ASCOF -- A MODULAR MULTILEVEL SYSTEM 
FOR FRENCH-GERMAN TRANSLATION 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt 
Universitltt des Saarlandes 
D-6600 Saarbr'dcken 
West Germany 
This paper is an overview of ASCOF, a modular multilevel system for French-German translation. In 
ASCOF, the classical divisions of the translation process (analysis, transfer, synthesis) have been 
adopted. 
The analysis is realized by three phases: (1) the morphological analysis, (2) the identification of 
non-complex syntactic phrases and the macrostructure of the sentence, and (3) the determination of 
the structure of complex syntactic phrases and the syntactic functions, in which syntactic and semantic 
criteria are used. Semantic criteria are stored in a semantic network. The syntax-oriented parts of 
the system interact with this semantic network during the identification of the syntactic functions. The 
iexicai transfer operates on the standardized output tree of the analysis. The structural transfer and 
the syntactic synthesis are achieved by transformational grammars; the morphological synthesis, at 
least, generates the word form of the target language (German). 
1 PROJECT HISTORY AND STATUS 
ASCOF (Analysis and Synthesis of French by means of 
COMSKEE) is a computer system for the processing of 
natural language with the purpose of translating written 
French texts into German texts. This system has been 
under development since 1981 at the University of the 
Saarland at Saarbrticken, West Germany (Project C of 
SFB 100). At present (1984-1985), the research team, 
which has drawn upon the experience and findings of 
previous studies, consists of six members. 
\[SFB 100 is a research group in which different linguis- 
tic and computer science-oriented projects cooperate. 
The SFB 100 was founded in 1973 and is financed by the 
DFG (German Research Foundation). Other projects of 
the SFB 100 have been developing the programming 
language COMSKEE and the systems SUSY and SUSY II. 
ASCOF is an independent system especially conceived for 
French and German. General descriptions of ASCOF are 
given in F6neyrol, Ritzke, and Stegentritt (1984) and 
Stegentritt (1983). Detailed descriptions of the various 
problems and their solutions are discussed in F6neyrol 
and Stegentritt (1982) and in Ritzke (1982).\] 
2 APPLICATION ENVIRONMENT 
The system is programmed in COMSKEE (Computing 
and String Keeping Language; cf. Mueller-von Brochow- 
ski et al. (1981), Messerschmidt (1984).) For the 
computer scientist, COMSKEE is a procedural (imper- 
ative) format-free, block-oriented programming language 
such as ALGOL and Pascal, yet comprising some of the 
qualities of functional languages (such as LISP or 
PROLOG). 
For the linguist, COMSKEE is a powerful device espe- 
cially due to its dynamic data types - dictionary, set, 
sentence, and string - and its dynamic operations - such 
as positional and contextual substring access and assign- 
ment. 
The system runs on a SIEMENS 7561 under the system 
BS 2000. ASCOF has been conceived as a completely 
automatic translation system. As yet, we have been less 
concerned with end-user application than with funda- 
mental research. For this reason, we have focused 
primarily upon linguistic and computer science problems, 
rather than upon processing speed and the like. 
Copyright1985 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided that 
the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To copy 
otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/85/020137-154503.00 
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 137 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
3 GENERAL TRANSLATION APPROACH 
3.1 
In ASCOF the "classical" divisions (cf. Vauquois 1975) 
have been adopted: analysis, transfer, and synthesis. The 
result of the sentence analysis is represented as a stand- 
ardized tree structure, which then serves as input for the 
transfer and synthesis of the target language. 
3.2 
The ASCOF analysis takes place in three steps based on 
different grammar and algorithm types. The morphologi- 
cal analysis PHASE I is carried out by an algorithm that 
realizes actually a mere pattern matching; in PHASE II 
context-free grammars identify non-complex syntactic 
phrases and the macrostructure of the sentence. A 
reduction in the homographies of word classes is simul- 
taneously achieved for the complete sentence. PHASE III 
determines the syntactic functions within the sentence, 
using syntactic and semantic criteria, and carries out the 
semantic disambiguation of lexemes. This phase of anal- 
ysis is performed by algorithms similar to ATN, repres- 
enting an interactive system. \[The term interactive system 
might be problematic in this context as this term often 
denotes components interacting with the user. Here we 
are concerned with the process communication between 
different components.\] Consequently, the ASCOF analy- 
sis does not constitute a one-pass parser but a system of 
parsers (cf. Figure la). The strategy applied resembles 
that of cascaded ATNs (Woods 1980) and was chosen 
for the following reason: the complexity and length of 
the sentences to be analyzed require - for reasons of 
efficiency - parsing strategies appropriate to the different 
problems, that is, context-free grammars for PHASE II, 
which works exclusively with syntactic information, and 
formalisms similar to ATN for PHASE III, where syntactic 
and semantic information is combined. A similar combi- 
nation of syntax and semantics often occurs in modern 
parsers of various orientations, e.g., in the determinism 
parser put forward by Marcus (1980, chap. 10). 
This approach not only allows a step-by-step realiza- 
tion of the test phases required for the development but 
also provides the user with alternative options for the 
output owing to the different depths of analysis in 
PHASE II and III. 
3.3 
Beyond the phase of analysis, ASCOF includes a phase of 
transfer and synthesis, where the words of the source 
language are exchanged for those of the target language 
and where simultaneously structures are altered in the 
tree structure if necessary. The changes of structure are 
carried out by a transformational grammar. The gram- 
mar operates on trees; grammar and algorithm are sepa- 
rate from each other and the algorithm interprets the 
externally stored rules of the grammar. \[The documenta- 
tion of the transformational component is put forward by 
Reding (1985); for the discussion of transformation 
grammars in machine translation, cf. Vauquois (1975), 
Boitet, GuiUaume and Qu6zel-Ambrunaz (1982), and 
Huckert (1979).\] 
On the leaves of the output tree, produced by the 
syntactic synthesis, a further algorithm operates, which 
interprets a set of morphological rules in order to gener- 
ate the correct word forms of the target language. The 
transfer and synthesis components of ASCOF are shown 
in Figure lb. 
The separation of grammar and algorithms allows the 
application of the above-mentioned components in other 
languages as well, provided that the grammar is replaced. 
4 LINGUISTIC TECHNIQUES AND 
COMPUTATIONAL REALIZATION 
4.1 
The most sophisticated phase within ASCOF concerns 
analysis (French); the synthesizing phase (German) has 
not yet been developed to such an elaborate extent. 
This paper consequently concentrates on the 
description of the anMyzing phase. Much space is 
devoted to segment analysis, the interaction of the 
complement analysis and the analysis of complex noun 
phrases, which is discussed in detail and illustrated by 
examples. 
4.2 
PHASE I of the analysis consists of the sentence/text 
input and the morphological analysis. Each word form is 
assigned the set of possible categories as well as the 
morpho-syntactic information. A full form and a stem 
dictionary (both approximately 47,000 entries) as well as 
a suffix dictionary (inflectional suffixes) are available. 
Unknown word forms undergo a derivational analysis 
(based on Stegentritt 1978). 
