Creating a Universal Networking Language Module 
within an Advanced NLP System 
Igor BOGUSLAVSKY, Nadezhda FRII), l,eonid IOM1)IN, Leonid KREII)LIN, hina SAGALOVA, 
Victor SIZOV 
Coml)utational l~inguistics Laboratory 
Institute for Information Transnlissioll Problems of lho P, ussian Academy of Sciences 
19 Bol'shcti Karetnyj, 101447 Moscow, Russia 
J)~nadya ionldin,leny{~lova~)~).ru 
Abstract 
A multifunctional NIA 9 environment, 
\[!'I'AI~-3, is presented. The environment has 
several NI,I ~ applications, inchtding a machine 
translation system, a natural  interface 
to SQI, type databases, synonymous 
l~araphrasing of sentences, syntactic error 
correction module, and a computer-assisted 
 learning tool. Ihnphasis is laid on a 
new naodtile of tile processor responsible for tlio 
intorl\]lcc with the Universal Networking 
l A.lllgtlagC, il roCOlll plodtlcl by the UN 
Universily inlended for the facilitation of 
nnlltihlnguage, multiethnic access 1o 
communication networks such as WWW. The 
UNL module of ETAP-3 naturally combines the 
two major al)proaches accepted in machine 
translation: the lransfer-based approach and the 
interlingua apl)roach. 
1. Introductory Renmrks 
ETAP-3 is a multilmrposc NIA ~ environmmlt 
that was conceived in the 1980s and has been 
worked out in the Institute for lnl~mnation 
Transmission Problems, Russian Academy of 
Sciences (Apresian et al. 1992, l?,oguslavsky 
1995). The theoretical foundation of ETAP-3 is 
tile Meaning ¢=> Text linguistic theory by Igor' 
Mel'6uk and the Integral Theory of Language by 
Jurij Apresian. 
Eq'AP-3 is a non-comlnercial environment 
primarily oriented at linguistic research rather 
than creating a marketable software product. The 
main focus of the research carried ()tit with 
I';TAP-3 is COlnputational modelling of natural 
s. This attitude explains our effort to 
develop the models in a way as linguistically 
sound as possible. We strive to incorporate into the 
system much linguistic knowledge irrespective of 
whether this knowledge is essential for better text 
processing (e.g. machine translation) or not. In 
particular, we want our parser to produce what wc 
consider a correct syntactic representation of tim 
sentence - first of all because we believe that this 
interpretation is a true fact about tile natural 
. We have had inany occasions to set that 
in the long run the iheorctical soundness and 
completeness of linguistic knowledge incorporated 
in an NIA ) application will pay. 
All NLP applications in F, TAP-3 are largely 
based on an original system of three-wdue logic 
and use an original formal  of linguistic 
descriptions, I~'Oi>&;T. 
2. ETAP-3: Modules, Features, Design, 
hnplementatioll 
2.11 ETAP-3 Modules 
The m~tjor NI,P modules of ETAP-3 are as 
lollows: 
• High Quality Machine Translation System 
• Natural Language Interface to SQL Type 
Databases 
• System of Synonymous Paraphrasing of 
Sentences 
• Syntactic Error Correction Tool 
• Computer-Aided Language LeamingTool 
• Tree \]?,ank Workbencll 
Another module, a new UNL converter 
responsible lot the interface with the Universal 
Networking Language, a recent product designed 
' Tile research reported here was in part supl)oricd by a grant (No 99-06-80277) fronl tile Russian Foundation 12)1" 
Fundamental Research, whose assistance is gratefully acknowledged. 
83 
by the UN University, is discussed in detail in 
Section 3. 
2.1.1. ETAP-3 MT System 
The most important module of ETAP-3 is 
the MT system that serves five  pairs: 
(1) English-Russian, (2) Russian - English, (3) 
Russian - Korean. (4) Russian - French, and (5) 
Russian - German. 
