Multilinguality in ETAP-3: Reuse of Lexical Resources 
Igor BOGUSLAVSKY 
Universidad Politecnica de Madrid 
28660 Boadilla del Monte, Madrid, Spain 
igor@opera.dia.fi.upm.es  
 
Leonid IOMDIN 
Institute for Information Transmission 
Problems, Russian Academy of Sciences 
19, B. Karetnyj 
Moscow, GSP-4, Russia  
iomdin@cl.iitp.ru 
 
Victor SIZOV 
Institute for Information Transmission Problems, Russian Academy of Sciences 
19, B. Karetnyj 
Moscow, GSP-4, Russia  
sizov@cl.iitp.ru 
 
Abstract 
The paper presents the work done at the Institute 
for Information Transmission Problems (Russian 
Academy of Sciences, Moscow) on the 
multifunctional linguistic processor ETAP-3. Its 
two multilingual options are discussed – machine 
translation in a variety of language pairs and 
translation to and from UNL, a meaning 
representation language.  
For each working language, ETAP has one 
integral dictionary, which is used in all 
applications both for the analysis and synthesis 
(generation) of the given language. In difficult 
cases, interactive dialogue with the user is used for 
disambiguation. Emphasis is laid on multiple use 
of lexical resources in the multilingual 
environment.  
1 General Information on ETAP  
The multifunctional ETAP-3 linguistic 
processor, developed by the Computational 
Linguistic s Laboratory (CLL) in Moscow (see e.g. 
Apresjan et al. 1992a,b, 1993, 2003), is the product 
of more than two decades of laboratory research 
and development in the field of language 
modeling. The most important features of the 
processor are as follows. 
(1) ETAP-3 is based on the general linguistic 
framework of the Meaning ⇔ Text theory, 
proposed by Igor Mel’cuk (e.g. Mel’cuk, 1974) 
and complemented by the theory of systematic 
lexicography and integrated description of 
language proposed by Jurij Apresjan [Apresjan 
1995, 2000].  
(2) ETAP-3 has a declarative organization of 
linguistic knowledge.  
(3) One of the major components of ETAP-3 is 
the innovative combinatorial dictionary. Apart 
from syntactic and semantic features and 
subcategorization frames, the dictionary entry may 
have rules of 8 types. Many dictionary entries 
contain lexical functions (LF).  
(3) ETAP-3 makes use of a formalism based on 
three-value predicate logic, in which all linguistic 
data are presented.  
(4) The ETAP-3 processor has a modular 
architecture. All stages of processing and all types 
of linguistic data are organized into modules, 
which warrants their reusability in many NLP 
applications both within and beyond ETAP-3 
environment.  
At the moment, the ETAP-3 environment 
comprises the following main options: 1) a rule -
based machine translation system; 2) a Universal 
Networking Language (UNL) translation engine; 
3) a system of synonymous paraphrasing of 
sentences; 4) a workbench for syntactic annotation 
of text corpora; and 5) a grammar checker. All the 
applications make use of the same dictionaries, but 
only the first and the second are multilingual. In 
Section 2 we will discuss multilingual lexical 
resources used in machine translation, and in 
Section 3 – in the UNL module.  
2 Multilinguality in ETAP 
2.1 Structure of the Dictionary Entry 
To support multilinguality, the dictionary entry 
of the ETAP dictionary has several sub-zones. 
There is one general zone and several zones 
oriented towards various languages. The general 
zone stores all types of monolingual information: 
part of speech, syntactic features, semantic 
features, subcategorization frames, lexical 
functions, syntactic and pre-syntactic rules, 
generation rules, and some other data. Each bi-
lingual sub-zone serves for establishing 
correspondence between the given language and 
another one (see Fig. 1).  
For example, the Russian zone of an English 
dictionary entry contains all the information 
needed to translate English words into Russian, the 
Arabic zone provides translation into Arabic, etc. 
Conversely, the information needed to translate 
Russian words into English is stored in the English 
zone of the Russian dictionary entries. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Fig. 1 
 