4.3 
Within ASCOF's PHASE II, we can distinguish two differ- 
ent parts, both realized by means of context-free gram- 
mars; in the first part, categorial (word class) ambiguities 
are resolved through a CFG working on the string of 
word classes issued from PHASE I. By applying this 
grammar, we obtain one - or possibly more - string(s) of 
unequivocal word classes. \[A special word class problem 
occurring in French, concerning the ambiguity of (-ant) 
forms, is discussed in F6neyrol (1982).\] As the consulted 
grammar itself represents a rudimentary syntactic analy- 
sis, here we arrive at an initial decomposition of the 
sentence into so-called simple syntactic units, such as 
nominal, prepositional, adjectival, adverbial, and verb 
units; (coordinating and subordinating) conjunctions, 
introducing words (relative pronouns and others), and 
commas form proper units, too. 
For a sentence such as 
(1) la directive du Conseil du 20 juillet, qui, ~ l'article 4, 
pr6voit une augmentation du prix du mais de 3%, 
touche les r6gions du sud et les pays africains qui 
d6pendent de ce produit d'importation. 
138 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
PHASE I 
PHASE II 
PHASE III 
Input of the sentence/text 
dictionary check morphological 
analysis 
Analysis of the macrostructure of 
the sentence by context-free 
grammars 
- reduction of the homographies 
- identification of non-complex 
syntactic phrases 
- sentence segmentation 
,L 
Analysis of the verb phrases 
(ATN) 
Interactive system 
Complement 
analysis 
Analysis of 
the complex 
noun 
phrases 
determination of the reading, 
semantic disambiguation 
DATA BASE 
dictionary for 
analysis 
47,ooo entries 
= dictionary 
information 
cf. grammar 
Semantic infor- 
mation, 
Semantic network 
Figure la. Analysis components in ASCOF. 
Computational Linguistics, Volume 1 l, Numbers 2-3, April-September 1985 139 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
TRANSFER 
SYNTHESIS 
Analysis 
Lexical 
Transfer 
French ---> German 
Structural 
Transfer 
Syntactic 
Synthesis 
Morpholoqical 
Synthesis 
J 
< 
DATA BASE 
Dictionary 
French - 
German 
Transforma- 
tional 
Grammar 
Transforma- 
tional 
Grammar 
Morphological 
Grammar 
Figure lb. Transfer and synthesis. 
the result of this first part of PHASE II will be as 
follows: 
(2) nnog pg pg com rel com pg com verb nnog pg pg pg 
com verb nnog pg coord nnog adj rel verb pg pg eos 
\[The abbreviation "eos" means "end of sentence", 
"nnog" means "nominal noun phrase", "rel" means 
"relative word"; the other abbreviations need no 
comment.\] This string of simple syntactic units serves as 
input for the second part of PHASE II, concerned with 
the segmentation of the sentence, that is, its decompos- 
ition into main and/or subordinate clause(s); the follow- 
ing shows in more detail how this part of ASCOF works. 
\[Some of the problems appearing in the segmentation of 
French sentences as well as a description of another type 
of segmentation module used earlier are described in 
F6neyrol (1983).\] 
The operation of segmentation itself takes place in two 
steps, each of which is realized by a CFG; in the first 
step, we apply the CFG IDENTIFICATION to the string 
under (2) in order to "transform" the sequence of simple 
syntactic units into so-called Segmentation Units (SU) 
\[we use the following conventions: "( )" for facultativi- 
ty, " I " for alternatives, and "*" for repetition\]: 
IDENTIFICATION: (TV1, NTV1, S1, R1) 
TV1 :nnog; com; rel; verb; coord; adj; pg; cos 
NTV1 :SUNNOG; SUN; SUCOM; SUCOORD; SURELS; 
SUVERB; SUEND 
R1 1 : S1 ~ (SUNNOG \[ SUN I SUCOM 
\[ SUCOORD \[ SURELS \[ SUVERB )* SUEND 
2 : SUNNOG * nnog(SUN)* 
3 : SUN -~ (Pg \[ adj)* 
4 : SUCOM -~ com 
5 : SUCOORD * coord 
6 : SURELS -~ rel (SUNNOG I SUN\[ SUCOM 
\[ SUCOORD )* SUVERB 
7 : SUVERB ~ verb 
8 : SUEND ~ cos 
Taking a look at some rules of IDENTIFICATION, one 
should note that, for example, in rule 2 SUNNOG is built 
up without taking account of the internal structure of the 
complex noun phrase; the actual relations inside the 
SUNNOG, l'augmentation du prix du ma~'s de 3% - in 
fact: (l'augmentation (du prix (du mars)) (de 3%)) - are 
not of importance in this part of ASCOF; it is only neces- 
sary to identify the complex noun phrase as SUNNOG. 
\[The accurate description of complex nominal phrases is 
140 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
the main task of NOMAL; for this section, see 4.4.3.\] 
Rule 6 builds up the relative clause, the minimal invento- 
ry consisting of relative word (rel) and verbal phrase 
(verb) with the bracketed facultative elements between 
them. 
Th.e result of IDENTIFICATION for sentence (1) via 
structure (2) is 
(3) SUNNOG SUCOM SURELS SUNNOG SUCOM 
SUVERB SUNNOG SUCOORD SUNNOG SURELS 
SUN SUEND 
These Segmentation Units obtained from IDENTIFICA- 
TION themselves constitute the terminal vocabulary of a 
second CFG GROUPING. \[Both CFGs (IDENTIFICA- 
TION and GROUPING) here only contain the categories 
occurring in the example (1).\] Its task consists of moving 
from the terminal Segmentation Units to the axiom TXS 
(Text Sentence) by way of a left-to-right, bottom-up 
parsing, providing a decomposition of the sentence into 
clauses and indicating their mutual relations. 
List of abbreviations: 
TXS Text Sentence 
CS Complex Sentence 
S Sentence 
NOV NOGC + VERBAL 
NOGC Complex Noun Group 
SSBCL Sequence of Subordinate Clause(s) 
CSBCLC (Comma) SBCL (Comma) 
SBCL Subordinate Clause 
POST Postverbal domain 
NEUTC Complex Neuter Unit 
NEUT Neuter Unit (/-- NOG) 
GROUPING: (TV1, NTV1, TXS, R1) 
TV1 :SUNNOG; SUCOM; SURELS; SUVERB; 
SUCOORD; SUN; SUEND 
NTV1 :CS; S; NOV; NOGC; VERBAL; POST; NEUTC; 
NEUT; SSBCL; CSBCLC; SBCL 
R1 1 : TXS ~ CSSUEND 
2 : CS -~ S(SUCOMS ISUCOORDS)* 
3 : S -~ (SSBCL)(SUCOM)NOV 
4 : NOV -~ NOGC VERBAL 
5 : NOGC ~ NOG(SSBCL) 
6 : NOG ~ SUNNOG ( SUCOM SUNNOG 
I SUCOORD SUNNOG )* 
7 : VERBAL--,- SUVERB ( POST ) 
8 : POST --~ NOGC I NEUTC 
9 : NEUTC --~ NEUT ( SSBCL ) 
10 : NEUT -~ SUN (SUCOM SUN I SUCOORD 
SUN )* 
11 : SSBCL ~ CSBCLC I SBCL ( SUCOM SBCL 
I SUCOORD SBCL I SUCOM 
VERBAL )* 
12 " CSBCLC-" ( SUCOM ) SUBCL ( SUCOM ) 
13 : SBCL ~ SURELS (POST) 
AS a single correct result of GROUPING, we obtain for 
our example the structure represented in Figure 2. 