By far the most advanced are the first two 
of these pairs. The system disposes of 50,000- 
strong so-called combinatorial dictionaries of 
Russian and English that contain syntactic, 
derivational, semantic, subcategorization, and 
collocational information. The system relies on 
comprehensive grammars of the two s. 
For the other  pairs smaller scale 
prototypes are available. 
ETAP-3 is able to present multiple 
translations when it encounters an ambiguity it 
cannot resolve. By default, the system produces 
one parse and one translation that it considers 
the most probable. If the user opts for multiple 
translation, the system remembers the 
unresolved ambiguities and provides all 
mutuany compatible parses and lexical choices. 
To give one example from the real output: the 
sentence They made a general remark that .... 
when submitted to the multiple translation 
option, yielded two Russian translations that 
correspond to radically different syntactic 
structures and lexical interpretations: (a) Oni 
sdelali obshchee zameehanie, chto... (= They 
made some comn-lon renlark that ...) and (b) Oni 
vynudili generala otmetit; chto... (= They forced 
some general to remark that ...). 
2.1.2. Natural Language Interface to SQL Type 
Databases 
This ETAP-3 module translates freely 
worded human queries to a database from 
Russian or English into SQL expressions. It can 
also produce the reverse generation of a NL 
query from an SQL expression. 
2.1.3. System of Synonymous Paraphrasing 
The module is designed for linguistic 
experiments in obtaining nmltiple meaning- 
retaining paraphrases of Russian and English 
sentences. The paraphrasing is based on the 
concept of lexical functions, one of the 
important innovations of the Meaning ¢=> Text 
theory. The following example shows the kind 
of paraphrases that can be produced by the 
module: 
(1) The director ordered John to write a report - 
The director gave John an order to write a report 
- John was ordered by the director to write a 
report - John received an orcler fonn the director 
to write a report. 
It is a very promising direction of linguistic 
research and developlnent that can be applied in a 
wide range of activities, including  
learning and acquisition, authoring, and text 
planning. Besides that, lexical functions are used 
for ensuring adequate lexical choice in machine 
translation and in the UNL module. 
2.1.4. Syntactic Error Correction Tool 
The module operates with Russian texts in 
which it finds a wide range of errors in 
grammatical agreement as well as case 
subcategorization and offers the user the correct 
version. 
2.1.5. Computer-Aided Language Learning Tool 
The module is a standalone software 
application constructed as a dialogue type 
computer galne intended for advanced students of 
Russian, English, and German as foreign 
s who wish to enrich their vocabulary, 
especially to master the collocations of these 
natural s and their periphrastic abilities. 
The tool relies on the apparatus of lexical 
limctions. It can also be used native speakers of the 
three s interested in increasing their 
command of the vocabulary (such as journalists, 
school teachers, or politicians). 
2.1.6. Tree Bank Workbench 
This is the module that utilizes the ETAP-3 
dictionaries, its morphological analyzer and the 
parser to produce a first-ever syntactically tagged 
corpus of Russian texts. It is a mixed type 
application that combines automatic parsing with 
human post-editing of tree structure. 
2.2. Major Features 
The following ate the most important features of 
the whole ETAP-3 environment and its modules: 
* Rule-Based Approach 
• Stratificational Approach 
• Transfer Approach 
• Syntactic Dependencies 
• Lexicalistic Approach 
• Multiple Translation 
• Maximum Reusabilty of Linguistic Resources 
84 
Ill tile current version of ETAP-3, its 
modules that process NL senteuces are strictly 
rule-based. However, ill a series of recent 
experiments, tile MT module was supplenlenled 
by all example-based component of a translation 
menlory type and a statistical component that 
provides semiautonmtic extraction of translatiou 
equivalents t¥om bilingual text corpora (see 
lomdin & Streiter 1999). 
ETAP-3 shares its stratificational feature 
with many other NLP systems. It is at tile level 
of tile normalized, or deep syntactic, structure 
that tile transfer flom tile source to tile target 
 takes place in MT. 
