VIRTEX - a German-Russlan Translation Experiment 
B. Busehbeck, R. Hensehel, I. H6ser, G. Klimonow, A. K~stner, I. Starke 
Zentralinstitut f~r Sprachwissenschaft 
Akademie der Wissensehaften dsr DDR 
Prenzlauer Promenade 149-152, Berlin, 1100, DD~ 
Machine Translation from German into 
Russian did not gain the attention it 
deserves so far. Especially such typical 
Russian phenomena like the aspect of verbs 
complicate the translation process. The 
known MT approaches do not deal with such 
problems in a satisfactory way. To over- 
come this lack, we started the German.- 
Russian translation experiment VIRTEX. We 
focussed our attention on the crucial 
point -- a high quality translation of 
verbal groups (VIRTEX = Verbs-into- 
Russian Translation Experiment). That 
means, VIRTEX is made to solve on princi- 
ple the linguistic difficulties connected 
with verb translation: 
- recognition of German analytical tense 
forms, passive constructions, reflexive 
forms and verbs with detached prefixes 
- analysis of complex German verb groups 
(modal and modal-like constructions) 
- selection of Russian equivalents accor- 
ding to valency frames including seman ~ 
tic features of verb complements 
- choice of the appropriate Russian aspect 
form 
passive-active transformation in case of 
missing Russian passive form 
adequate translation of modal construc- 
tions. 
The used vocabulary, above 200 sample 
verbs, is selected to represent these 
phenomena completely. 
To translate verbal groups adequately, 
£t is necessary to have some context. 
Tberefore we accept as input simple sen- 
tence 9 with verb complements and adver- 
bials. That means we have also a set of 
nouns, pronouns, adverbs and some other 
words, The restricted syntax model is made 
for short main clauses; it excludes rela-- 
tire and other subordinated clauses, at- 
tributes and infinitive clauses with 
complements. VIRTEX can be classified as 
an NT system between second and third 
generation with a syntactic interface 
structure completed by some semantic in- 
formation. The system's architecture is 
the classical one. Analysis, Transfer and 
Synthesis are sequentially processed sepa- 
rate routines. For each of them an extra 
lexicon has been implemented. The input- 
work is supported by special tools includ- 
ing an automatic generation of the inflec- 
tional types. While the procedures are 
written in LISP, the dictionaries and 
their editors are made in dBASE. The 
format restrictions of dBASE are compen- 
sated by a special management of the en- 
tries and a corresponding way of LISP 
access to them. VIRTEX runs on PC-XT. 
Some steps of the system will shortly 
be illustrated. 
ANALYSIS 
The lexical .entries for German syntac- 
tic analysis are taken from the analysis 
dictionary which is combined with a system 
of inflection lists; an analysis of in- 
flectional endings is embedded in the 
dictionary access. A word form entry con- 
tains word class information, morphologi- 
cal features and values, some semantic 
information and the name of the lexeme. 
While the system is running, the lists of 
information and features are converted 
1 321 
into numbers. 
For syntactic parsing an ATN-grammar is 
used to recognize nominal and prepositio ~ 
nal phrases, adverbials and verbs or 
verbal groups. A special embedded program 
analyzes the verb constituents. Homonymous 
morphological and word class features are 
disambiguated as far as possible through- 
out the parsing process. 
TRANSFER 
All transfer rules are written within 
our tree transformation'language GRACOLI. 
This is a universal software tool deve- 
loped to handle MT transf,r problems. 
GI~ACOLI manages the input and processing 
of tree rewriting rules. One rule consists 
of a left side and a right side tree. 
According to the usual notation of syntac- 
tic tress, every tree node can be labelled 
by a finite set of attribute-value-pairs. 
So we have got a universal rule format for 
structure changing with assignment of new 
labels depending on the occurrence of the 
left side structure and left side label 
conditions. It includes all simplier 
transfer actions like label tests, 
structure tests, lexical transfer etc. The 
compilation of GRACOLI rule sequences in 
GRACOLI grammars allows high evidence and 
easy correction of implemented transfer 
algorithms. 
The transfer lexicon comprises all 
lexeme-dependent translation information. 
As for the verbs in case of lexical ambi- 
guity, the equivalents are listed in an 
order suitable for selecting the 
appropriate one by checking 
(I) characteristics of valency frames, 
(2) semantic features of verb comple- 
ments or 
(3) other conditions. 
ad I) E.g.. the verb eintrsten may be 
disambiguated by the prepositions: 
eintre~en f~r - v v~tuDatj za 
'to stand up for'; 
sintreten in - vstupat~ v 
' to join sth. ' 
ad 2) One or more semantic features are 
attributed to nouns. They may be used for 
disambiguation, e.g.: 
absehlie~en (accusative inanimated), 
-> ~ 'to lock' 
absehlieBen (accusative process) 
-> zakanchivat i 'to terminate' 
ad 3) Concrete noun lexemes can also be 
used as fillers of verb slots, e.g. 
abschlieBen (accusative Vertra~ 
'contract') 
-> zakluchatj (do_govor) 
' to conclude' . 