It can be easily demonstrated that some other deriva- 
tion attempts will not succeed; thus, for instance, if we 
try to apply rule 2 of GROUPING in order to arrive at a 
coordination of main clauses through the conjunction et, 
the second presumed main clause will not be completable 
since a verb phrase is missing (in contrast to a sequence 
such as . . . et les pays africains (relative clause) souffri- 
ront de la famine.). For another case of an ineffectual 
attempt, we can note rule 11, which would initiate a 
coordinate Sequence of Subordinate Clauses (SSBCL),; 
here this would lead to a missing verb phrase belonging 
to the first SUNNOG of the sentence. 
Each correct augmentation result - such as the one in 
Figure 2 - forms the basis upon which the subsequent 
PHASE III of ASCOF will then operate clause by clause. 
\[We are at present working on the combination of the 
different distinct parts (word class disambiguation/ 
decomposition into simple syntactic units; 
IDENTIFICATION and GROUPING); this conception will 
ultimately lead to a single CFG achieving PHASE II of 
ASCOF.\] 
4.4 
4.4.1 
The task of analyzing verb sequences, the first step of 
PHASE III (cf. Figure la), is to group together isolated 
verb elements (finite verbs, participle I, participle II, 
infinitive) within a segment to assign a structural 
description to these phrases (e.g., to determine VOICE 
and TENSE) and - ultimately - to interpret those phrases 
as nodes of a tree structure. 
The grammar of the verbal analysis is conceived as a 
two-step ATN (cf. Biewer 1985). 
The transitions between the different states of the 
ATN are guided primarily by a subclassification of verb 
elements (e.g., participle II of avoir). The conception of 
the ATN enables the processing of an unlimited number 
of infinitive phrases; the integration of these infinitive 
phrases into the tree structure is governed by categorial 
information - recorded in dictionaries - as well as by 
tests concerning the syntactic context. Nevertheless, the 
interpretation of non-complex infinitive phrases as nodes 
of a dependency tree does not exclude ambiguities that 
cannot be resolved at this stage of the verbal analysis. 
Accordingly, it is a set of hypothetical structures, gained 
by purely syntactic and surface-related data, that forms 
the output of the verbal analysis. 
4.4.2 
When the analysis of verb phrases is completed, the 
sentence is structured in such a way that parts of main 
and subordinate clauses and their interrelations are iden- 
tified. Furthermore, non-complex syntactic (one-nu- 
clear) noun and prepositional phrases as well as verb 
sequences are determined. An interactive component 
operates on this input performing (1) the complement 
analysis, (2) the analysis of complex (multinuclear) noun 
and prepositional phrases, and (3) the disambiguation of 
Computational Linguistics, Volume 11, Numbers 2~3, April-September 1985 141 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
:l r..) 
.8 
/ pO ~0~ O' //~ ~ ~ 
u u ~ 
U 
'(.) 
m~ 
Figure 2. Tree-structure of the segmented phrase. 
142 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F~neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
lexical items, all according to syntactic and semantic 
criteria. 
In contrast to many other systems of machine trans- 
lation or linguistic data processing, semantics and syntax 
are here equally treated, neither having priority over the 
other. Another characteristic feature is the interaction 
between each step of analysis: the complement analysis 
determines the head of a complement (SUB, K1, K2 in 
Figure 3), and the internal structure of the complement 
itself (hypotactic and paratactic relations) is determined 
by the complex nominal analysis (NOMAL), the findings 
of which are then returned to complement analysis (cf. 
Figure 3). This interactive procedure is guided by the 
syntactic-semantic features of the dominating verb. The 
required syntactic features have already been calculated 
in the previous analysis; in contrast, the semantic infor- 
mation is memorized in the data base of the semantic 
network and must be taken from there (cf. 4.4.4). 
ASCOF analysis offers, as output, syntactic-functional 
structure trees representing a synthesis of constituents 
and dependencies. 
In the complement analysis, the dictionary entries of 
the verb - the verb being the central node of the 
sentence - determine those noun or prepositional phrases 
that belong to the verbal frame. The remaining phrases 
are parts of complex noun or prepositional phrases or of 
adverbials. For the preverbal and postverbal fields, 
complement analysis operates separately, reflecting the 
dissimilar structure of both fields. The component SUB, 
operating in the preverbal field, defines subjects of non- 
dominated verbs as well as possibly occurring pronominal 
complements; dominated verbs (e.g., causative 
constructions) are analyzed by another component, not 
described in this paper (INFSUB, cf. Stegentritt (1982, 
1984)). In the postverbal field, the components K1 and 
K2 are activated for the complement analysis. In every 
case, only the first constituent of the complement is 
determined. Then the component NOMAL (cf. 4.4.3) is 
called and investigates the structure of the identified 
complement. 
The component determining the subject (SUB) is 
bipartite; in addition to the preverbal pronominal 
complements, the first part also defines the subject, 
provided that it is a pronominal subject. If there is no 
pronominal subject, part two defines a nominal subject. 
in the first part, the analysis progresses from right to left, 
starting at the finite verb/auxiliary. All preverbial 
elements are processed subsequently, until a pronominal 
subject is found. 
Figure 4 represents the first phase of SUB as an ATN. 
The different paths correspond to the possible distrib- 
utions of syntactic phrases in the preverbal field of the 
French language. If the verb is directly preceded by a 
comma, it is highly probable that this comma indicates 
the end of an insertion, such as in sentence (4). In this 
case, the analysis is continued up to the next comma, 
with no consideration of the insertion. The edges and 
states of the ATN that had been activated are listed in 
(4a). 
H r,. 
t~ 
Figure 3. Interaction of different components in Phase III. 
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
f~ t-i 
f~ 
0 
143 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
144 
8 
~1 I~ ~1 -/ / ~'~1 I~ =1 / / 
0 
,e 
Z 0 
0 
Z 
m 
I 
E~ 
I--4 
ee 
I-I 
Z 0 
I--t 
0 Ul 
Figure 4. ATN grammar of the preverbal field. 
Computational Linguistics, Volume 1 i, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F~neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
(4) ... qui, ~ l'article 4, pr6voit ... 
(4a) 1 - COMMA - 2 - ( not COMMA)*4 - 2 - COMMA 
-1 
In the first phase, combinations of clitics as noted in (5) - 
(7) may be identified. 
