On UNL as the future "html of the linguistic content" & the reuse of

exi sti ng NLP component s in UNL- rel at ed appl i cat i ons wi th the
exampl e of a UNL- French deconvert er
Gilles St~RASSET
GETA, CLIPS, I MAG
385, av. de la biblioth~que, BP 53
F-38041 Grenobl e cedex 9, France
Gilles.Serasset@ imag, fr
Christian BOI TET
GETA, CLIPS, I MAG
385, av. de la biblioth~que, BP 53
F-38041 Grenobl e cedex 9, France
Christian. Boitet @i mag. fr
Abstract
Aft er 3 years of speci fyi ng the UNL (Universal Net worki ng Language) l anguage and
prot ot ypi ng deconvert ers I from more than 12 languages and enconvert ers for about 4, the
UNL proj ect has opened to the communi t y by publishing the speci f cat i ons (v2.0) of the UNL
language, i nt ended to encode the meani ng of NL utterances as semantic hypergraphs and to be
used as a "pivot" representation in multilingual information and communi cat i on systems.
A UNL document is an html document with special tags to delimit the ut t erances and t hei r
renderi ng in UNL and in all natural languages current l y handled. UNL can be viewed as t he
future "html of the linguistic content". It is only an interface format, leading as well to the reuse
of existing NLP component s as to the devel opment of original tools in a variety of possible
applications, from automatic rough enconversi on for i nformat i on retrieval and i nf or mat i on
gathering translation to partially interactive enconversi on or deconversi on for hi gher quality.
We illustrate t hese points by describing an UNL-French deconvert er organi zed as a specific
"localizer" fol l owed by a classical MT transfer and an existing generator.
Keywords
UNL, interlingua, pivot, deconversion, UNL~French localization, transfer, generation.
Introduction
The UNL project of net wor k- or i ent ed
multilinguat communi cat i on has proposed a
standard for encodi ng the meani ng of natural
l anguage ut t erances as semantic hyper gr aphs
intended to be used as pivots in mul t i l i ngual
i nformat i on and communi cat i on systems. In t he
first phase (1997-1999), mor e than 16 partners
representing 14 languages have worked to bui l d
deconvert ers t ransformi ng an (i nt erl i ngual )
UNL hyper gr aph into a natural l anguage
ut t erance.
In this project, the strategy used to achieve this
initial obj ect i ve is free. The UNL- Fr ench
deconvert er under devel opment first performs a
"localization" operation within the UNL format,
and then classical t ransfer and generation steps,
using the Ari ane-G5 envi ronment and some
UNL- s peci f c tools.
The use of classical transfer and generat i on
steps in the cont ext of an interlingual proj ect
may sound surprising. But it reflects many
interesting issues about the status of the UNL
language, desi gned as an interlingua, but
diversely used as a linguistic pivot (disambi-
guated abstract English), or as a purely semantic
pivot.
Aft er introducing the UNL language, we present
the archi t ect ure of the UNL- Fr ench deconvert er,
whi ch "generates" from t he UNL interlingua by
first "localizing" the UNL form for French,
within UNL, and then appl yi ng slightly adapt ed
but classical t ransfer and generation techniques,
i mpl ement ed in the Ari ane-G5 environlnent,
suppl ement ed by some UNL-speci fi c tools.
Then, we discuss the use of the UNL l anguage
as a linguistic or semantic pivot for hi ghl y
multilingual i nformat i on syst ems.
1 The UNL project and language
1.1 The project
UNL is a project of multilingual personal
net worki ng communi cat i on initiated by the
University of United Nations based in Tokyo.
The pivot paradi gm is used: the represent at i on
I The terms <<deconvcrsion, and <~enconvcrsion, are specific to tile UNL proiect and are defined at paragraph 2.
768
of an ut t erance in the UNL interlingua ( UNL
st ands for "Universal Net worki ng Language") is
a hyl)ergraph where normal nodes bear UWs
CUni versal Words", or interlingual accept i ons)
with semant i c attributes, and arcs bear semantic
relations (deep cases, such as agt, obj, goal, etc.).
Hyper nodes group a subgraph defi ned by a set
of connect ed arcs. A UW denot es a set of
interlingual accept i ons (word senses), al t hough
we often l oosel y speak of "the" word sense
demoted by a UW.