2.2 Default and Specific Translation 
The information stored in a bi-lingual zone 
consists of two parts: a default translation and 
lexical translation rules. Default translation is a 
single word that translates the given word in non-
specific contexts (it is introduced by a special 
label: TRANS). Any other type of translation is 
carried out by means of rules. If the word is 
translated by a phrase consisting of several words, 
the rule shows how the words in the phrase are 
connected to each other and how this phrase is 
incorporated into the sentence. For example, in the 
entry bachelorship we find a reference to one of 
the standard translation rules (TRADUCT2.42). 
The slots of the rule are filled with specific lexical 
items, grammatical features or syntactic relations.  
TRAF:TRADUCT2.42 
LR1:STEPEN’,LR2:BAKALAVR,T2:SG, 
T3:QUASIAGENT 
The rule says that bachelorship  should be 
translated into Russian with a phrase consisting of 
two words – stepen’ (‘degree’) and bakalavr 
(‘bachelor’). These words should be connected by 
the quasiagent(ive) syntactic relation, and the 
number feature of bakalavr should be singular.  
If the word is translated in a specific way in a 
specific context or in specific phrases, the rule 
describes this context and the resulting structure. 
When a word is translated, normally first the 
translation rules in its dictionary entry are tried. If 
no rule applies in the given sentence, then the 
default translation is used.  
2.3 Multiple Translation 
The default option of ETAP produces a single 
translation of the sentence – the one that 
corresponds to the first lexico-syntactic structure 
obtained by the parser. The option of multiple 
translation produces much more. First, it generates 
all lexico-syntactic structures that are compatible 
with the grammar and the dictionary. Since these 
structures are disambiguated both syntactically, 
and lexically, this set of structures contains all 
lexical variants for the source sentence. Then, for 
each structure all possible translation variants are 
tried. As is known, even disambiguated words can 
be translated into another language in different 
ways and it is not always possible to formulate a 
rule that could select an appropriate variant. For 
example, English adjuration can be translated into 
Russian as mol’ba and as zaklinanie, adventurer – 
as avantjurist and as iskatel’ prikljuchenij 
(literally, ‘adventure seeker’), alarm – as trevoga 
and as avarijnyj signal (‘alarm signal’). In all these 
cases, we are most probably dealing with a single 
meaning of the English word and yet translation 
variants are not fully synonymous. Since we 
cannot choose among these variants by means of 
rules and at the same time do not want to lose any 
of them, we have to treat them as alternative 
translations to be activated in the “Multiple 
translation” option. As mentioned in the previous 
section, there are two types of translation devices 
in the bilingual zones of the dictionary: a default 
translation (a single word) and rules. In both cases, 
it is possible to provide alternative translations. For 
example, in the entry for adjuration alternative 
translations are listed in the default part since both 
of them are single words: 
 
ADJURATION 
… 
TRANS: MOL’BA / ZAKLINANIE 
 
If the user selects the “Single translation” option, 
only the first of these variants will be used. If 
ENGLISH WORD 
General information: 
- part of speech 
- syntactic features 
- semantic features 
- subcategorization frame 
- …  
Russian zone 
Arabic zone 
… 
UNL zone  
he/she wishes to get all possible translations and 
activates the “Multiple translation” option, both 
alternatives will be produced.  
In the adventurer entry, the alternative translation 
iskatel’ prikljuchenij should be introduced by a 
rule, since it is not a single word but a phrase. Such 
rules are supplied by a special marker, OPT(ional), 
which shows that the translation is alternative.  
 