(5) il ne le lui propose pas 
(5a) 1 - JUMP - 9 - LUI - 10 - PRN(acc) - 12 - JUMP 
- 5 - NEG - 6 - PRN-Subject - 8 - STOP 
(6) il me le donne 
(6a) 1 - JUMP - 9 JUMP - 10 PRN(acc) - 11 - 
PRN(dat) - 12 - JUMP - 5 JUMP - 6 PRN-Subject 
- 8 - STOP 
(7) il m'en parle 
(7a) 1 - EN - 3 - JUMP - PRN(dat) - 5 - JUMP - 6 - 
PRN-Subject - 8 - STOP 
If the network comes to the end of a clause or meets a 
configuration not covered by the rules, the second phase 
of analysis - starting at the beginning of the clause in the 
reverse direction of analysis - is called to search for the 
first nominal phrase (i.e., the leftmost). This is, however, 
a mere candidate for subject function, the form of a 
syntactic phrase (NNOG) not yet indicating the actual 
function of this phrase. As an illustration, compare (8). 
(8) Apr~s les changements de cours pour le froment, 
l'avoine et la seigle, la directive (...) 
touche (...) 
As the set of rules in this second phase supplies the first 
noun phrase (NNOG, = l'avoine in (8)) as a subject, the 
phrase in question must be checked as to whether or not 
it is part of a coordination of prepositional phrases in 
which the preposition can be deleted. This subject-con- 
trol mechanism presupposes that the coordinating 
conjunctions et or ou always indicate the end of a coordi- 
nation. It follows that the subject candidate cannot be 
the subject if the following structure is given: 
(9) CANDIDATE-COORD-NNOG-COMMA-NNOG-VERB 
In this structure, however, the final preverbal NNOG (in 
(8), la directive) is the correct subject. In structure (10): 
(10) CANDIDATE-COORD-NNOG-VERB 
the subject candidate is confirmed. The head of the 
subject having been identified, phase NOMAL is called to 
examine a possibly complex internal structure of the 
subject and further potentially complex nominal sentence 
elements in the preverbal field - as, for example, the 
adverbial insertion ~ l'article 4 in sentence (4) (cf. 4.4.3). 
The complement analysis in the postverbal field is 
carried out by another pair of components (K1 and K2), 
also cooperating with NOMAL. K1 is the less sophisti- 
cated complex, identifying nothing but direct comple- 
ments - which it recognizes merely by form - and 
prepositional complements obligatorily demanded by the 
verb. Yet because of the various functions prepositional 
phrases may assume, prepositional complements at this 
stage are to be identified only if no more than one PG 
within the segment is a potential candidate. Applied to 
our example (1) 
(1) la directive du Conseil du 20 juillet, qui, ~ l'article 4, 
pr6voit une augmentation du prix du mais de 3%, 
touche les r6gions du sud et les pays africains qui 
d6pendent de ce produit d'importation. 
the results are: 
- for prgvoir: une augmentation as a direct object 
- for toucher: les rdgions as a direct object. 
Again, the determination of the coordinative relationship 
between rdgions and pays is given within NOMAL (cf. 
4.4.3). 
Yet, whenever the prepositional complement is 
optional or when more than one candidate for an obliga- 
tory prepositional complement is available, semantic 
criteria must be observed in addition to formal and 
syntactic criteria in order to correctly determine the 
complements. 
The semantic restrictions that must be respected when 
a certain phrase is attributed the function of a comple- 
ment are denoted in the verb entry in the dictionary. 
According to these (semantic) restrictions, the algorithm 
searches for the required criteria, namely among the 
nuclear lexemes of the potential complement phrase 
under consideration. The semantic information for these 
lexemes, however, is not stored with their dictionary 
entries (e.g., nouns) themselves, but is represented by a 
semantic network (cf. 4.4.4). 
The strategy used by ASCOF in K2 shall be illustrated 
by the three verbs in sentence (1). 
Prdvoir does not demand the attribution of a further 
complement, the verbal frame being saturated as soon as 
une augmentation is identified as a direct complement. 
The verb toucher raises other problems; toucher may 
occul" 
a. without any complement: le tireur d'dlite a touchg. 
b. with a direct complement: il touche ma main. 
c. with an h-complement: le chercheur touche au but. 
d. with a direct complement and an optional de- comple- 
ment: il m'a touchd du doigt. 
As to sentence (1), the direct complement les rdgions 
having been located, it must still be examined whether 
one of the following de-phrases is a complement; that is, 
whether verbal frame (b) or (d) is valid. A de-phrase, 
however, can for instance be the complement of toucher 
if the noun of this phrase is semantically labeled as "part 
of the body". Consequently, the semantic value of the 
nuclear lexemes of the de-phrase following les rdgions 
must be examined in this respect. The results obtained 
through consultation of the semantic network show that 
the nuclear lexeme (sud) of this phrase does not satisfy 
the required condition. It follows that verbal frame (b) is 
actualized in this sentence. Simultaneously, this analysis 
defines the reading of the verb. 
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 145 
Axel Biewer, Christian F~neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
As to the last verb of sentence (1), ddpendre, no direct 
complement can be found. Thus the reading 'to be 
dependent on' versus the reading 'to take down' is imme- 
diately determined. This reading of d6pendre demands an 
obligatory de-valency. Yet no restrictions can be formu- 
lated for the semantic quality of the nuclear lexeme of 
the de-phrase. This implies that only formal and syntac- 
tic criteria can be considered. In the case of the first 
phase, de ce produit, the formal and syntactic criteria hold 
true (a preposition introducing a phrase, an article) and 
thus allow this phrase to be identified as a complement 
and handed over as a head-phrase to the analysis of the 
complex noun phrases. A second run of K2 checks the 
following phrase, d'importation, which is rejected by 
those criteria identifying this phrase as a compound in 
NOMAL. The criteria are: no proper noun, number = 
singular, no article, no attribute. One single attribution is 
performed accordingly. 
An interesting variant of sentence (1) is: 
(11) ... drpendent de l'effet de l'importation. 
where both phrases may function as a complement. If 
the second phrase (de l'importation) is identified as a 
complement, the first postverbal phrase can only be attri- 
buted the function of an adverbial. In a concluding 
control phase for those phrases scheduled as adverbials, 
the structure variant of de l'effet as an adverbial and de 
l'importation as a complement can be erased, effet not 
being apt to assume the function of an adverbial in this 
phrase, in contrast to prepositional phrases, such as de 
cette mani~re or de cette faqon. 