The form of lexical entries allows to 
adapt them as GRACOLI rules. With the held 
of the transfer lexicon lexieal and valen- 
cy frame transfer is managed simultaneous~ 
IF. Passive sentences get a special pre~ 
processing before the lexicon is con- 
sulted. After this, two distinguished 
GRACOLI procedures for active and passive 
voice select the aspect form according to 
a hierarchy of criteria. Three groups of 
such criteria are distinguished: 
(I) a lexieal criterion, 
(2) syntactic-semantic criteria and 
(3) German tense criteria. 
ed I) For certain German verbs, the trans- 
fer lexicon specifies which of the two 
aspects of the Russian equivalent has to 
be chosen, e.g. 
Pr~fun~ bestehen 
-> sdat i ~kzamen (perfective aspect) 
'to pass an examination' 
sich bemhhen um 
-> dobivat s a (imperfective aspect) 
'to try hard'. 
ad 2) In some eases verbal aspect depends 
on the valency frame of the German verb 
or/and on semantic features of comple- 
ments: 
EF schrieb an einem Such 
-) On pisgl knl_~ 'he wrote a book' 
(preposition an, case dative 
--> imperfective aspect). 
The reoccurrence of verbs with certain 
types of adverbials influences the verbal 
aspect choice, too. Adverbs like jahrelang 
322 2 
'for years', ~ 'o, ften', z n_E99_e_h~hmend 
'more and more' demand the imperfeetive 
aspect, but ~l~ts!ich. 'suddenly', unerwar- 
te__~t 'unexpected', d e~ 'soo~' the 
perfective one. If there occur adverbials 
of both types within the sentence, the 
choice of Imperfeetive aspect is prefer- 
t'ed~ 
ad 3) If none of the aforesaid criteria 
applies, the tense of the finite German 
verb determines the verbal aspect: 
future, perfect, pluperfect 
=-> perfective aspect 
present.. --> imperfective aspect 
In case of'preterite, perfect and future 
semantic verb subclassification as well as 
definiteness of the di.rect object are 
additionally taken into account. 
A separate GRACOLI grammar is also used 
to carry out the structural transfer of 
modal constructions because the descrip- 
tion of the complex conditions of these 
constructions does not fig in the frame of 
the lexical entries of the transfer dic- 
tionary. This shall be demonstrated by the 
translation of the German verb k~nnen 
'can', ' to be able to'. It is necessary to 
select a limited number of translation 
patterns because of the wide variety of 
corresponding Russian constructions. Thus 
two Russian constructions were taken into 
consideration. Their choice depends mainly 
on the kind of the subject and the voice 
of the infinitive. 
(I) The impersonal construction with the 
predicative mozhn_qo 'it is possible' is 
selected if one of the following condi- 
tions to the subject is fulfilled: 
- indefinite-personal pronoun man or 
- a nounphrase which belongs to the sub- 
class 'process', or 
- a formal is, or 
there is no subject, 
for example: 
Es kan n (ibm) geholfen.werden. 
'It is possible to help him.' 
-> Emu me,hag omoch' 
(2) The personal construction with a fi- 
nlte form of the. modal verb moth& 'can', 
'to be able to' is ~ preferred if: 
- none of the above given conditions is 
applicable, and if 
- the subject is a personal ODe, 
for example: 
EP kann die Aufsabe l~sen 
'He is able to solve the task' 
-> On moshet reshitj zadachu. 
SYNTHESIS 
The transfer result is a constituent 
structure containing the Russian equiva- 
lents and information for syntactic and 
morphological synthesis. Since the valen- 
cy frame transfer is already done, only 
some tasks remain for syntactic synthesis. 
The Russian verbs!must be expanded if they 
demand analytical fcrms, that holds for 
future, subjunctive and passive construc- 
tions. Adverbial groups which are handled 
as only one constituent will get their 
explicit form now. Gender agreement for 
Russian preterite and reflexive forms is 
formed here. Furthermore, ~usaian word 
order processing is done. 
Input of the morphological synthesis 
are the results of the previous steps. 
Information of the Russian dictionary 
entries is combined with the inflectional 
lists. 
The experimental system VIRTEX proposes 
a solution for the aspect problem, and is 
therefore not only a new one in the long 
list of existing MT systems, but a sub- 
stantial contribution to MT research. 
VIRTEX will be demonstrated at COLING ' 90. 
3 323 
