not two, 
in reading aloud: Lexical analogies do the work 
of non-lexical rules. Quarterly Journal of Experi- 
mental Psychology, 33A:397-413. 
A. J. Marcel. 1980. Surface dyslexia and beginning 
reading: A revised hypothesis of the pronuncia- 
tion of print and its impairments. In M. Colt- 
heart, K. E. Patterson, and J. C. Marshall, ed- 
itors, Deep Dyslexia, pages 227-258. Routledge 
and Kegan Paul, London, UK. 
Y. Marchand and R. I. Damper. 2000. A multi- 
strategy approach to improving pronunciation by 
analogy. Computational Linguistics, 26:195-219. 
R. McCarthy and K. Warrington. 1986. Phonolog- 
ical reading: Phenomena and paradoxes. Cortex, 
22:359-380. 
N. McCulloch, M. Bedworth, and J. Bridle. 1987. 
NETspeak - a re-implementation of NETtalk. 
Computer Speech and Language, 2:289-301. 
K. E. Patterson and J. Morton. 1985. From orthog- 
raphy to phonology: An attempt at an old inter- 
pretation. In K. E. Patterson, J. C. Marshall, and 
M. Coltheart, editors, Surface Dyslexia: Neuro- 
psychological and Cognitive Studies of Phonolog- 
ical Reading, pages 335-359. Lawrence Erlbaum 
Associates, London, UK. 
V. Pirrelli and S. Federici. 1995. You'd better say 
nothing than something wrong: Analogy, accu- 
racy and text-to-speech applications. In Proceed- 
ings of ~th European Conference on Speech Com- 
munication and Technology, Eurospeech'95, vol- 
ume 1, pages 855-858, Madrid, Spain. 
V. Pirrelli and F. Yvon. 1999. The hidden dimen- 
sion: A paradigmatic view of data-driven NLP. 
Journal of Experimental and Theoretical Artificial 
Intelligence, 11(3):391-408. 
D. C. Plaut, J. L. McClelland, M. S. Seidenberg, 
and K. E. Patterson. 1996. Understanding nor- 
mal and impaired word reading: Computational 
principles in quasi-regular domains. Psychological 
Review, 103(1):56-115. 
J. A. Reggia, P. M. Marsland, and R. S. Berndt. 
1988. Competitive dynamics in a dual-route con- 
nectionist model of print-to-sound transforma- 
tion. Complex Systems, 2:509-547. 
M. S. Seidenberg and J. L. McClelland. 1989. 
A distributed, developmental model of word 
recognition and naming. Psychological Review, 
96(4):523-568. 
T. J. Sejnowski and C. R. Rosenberg. 1987. Parallel 
networks that learn to pronounce English text. 
Complex Systems, 1:145-168. 
T. Shallice, E. K. Warrington, and R. McCarthy. 
1983. Reading without semantics. Quarterly 
Journal of Experimental Psychology, 35A:111- 
138. 
K. P. H. Sullivan and R. I. Damper. 1993. Novel- 
word pronunciation: A cross- study. 
Speech Communication, 13:441-452. 
F. Yvon. 1996. Grapheme-to-phoneme conversion 
using multiple unbounded overlapping chunks. 
In Proceedings of Conference on New Methods 
in Natural Language Processing (NeMLaP-2'96), 
pages 218-228, Ankara, Turkey. 
F. Yvon. 1997. Paradigmatic cascades: A linguisti- 
cally sound model of pronunciation by analogy. In 
Proceedings of 35th Annual Meeting of the Asso- 
ciation for Computational Linguistics, pages 429- 
435, Madrid, Spain. 
M. Zorzi, G. Houghton, and B. Butterworth. 1998. 
The development of spelling-sound relationships 
in a model of phonological reading. Language and 
Cognitive Processes, 13:337-371. 