ETAP-3 makes use of syntactic dependency 
trees for sentence structure representation 
instead of constituent, or phrase, structure. 
Tile ETAP-3 system takes a lexicalistic 
stand ill tile sense that lexical data are 
considmed as important as gl'ammar 
infornlation. A dictionary entry contains, in 
addition to tile lemma name, information on 
syntactic and semantic features of tile word, its 
subcategorization flame, a default translation, 
and rules of various types, and wdues of lexical 
functions for which tile lemma is tile keyword. 
The word's syntactic t'eatures characterize its 
ability/nou-ability to participate ill specific 
syntactic constructions. A word can have several 
syntactic features selected from a total of more 
than 200 items. Semantic features arc needed to 
check tile semantic agreement between the 
words ill a sentence. Tile subeategorization 
frame shows the surface marking of tile word's 
arguments (in terms of case, prepositions, 
conjtmctions, etc.). Rules are an essential part of 
the dictionary entry. All the rules operating in 
ETAP-3 are distributed betwecn tile granmmr 
and tile dictionary. Grammar rules me more 
general and apply lo large classes of words, 
whereas tile rules listed or simply referred to in 
the dictionary are restricted ill their scope and 
only apply to small classes of words or even 
individual words. This organization of tile rules 
ensures the self-tuning of tile system to tile 
processing of each particular senteuce. In 
processing a sentence, only those dictionary 
rules are actiwlted that are explicitly referred to 
ill the dictionary entries of tile words making up 
tile sentence. A sample dictionary enlry 
fl'agment for tile English noun chance illustrates 
what was said above: 
\[1\] CHANCE1 
\[21 POR:S 
\[3\] SYNT:COUNT,PREDTO,PREDTHAT 
\[4\] DES:'FACT','ABSTRACT' 
\[5\] D1.1 :OF,'PERSON' 
\[6\] D2.1 :OF,'FACT' 
\[7\] D2.2:TO2 
\[8\] D2.3:THAT1 
\[9\] SYN 1 :OPPORTUNITY 
\[10\] MAG N: GOOD 1~FAIR 1~EXCELLENT 
\[11\] ANTIMAGN: SLIGHT/SLIM/POOR/LITTLE1/ 
SMALL 
\[12\] OPER1 :HA VE/STAND1 
\[13\] REAL1-M: TAKE 
\[14\] ANTIREAL1 -M:MISSl 
\[15\] INCEPOPER1 :GET 
\[16\] FINOPER1 :LOSE 
\[17\] CAUSFUNC1 :GIVE<TOI>/GIVE 
\[18\] ZONE:R 
\[19\] TRANS:SHANS/SLUCHAJ 
\[20\] REG:TRADUCT2.00 
\[21\] TAKE:X 
\[22\] LOC:R 
\[23\] R:COMPOS/MODIF/POSSES 
\[24\] CHECK 
\[25\] 1.1 DEP-LEXA(X,Z,PREPOS,BY1) 
\[26\] N:01 
\[27\] CHECK 
\[2811.1 DOM(X,*,R) 
\[29\] DO 
\[30\] 1 ZAMRUZ:Z(PO1) 
\[31\] 2 ZAMRUZ:X(SLUCFtAJNOST') 
\[32\] N:02 
\[33\] CHECK 
\[34\] 2.1 DOM(X,*,*) 
\[351 DO 
\[36\] I ZAMRUZ:Z(SLUCHAJNO) 
\[37\] 2 STERUZ:X 
\[38\] TRAF:RA-EXPANS.16 
\[39\] LA:THAT1 
\[40\] TRAF:RA-EXPANS.22 
lane \[12\] - part of speech: a noun. 
Line 113\] - tile list of syntactic features. 
Lille \[4\] - tile list of senmntic features. 
Lines \[5\] - \[8\] - tile subcategorization fiame. 
