TransType: a Computer--Aided Translation Typing System 
Philippe Langlais and George Foster and Guy Lapalme 
RALI/DIRO -- Universit@ de Montr@al 
C.P. 6128, succursale Centre-ville 
H3C 3J7 Montr4al, Canada 
Phone:: +1 (514) 343-2145 
Fax: +1 (514) 343-5834 
email: {felipe, foster, lapalme}©iro, umontreal, ca 
Abstract 
This paper describes the embedding of a sta- 
tistical translation system within a text editor 
to produce TRANSTYPE, a system that watches 
over the user as he or she types a translation and 
repeatedly suggests completions for the text al- 
ready entered. This innovative Embedded Ma- 
chine Translation system is thus a specialized 
means of helping produce high quality transla- 
tions. 
1 Introduction 
TRANSTYPE is a project set up to explore an 
appealing solution to the problem of using In- 
teractive Machine Translation (IMT) as a tool 
for professional or other highly-skilled transla- 
tors. IMT first appeared as part of Kay's MIND 
system (Kay, 1973), where the user's role was 
to help the computer analyze the source text 
by answering questions about word sense, el- 
lipsis, phrasal attachments, etc. Most later 
work on IMT, eg (Blanchon, 1991; Brown and 
Nirenburg, 1990; Maruyama and Watanabe, 
1990; Whitelock et al., 1986), has followed in 
this vein, concentrating on improving the ques- 
tion/answer process by having less questions, 
more friendly ones, etc. Despite progress in 
these endeavors, systems of this sort are gen- 
erally unsuitable as tools for skilled trans\]\[ators 
because the user serves only as an advisor, with 
the MT components keeping overall control over 
the translation process. 
TRANSTYPE originated from the conviction 
that a better approach to IMT for competent 
translators would be to shift the focus of in- 
teraction from the meaning of the source text 
to the form of the target text. This would re- 
lieve the translator of the burden of having to 
provide explicit analyses of the source text and 
allow him to translate naturally, assisted by the 
machine whenever possible. 
In this approach, a translation emerges from 
a series of alternating contributions by human 
and machine. The machine's contributions are 
basically proposals for parts of the target text, 
while the translator's can take many forms, in- 
cluding pieces of target text, corrections to a 
previous machine contribution, hints about the 
nature of the desired translation, etc. In all 
cases, the translator remains directly in control 
of the process: the machine must respect the 
constraints implicit in his contributions, and he 
or she is free to accept, modify, or completely 
ignore its proposals. 
So TRANSTYPE is a specialized text editor 
with an embedded Machine translation engine 
as one of its components. In this project we 
had to address the following problems: how to 
interact with the user and how to find appro- 
priate multi-word units for suggestions that can 
be computed in real time. 
2 The TransType model 
2.1 User Viewpoint 
Our interactive translation system is illustrated 
in figure 1 for an English to French translation. 
It works as follows: a translator selects a sen- 
tence and beg!ns typing its translation. After 
each character typed by the translator, the sys- 
tem displays a proposed completion, which may 
either be accepted using a special key or rejected 
by continuing to type. This interface is simple 
and its performance may be measured by the 
proportion of characters or keystrokes saved in 
typing a translation. Note that, throughout this 
process, the translator remains in control, and 
the machine must continually adapt its sugges- 
tions to the translator's input. This differs from 
the usual machine translation set-ups where it is 
the machine that produces the first draft which 
46 
• . • . :, • ,, ------~ 
..... Fich:ier :::= ptions 
• :'. " - :.'. ... " ..: .... ".11... It" I: am:pleased:to: takepart:m this debate today. 
Usingitoday'S technologies,it:is possiblefOrall ~adiaqs to 
.... . a ..... borrowing:. 
..... • ......... :::i~ : 
d~batl.. 
GraCel ~i~la t~chnOIogiemoderne, tousles Can adiehs peuVent se 
prononcer:sur:lesquestions de:depenses: et~:d em runts:de/EZra 
Figure 1: Example of an interaction in TRANSTYPE with the source text in the top half of the 
screen. The target text is typed in the bottom half with suggestions given by the menu at the 
insertion point. 
then has to be corrected by the translator. 