ADVENTURER 
… 
TRANS: AVANTJURIST 
TRAF:TRADUCT2.42 
OPT:1 
LR1:ISKATEL’2,LR2:PRIKLJUCHENIE,T2:PL, 
T3: ATTRIB 
 
This is another instance of the same rule that we 
saw above in the bachelorship example: the only 
difference is that it introduces different words, 
connects them with a different syntactic relation 
(attributive) and generates a different number 
feature. The marker OPT:1 shows that the 
translation introduced by this rule is less common 
than the default translation avantjurist and should 
be presented to the user after it. Should it be 
otherwise, the rule would have the marker OPT:0 
and have a priority over the default translation.  
2.4 Interactive selection of the translation 
equivalent  
It is well known that ambiguity of linguistic 
units is one of the most difficult problems in NLP. 
In ETAP there is no single stage of processing that 
expressly deals with disambiguation. The sentence 
is gradually disambiguated at different stages of 
processing on the basis of restrictions imposed by 
the linguistic knowledge of the system. However, 
in many cases this knowledge is not sufficient for 
complete disambiguation, since the understanding 
of a text by humans is not based on their linguistic 
knowledge alone. To cope with this problem, we 
are developing an interactive option that at certain 
pivotal points of text processing is expected to ask 
for human intervention and use human assistance 
to resolve those ambiguities that are beyond the 
scope of linguistic knowledge of the system 
(Boguslavsky et al 2003). It should be stressed that 
the interactive tool is only resorted to if an 
ambiguity cannot be resolved automatically and 
therefore requires human intervention. This work 
is in line with the approach proposed in a series of 
publications by the GETA group (Blanchon, 1995, 
1996, 1997, Boitet & Blanchon, 1995).  
As mentioned above, the dialogue with the user 
is activated at different stages of the processing 
depending on the tasks solved at each stage. 
During the parsing, which results in the 
construction of the lexico-syntactic structure of the 
sentence, all lexical and syntactic ambiguity should 
be resolved. However, this is done entirely within 
the processing of the source language text and 
represents monolingua l ambiguity. It is not directly 
relevant for our topic of multilinguality. Of 
relevance here are cases of the so-called 
translational (or transfer) ambiguity (Hutchins, 
Somers, 1992: 87). The source language words can 
be unambiguous for the native speakers of this 
language but can be translated by a number of 
different target language expressions. In this sense, 
they are ambiguous from the viewpoint of the 
target language and have to be dealt with at the 
translation stage. An example is the English verb 
wash with respect to Russian. It translates 
differently depending on the type of object that is 
being washed: if it is something made of cloth, for 
example clothes, a special verb has to be chosen. If 
the dictionary provides semantic information on 
what objects are made of, the correct choice of the 
verb can in principle be made automatically. Cf., 
however, cases like We must wash it where such 
information is definitely missing.  
This must be viewed as a relatively inoffensive 
case, though, because most sentences will be 
translated correctly with the help of a simple rule 
(and if not, the mistake is not too important). There 
are many words for which it is much more difficult 
to write a disambiguation rule. A notorious 
example is English blue that corresponds to two 
Russian adjectives, one meaning ‘light blue’ and 
the other – roughly – ‘dark blue’. The only way to 
translate this word correctly in most of the contexts 
is to get assistance from the user. The dialog with 
the user is based on the information stored in the 
dictionary and activated at the appropriate 
moment.  
This is how the interactive disambiguation 
currently works. The sentence to be translated is 
entered in the upper window of the ETAP 
environment (Fig. 2) 
 
  
Fig. 2 
When it comes to translating the word blue, the 
system finds that there are two options and no way 
to choose among them and activates the dialogue 
(Fig. 3). 
 
 Fig. 3 
 
In the dialogue box each option is provided with 
a short comment and/or example that helps the user 
choose among them. The user has to click the 
appropriate option (in Fig. 3 ‘light blue’ is 
selected) and the system moves on. The result of 
the translation of this sentence is shown in Fig. 4.  
 
 Fig. 4 
 
Should we have selected the other option in the 
dialogue in Fig. 3, the result would have been 
different (Fig. 5).  
 
 Fig. 5 
 
It is important to note that the interactive 
disambiguation mode fully corresponds to the 
multiple translation possibilities discussed in the 
previous section. In particular, the dialogue takes 
into account all types of alternative translations 
irrespective of the way they are presented in the 
dictionary. It can be lexical or syntactic ambiguity 
that manifests itself in different lexico-syntactic 
structures of the source sentence, one-word 
translation variants within the same lexical 
meaning (of the adjuration type discussed above) 
or more complex phrases that translate a source 
word (of the adventurer type above).  
 