Lines \[9\] - \[17\] - the list of lexical functions used 
to describe restricted lexical co-occurrence. 
Lille \[18\] - marks the end of the application- 
independent infomlation and beginning of the 
information used in tile English-Russian 
translation. 
Lille \[ 19\] - default translation into Russian. 
lanes \[20\] - \[37\] - a rule for translating tile phrase 
t) 3 , chance in different contexts. 
Lines \[38\] - \[39\] - a reference to the rule which 
introduces a semantically empty conjunction (that: 
a chance that we obtain a grant). 
Line \[40\] - a reference to the rule which 
introduces particle to (a chance to win). 
85 
2.3. General Architecture of the ETAP-3 
environment. 
To give a general idea of how the ETAP-3 NLP 
operates, we show here the layout of the MT 
module (Fig. I). In a way, all the other modules 
can be viewed as this module's deriwttives. 
OBJECTS STAGES 
Source son\[ollCO 
MorphS source 
SyntS source 
NormS source 
NorlnS target 
SyntS target 
MorphS target 
Target sentence 1 
Morphological analysis 
Pal'sing 
Norlnalization 
Transfer 
Expansioll 
Syntactic synthesis 
Morphological synthesis 
Fig.1 
DICTIONARIES 
SOUI'CO 
morplaological 
dicl:ionarv 
SOllI'CO 
Combinatorial 
Dictionary 
.~ t~ Target 
Colnbinatorial 
~ I DictionarYTarget 
-~ lnorphological 
dictionary 
2.4. hnplementation 
The ETAP-3 environment has been implemented 
on a PC under Windows NT 4.0 environment. 
The environment has a number of auxiliary 
tools, including a sophisticated lexicographer's 
toolkit that allows the developers and the users 
to effectively maintain and update the ETAP-3 
dictionaries. 
3. The UNL Interface 
3.1 Aims and scenario 
The UNL project has a very ambitious goal: 
to break down or at least to drastically lower the 
 barrier for the Internet users. With time 
and space limitations already overcome, the 
Internet community is still separated by 
 boundaries. Theoretically, this seems 
to be the only major obstacle standing in the way 
of international and interpersonal 
communication in the information society. This 
is why the problem of the  barrier on 
the Interact is perceived as one of the global 
problems of mankind, and a project aiming to 
solve this problem has been initiated under the 
UN auspices - by the Institute of Advanced 
Studies of the United Nations University. 
Started in 1996, the project curremly 
embraces 15 universities and research 
institutions fiom Brazil, China, Egypt, France, 
Germany, India, Indonesia, Italy, Japan, Jordan, 
Latvia, Mongolia, Russia, Spain, and Thailand. 
In the following years more groups are expected 
to join, so that in the long run all s of 
the UN member states will be covered. 
The idea of the project is as follows. An 
interlingua has been developed which has 
sufficient expressive power to represent 
relevant information conveyed by natural 
s. This interlingua entitled Universal 
Networking Language (UNL) has been 
proposed by H. Uchida (UNU/IAS). For each 
natural , two systems should be 
developed: a "deconverter" capable of 
translating texts from UNL to this NL, and an 
"enconverter" which has to convert NL texts 
86 
into UNL. it sholtld be emphasized that the 
procedure of producing a UNL text ix not 
supposed to be fully autolnatic. It will be an 
interactive process with the labor divided 
between the COlnputer and a human expert 
("writer") in UNI+. 
This paradigm makes UNL radically 
different from conventional machine lranslation. 
Duo to the interactive oncoilversion, the UNL 
expression, which serves as inpul for generation, 
can be nlado as good as Clio wishes. The UNL 
writer will edit the rough result proposed by the 
OllConvorlor, corfect its errors, eliminate the 
renlaining ambiguities. He/she can run a 
deconvorior of his own  to lest the 
wtlidity of the UNL expression obtained alld 
then refine it again tin one is fully satisfied with 
the l'inal result. 