The first version of TRANSTYPE (Foster et 
al., 1997) only proposed completions for the cur- 
rent word. We are now working on predictions 
which extend to the next several words in the 
text. The potential gain from multiple-word 
predictions (Langlais et al., 2000) can be ap- 
preciated in the one-sentence translation task 
reported in table 1, where a hypothetical user 
saves over 60% of the keystrokes needed to pro- 
duce a translation in a word completion sce- 
nario, and about 75% in a "unit" completion 
scenario 
2.2 System Viewpoint 
The core of TRANSTYPE is a completion engine 
which comprises two main parts: an evaluator 
which assigns probabilistic scores to completion 
47 
This bill is very similar to its companion bill which we dealt with yesterday 
in the house of commons 
word-completion task. unit-completion task 
pref. completions pref. completions 
C+ /loi • c/e pro jet de loi ce 
pro jet 
de 
Ioi 
est 
tr~s 
semblable 
au 
pro jet 
de 
Ioi 
que 
nous 
avons 
examin4 
hier 
la 
chambre 
des 
communes 
ce+ /lOi" C/' 
p+ /est p/rojet 
d+ /tr~s d/e 
I+ /tr~s I/oi 
e+ /de e/st 
t+ /de • t/r~s 
se+ /de • s/es 
au+ /loi • a/vec 
p+ /loi p/rojet 
d+ /loi • d/e 
I+ /nous • I/oi 
qu+ /nous • q/ui • 
+ /nous 
av+ /nous 
ex+ /hier 
+ /hier 
se/mblable 
qu/e 
•  v/ons 
• e/n. ex/amin~ 
~+ /bier • ~/ la 
+ /chambre 
de+ /communes • dle • 
+ /communes de/s 
e+ 
t+ 
se+ 
a+ 
/de e/st 
/de. t/r~s 
/de • s/es se/mblable 
/loi • a/u projet de loi sur 
qu+ /nous • q/ui • qu/e 
+ /nous 
av+ /nous. a/vec, av/ons 
exa+ /& la chambre des communes 
e/n. ex/istence, exa/min~ 
h-F /& la chambre des communes 
h/let 
+ /& la chambre des communes 
106 char. 23 20 accept. 14 11 accept. -t- 1 correc. 
43 keystrokes 26 keystrokes 
Table h A one-sentence session illustrating the word- and unit- completion tasks. The first col- 
umn indicates the target words the user is expected to produce. The next two columns indicate 
respectively the prefixes typed by the user and the completions made by the system under a word- 
completion task. The last two columns provide the same information for the unit-completion task. 
The total number of keystrokes for both tasks is reported in the last line. + indicates the accep- 
tance key typed by the user. A Completion is denoted by a/fl where a is the typed prefix and fl 
the completed part. Completions for different prefixes are separated by • . 
hypotheses and a generator which uses the eval- 
uation function to select the best candidate for 
completion. 
2.2.1 The evaluator 
The evaluator is a function p(t\[t', s) which as- 
signs to each target-text unit t an estimate of 
its probability given a source text s and the to- 
kens t' which precede t in the current translation 
of s. Our approach to modeling this distribu- 
tion is based to a large extent on that of the 
IBM group (Brown et al., 1993), but it diflhrs in 
one significant aspect: whereas the IBM model 
involves a "noisy channel" decomposition, we 
use a linear combination of separate predictions 
from a  model p(t\[t') and a transla- 
tion model p(t\[s). Although the noisy channel 
technique is powerful, it has the disadvantage 
that p(s\[t', t) is more expensive to compute than 
p(t\[s) when using IBM-style translation models. 
Since speed is crucial for our application, we 
chose to forego it in the work described here. 
Our linear combination model is fully described 
in (Langlais and Foster, 2000) but can be seen 
as follows: 
48 
p(tlt',s ) = p(tlt' ) A(O(t',s)), (1) 
 
+ p(tls)\[1-~(O(t',s))! 
translation 
where .~(O(t',s)) e \[0,1\] are context- 
dependent interpolation coefficients. O(t~,s) 
stands for any function which maps t~,s into a 
set of equivalence classes. Intuitively, ),(O(t r, s)) 
should be high when s is more informative than 
t r and low otherwise. For example, the trans- 
lation model could have a higher weight at the 
start of sentence but the contribution of the lan- 
guage model can become more important in the 
middle or the end of the sentence. 
2.2.2 The  model 
We experimented with various simple linear 
combinations of four different French  
models: a cache model, similar to the cache 
component in Kuhn's model (Kuhn and Mori, 
1990); a unigram model; a trielass model (Der- 
ouault and Merialdo, 1986); and an interpolated 
trigram (Jelinek, 1990). 
We opted for the trigram, which gave signifi- 
cantly better results than the other three mod- 
els. The trigram was trained on the Hansard 
corpus (about 50 million words), with 75% of 
the corpus used for relative-frequency parame- 
ter estimates, and 25% used to reestimate inter- 
polation coefficients. 
2.2.3 The translation model 
Our translation model is based on the linear in- 
terpolation given in equation 2 which combines 
predictions of two translation models -- Ms and 
Mu -- both based on an IBM-like model 2 (see 
equation 3). Ms was trained on single words 
and Mu was trained on both words and units. 
p( tls) = Z pt( tls) ,+ (1 - Z).p2 ( (s ) ) 
word unit 
(2) 
where Ps and Pu stand for the probabilities 
given respectively by Ms and M~. ~(s) repre- 
sents the new sequence of tokens obtained after 
grouping the tokens of s into units. 