4.4.3 
The task of the analysis of complex noun phrases (multi- 
nuclear noun- and prepositional phrases) is to identify 
their boundaries and to describe their internal structures, 
i.e., to determine the syntactic-functional relations 
between the various parts of the complex. \[For more 
details, cf. Ritzke (1985).\] 
The following syntactic-functional relations are 
defined for the analysis: 
- head-function of the central phrase of a complex (la 
maison du pkre) 
- paratactic relations of phrases depending on the head: 
coordination (CO): the coordinated phrase has the 
same syntactic function as the phrase with which it is 
coordinated (l~tude de la dgtermination et de la classi-. 
fication). 
apposition (AP): a paratactic relation on the level of 
syntactic analysis (du froment tendre, plante comestible 
qui . . . ) 
- hypotactic relations of phrases depending on the head: 
prepositional complement/object (PO): signalizes a 
very close relation between two phrases; the preposi- 
tion of the second phrase is mostly synsemantic; the 
PO and the inflectional case of the two languages 
(French, German) frequently correspond to each 
other (la voix de son mattre- die Stimme seines Herrn). 
\[For more details concerning the function of prep- 
ositions and their analysis, cf. Ritzke ( 1981).\] 
prepositional complement/part of a compound (PC): 
impfies an even closer unity of the two phrases; PC 
and the governing phrase, in contrast to other struc- 
tural units, represent primary units often functioning 
as a one-nucleus phrase; this is of great importance to 
the language pair French and German, the complex 
phrase in French being equivalent to a German one- 
word compound (e.g., la notation de base - Basisnota- 
tion) 
prepositional attribute (PA): signals a relatively free 
relation to position and introductory preposition, 
which is always autosemantic; for this reason, the PA 
may assume many different semantic values (e.g., 
temporal or local attribute, etc.) un groupe simple g~ 
l'intdrieur du syntagme). 
Basic elements for the analysis of complex noun 
phrases are syntactic phrases having only one nucleus. 
They may occur in the form of noun phrases, preposi- 
tional phrases, or pronouns functioning as nouns. 
Every one-nucleus noun - or prepositional - phrase of 
these basis elements is directly relevant to the analysis of 
complex noun phrases as a potential element of a 
complex phrase. Every non-nominal phrase (e.g., verb 
phrase, etc.) is indirectly relevant as an indicator of the 
boundary of the complex. 
As for French, the identification of the left boundary 
of a complex noun phrase presents no problem since the 
phrase furthest left of a complex is its central phrase (its 
head) in the majority of cases. The head represents the 
syntactic function of the whole complex on the sentence 
level and at the same time marks the boundary of the left 
side. Further phrases are located on the right side of the 
head, the analysis being directed from left to right. 
The head of a complex noun phrase is identified and 
determined in syntax and function by means of the anal- 
ysis of verb complements. The complement analysis 
forms an interacting system with the analysis of nominal 
complexes - as explained above. The potential head of a 
potential complex noun phrase, identified by complement 
analysis, is a signal for the call of the complex nominal 
analysis. 
The search for further phrases on the right side of the 
head begins at the head. Here, two phrases are checked 
to identify a boundary. These phrases consist of the 
head on the left and a dependent phrase on the right 
which is connected with this head. The syntactic func- 
tion is simultaneously identified. Thereupon, comple- 
ment analysis is called again to determine further 
complements or non-complements, respectively. 
Non-complements (adverbials) may also have a 
complex internal structure. Accordingly, after the analy- 
sis of complex-structured complements, non-comple- 
ments are analyzed as to their possibly complex structure. 
The preverbal field is investigated first. If, for example, 
the complex structure of a subject has been identified 
there, a further complex nominal analysis is undertaken 
146 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
for the first phrase of the sentence - provided that this 
phrase is identified as adverbial and that other phrases 
follow in the field between its position and the head of 
the subject. This is done in order to clarify whether this 
phrase is a central phrase of an adverbial complex. 
A similar approach is applied for the postverbal field: 
first the complements, then the non-complements, are 
analyzed as to their possibly complex internal structures. 
In principle every phrase can be related to another 
phrase in its direct neighborhood. This phrase again may 
dominate a subsequent phrase, etc., so that highly recur- 
sive structures may appear within a nominal complex 
(right-branches cascades). 
At the same time two related phrases may be discon- 
nected by other phrases, when several phrases of differ- 
ent syntactical relations are dependent upon the same 
phrases. In addition to the potentially recursive struc- 
tures, discontinuous relations must be detected as well. 
The algorithm NOMAL has been conceived especially 
for the solution of those recursions and discontinuities 
within the complex nominal syntax (cf. Figure 5). 
NOMAL is a procedural (sub)system similar to an 
ATN. Each syntactic function has its own procedure. 
The procedures are called successively by the main 
program whenever a pair of phrases - the head and the 
actual phrase - is examined. The identification of the 
relation that is - or is not - established between both 
phrases is governed by rules containing conditions for the 
head and the actual phrase. Each procedure is provided 
by a proper set of rules; the monitoring of the rule condi- 
tions is carried out by test operations on edges. 
If the testing of a rule attains a positive result - that is, 
if the conditions in this rule for head and actual phrase 
hold true - the phrases are unified and their relation 
interpreted according to the procedure where the rule 
was found. Valid conditions in the procedure PO, for 
example, have the effect that the relation between the 
phrases is identified as a prepositional comple- 
ment/object. The approach is analogous for the remain- 
ing procedures. 
(INPUT) 
PROC: APPOSITION (OUTPUT) 
PROC: COORDINATION 
PREPOSITIONAL 
COMPLEb~NT/OBJECT 
PROC: PREPOSITIONAL 
ATTRIBUTE 
PROC : PREPOSITIONAL 
COMPLEMENT/PART 
OF A COMPOUND 
Figure 5. The subsystem NOMAL. 
Computational Linguistics, Volume I l, Numbers 2-3, April-September 1985 147 
Axel Biewer, Christian Frneyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
The actual phrase having thus been attributed a func- 
tion, it b_ecomes the head itself in the next run; the actual 
phrase will then be the subsequent phrase. Analogously, 
further positive results gradually move the actual phrase 
to the head of the next run; the head is deferred to the 
right, thus allowing the correct analysis of recursive 
cascades. 
If none of the procedures attains a positive result, the 
new run is begun with another head, that is, with the 
phrase to the left of the head last tested. If there is still 
no positive result, the head keeps moving analogously to 
the left, until all the possibilities have been tried. The 
actual phrase is identical for all runs, thus allowing the 
analysis of relations between two discontinuous phrases. 
In every case, a boundary is reached when none of the 
runs offers a positive result. The former actual phrase 
then no longer belongs to the complex but rather indi- 
cates its boundary. 
The conditions denoted in the rules consist of 
morpho-syntactic and semantic information. Morpho- 
syntactic information (e.g., an article, the form of an 
introductory preposition) can be read directly, the previ- 
ous step of the analysis having made it available. In 
contrast, the semantic information must be deduced from 
the semantic network, which serves as a data base for the 
syntactic-functional analysis (cf. 4.4.4). 
The importance that evaluating rule conditions holds 
for the analysis shall be demonstrated in the following by 
the complex noun phrases in example (1). 
The nominal complex Les rdgions du sud et les pays 
africains: 
Formal indicators can be checked at this point for the 
determination of the paratactic relation: in our exam- 
ple, the coordinating conjunction et. Commas and the 
sequences of comma and conjunction may be evalu- 
ated as well. Yet the mere checking of such indicators 
is generally not sufficient for determining the proper 
parts of a coordination. 
In order to determine the correct parts, the seman- 
tics of the nouns must also be checked, that is, the 
closeness of the semantic relation between the candi- 
dates in question must be checked in the example. 