3 UNL module in ETAP 
One of ETAP-3 options is translation between 
Russian and the Universal Networking Language 
(UNL), put forward by H. Uchida of the United 
Nations University. Full specification of UNL and 
references to publications can be found at 
http://www.undl.org. 
UNL is a formal language intended to represent 
information in a way that allows the generation of 
a text expressing this information in a large 
number of natural languages. A UNL expression is 
an oriented hyper-graph that corresponds to a NL 
sentence in the amount of information conveyed. 
The arcs are interpreted as semantic relations like 
agent, object, time, place, manner, etc. The nodes 
are special units, the so-called Universal Words 
(UW), interpreted as concepts, or groups of UWs. 
The concepts are built on the basis of English. 
When needed, English concepts can be modified 
by means of semantic restrictions in order to match 
better with the concepts of other languages. The 
nodes can be supplied with attributes which 
provide additional information on their use in the 
given sentence, e.g. @imperative, @generic, 
@future, @obligation. 
3.1 Architecture  
Since ETAP-3 is an NLP system based on rich 
linguistic knowledge, it is natural to maximally re-
use its knowledge base and the whole architecture 
of the system in all applications. Our approach to 
UNL (described in Boguslavsky et al. 2000) is to 
build a bridge between UNL and one of the 
internal representations of ETAP, namely 
Normalized Syntactic Structure (NormSS), and in 
this way link UNL with all other levels of text 
representation, including the conventional 
orthographic form of the text.  
The level of NormSS is best suited for 
establishing correspondence with UNL, as UNL 
expressions and NormSS show strong similarities. 
The most important of them are as follows: 
a) 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 lexical items is not 
decomposed into semantic primitives, and the 
relations between lexical items are language 
independent. 
b) The nodes of both UNL expressions and 
NormSSs are terminal elements (UWs in UNL vs. 
lexical items in NormSS) and not syntactic 
categories. 
c) The nodes carry additional characteristics 
used in particular to convey grammatical 
information (attributes). 
d) The arcs of both structures are non-
symmetrical dependencies. 
At the same time, UNL expressions and 
NormSSs differ in several important respects:  
a) All nodes of NormSSs are lexical items, while 
a node of a UNL expression can be a sub-graph. 
b) Nodes of a NormSS always correspond to one 
word sense, while UWs may either be broader or 
narrower than the corresponding English words. 
c) A NormSS is a tree, while a UNL expression 
is a hyper-graph, which is a much more 
complicated object. Its arcs may form loops and 
connect sub-graphs.  
d) The relations between the nodes in a NormSS 
are purely syntactic and are not supposed to 
convey a meaning of their own, while UNL 
relations denote semantic roles.  
e) Attributes of a NormSS mostly correspond to 
grammatical elements, while UNL attributes often 
convey a meaning that is expressed in English or 
other natural languages by means of lexical items 
(e.g. modals).  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Fig. 6 
 
UNL Structure  
English Normalized 
Syntactic Structure 
Russian Normalized 
Syntactic Structure 
Russian Surface 
Syntactic Structure  
Russian Morphological 
Structure 
Russian Sentence 
English Surface 
Syntactic Structure  
English Morphological 
Structure 
 