Both models are based on IBM translation 
model 2 (Brown et al., 1993) which has the 
49 
property that it generates tokens independently. 
The total probability of the ith target-text to- 
ken ti is just the average of the probabilities 
with which it is generated by each source text 
token sj; this is a weighted average that takes 
the distance from the generating token into ac- 
count: 
is1 
p(tils) = ~p(tilsj) a(jli, Is\[) 
j=O 
(3) 
where p(ti Is j) is a word-for-word translation 
probability, Isl is the length (counted in tokens) 
ofthe source segment s under translation, and 
a(jli , Is\]) is the a priori alignment probability 
that the target-text token at position i will be 
generated by the source text token at position 
j; this is equal to a constant value of 1~(Is I + 1) 
for model 1. This formula follows the conven- 
tion of (Brown et al., 1993) in letting so des- 
ignate the null state. We modified IBM model 
2 to account for invariant entities such as En- 
glish forms that almost invariably translate into 
French either verbatim or after having under- 
gone a predictable transformation e.g. numbers 
or dates. These forms are very frequent in the 
Hansard corpus. 
2.3 The Generator 
The task of the generator is to identify units 
matching the current prefix typed by the user, 
and pick the best candidate using the evalua- 
tion function. Given the real time constraints 
of an IMT system, we divided the French vocab- 
ulary into two parts: a small active component 
whose contents are always searched for a match 
to the current prefix, and a much larger passive 
part which comes into play only when no candi- 
dates are found in the active vocabulary. Both 
vocabularies are coded as tries. 
The passive vocabulary is a large dictionary 
containing over 380,000 word forms. The ac- 
tive part is computed dynamically when a new 
sentence is selected by the translator. It relies 
on the fact that a small number of words ac- 
count for most of the tokens in a text. It is 
composed of a few entities (tokens and units) 
that are likely to appear in the translation. In 
practice, we found that keeping 500 words and 
50 units yields good performance. 
3 Implementation 
From an implementation point of view, the core 
of TransType relies on a flexible object ori- 
ented architecture, which facilitates the integra- 
tion of any model that can predict units (words 
or sequence of words) from what has been al- 
ready typed and the source text being trans- 
lated. This part is written in C÷+. Statisti- 
cal translation and  models have been 
integrated among others into this architecture 
(Lapalme et al., 2000). 
The graphical user interface is implemented 
in Tcl/Tk, a multi-platform script  well 
suited to interfacing problems. It offers all the 
classical functions for text edition plus a pop-up 
menu which contains the more probable words 
or sequences of words that may complete the 
ongoing translation. The proposed completions 
are updated after each keystroke the translator 
enters. 
4 Evaluation 
We have conducted a theoretical evaluation of 
TransType on a word completion task, which 
assumes that a translator carefully observes 
each completion proposed by the system, and 
accepts it as soon as it is correct. Under 
these optimistic conditions, we have shown that 
TransType allows for the production of a trans- 
lation typing less than a third of its characters. 
In order to better grasp the usefulness of 
TRANSTYPE, we also performed a more prac- 
tical evaluation by asking ten translators to 
use the prototype for about one hour to trans- 
late isolated sentences. We first asked them to 
translate without any help from TRANSTYPE 
and then we compared their typing speed with 
TRANSTYPE suggestions turned on. Overall, 
translators liked the concept and found it very 
useful; they all liked the suggestions although 
it seemed to induce a literal style of transla- 
tion. We also asked them if they thought that 
TRANSTYPE improved their typing speed and 
the majority of them said so; unfortunately the 
figures showed that none of them did so ... The 
typing rates are nevertheless quite good, given 
that the users were new to this environment and 
this style of looking at suggestions while trans- 
lating. But interestingly this practical ew~lua- 
tion confirmed our theoretical evaluation that a- 
translation can be produced with TRANSTYPE 
by typing less than 40% of the characters of a 
translation. Results of this evaluation and com- 
parisons with our theoretical figures are further 
described in (Foster et al., 2000). 
This experiment made us realize that this 
concept of real-time suggestions depends very 
much on the usability of the prototype; we had 
first developed a much simpler editor but its 
limitations were such that the translators found 
it unusable. So we are convinced that the user- 
interface aspects of this prototype should be 
thoroughly studied. But the TRANSTYPE ap- 
proach would be much more useful if it was 
combined with other text editing tasks related 
to translation: for example TRANSTYPE could 
format the translation in the same way as the 
source text, this would be especially useful for 
titles and tables; it would also be possible to 
localize automatically specific entities such as 
dates, numbers and amounts of money. It would 
also be possible to check that some translations 
given by the user are correct with respect with 
some normative usage of words or terminologi- 
cal coherence; these facilities are already part of 
TRANSCHECK, another computer aided transla- 
tion tool prototype developed in our laboratory 
(Jutras, 2000). 
