DIVe~OPING A READING MACHINE FOR ~ BLIND 
M. Boot 
Inst. of Applied and Computational Linguistics. WilhelmADa- 
park 11/12, 3581 NC Utrecht 
The develo~aent of a readi~ machine for the blind 
implies the solution of problems on such diverse fields as 
linguistics, micro computing and ergonomics. Because of the 
stats of the art in Computational Lin&qaist$cs, however, the 
lin~qaistic problems turn out to be the major draw bask in 
this field of scientific endeavor. That is the reason why the 
paper for the greater part is devoted to the description of a 
new model for automated phonemization. This model is applied 
to Dutch. The model was developed for words only. Thus, the 
reading manhins as it stands now is able to pronounce a series 
of words e 
Therefore, the texts read into the computer are treated 
as a series of single words by the reading machine. The prob- 
le~ of prosody e:ce not tackled uptil! news On the other hand, 
all problems concerning assimilation in ths words have been 
8olvedo The computer program that performs this te~k is call- 
ed YONGRAF. It was developed at the University of Utrecht. The 
computer program FONGRAF is able to perform a transcription 
of written text into the phonematic foxiest according to the 
principles of phonematic transor£ption. The paper fooueses on 
the design of the prepare and enswers questions concerning the 
relation between the technical part (implamentation) and the 
linguistic considerations behind the computer progj~em. 
It is argued that in the past computer programs perform- 
ing this linguistic task were principally designed from the 
- 46- 
implementation point of view. This has led to computer pro- 
grams wi~h a strong ad hoc kind of problem solving part in it. 
Therefore,these computer prod, rams turn out to be not adaptable 
tO new situations and unforeseen mistakes.ln this paper it is 
argued that for the solution of linguistic problems of this 
kind a pattern mato~-tng computer program has to be developed. 
It should go without saying that a computer designed for 
lin6~istio purposes should be firmly based on the linguistic 
analysis of the task. As far as the search for regularities in 
the phonetic interpretation of written text is concerned we 
used the phonological theory as an important aid. The phono- 
logical description of the Dutch language was used as the most 
important source for the definition of the pattern marcher in 
the computer program FONGRAP. Many of the observed regularit- 
ies regarding the Dutch phoneme distribution and phonologloal 
rules concerning the phonetic interpretation of the phonologic- 
al forms of Dutch morphemes are particularly useful to our 
problem in that they state the surroundings which affect a 
particular phoneme. For instance the assimilation rule that a 
consonant becomes voiceless or voiced according to the "voice" 
of the following consonant. This kind of rules applies even in 
surroundings where the syllable boundaries are involved. This 
is the reason why we consider the application of hyphenation 
programs to be out of place as far as the solution of the 
phonematization problem for Dutch words is concerned, This also 
is the main reason why we developed a pattern matching comput- 
er program. A further advantage of the pattern matching pro- 
gram is that it is easy to implement "new" regularities. With 
the notion "new" regularities we refer to rules and regularit- 
ies not described b~ normal phonology. Those regularities 
often are caused by the fact that a computer is too "literal". 
Native speakers will, e.g. not confuse any letter-sequence 
oi_~r or isc_._hh as in hooi_._~k and macaronlsch_.__otel with the suffix- 
es oi__~r and ise~ho A computer, however, does not have this 
linguistic knowledge. Thus, one has to design means to implem- 
- 47 - 
ent this knoWledge. This is done by the definition of patterns. 
In the paper we shortly refer to the eolut£on of prcnunc£at£on 
8rabidity caused by semantic reasonm. The computer pro~ea 
FONGRAP w~s tested with the help of a variety of corpus~s 
consisting of natural le~guage texts, The results of these 
tests are reportedo 
- 48 - 