This is accomplished by retrieving operations within 
the semantic network using the nominal lexemes of the 
different phrases. The common generic term 
(geographic place), which is reached by rdgion and 
pays, satisfy the two conditions: coordination indicator 
et and greatest semantic closeness. They are parts of 
the coordination, the phrase du sud is recognized to be 
a prepositional complement/object of the first phrase 
by the procedure PO. The result is the following 
structure: 
((les r6gions (du sud)(PO)) (les pay africains)(CO)) 
The nominal complex la directive du Conseil du 20 juillet: 
For hypotactic relations, too, the examination of 
semantic criteria is necessary over and above that of 
the syntactic conditions. The search for a certain 
preposition at the beginning of a clause - in our exam- 
pie, de as an indicator of object references - must be 
performed by examining the semantic closeness of the 
nuclei (nouns) of the phrases. In the given example, 
the direct time indication, which suggests an attribu- 
tive relation, is thus identified. In the network, a 
temporal attribution is closer to the information of 'act 
of legislation' (connected with the lexeme directive) 
than to 'institution' (connected with the actual reading 
of the lexeme Conseil). Accordingly, valid conditions 
exist for the. relation of prepositional 
complement/object between la directive and du 
Conseil and for the relation prepositional attribute 
between la directive and du 20 juillet. The last relation 
is discontinuous. The result of the analysis is the 
following structure: 
((la directive (du Conseil)(PO))(du 20 juillet)(PA)) 
The nominal complex une augmentation du prix du mars 
de3%: 
The last non-complex phrase, de 3 %, of this complex 
phrase holds a - likewise discontinuous - relation to 
the first phrase, une augmentation, whereas the 
remaining de-phrases form a cascade and each 
depends upon its preceding phrase (the relation of 
prepositional complement/object). 
The phrase de 3%, an indication of quantity, is 
more closely related to the action, augmentation 
(action/state of affairs: increase), by the correspond- 
ing edge configuration in the network than to prix 
(notion of a measure) or to ma~s (plant, agricultural 
product). Accordingly, a relation is established 
between de 3% and une augmentation, disregarding 
the interposed phrases. The result is the following 
structure: 
((une augmentation((du prix (du mais)(PO))(PO)) 
(de 3% )(PA))) 
The nominal complex de ce produit d'importation: 
In order to determine the relation between the two 
phrases, formal criteria may be more readily applied 
than in the previous cases. The second phrase is char- 
acterized by the fact that neither a determinative nor 
an attributive element is contained or follows. More- 
over, the nominal nucleus neither is in the plural nor is 
it a proper name. The phrase is introduced by the 
preposition de. All the characteristics are rule condi- 
tions for an actual phrase in the procedure preposi- 
tional complement/part of a compound. These 
conditions being fulfilled and no other candidate being 
present, the relation between the phrases may be 
determined in this particular case without semantic 
examinations. The result is the bracketing: 
(de ce produit (d'importation)(PC)) 
As it is difficult in many cases, mainly in the postver- 
bal field, to determine a function with certainty after a 
single run of the analysis of complex noun phrases, deci- 
sions are made only if there is no doubt (e.g., if there is 
148 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F6neyrol, Johannes ~Ritzke, Erwin Stegentritt ASCOF 
only one potential candidate or if candidates can be 
excluded because of semantic incompatibility, etc.). 
Otherwise a solution must be sought through repeated 
interaction with its complement analysis, which must 
then be given priority. Yet, given the basic alternative of 
the PP-attachment (complement, adverbial, part of a 
complex noun phrase), a reliable solution cannot always 
be found and several solutions may be admitted in such 
doubtful cases. 
4.4.4 
Conceding that structural descriptions within machine 
translation - functioning as the input for a transferring 
component and thus defining a sort of "interlingua" - 
cannot be established without semantic information, the 
appropriate form for representing this information as well 
as its integration into the analysis process appears to be 
controversial. 
With regard to the various ways of integrating seman- 
tic information, and above all to the proper moment of its 
integration, there are two extremely different points of 
view. On the one hand, syntax may be accorded absolute 
priority, in such a way that all syntactic and structural 
descriptions of a given input sentence are generated and 
only then semantically interpreted and filtered according 
to the semantic information. On the other hand, the 
analyzing algorithm can be guided by semantic informa- 
tion, the "level" of syntactic relations only being 
consulted as a secondary control mechanism. 
Most machine translation systems favor the first point 
of view, partly owing to the history of this science. Yet 
then, of course, the problem with highly complex 
sentences is how to deal with syntactic ambiguities. If, 
for example, an input sentence is attributed a number of 
tree structure by the syntax-driven analysis, part of the 
syntactic relations will necessarily be represented redun- 
dantly; if, on the other hand, the input sentence is attri- 
buted a chart or chart-like structural description by the 
syntactic analysis, much more time must be sacrificed to 
the semantic component - as a consequence of the more 
complex input structure. 
Semantically driven parsers, whose equivalents can 
mainly be found in language-oriented AI research, guar- 
antee a highly efficient parsing, but are extremely 
dependent upon their discourse area, due to their consid- 
erably limited "expectancy", and generally operate on 
non-complex sentences, linguistically speaking. 
In order to isolate correct structural descriptions as 
soon as possible, cross-connections between syntax and 
semantics are established by the ASCOF system (cf. 
4.4.2). This is to say that semantic information is made 
available for the identification of syntactic-functional 
relations and that, vice versa, syntactic and functional 
relations can be used for semantic interpretation (disam- 
biguation). 
The flow of information between syntax and semantic 
in ASCOF analysis is therefore not based on a phase 
model (e.g., the sequential operations of separate 
modules). The semantic information in the ASCOF 
system is stored in a semantic network. (For different 
conceptions of semantic networks, cf., above all, Findler 
(1979).) 
The implementation of the network structure is based 
on the COMSKEE data structure "dictionary", which is a 
string-indexed array (of variable length) over strings (of 
variable length) that can be held externally. "Of variable 
length" means that a string can be of any size so that its 
length is only limited by the memory space of the 
machine. 
ARRAY STRING indentifier ? 
the entry 
can also be 
a composed 
data type 
(a record) 
the index (keyword)~ 
can take an arbitrary I 
value of the data type I 
string / 
the (optional) // 
EXTERN-markino 
means that the 
dictionary is 
stored in an 
external mass 
storage 
Figure 6. Declaration of the COMSKEE data structure "dictionary". 
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 149 
Axel Biewer, Christian F6neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
A module that interprets this structure as a network 
operates on the "dictionary", as shown in Figure 7a. 
\[Figures 7a-7c. Structures of the dictionary entry.\] 
7a. mdex 
household 
utensil 
artifact 
entry 
ISA : artifact 
l 
ISA : concrete object 
A part of the entry can be the index (keyword) of 
another entry. Thus it is not necessary to store explicitly 
the information that a person has a head; rather this 
information can be deduced on demand by following an 
ISA path, which means in our implementation, interpret- 
ing an array element such as: 
7b. index entry 
A ISA: B, 
as a pointer from A to B labeled by ISA, as in 
ISA 
7c. A > B 
All directly addressable nodes in the network repre- 
sent basic forms that point to semantically unambiguous 
readings following denoting edges. Consequently, they 
represent the FORM-CONTENT relation. For example, 
the basic lexical form cuisini~re points in the semantic 
readings of 'female cook' and 'oven', following the 
denoting edges, as Figure 8 shows: 
As also shown in Figure 8, several basic lexical forms 
may refer to the same semantic reading when following 
the denoting edges. The relation defined by denoting 
edges thus takes into account the linguistic phenomena of 
synonymy and polysemy. 