5 Conclusion 
We have presented an innovative way of em- 
bedding machine translation by means of a pro- 
totype which implements an appealing interac- 
tive machine translation scenario where the in- 
teraction is mediated via the target text under 
production. Among other advantages, this ap- 
proach relieves the translator of the burden of 
source analyses, and gives him or her direct con- 
trol over the final translation without having to 
resort to post-edition. 
Acknowledgements 
TRANSTYPE is a project funded by the Natu- 
ral Sciences and Engineering Research Council 
of Canada. We are greatly indebted to Elliott 
Macklovitch and Pierre Isabelle for the fruitful 
orientations they gave to this work. 

References 
Herv6 Blanchon. 1991. Probl~mes de 
d@sambigffisation interactive et TAO per- 
sonnelle. In L 'environnement Traductionnel, 
Journ@es scientifiques du R@seau th@matique 
de recherche "Lexicologie, terminologie, 
traduction", pages 31-48, Mons, April. 
Ralf D. Brown and Sergei Nirenburg. 1990. 
Human-computer interaction for semantic 
disambiguation. In Proceedings off the Inter- 
national Conference on Computational Lin- 
guistics (COLING), pages 42-47, Helsinki, 
Finland, August. 
Peter F. Brown, Stephen A. Della Pietra, Vin- 
cent Della J. Pietra, and Robert L. Mercer. 
1993. The mathematics of machine transla- 
tion: Parameter estimation. Computational 
Linguistics, 19(2):263-312, June. 
A.-M. Derouault and B. Merialdo. 1986. Nat- 
ural  modeling for phoneme-to-text 
transcription. IEEE Transactions on Pattern 
Analysis and Machine Intelligence (PAMI), 
8 (6): 742-749, November. 
George Foster, Pierre Isabelle, and Pierre Pla- 
mondon. 1997. Target-text Mediated Inter- 
active Machine Translation. Machine Trans- 
lation, 12:175-194. 
George Foster, Philippe Langlais, Guy 
Lapalme, Dominique Letarte, Elliott 
Macklovitch, and S@bastien Sauv@. 2000. 
Evaluation of transtype, a computer-aided 
translation typing system: A comparison of 
a theoretical- and a user- oriented evaluation 
procedures. In Conference on Language 
Resources and Evaluation (LREC), page 8 
pages, Athens, Greece, June. 
Frederick Jelinek. 1990. Self-organized lan- 
guage modeling for speech recognition. In 
A. Waibel and K. Lee, editors, Readings in 
Speech Recognition, pages 450-506. Morgan 
Kaufmann, San Mateo, California. 
Jean-Marc Jutras. 2000. An automatic reviser: 
The TransCheck system. In Applied Natu- 
ral Language Processing 2000, page 10 pages, 
Seattle, Washington, May. 
Martin Kay. 1973. The MIND system. In 
R. Rustin, editor, Natural Language Process- 
ing, pages 155-188. Algorithmics Press, New 
York. 
Roland Kuhn and Renato De Mori. 1990. 
A cache-based natural  model for 
speech recognition. IEEE Transactions on 
Pattern Analysis and Machine Intelligence 
(PAMI), 12(6):570-583, June. 
Philippe Langlais and George Foster. 2000. Us- 
ing context-dependent interpolation to com- 
bine statistical  and translation 
models for interactive machine translation. 
In Computer-Assisted Information Retrieval, 
Paris, April. 
Philippe Langlais, George Foster, and Guy 
Lapalme. 2000. Unit completion for a 
computer-aided translation typing system. In 
Applied Natural Language Processing 2000, 
page 10 pages, Seattle, Washington, May. 
Guy Lapalme, George Foster, and Philippe 
Langlais. 2000. La programmation orient@e- 
objet pour le d~veloppement de modules de 
langages. In Christophe Dony and Houari A. 
Sahraoui, editors, LMO'O0 - Langages et 
modules ~ objets, pages 139-147, Mont St- 
Hilaire, Qu@bec, 27 Janvier. Hermes Science. 
Conference invit@e. 
Hiroshi Maruyama and Hideo Watanabe. 1990. 
An interactive Japanese parser for machine 
translation. In Proceedings of the Interna- 
tional Conference on Computational Linguis- 
tics (COLING), pages 257-262, Helsinki, Fin- 
land, August. 
P. J. Whitelock, M. McGee Wood, B. J. Chan- 
dler, N. Holden, and H. J. Horsfall. 1986. 
Strategies for interactive machine transla- 
tion: the experience and implications of the 
UMIST Japanese project. In Proceedings of 
the International Conference on Computa- 
tional Linguistics (COLING), pages 329-334, 
Bonn, West Germany. 