Because English is known by all UNL
devel opers, the syntax of a tlormal WW is:
" < E n g l i s h w o r d o r c o m p o u n d > ( < l i s t
o f r e s t r i c t i o n s > ) ", O. Z. " l o o k f o r
( i c l > a c t i o n , a g t > h u m a n , o b j > t h i n g ) "
Goi ng fronl a text to t h e correspondi ng "UNL
text" or interactively constructing a UNL text is
called "enconversi oi f' , while produci ng a t ext
f i om a sequence of UNL graphs is called
"deconversion".
This depart ure fi'om the standard terms of
analysis and generation is used to stress that this
is not a classical M]: projecl, bu! that UNL is
planned to be the source format preferred f or
represent i ng textual inl:ormation in t i l e
envi saged multilingual net work envi ronment .
Tile schedul e of tile project, beginning with
deconver si on rather than cnconvcrsi on, also
reflects that di fference.
14 hmguages have been tackled during the first
3--year phase o f the prqject (1997-1999), while
many more arc to be added in tile second
phase. Each group is fi-ee to reuse its own
soft ware lools and/or lingware resources, or to
devel op directly with tools provi ded by tile
UNL Cent er (UNU/ IAS).
Emphasi s is on a very large lexical coverage, so
that all groups spend most of their time on tile
UNL- NL lexicons, and devel op tools and
met hods for effi ci ent lexical devel opment . By
contrast, gramnmrs have been initially limited to
those necessary for deconversi on, and will then
bc gradual l y expanded to allow for mor e
naturalness m formul at i ng text to be
enconver t ed.
1.2 The UNL component s
1. 2. 1 Uni v e r s al Wor ds
Tile nodes of a UNL utterance are called
Universal Words (or Uws). The syntax of a
normal UW consists of 2 parts :
a headword,
a l i st o f restrictions
Because English is known by all UNL
devel opers, tile headword is an English word or
compound. The restrictions are given as all
attribute value pail" where attributes are semantic
relation labels (as the ones used in the graphs)
and wllues are other UWs (restricted or not).
A UW denot es a col l ect i on of interlingual
accept i ons (word senses), although we of t en
l oosel y speak of "the" word sense denoted by an
UW. For exampl e, the unrestricted UW " l o o k
for" denot es all the word-senses associated to
tile English compound word "l ook for". Tile
r e s t r i c t e d U W " l o o k f o r ( i c l > a c t i o n ,
a g t > h u m a n , o b j > t h i n g ) " represents all tile
word senses of the English word "l ook f o r "
that are an action, perl%rmed by a human that
affects a thing. In this case this leads to the word
sense: "l ook f o r - to try to find".
1. 2. 2 UNL h y p e r g r a p h
A UNL expressi on is a hypergraph (a graph
where a node is si mpl e or recursively contains a
hypergraph). Tile arcs bear semantic relation
labels (deep cases, such as agt, obj, goal, etc.).
score( icl> ev ent ag t> hu m an, tld> sp ort) [
I @ entry . @ p ast. @ com p lete [
7 i \
a g t ....... ' / i n s \ ~
i R ona~ do 1 b ~ / \ p i t
head( p ~ ol> b ody ) ~ "\\
/ , ~
corner
/
g oal i~ cl> thing ) [ . 4 1 o b j m o d
[
~ l e f t :
Figm'e I. 1: A UNL graph deconvertible as "Ronaldo
has headed the ball into the left corner of the net"
In a UNL graph, UWs appear with attributes
describing what is said from tile speaker' s poi nt
of view. This i ncl udes phenomena like speech
acts, truth wllues, time, etc.
Hyper nodes may also be used ill UNL
expressi ons.
a g t . . . . . . i
I d r i v e r . ~ P l ]
a o j
I r e c k l e s s ]
01 . @ entry
[ drink ]
\
d r i v e ]
Figure 1.2: A UNL Io,pergraph that may be
deconverted as "Reckless drivers drink and drive"
Graphs and subgraphs nmst contain one special
node, called the ent ry of tile graph.