ISA edges, which define a special implicit relation, are 
important for two reasons. First, they allow the defining 
of a sort of semantic distance between linguistic units, 
and second, they economically administer information 
deducible by inheritance paths. 
Semantic distance in the linguistic sense may be of 
importance for the correct identification of coordinative 
structures. The complex noun phrase in 
(12) La vente des r6frig6rateurs et des cuisini6res 
may - on a purely syntactic level - be attributed two 
structures (des can be interpreted as a preposition as well 
as an indefinite article), cf. Figures 9a and 9b: 
Based on the information as illustrated in Figure 8, it 
is now possible to carry out a lexeme disambiguation of 
cuisiniere (namely, as (cuisini~re 2)) as well as to give 
structure 9a priority over structure 9b, since 
(cuisini6re 2) is semantically more closely related to 
rdfrig~rateur than (cuisini~re 1) and, furthermore, since 
(cuisini6re 1) and (cuisini~re 2) are more closely 
related to the node 'r6frig6rateur' than to 'vente'. 
Because of the relations expressed in Figure 8, structure 
9c can be assigned to sentence 12: 
As made clear by this example, syntactic-functional 
and semantic decisions are made almost simultaneously 
in ASCOF. A system admitting a two-phase model for 
the informational flow between syntax and semantics 
would not only have to organize the ambiguity of struc- 
tures in sentence (12) but - in order to undertake a 
lexeme disambiguation of cuisinidre - would also have to 
construct exactly those syntactic relations directly avail- 
able in an interactive system like ASCOF 
For lexeme disambiguation as well as for the identifi- 
cation of syntactic-functional relations, linguistics units 
that may function as predicates are of major importance. 
These are primarily adjectives and verbs. Each "verbal 
node" following denoting edges refers to a set of verbal 
frames whose complement slots are described by syntac- 
tic and semantic conditions. 
In detail, each complement slot is attributed the 
following information. 
o A complex of conditions that must be fulfilled by a 
syntactic phrase in order to function as the correspond- 
ing complement: 
- syntactic conditions 
- semantic ,conditions (to be fulfilled by the head of 
the phrase in question) 
e.g., for toucher in 4.4.2, the formal-syntactic condi- 
tions for the prepositional object are: preposition = 
de; the semantic condition for the relevant head is : 
ISA: part of the body. 
- the necessary occurrence of the argument 
- the possibility of rejecting competitive verbal frames 
In the following, the procedure and the efficiency of 
the semantic network within the syntactic-functional 
analysis shall be explained with the aid of a represen- 
tative example. 
The French verb fumer, for instance, is assigned three 
verbal frames; they correspond to the readings of 'to 
smoke a cigarette, to give off smoke, to smoke a trout', 
cf. Figure 10. 
As in Figure 10, a network relation between entity 
nodes and predicate nodes is established according to 
semantic conditions defining the restrictions for the 
different complement slots. In Figure 10, this allows the 
interpretation of ruiner as 'to give off smoke' and the 
rejection of the other interpretations (cf. the action 
commands within verbal frames). This is to say that 
once cheminde is identified as a subject, there is no need 
to look for a direct object. Furthermore, Figure 10 
demonstrates that "semantic markers" alone do not guar- 
antee correct semantic interpretations. It is by no means 
sufficient to know that poisson is a concrete object in 
order to interpret fumer in 
(13) La cuisini~re fume du poisson 
in the sense of 'to smoke-dry'; because the meaning of 
(fumer 3) 'to smoke-dry' and (fumer__2) 'to smoke' - 
150 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 
Axel Biewer, Christian F~neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
-- " ... -tS,~ tO~.~(CUISINIERE i) ISA> .. ISA>(PERSON)~ ~ (CONCRETE 
OBJ ) 
"CUISINIERE" ~ ~ %~,~/ 
o 
~)~A'~.-k (CUI S INIERE 2) > UT~.NSIL) 2 ........ ISA (HOUSEHOLD ISA c ~TrFAr.,~ / 
,, RE FR i GERATEUR ,~OT~ ~. .~ /'7 
~O~..~ (REFRI GEP~ATEUR ) ~ / 
"FRIGIDAIRE" ~ 9~N ~ 
Figure 8. Example of the network structure. 
NNOG NNOG I /\ 
VENTE VENTE CUI S INIERE 
REFRIGERATEUR CUISINIERE REFRI GERATEUR 
Figures 9a-9b. Possible phrase-structures of (12). 
NNOG 
I 
VENTE /\ 
REFRIGERATEUR CUISINIERE 2 
Figure 9e. Correct interpretation of (12). 
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 151 
Axel Biewer, Christian F~neyrol, Johannes Ritzke, Erwin Stegentritt ASCOF 
as in (14) - is not discernible by the semantic marker 
/concrete/. 
(14) La cuisini~re fume une cigarette 
In a componential model, every animal would have to be 
marked /can be smoked (dried)/, whereas in the 
network model this information may be deduced by 
following a "function" arc from 'animal' to 'viande'. 
Moreover, with the complement identification, semantic 
conditions attributed to the different complement slots 
may be considered for lexeme disambiguation. For 
example, the reading 'oven' for the noun cuisini~re in 
(13) and (14) can be rejected on grounds of the semantic 
conditions attributed to the subject slots. 
At the end of this part of the analysis, the syntactic- 
functional structure of the sentence is produced (i.e., 
main and sub-/coordinate clauses, the verb predicate and 
its complements, adverbials, as well as the internal struc- 
ture of noun and prepositional phrases are identified). 
The analysis yields no interlingual structures but rather 
structures more closely related to the source language. 
These are represented by canonical trees. At the same 
time, due to the interaction between syntax and seman- 
tics, many of the lexical ambiguities are resolved. Before 
starting ASCOF's final two phases of transfer and synthe- 
sis, any remaining ambiguities must be cleared up espe- 
cially by means of further semantic information. 
4.5 
4.5.1 
The phases of the system subsequent to the analysis 
include a transformation component (transfer and 
syntactic synthesis) and a morphological synthesis. In 
the transfer phase, the source language lemmata are first 
replaced by their target language equivalents. Here it is 
not sufficient to merely exchange lexemes since target 
language differences of a lexical nature arising from 
syntax and/or semantics need to be borne in mind. 