English Sentence  
 
f) A NormSS contains information on the word 
order, while a UNL expression does not say 
anything to this effect. 
These differences and similarities make the task 
of establishing a bridge between UNL and 
NormSS far from trivial but feasible. Between the 
two types of NormSS readily available in ETAP – 
the Russian and the English one – we have chosen 
the latter, since it is the English concepts that 
serve for UNL as building blocks.  
The architecture of the UNL module of ETAP-3 
is given in Fig. 6. 
3.2 UNL vs. English vs. Russian 
As shown in Fig. 6, the interface between UNL 
and Russian is established at the level of the 
English NormSS. It ensures the maximum reuse 
of ETAP’s English-to-Russian machine 
translation facility.  
In the simple case, this scenario suggests that 
the UNL – Natural Language link can be localized 
within the English dictionary. This dictionary will 
only provide an English correspondence to UNL, 
which in most cases is not very difficult, and all 
the rest will be taken care of by the translation 
engine of ETAP. In this case, direct link between 
Russian and UNL is not needed at all, as long as 
ETAP covers the English-to-Russian 
correspondence.  
However, the situation is not that simple. If we 
try to look at one language (Russian) through the 
perspective of another one (English), we 
encounter well-known problems. Let us illustrate 
the issue with an example. In Russian, there is no 
neutral equivalent of the English non-causative 
verb to marry as represented in sentences like 
John married Ann in June. The expression that 
exactly corresponds to this English verb – vstupat’ 
v brak (‘to contract a marriage’) – is an official 
term and is not used in everyday life. Instead, 
Russian speakers make use of two different 
expressions: zhenit’sja, if the agent of the action is 
a male, and vyxodit’ zamuzh, if it is a female. 
Since the English and the Russian words differ in 
their meaning, they correspond to different UWs. 
The UW for English to marry looks like (1), while 
Russian expressions have UNL equivalents with a 
more narrow meaning – (2) and (3), respectively 
(for simplicity’s sake, only the relevant fragments 
of the UWs are given):  
(1) marry(agt>human) 
(2) marry(agt>male) 
(3) marry(agt>female)  
(Here agt stands for “agent”).  
Suppose the UNL expression that we receive at 
the input of our generator contains UW (2). Since 
we have to pass through English, we must first 
translate this concept into English and then 
translate the English word into Russian. But 
English has no direct equivalent of (2). It only has 
a word with a more general meaning – to marry. 
If our objective were to get the English text, this 
word would be perfectly in place. But since our 
target language is Russian, we cannot stop here 
and have to make a difficult choice between two 
different Russian equivalents.  
This is exactly the problem that faces any 
translator from English into Russian, human or 
machine. Sometimes such a problem can be easily 
solved with the help of the context, sometimes it 
is less easy to solve or even unsolvable. For 
example, in the case of blue vs. goluboj – sinij 
discussed in 2.4 the context would hardly help to 
choose an appropriate Russian translation. 
However, in our example (2) the UNL source 
expression provides unambiguous information 
that allows avoiding this problem altogether, since 
the UW has only one correlate in Russian. If we 
pass from UNL to English and lose sight of the 
UNL source, we will lose the control of the 
semantic information and the quality of the output 
will deteriorate. This should not be permitted. Our 
solution to this problem is presented in 3.3. 
In view of the above, it may seem that a better 
idea would be to sacrifice the benefit of reuse and 
establish a direct link between UNL and Russian.  
However, the architecture shown in Fig. 6 has 
two more advantages that seem crucial.  
First, this architecture allows us to make the 
UNL module of ETAP multilingual, that is to link 
UNL not only with Russian but also with English. 
In view of this perspective, it is reasonable to 
produce a full-fledged English NormSS that is 
much closer to UNL than the Russian one.  
Second, the stock of the UNL concepts is 
continuously growing through the contributions 
coming from diverse languages. The UNL 
dictionaries of different languages grow at 
different rates and in different directions. Very 
often, the generator of language L1 receives the 
UNL input produced by the UNL group of 
language L2 that contains UWs that are absent 
from the UNL-to-L1 dictionary. This happens 
particularly often with the so called multi-word 
UWs of the type  
(4) International Research and Training 
Institute for the Advancement of Women 
(pof>General Assembly {(pof>United 
Nations)}).  
If our only source of lexical knowledge were 
the UNL – Russian dictionary, we would not be 
able to interpret such UWs, had they not been 
introduced in this dictionary in advance.  
Our UNL-to-English architecture provides a 
universal solution to all difficulties of this kind. If 
the UW is not listed in the UNL dictionariy of 
ETAP, it is analyzed by means of the ETAP 
English dictionary and, if it is a multi-word 
expression, the English parser, which results in a 
reasonably good representation of the UW.  
Moreover, it is often possible to correctly 
translate a UW that is absent from ETAP’s UNL 
dictionary even if its headword is ambiguous. For 
example, if we receive UW  
(5) open(mod<thing)  
and do not find it in our UNL dictionary, we can 
replace it with the English word that stands in the 
position of the headword, that is open. However, 
this headword is ambiguous. In ETAP’s English 
dictionary there are three entries for open - the 
adjective, the verb and the noun. A simple rule 
allows selecting the correct entry on the basis of 
the UW restriction: (mod<thing) means that the 
headword serves as a modifier of things. Hence, 
its English correlate is an adjective and not a verb 
or a noun.  
3.3 UNL dictionary vs. English dictionary vs. 
Russian dictionary  
The UNL-related information is distributed 
among the three ETAP dictionaries: UNL, English 
and Russian. The general idea is to combine (a) 
the idea of having the English NormSS as an 
intermediate level between UNL and the Russian 
NormSS and as a source of Russian and English 
generation and (b) the requirement of adequately 
treating cases of non-isomorphism between the 
English and the Russian concepts.  
As shown in section 2.1, the ETAP dictionary 
entry contains several bilingual sub-zones, 
according to the number of working languages. In 
particular, the Russian dictionary has sub-zones 
for English and UNL, the English dictionary – for 
Russian and UNL and the UNL dictionary – for 
English and Russian.  
Let us consider two cases: (1) the Russian and 
the English words are synonymous (as, for 
example, to divorce and razvodit’sja) and (2) they 
are not synonymous (as, for example, to marry 
and zhenit’sja).  
The relevant fragments of the dictionary entries 
(with some simplifications) are as follows. 
 UNL dictionary: 
NAME: divorce(agt>human) 
ZONE:EN 
TRANS: divorce 
ZONE:RU 
<none> 
NAME: marry(agt>human) 
ZONE:EN 
TRANS: marry 
ZONE:RU 
<none> 
NAME: marry(agt>male) 
ZONE:EN 
<none> 
ZONE:RU 
TRANS: zhenit’sja  
 English dictionary 
NAME: divorce 
ZONE: RU 
 TRANS: razvodit’sja  
ZONE:UNL 
 TRANS: divorce(agt>human) 
NAME: marry 
ZONE: RU 
 TRANS: zhenit’sja / vyxodit’ zamuzh 
ZONE:UNL 
 TRANS: marry(agt>human) 
 Russian dictionary 
NAME: razvodit’sja  
ZONE: EN 
 TRANS: divorce 
ZONE:UNL 
 TRANS: divorce(agt>human) 
NAME: zhenit’sja  
ZONE: EN 
 TRANS: marry 
ZONE:UNL 
 TRANS: marry(agt>human) 
 