1. 2. 3 De n o t i n g a UNL g r a p h
These hyper gr aphs are denot ed using the UNL
language per se. In the UNL hmguagc, an
769
expression consists in a set of arcs, connect i ng
the different nodes. As an exampl e, the graph
present ed in figure 1.1 will be denot ed as:
agt(score(...).@entry.@past.@complete,
Ronaldo)
o b j ( s c o r e ( _. ) . @e n t r y. @p a s t . @c o m p l e t e ,
goal(icl>thing))
ins(score(...) . @e n t r y. @p a s t . @c o m p l e t e ,
head(pof>body))
plt(score(...) . @e n t r y. @p a s t . @c o m p l e t e ,
corner)
obj (corner, g o a l ( i c l > t h i n g ) )
m o d ( c o r n e r , left)
Hyper nodes are denot ed by numbers. The
graph cont ai ned by a hypernode is denot ed as a
set of arcs col ored by this number as in:
agt (:Ol.@entry, driver. @pl)
aoj (reckless, driver.@pl)
and:Ol (drive, d r i n k . @e n t r y)
Entries of the graph and subgraphs are denot ed
with the ". @ent r y" attribute.
2 I n s i d e t he F r e n c h d e c o n v e r t e r
2. 1 Ov e r v i e w
Deconver si on is the process of t ransformi ng a
UNL graph into one (or possi bl y several)
ut t erance in a natural language. Any means
may be used to achi eve this task. Many UNL
proj ect partners use a speci al i zed tool called
DeCo but, like several other partners, we choose
to use our own tools for this purpose.
One reason is that DeCo realizes the
deconversi on in one step, as in some transfer-
based MT syst ems such as METAL [17]. We
prefer to use a more modul ar architecture and
to split deconversi on into 2 steps, transfer and
generation, each di vi ded into several phases,
most of them written in Arl ene-G5.
Anot her reason for not using DeCo is that it is
not well suited for the morphol ogi cal gene-
ration of inflected languages (several t housands
rules are needed for Italian, tens of t housands
for Russian, but onl y about 20 rules and 350
affixes suffi ce to build an exhaust i ve GM f or
French in Sygmor). Last, but not least, this
choi ce allows us to reuse modul es al ready
devel oped for French generation.
This strategy is illustrated by figure 2.1.
/ ~ ; ~ , ; o_;.,,.~ ,', T ransfer
~ . . . . . . v " Ge I~ ati0n her
/ \,4v
F r enc h u t t er anc e
Fig. 2.1:2 possible deconversqon strategies
Usi ng this approach, we segment the decon-
version process into 7 phases, as illustrated by
fi gure 2.2.
The third phase (graph-t o-t ree) produces a
decorat ed tree which is fed into an Ari ane-G5
TS (structural transfer).
V alklatiolff l, exicaI l'lansl~.'r (h~,li~h 1o t ree
[ .ocalization COllversion
, Z ~ "UNL Tree"
l'araphra~c choice
UMA stru ctu re?N\
Syntactic ~gcnerali(ln
,t
UMC s t r u c t u r e ~
Morl~ lm l.g ic[ll generation
' t
French u tterance
Fig. 2.2: architecture of the French deconverter
2 . 2 T r a n s f e r
2. 2. 1 Val i dat i on
When we recei ve a UNL Graph for decon-
version, we first check it for correctness. A UNL
graph has to be connect ed, and the di fferent
features handled by the nodes have to be
defi ned in UNL.
If the graph proves incorrect, an explicit error
message is sent back. This validation has to be
per f or med to i l aprove robust ness of the
deconvert er, as there is no hypot hesi s on the
way a graph is created. When a graph proves
valid, it is accept ed for deconversi on.
2. 2. 2 Loeal &at i on
In order to be correctly deconvert ed, tile graph
has to be slightly modified.
2. 2. 2. 1 Lexi cal localization
Some lexical units used in the graph may not be
present in the French deconversi on dictionary.
This pr obl em may appear under di fferent
ci rcumst ances. First, the French di ct i onary
(which is still under devel opment ) may be
i ncompl et e. Second, the UW nmy use an
unknown notation to represent a known French
word sense, and third, the LAV may represent a
non-French word sense.
We sol ve these probl ems with the same method :
Let w be a UWi n the graph G. Let D be the
French di ct i onary (a set of UWs). We substitute
w in G by w' such t hat : w' e D and
V x e D d(w, w' , G) = d(w, x, G). where d is a
pseudo- di st ance funct i on.
770
If different French UWs are at the same pseudo-
distance of w, w' is chosen at random among
these UWs (default in non-interactive mode).