Anolhor ilnl)oriant distinction l'roill MT 
systonis is thai lhe inlorlirigua roprosenhilion of 
texts will be created and stored irrespectively of 
ils goiloration into particular s. UNL 
Call be soon as all independent i-ileal-iS of iYloanillg 
ropreselllation. UNL doctlmonts Call 13e 
processed by indexing, retrieval and knowledge 
extraction tools without being converted to 
llattll'al lallguages. Gellcration \viii only be 
needed when the document has roached the 
htllll;_lll HSOl +. 
A doconvoftor and an enconvoi'tor for each 
lliligtlagC form ii IAlnguago Server residing in the 
hilernot. All  scrvolS will be cotlnoclod 
in the IJNL network. They \viii allow ally 
IlliOiTiOt user to doconvorl a UNI, docunleili 
found on the web into his/her native , as 
well as to produce UNI, represelltatiOllS of the 
texts he/she wishes to nlako available to 
inultiethnic public. 
3.2 UNL  
We cannot describe the UNL  here 
in all details: this topic deserves a special paper 
which will hopefully be written by the author of 
the  design - l)r. Hiroshi Uchida. We 
will only characterize it to the extent necessary 
for the description of our deconversion module. 
Full specification of UNL can be found at 
/lllp://WWW. tml. tax. ttnu. edu/. 
UNL is a comlmter  intended to 
represent infolmation in a way that allows to 
generate a text expressing this information in a 
very large number of nahtral s. A UNL 
expression is an oriented hyper-graph that 
corresponds to a NL sentence in the amount of 
information conveyed. The arcs of t11o graph are 
interpreted as senmntic relations of the type 
agent, ob.ject, lime, place, inslrttment, manlier, 
etc. The nodes of the graph are special units, the 
so-called Universal Words (UW) interpreted as 
concepts, or groups of UWs. The nodes can be 
supplied with attributes which provide 
additional information on their use in lhc given 
sentence, e.g. @imperative, @generic, @future, 
@obligation. 
Each UW is represented as an t~,nglish 
word that can be optionally supplied with 
semantic specifications to restrict its meaning. 
In most cases, these specifications locate the 
concept in the knowledge base. It is done in the 
following way: UW A(icl>B) ix interpreted as 
'A is subsumed under the category B'. For 
example, the UW coach used without any 
restrictions denotes anything the English coach 
can denote, ll' eric wants to be more precise, one 
can use restrictions: coach(icl>transl)ort ) 
denotes a bus, coaclz (icl>lmman) denotes a 
trainer and coach (icl>do) denotes the action of 
training, in a sense, the apparatus of restrictions 
allows to represent UWs as disambiguated 
l';nglish words. On ltle other hand, restrictions 
allow to denote concepts which are absent in 
I~;nglish. For cxmnple, in Russian there is a large 
group of motion words, whose meaning 
incorporates the idea of the mode of locomotion 
or tral/sportation: priletel' 'come by flying', 
prO@,/' 'come by ship', l)ril)olzti 'come by 
crawling', l)ril)eJlal ' 'come running', elc. 
l!nglish has no neutral words to denote these 
concepts. Still, on the basis of English one can 
constrttct lJWs that approximate required 
concepts, e.g. conw(met>shil) ) is interpreted as 
'come and the method o1' coming ix a ship'. 