Suppose we have to process a UNL expression 
that contains UW “divorce(agt>human)”. Since 
this concept corresponds to both English and 
Russian words, we can do safely without any 
information on the Russian word in the UNL 
dictionary and obtain the NormSS with English to 
divorce taken from the English zone of the UNL 
entry. This NormSS allows generating both 
English and Russian texts by means of the 
standard ETAP transfer and generation facilities.  
Let us consider the source UNL expression that 
contains UW “marry(agt>human)”. It may have 
come from the language that, like English, 
German or Spanish, but unlike Russian or Polish, 
does not distinguish between the male -marriage 
and the female -marriage. The UNL dictionary 
entry for this UW will have the English translation 
but no Russian one, since Russian has no direct 
correlate for this concept. The problem of finding 
an appropriate Russian term is shifted to the level 
of the NormSS. At this level, we will have to find 
an equivalent of English to marry, just as if we 
translated from English and not from UNL. In this 
case, the UNL source does not help us make a 
choice between two types of marriage. What does 
help is the mechanism of the interactive resolution 
of translational ambiguity described above, in 2.4.  
Finally, let us examine the most interesting case 
- a UNL expression with UW “marry(agt>male)”. 
The dictionary entry of this UW is symmetric to 
the entry of “marry(agt>human)”: it contains a 
Russian correlate but no English one. In this 
situation, both English and Russian generations 
are not quite straightforward. As there is no direct 
English equivalent of this UW, the translation 
should be found by means of the UNL Knowledge 
Base (Uchida, 2003). In the absence of the 
operational version of KB, the general solution for 
processing an unknown UW is to extract the 
headword of the UW (marry) and treat it as an 
English word (cf. above, 3.2). This solves the 
problem of the generation of the English text. As 
for Russian, zhenit’sja indicated in the Russian 
zone of the UW entry is attached as a feature to 
the English node marry. At the stage of transfer 
from NormSS-English to NormSS-Russian, this 
feature will be lexicalized and replace the word 
marry.  
4 Conclusion 
The organization of lexical resources of the 
ETAP system allows reusing the dictionaries in 
diverse applications, such as machine translation 
in various language pairs and translation to and 
from UNL. In all the applications, there are three 
modes of operation supported by the dictionaries: 
automatic production of a single (most probable) 
translation, automatic production of all possible 
translations and the interactive translation with the 
dialogue-based disambiguation.  