2. 2. 2. 2 "Cultural" localization
Some crucial information may be missing,
dependi ng on the language of the source
utterance (sex, modality, number, determination,
politeness, ki nshi p. . . ).
It is in general impossible to solve this pr obl em
fully aut omat i cal l y in a perfect manner, as we
do not know anyt hi ng about the document , its
c:ontext, and i t s intended usage: FAHQDC 2 is no
more possible than FAHQMT on arbitrary texts.
We have to rely on necessarily i mperfect
heuristics.
ttowever, we can specialize tile general French
deconvert er to produce specialized servers f or
different tasks and different (target)
sublanguages. It is possible to assign priorities
not only to various parts of the dictionaries
(e.g., specialized vs. general), but also to
equivalents of the same UW within a given
dictionary. We can then define several user
profiles. It is also possible to build a memory of
deconvert ed and possibly postedited utterances
for each specialized French deconversi on
server.
2. 2. 3 Lexi cal Transf er
After the localization phase, we have to per f or m
the lexical transfer. It would seem natural to do
ill within Ariane-G5, after converting the graph
into a tree. But lexical transfer is cont ext -
sensitive, and we want to avoid the possibility of
transferring di fferent l y two tree nodes
corresponding to one and the same graph node.
Each graph node is replaced by a French lcxical
unit (LU), along with some variables. A lexical
unit used in tile French dictionary denotes a
derivational family (e.g. in English: de s t r oy
d e n o t e s d e s t r o y , d e s t r u c t i o n , d e s t r u c t i b l e ,
d e s t r u c t i v e . . . . i n F r e n c h : d 6t r u i r e f o r d 6t r u i r e ,
d e s t r u c t i o n , d e s t r u c t i b l e , i n d e s t r u c t i b l e ,
d e s t r u c t i f , d e s t r u c t e u r ) .
There may be several possible lexical units f or
one UW. This happens when there is a real
s ynonymy or when different terms are used in
different domains to denote the same word
s e n s e 3. I n that case, we currently choose tile
lexical unit at random as we do not have any
i nformat i on on tile task the deconverter is used
for.
Tile same problem also appears because of tile
slrategy used to build the French dictionary. In
order to obtain a good coverage from the
beginning, we have underspecified tile UWs and
linked them to dift' ercnt lexical units. This way,
we considered a UW as tile denotation of a set
of word senses in French.
Hence, we were able to reuse previous
dictionaries and we can use the dictionary even
if it is still under devel opment and incolnplete.
In our first version, we also solve this pr obl em
by a random selection of a lexical unit.
2. 2. 4 Graph to tree conversi on
The subsequent deconversion phases are
performed in Ariane-G5. Hence, it is necessary
to convert the UNL hypergraph into an Ari ane-
G5 decorated tree.
The UNL graph is directed. Each arc is labelled
by a semantic relation (agt, obj, ben, con. . . ) and
each node is decorated by a UW and a set of
features, or is a hypernode. One node i s
distinguished as the "entry" of the graph.
An ARI ANE tree is a general (non binary) tree
with decorations on its nodes. Each decorat i on
is a set of wlriable-value pairs.
The graph-to-tree conversion algorithln has to
lnaintain the direction and labelling of the
graph along with the decoration ot' the nodes.
Our algorithm splits tile nodes that are the target
of more than one arc, and reverses the di rect i on
of as few arcs as possible. An example of such a
conversion i s shown in figure 2.3.
! a ]
[5E3
/ \
J x,
x y
I ~ ! I c e
z t
b ( x
I
I
= > l
d : z +~
c : Y i
: 1
Fig. 2.3: example graph to tree convel:vion
Let Z be the set of nodes of G, A the set of
labels, T the created tree, and N is the set of
nodes of T.
Tile graph G= { (a,b,l) l a c Y. , b6 Z, I ~ A} i s
defined as a set of directed labelled arcs. We use
an association list A = { (n,;,n.r) I ,,,+ ~ r,, U. r E
N }, where we memori ze the cor r espondence
between nodes of the tree and nodes of the
graph.
2 fully autonmtic high quality dcconvcrsion.
3 strictly speaking, tile same collection o f intcrlingual
woM senses (acccptions).