IIere is an example of a UNL expression 
for the sentence 
(2) Howevel, hmgua,q,e dll/ferences are a barrier 
to the smoot/L/low of in.fomnation in our society. 
l';ach line is an expression of the kind 
rehttion(UWl, UW2). For simplicity, UWs are 
not supplied with restrictions. 
aoj (barrier. @entry. @present. @indef. @however, 
difference. @pl) 
rood(barrier. @entry. @present. @indef. @ however, 
Ilow. @dcl) 
mod(differencc.@pl, ) 
aoj(smoofli, flow. @del) 
meal(flow. @def, in fol'nmtion) 
scn(fk+w. @dcl, society) 
pos(society, we) 
P, ehttions used: ao i a relation that holds 
between a thing and its state, nmd - a relation 
87 
between a thing and its modifier, scn- a relation 
between an event or a state and its abstract 
location, pos - a relation between a thing and its 
possessor. Attributes: @entry - denotes the top 
node of the structure, @present - present tense, 
@def - definite NP, @pl - plural, @however - a 
modal meaning corresponding to English 
however. 
3.3. UNL - Russian deeonversion by 
means of ETAP-3 
As was shown in Section 1, ETAP-3 is a 
transfer-based system where the transfer is 
carried out at the level of the Normalized 
Syntactic Structure (NormSS). This level is best 
suited for establishing correspondence with 
UNL, as UNL expressions and NormSS show 
striking similarities. The most important of theln 
are as follows: 
1. Both UNL expressions and NormSSs 
occupy an intermediate position between the 
surface and the semantic levels of 
representation. They roughly correspond to 
the so-called deep-syntactic level. At this 
level the meaning of the lexical items is not 
decomposed into the primitives, and the 
relations between the lexical items are 
 independent; 
2. The nodes of both UNL expressions and 
NormSSs are terminal elements (lexical 
items) and not syntactic categories; 
3. The nodes carry additional characteristics 
(attributes); 
4. The arcs of both structures are non- 
symmetrical dependencies. 
At the same time, UNL expressions and 
NormSSs differ in several important respects: 
1. All the nodes of NormSSs are lexical items, 
while a node of a UNL expression can be a sub- 
graph; 
2. Nodes of a NormSS always correspond to one 
word sense, while UWs may either be broader or 
narrower than the corresponding English words: 
2.1. they can cover a meaning area that 
corresponds to several different word senses at a 
time (see above); 
2.2. they can correspond to a fi'ee word 
combination (e.g. computer-based or high- 
quality); 
2.3. they can correspond to a word form 
(e.g. best which a form of good or well); 
2.4. they can denote a concept that has no 
direct correspondence in English (see above). 
3. A NormSS is the simplest of all connected 
graphs - a tree, while a UNL expression is a 
hyper-graph. Its arcs may form a loop and 
connect sub-graphs; 
4. The relations between the nodes in a NormSS 
are purely syntactic and are not supposed to 
convey a meaning of their own, while the UNL 
relations denote semantic roles; 
5. Attributes of a NormSS mostly correspond to 
grammatical elements, while UNL attributes 
often convey a meaning that is expressed both 
in English and in Russian by means of lexical 
items (e.g. modals); 
6. A NormSS contains information on the word 
order, while a UNL expression does not say 
anything to this effect. 
The NormSS of tile sentence (2) looks as 
follows: 
be, present 
I~' prcdic "",ik 
however difference, pl barrier, i,l~t~l 
compos ~ l-compt 
flow, def 
lanvuave / I \attrib ll\]~-Colnp\]~ 
smooth information in 
prepos~ 
society 
Fig. 2 modif~ 
our 
As UNL makes use of English lexical labels, it 
is expedient to bridge the gap between UNL and 
Russian via English NormSS which actually 
serves as an Intermediate Representation (IR). 
in this case tile UNL - Russian interface will be 
the simplest. After the English NormSS has 
been reached, conventional ETAP English-to- 
Russian machine translation mode of operation 
can be used. 
The UNL-to-Russian module carries out the 
following three steps: 
1. Transfer from UNL to the intermediate 
representation (IR). 
88 
2. Transfer fronl tile IR to tile Russian 
ilOllnalized syntactic structure (NorlriSS-1)@ 
3. (\]eneration of a P, ussian sentence from the 
NornlSS-R. 
Tile archilecture of tile UNL-Russian 
deconverter is shown in Fig. 3. 
