PROPOSALS FOR A HIERARCHY OF FORMAL TRANSLATION MODELS 
Klaus-Jurgen Engelberg 
Universitat Konstanz, Philosophische Fakult~t, Fachgruppe 
Sprachwissenschaft, BRD 
The present deplorable state-of-the-art in the field of 
machine translation seems greatly due to a fundamental laok 
of formal translation models needed in natural language pro- 
cessing. 
From the methodological point of view it appears difflo- 
ult to delineate a borderline between translation theory and 
modern theoretical lingu~stlos (availing itself of model 
theoretical semantics) or full natural language understanding 
systems as developed in Artificial Intelligence research. It 
seems plausible to postulate that any prospective translation 
theory should draw on ideas from both fields. Unfortunately, 
problems discussed in painstaking detail in linguistics like 
differences in quantifier scope appear to be of lesser concern 
to a translator (since these ambiguities may well remain pres- 
ent in the target language) , neither seems a full or deep 
understanding necessary in many cases, standard syntactic 
phrasing may suffice. More specifically, we regard the pro- 
blems of disambi~ation, mandatory insertion of lexical items 
not conventionally implied in the source language and corefer- 
ence/anaphora resolution as the crucial problem areas of 
machine translation. 
In this paper, we will endeavour - in this preliminary 
draft only in a very sketchy manner - to set up a hierarchy of 
formal translation models ordered according to their inoreas- 
- 90 - 
ing systematic disambiguation power for certain types of 
texts. 
Quite analogous to comple.xity considerations in mathe- 
matics, the power of a translation system is assumed to be 
measured by the amount of storage needed for the lexioal com- 
ponent (A~-people mAght call this long-term-memory) and/or 
for the transient or dynamic data (short-term-memory) built up 
duming the interpreting process of a particular text. Any 
model will be capable %o translate only certain restricted 
types of texts in a systematic manner and with satisfactory 
results, but the idea i8 that any model will also contain 
components of lower levels of complexity. This is to make sure 
that in oases in which disambigustion on purely syntactic 
grounds is possible no such process via "deep" semantic re- 
presentations will be attempted for this particular case. The 
rationale, of course, will be to utilize ever lexger portions 
of contextual (or rather co-textual) information for these 
ends. As the reader will notice, powerfull translation system 
have to incorporate more and more knowledge-of-the world into 
the database, as becomes apparent from the famous examples 
The soldiers shot the women. They fell down. 
Les soldate abbatirent les femmes. Ils/el~es? tomberent. 
Syntactic methods 
Level Synl : Word-to-word translation 
TS out for appsLrent reasons! (although a full bilingual dict- 
ionary would require a considerable amount of storage space 
in a oomputer) 
Level Sy~: Constituent preserving translation 
These models utilize the immediate syntactical context (e.g. 
valency of verbs) for dlsambiguatlon purposes. In such a sys- 
tem a rule may look llke 
x slob erinnern ~ x remember, but, x erinnern y ~x re- 
mand y 
- 91 - 
At any rate, a valency oriented lexicon would be helpful in 
the following models, too. The search strategy would be long- 
est match first. 
Level Sy~3: Tree-to-tree translation 
Unbouded translations allow for reordering of arbitrarily 
long portions of a sentence. We think it reasonable to assume 
that a quarter-century of Generative Grammar research in 
Linguistics will have produced enough theoretical and practlo- 
.1 apparatus to deal with any type of tree-restructurlng that 
may be needed in direct syntactic translations between natural 
languages (also of. the French system GETA). 
Semantic methods 
Level Seml : Case.grammar oriented translations 
There are several MT systems that impose heavy restrictions 
on the possible arguments of verbs by encoding semantic feat- 
ures in the lexicon (e.g. METEO in Canada). By this, of cour- 
se, disambi~ation can take place only within the limits of a 
single sentence or clause. 
Level Sere2: Translations using coherence relations 
The basis of this approach is the assumption that there exist 
finitely many determined and computable coherence relations 
between two subsequent sentences and/or clauses in certain 
types of texts. (sometime called the cohesive-ties-approach). 
They may be even indications of these relations at the sur- 
face level of the discours e.g. ~whereas°suggesting CONTRAST 
or °then" suggesting TI~E-SEQUENCE, other relations may be 
ELABORATION, EFFECT, CAUSE (Hirer /1981/). Processing of these 
texts could be done by semantic finite state automata that 
would accept only highly constrained discourses in which no 
abrupt shifts of focus would be allowed. At last at this level 
of complexity It seems necessary to assume that the vocabulary 
should be organized - in addition to the usual lexioographlc 
- 92 - 
order - as a sort of semantic network containing all types of 
sense relations like super-subset relation, antonymy, conver-- 
seness, time-sequence - existing even between several places 
verbs. 
Level Se~: Translations using story trees 
These models dynamically build up a tree-like maorestructu~e 
for a text in which arbitrary deep embeddings of themes and 
sub-themes are represented. In this approach, coherence re- 
lations between entire portions of text or paragraphs could 
be established - thus allowing for ooreferen across long 
distances in a text (vide Rumelhart /1975/), This process 
may be facilitated by what Y. Wilks chose to call "paraplates" 
in the database. 
Level Sere4: Translations uslng semantic networks 
This model is designed for not so orderly texts as assumed in 
the previous levels. A semantic network as the dynsmlc macro- 
structure of & text would allow for multiple views or thema- 
tic structures associated with a portion of a text. To make 
this effective, a very rloh fabric of various types of assoc- 
iative links would be needed in the database. 
Level Sere5: 1~rsmevbased translations 
"Frames" or "scripts" have been widely discussed in the AI 
con:mx~ity In the past 10 years or so. The idea seems to bee 
to aggregate all sorts of information objeot~oentred linked 
with a particular "stereotypical situation" into a structured 
entity - called "frame'. This approach would,'in principle, 
allow one - by default reasoning - to recover information not 
explicitly mentioned in the texts In particular, this may be 
helpful when translating into a western language from Russian, 
in which the deftni%e/tndefihite or known/unknown distinction 
in nouns is lacking. Consider the translation problems in the 
following example (drawing on Schank's favou~ite soript)s 
- 93 - 
Petr posel v restoran. Oficiant podal emu menJu. =7 
Peter went to a restattt~lt. Th_.~e waiter handed him the 
menus 
Soripts could account for associations induced by "spatial- 
- temporal contiguities" as present in this example. 
Doubts as to the feasibility of MT based on frames - 
except possibly in very restricted areas of discourse - have 
come from various quarters. First, the coding effort could 
turn out to be enormous. Second, a intricate problem seems 
to be how to find out which script is relevant to the current 
portion of text, 
- 94 - 