771
l e t e(; e s u c h t h a t e i s t h e e n t r y o f G
e r 6- n e w t r e e - n o d e (ed, e n t r y )
i n T +- e r ( ) ; N 6- {e , r ] ; A <-- { ( e c ; , e T ) }
w h i l e G :~ O d o
i f t h e r e i s ( a , b , l ) i n G s u c h t h a t
G ~- G \ ( a , b , l ) ;
b r 6- n e w t r e e - n o d e ( b , i) ;
A 6- A <J {(b,b,,));
l e t a, r e N s u c h t h a t (a,a, r) e A
i n a d d b r t o t h e d a u g h t e r s o f a, r;
e l s e i f t h e r e i s ( a , b , l ) i n G s u c h t h a t (b,br) 6
G e- G \ ( a , b , l ) ;
a T ( - n e w t r e e - n o d e ( a , l i);
A <--- A U {( a , a . r ) };
l e t brl,e N s u c h t h a t ( b , b r ) e A
i n a d d a,, t o t h e d a u g h t e r s o f b r ;
else e x i t o n e r r o r ( " n o n c o n n e c t e d g r a p h " ) ;
( a , a. r) e A t h e n
A t h e n
2.2.5 Structural transfer
The purpose of the structural transfer is to
transform the tree obt ai ned so far into a
Generating Multilevel Abst ract ( GMA) st ruct ure
[ 4 ] .
In this structure, non-interlingual linguistic
levels (syntactic functions, synt agmat i c
cat egori es. . . ) are underspeci fi ed, and (i f
present), are used onl y as a set of hints for t he
generation stage.
2. 3 Ge ne r a t i o n
2.3.1 Paraphrase choice
The next phase is in charge of the par aphr ase
choice. During this phase, deci si ons are t aken
regarding the derivation appl i ed to each lexical
unit in order to obtain the correct synt agl nat i c
cat egory for each node. Duri ng this phase, the
order of appearance and the syntactic f unct i ons
of each parts of the utterance is also deci ded.
The resulting structure is called Uni que
Multilevel Abst ract ( UMA) structure.
2.3.2 Syntactic and morphological generation
The UMA structure is still lacking the synt act i c
sugar used in French to realize the choi ces
made in the previous phase by generat i ng
articles, auxiliaries, and non connect ed
compunds such as ne. . . pas, etc.
The role of this phase is to create a Uni que
Multilevel Concret e ( UMC) structure. By
concrete, we mean that the structure ~s
projective, hence the correspondi ng French text
may be obt ai ned by a standard left to right
traversal of the l eaves and simple mor phol ogi cal
and graphemi c rules. The result of t hese phases
is a surface French utterance.
3 Di f f e r e nt uses of t he UNL l anguage
3 . 1 Hy p e r g r a p h s vs col ored graphs
As present ed in section 1.2.3, the syntax of the
UNL l anguage is based on the descri pt i on of a
graph, arc by arc. Some of t hese arcs are
"col oured" by a number. This col ouri ng is
current l y i nt erpret ed as hyper nodes ( nodes
containing a graph, rather than a classical UW).
This interpretation is arbitrary and i mposes
semantic constraints on a UNL utterance:
the subgraph (the set of arcs l abel ed with
the same colour) is connect ed,
arcs with different col ours cannot be
connect ed to the same node.
However, even if one uses the UNL l anguage
for a particular kind of application, a di fferent
interpretation may be chosen. By addi ng new
semantic constraints to UNL expressi ons, one
may restrict to the use of trees. On the cont rary,
by l ooseni ng semantic constraint, one may use
col or ed graphs instead of the more restrictive
hyper gr aphs.
This fl exi bi l i t y of UNL may lead to uses that
di ffer from the comput er sci ence poi nt of view
(di fferent structures leading to di fferent ki nds
of met hods and applications) as well as from the
linguistic point of view (di fferent ways to
represent the linguistic content of a utterance).
This ki nd of structure is very useful to represent
some ut t erances like "Chri st i an pulls Gi l l es'
leg". Usi ng a col ored graph, one can represent
the ut t erance with the graph shown in f i gur e
3.1, whi ch is not a hypergraph.