References  
Apresjan Ju.D., Boguslavskij I.M., Iomdin L.L., 
Lazurskij A.V., Mitjushin L.G., Sannikov, V.Z., 
Cinman, L.L. (1992) Lingvisticheskij processor 
dlja slozhnyx informacionnyx sistem. [A 
linguistic processor for advanced information 
systems.] Moskva, Nauka. 256 p. 
Apresjan Ju.D., Boguslavskij I.M., Iomdin L.L., 
Lazurskij A.V., Sannikov V.Z. and Tsinman 
L.L. 1992b. The Linguistics of a Machine 
Translation System. Meta , 37 (1): 97-112. 
Apresjan Ju.D., Boguslavskij I.M., Iomdin L.L., 
Lazurskij A.V., Sannikov V.Z. and Tsinman 
L.L. 1993. Systeme de traduction automatique 
{ETAP}. In: La Traductique. P.Bouillon and 
A.Clas (eds). Montreal, Les Presses de 
l'Universite de Montreal. 
Apresjan, Ju.D. 1995. Integral’noe opisanie 
jazyka i sistemnaja leksikografija [An 
Integrated Description of Language and 
Systematic lexicography.] Moscow, Jazyki 
russkoj kul’tury.  
Apresjan, Ju. D. 2000. Systematic Lexicography. 
Oxford University Press, London, 304 p.  
Apresian Ju., I. Boguslavsky, L. Iomdin, A. 
Lazursky, V. Sannikov, V. Sizov, L. Tsinman. 
2003. ETAP-3 Linguistic Processor: a Full-
Fledged NLP Implementation of the MTT. In: 
MTT 2003, First International Conference on 
Meaning – Text Theory. Paris, Ecole Normale 
Superieure, Paris, 279-288. 
Blanchon, H. Interagir pour traduire: la TAO 
personnelle pour redacteur monolingue. La 
Tribune des Industries de la Langues. Vol. 17-
18-19, 1995, pp. 28-34. 
Blanchon, H. A Customizable Interactive 
Disambiguation Methodology and Two 
Implementations to Disambiguate French and 
English Input. Proc. MIDDIM'96. Le col de 
porte, Isere, France. 12-14 Aout 1996. Vol. 1/1, 
1996, pp. 190-200. 
Blanchon, H. Interactive Disambguation of 
Natural Language Input: a Methodology and 
Two Implementations for French and English. 
Proc. IJCAI-97. Nagoya, Japan. August 23-29, 
1997. Vol. 2/2, 1997, pp. 1042-1047 
Boguslavsky I., N. Frid, L. Iomdin, L. Kreidlin, I. 
Sagalova, V. Sizov. 2000. Creating a Universal 
Networking Language Module within an 
Advanced NLP System. Proceedings of the 18th 
International Conference on Computational 
Linguistics (COLING 2000), 2000, 83-89.  
Boguslavsky I., L. Iomdin, V. Sizov. 2003. 
Interactive enconversion by means of the 
ETAP-3 system. In “Proceedings of the 
International Conference on the Convergence of 
Knowledge, Culture, Language and Information 
Technologies”, Alexandria, 2003. 
Boitet, C. & Blanchon, H. Multilingual Dialogue-
Based MT for monolingual authors: the LIDIA 
project and a first mockup. Machine 
Translation. Vol. 9(2), 1995, pp 99-132. 
Hutchins W. J., H. L. Somers. 1992. An 
Introduction to Machine Translation. Academic 
Press, London. 
Mel’cuk I. 1974. Opyt teorii lingvisticheskix 
modelej “Smysl – Tekst”. Moscow, “Nauka” 
Publishers.  
Uchida H. 2003. The UW Manual. 
http://www.undl.org. 