It follows fi'om tile previous discussion that the 
UNL - NormSS interface should solve the 
following five tasks: 
1. An appropriate English lexeme for every 
UW should be selected where it is possible; 
a Russian lexeme will be provided by tile 
ETAP English - Russian transfer dictionary. 
If no appropriate English word can be found 
for a UW, other means of expression should 
be found. 
2. UNL syntactic relations should be 
tl-anslated, either by means of I~q'AP 
relations or widl tile help of lexical items. 
3. UNL attributes should be translated, either 
by lneaus of granunatical features or with 
the help of lexical items (e.g. @however - 
however). 
4. UNL graph should be converted in a tree. 
5. Word order should be established. 
The first aild (parlly) the second tasks are 
soh, ed by uleaus Of the infornlatiou stored in the 
UW English and English conlbinalorial 
dictionaries. All lhe rest (tasks 2 io 5) is done by 
the rules written in the logical-based I~'OP, tZT 
formalism. 
Let us give one example lo ilhlstrate the 
transformation of UNL relations into NL words. 
UNL has a tim relation that holds between an 
event and its linle. As is known, lhe choice of 
approl)riaie words to express lhis relation is to a 
largo oxleni doterilliried by lexical properties of 
tile word denoting tilne; cf. oz._It Moll(lay, at 
midnight, idAl summe#; rhtri, e~ the it,at; etc. In 
ETAP-3 all these cases are treated as tile lexical 
function LOC denoting (tenlporal) locality (on 
lexical functions see 2.1.3). Tile values of all 
lexical fimctions are given in the lexicon in the 
entries of their arguments (see an example in 2.2 
above). While processing tile UNL expression, 
the tim relation is linked to the lexical ftluclioll 
LOC which allows to l'iud a correct preposition, 
both in English and in Russian. 
3.4. Current state and prospects for the 
future 
Tile module of Russian deconversiou is 
operational and can be tested at 
ccc ctrcctcre 
II 
Intermediate Representation cccnccisc cormaciced 
ccntactic ctrcctcrec 
cnccisc cc :rcace ccntactic i I ctrcctcre 
i 
Fig.3 
I Rcssian cormaciced \[ ccntactic ctrcctcre 
ll 
ii 
II 
Rcssian ccrcace 2 ccntactic ~' ~ctrcctcre 
F-----G cnccisccorpcocoeicac I Rcssian corpcococicac ctrcctcre I ctrcctcre ------T 
centence centence 
hitp://proling.iitp.ru/Ooco. We plan to put it to 
geUela\] rise by aulunlu 2000. Tile interactive 
enconvorsion n\]odulo will be our next concorll. 
As sllo,vn ill Fig. 3, the interface botweou 
UNL and Russian is established at tile level of 
the English NorlllS . At this point ETAP 
English-to-Russian nlachine Iranslation facility 
can be switched which carries through tile 
phases of transfer and Russian generation. This 
architecture allows to obtain English generalion 
for relatively cheap, as ETAP has a Russiau-to- 
English mode of operation as well. First 
experiments in this direction have been carried 
Otll which proved quite promising. 

References 

Aprcsian Ju.D., I.M.B%uslavsky, L.L.Iomdin et al. 
(1992). ETAP-2: The Linguistics of a Machine 
Translation System.//META, Vol. 37, No 1, pp. 
97-112. 

Boguslavsky 1.(1995). A bi-dircclional Russian-lo- 
English nlachino Iranslaiion system (ETAP-3). H 
Proceedings el' die Machine Trallslalion Sumnlii 
V. guxonlbourg. 

lomdin L.& O. Slroiter. (1999). Learning 1"1"o111 
Pa,'allel Corpora: Experiments in Machine 
Translation. // l)ialogue'99: Compulational 
IJnguislics and ils Applications International 
Workshop. Tarusa, Russia, June 1999. Vol.2, pp. 
79-88. 