772
01 . @ entry
a g t . . i [ p u ll. @ entry i
[ Chns~ lian ] I ~ , o b j
p o s
G es ]
Figure 3.1: this graph is not cut hypergral)h, it can
however be represented in UNL htnguage
When using normal hypergraphs, one coul d
only represent the utterance as shown in f i gur e
3.2.
a g t . . . . [ m a k e fu n of i
i Chns' ~ tan I
, i o b j
i i
Fi gure 3.2: this graph i s a valid hyperglztph
Heuce, keepi ng backward compat i bi l i t y with
other UNL based systems, one may devel op an
entirely new and more powerfl d kind of
application.
3. 2 Li n g u i s t i c vs s e n mn t i e p i v o t
The UNL language defines the i nt erface
structure to be used by applications (either a
hypergraph or a col ored graph). However, it
does not restrict the choi ce of the data to be
encoded.
Since t i l e beginning, two possible and wflid
apl~roaches has been ment i oned. Duri ng t he
ki ckoff meet i ng of t i l e UNL prelect, Pr. Tsujii
prolnoted t he use of UNL as a linguistic pivot.
With this approach, a UNL ut t erance should be
the encodi ng of the deep structure of a valid
English ut t erance that reflects the meani ng of
the source utterance. With this approach, t he
German sent ence "Hans schwi mt sehr g e r n "
should be encoded as shown in figure 3.3.
ag t. . _ - lik e. @ entry ~ . . .
[ H a-' ~ s [ ' "-. m a n
I o b j "- .
"A,
" ~- - a gt . . . . . . . . [ s~wim ] i much ,
Fi gmv 3.3: a linguistic encoding of "ltcms schwimt
sehr gern "
On the opposite, Hiroshi Uchi da promot es t he
use of UNL as a semantic pivot. With this
second approach, the same sent ence should be
encoded as shown in figure 3.4.
a g t / z s w i m . l @ e n t r y
/ / / I
~ i m a n
H w i l ~ l g l y
l i ned
Figure 3.4: a semantic encoding of "ltans schwimt sehr
gem"
Each approach has its advantages and
drawbacks and the choi ce between them can
onl y be made with an application in mind. The
linguistic approach leads to a better quality ill
the produced results and is an answer to hi ghl y
multilingual machi ne translation projects. With
this approach, the UNL graphs can onl y be
produced by people mast eri ng English or by
(partially) automatic enconvert ers.
With the semantic approach, subtle di f f er ences
in source utterances (indefinite, refl exi vi t y. . . )
can not be expressed, leading to a lower qual i t y.
However, using this approach, the UNL
encodi ng is much mor e natural and easy to
perform by a non English speaker (as t he
semantic relations and UWs are expressed at t he
source level). Hence, this approach is to be used
for multilingual casual communi cat i on where
users may express t hemsel ves by di rect l y
encodi ng UNL expressions with an appropri at e
editing tool.
C o n c l u s i o n
Worki ng oil tile French deconvel-ter has led to
im interestiug archi t ect ure where deconversi on,
in principle a "generation from interlingua", is
i mpl ement ed as t ransfer + generation f r om all
abstract structure ( UNL hypergraph) pr oduced
from a NL utterance. The idea to use UNL f or
directly creating document s gets here an
indirect and perhaps paradoxical support,
although it is cl ear that considerable progress
and innovative interface design will be needed
to make it practical.
However, the UNL l anguage proves fl exi bl e
enough to be used by very di fferent proiects.
Moreover, with deconvert ers cur r ent l y
devel oped for 14 languages, j oi ni ng the UNL
project is really attractive. Let ' s hope that this
effort will help breaking the language barriers.
Ac k n o wl e d g e me n t s
We would like to thank the sponsors of the UNL
project, especially UNU/IAS (T. Della Senta) &
ASCII (K.Nishi) and of the UNL-FR subproject,
especially UJF (C. Feuerstein), I MAG
(J. V oiron), CLIPS (Y. Chiaramella), and t he
773
French Ministery of Foreign Affai rs (Ph. Perez),
as well as the members of UNL Center,
especi al l y project leader H. Uchi da, M. L. Zhu,
and K. Sakai. Last but not least, ot her member s
of GETA have cont ri but ed in many ways to the
research reported here, in particular N. N6deau,
E. Blanc, M. Mangeot, J. Sitko, L. Fischer,
M. Tomoki yo, and K. Fort.
R e f e r e n c e s
[1] Bl anc l~. & Gui l l aume P. (1997)
Developing MT lingware through hlternet :
ARIANE and the CASH interface. Proc. Pacific
Association for Computational Linguistics 1997
Conference (PACLING'97), Ohme, Japon, 2-5
September 1997, vol. 1/1, pp. 15-22.
[2] Blanehon H. (1994) Persl)ectives of DBMT
f or monolingual authors on the basis of LIDIA-I, an
implemented mockup. Proc. 15th International
Conference on Computational Linguistics,
COLING-94, 5-9 Aug. 1994, vol. 1/2, pp.
115--119.
[3] Boitet C., R6d. (1982) " DSE- I " - - Le poi nt
sur ARIANE-78 ddbut 1982. Contrat ADI/CAP-
Sogcti/Champollion (3 vol.), GETA, Grenoble,
fdvrier 1982, 400 p.
[4] Boitet C. (1994) Di al ogue-Based MT attd
se(f exl)lahting documents as atl alternative to
MAHT and MT of controlled languages. Proc.
Machine Translation 10 Years On, 11 - 14 Nov. 1994,
Cranfield University Press, pp. 22.1--9.
[5] Boitet C. (1997) GETA' s MT met hodol ogy
attd its current development towards petwonal
networking communication attd speech translation in
the context of the UNL and C-STAR projects. Proc.
PACLING-97, Ohme, 2-5 September 1997, Meisei
University, pp. 23-57. (invited communication)
[6] Boi tet C. & Bl anehon H. (1994)
Multilingual Dialogue-Based MT f or monolingual
authotw: the LIDIA project arm a fil:s't mockup.
Machine Translation, 9/2, pp. 99--132.
[7] Boi tet C., Gui l l aume P. & Qu6zel-
Ambrunaz M. (1982) ARI ANE- 78, an
hltegrated environment f or atttomated translation attd
human revision. Proc. COL1NG-82, Prague, July
1982, pp. 19--27.
[18] Brown R. D. (1989) Augme nt at i on.
Machine Translation, 4, pp. 1299-1347.
[19] Ducrot J.-M. (1982) TI TUS I V. In
"Information research in Europe. Proc. ot' tile
EURIM 5 conf. (Versailles)", P. J. Taylor, cd.,
ASLIB, London.
[10] Kay M. (1973) The MIND system. In
"Courant Computer Science Symposium 8: Natural
Language Processing", R. Rustin, ed., Algorithmics
Press, Inc., New York, pp. 155-188.
[11] Maruyama H., Wat anabe H. & Ogl no S.
(1990) An Interactive Japanese Parser f or Machine
Translation. Proc. COLING-90, Helsinki, 20-25
aofit 1990, ACL, vol. 2/3, pp. 257-262.
[12] Mel by A. K., Smi t h M. R. & Pet er son
J. (1980) ITS : An Interactive Translation
System. Proc. COLING-80, Tokyo, 30/9-4/10/80,
pp. 424---429.
1113] Monei nme W. (1989) ( 159 p.
+annexes)7;40 vet's" l'arabe. Sp&'ification d'une
gdudration sfatldard de l'arabe. Rdalisation d'un
l)romO'l)e anglais'-ambe ?t partir d'un attalyseur
existant. Nouvelle thbse, UJF.
[114] Ni r enbur g S. & al. (1989) KB MT - 8 9
Project Report. Center for Machine Translation,
Carnegie Mellon University, Pittsburg, April 1989,
286 p.
[115] Nyberg E. H. & Mi t amura T. ( 1992)
The KANT system: Fast, Accurate, High-Quality
Translation in Practical Domains. Proc. COLING-
92, Nantes, 23-28 July 92, ACL, vol. 3/4, pp.
1069--1073.
[16] Qu6zel -Ambrunaz M. (1990) Ar i ane- G5
v.3 - Le moniteut: GETA, IMAG, juin 1990, 206 p.
[17] Sl oeum J. (1984) METAL: the LRC
Machine Translation O,stem. In "Machine
Translation today: the state of the art (Proc. third
Lugano Tutorial, 2-7 April 1984)", M. King, cd.,
Edinburgh University Press (1987).
[18] Wehrl i E. (1992) The IPS System. Proc.
COLING-92, Nantes, 23-28 July 1992, vol. 3/4, pp.
870-874.
774