Instead, the replacement process must take into consider- 
ation the information contained in the analysis tree, for 
example, the syntactic and semantic verb frame actuali- 
zation. Subsequent to the lexical transfer, the analysis 
structures are transferred to the structures appropriate to 
the target language (syntactic synthesis). This transfer 
component is implemented as a tree transformation algo- 
rithm that interprets externally stored rules (transforma- 
tion instructions). Linguistic data and algorithm are thus 
strictly separated from one another so that this compo- 
nent may also be used for other language pairs. The 
transformation algorithm runs through the analysis tree 
in preorder, tests for each node whether a package of 
rules exists for the given node label and - provided that 
the conditions of a rule are fulfilled - carries out the 
instructions that refer to some few elementary oper- 
ations. The conditions and the transformation 
instructions can refer to both subtrees and the attribute- 
value pairs associated with the node. (For example, the 
inheritance of the target language gender of the head of a 
NOG to attribute and determiner complex.) 
4.5.2 
The input for the morphological synthesis is the labeled 
tree taken from the syntactic synthesis. In contrast to the 
syntactic synthesis, the morphological synthesis operates 
only locally, in other words, the pre-terminal nodes are 
examined and processed isolated from each other. Tree 
transformations are thus no longer carried out. The basic 
forms of the lexemes and the morpho-syntactic informa- 
tion from the pre-terminal nodes serve as keys that call 
the appropriate rule of the morphological generative 
grammar. Grammar and algorithm are separate. The 
grammar for the morphological synthesis of German is 
able to generate German word forms, provided that the 
necessary information supplied from all of the preceding 
phases is complete and accurate. 
5 FUTURE PROSPECTS 
The ASCOF system has been implemented in the sections 
described in the present paper and has in part already 
been tested. Here we have a version that covers syntac- 
tic structures and vocabulary based on texts (EC bulle- 
tins) taken from the agricultural sphere. The vocabulary 
of this sphere is completely covered on the morpho-syn- 
tactic level. As yet, owing to the complexity of the 
semantics, the semantic data bases (semantic network) 
has been developed only as a prototype. The improve- 
ment of .the quality of the translations performed by 
ASCOF is basically dependent upon the development and 
elaboration of the semantic networks. 

REFERENCES 
Bates, Madeleine 1978 The Theory and Practice of Augmented Tran- 
sition Network Grammars. In Bolc, Leonard, Ed., Natural Language 
Communication with Computers. Springer, Berlin, Heidelberg, New 
Yor, W. Germany/USA: 191-259. 
Biewer, Axel 1985 Ein ATN zur Verbalgruppenanalyse des 
Franz~sischen. In Figge, Udo L. 1985: 13-26. 
Boitet, Christian; Guillaume, Pierre; and Qu6zel-Ambrunaz, Maurice 
1982 Implementation and Conversational Environment of ARIANE 
78.4. An Integrated System for Automated Translation and Human 
Revision. In Fix et al. 1985: 225-236. 
F6neyrol, Christian 1982 La forme en (-ant): un probl~me de 
l'analyse automatique du fran~ais. In Fix et al. 1985: 237-262. 
F6neyrol, Christian 1983 La segmentation automatique de la phrase 
dans le cadre de l'analyse du frant;ais. In Actes du Colloque Interna- 
tional Informatique et Sciences Humaines. 1981. LASLA. Universitd 
de Liege. Liege, Belgium: 353-368. 
F6neyrol, Christian; Ritzke, Johannes; and Stegentritt, Erwin 1984 
Esquisse d'une analyse fonctionnelle du fran~ais (ASCOF). In Acres 
de la Xle Conference Internationale de I'ALLC. Louvain-la-Neuve. 
Forthcoming. 
F6neyrol, Christian and Stegentritt, Erwin 1982 Komplementanalyse 
in komplexen S~itzen des Franz~sischen. In Stegentritt, Erwin, Ed., 
Maschinelle Sprachverarbeitung 1981. Vortage auf der 12. Jahresta- 
gung der GAL, Mainz 1981. Sektion Maschinelle Sprachverarbeitung. 
Sprachwissenschaft - Computerlinguistik 8. AQ, Dudweiler, W. 
Germany: 55-68. 
Figge, Udo L., Ed. 1985 Romanistik und Datenverarbeitung 1983. 
Akten des Deutschen Romanistentages, Berlin 1983. Sektion Romanis- 
tik und Datenverarbeitung. Sprachwissenschaft - Computerlinguistik 
11. AQ, Dudweiler, W. Germany. 
Findler, Nicholas 1979 Associative Networks, Representation and Use of 
Knowledge by Computers. Academic Press, New York, San Francis- 
co, London, USA/Great Britain. 
Fix, Hans; Rothkegel, Annely; and Stegentritt, Erwin, Eds. 1985 Spra- 
chen und Computer, Festschrift zum 75. Geburtstag yon Hans Eggers. 
Sprachwissenschaft - Computerlinguistik 9. AQ, Dudweiler, W. 
Germany. 
Huckert, Edgar 1979 Automatische Synthese des FranzOsischen aus einer 
logischen Basis. Sprachwissenschaft - Computerlinguistik 2. AQ, 
Dudweiler, W. Germany. 
Marcus, Mitchell P. 1981 A Theory of Syntactic Recognition for Natural 
Language. MIT Press, Cambridge, London, USA/Great Britain. 
Messerschmidt, Jan 1984 Linguistische Datenverarbeitung mit 
COMSKEE. Teubner, Stuttgart, W. Germany. 
Mueller-v. Brochowski et al. 1981 The Programming Language 
COMSKEE, 2nd Revised Report. Linguistische Arbeiten, Neue Folge, 
4. SFB 100, Saarbrt~cken, W. Germany. 
Reding, Helene 1985 TransformationskomponentefiirBaumstrukturen. 
Linguistische Arbeiten, Neue Folge, 11. SFB 100, Saarbr0cken, W. 
Germany. 
Ritzke, Johannes 1981 Probldme de l'analyse automatique des 
prdpositions du frangais. In Schwarze, Christoph, Ed., Analyse des 
prgpositions, llle Colloque franco-allemand de linguistique thgorique 
du 2 au 4 fdvrier 1981 gl Constance. Niemeyer, Ttibingen, W. 
Germany: 139-157. 
Ritzke, Johannes 1982 Zur automatischen Analyse adverbialer Rela- 
tionen. In Fix et al. 1985: 287-307. 
Ritzke, Johannes 1985 Automatische Analyse des FranzOsischen: 
Komplexe Nominalstrukturen. In Figge, Udo L. 1985: 37-54. 
Stegentritt, Erwin 1978 MORPHO H B. Automatische derivationelle 
Analyse des FranzOsischen. Sprachwissenschaft - Computerlinguistik 
1. AQ, Dudweiler, W~ Germany. 
Stegentritt, Erwin 1982 LAISSER + Infinitiv - Konstruktionen in der 
automatischen Analyse des Franzdsischen. In Fix et al. 1985: 309- 
327. 
Stegentritt, Erwin 1983 Apergu d'une analyse automatique du 
frangais. In Cignoni, Laura and Peters, Carol, Eds., Linguistica 
Computazionale, Vol. IlL Supplement Actes du Vile Congr~s de 
I'ALLC, Pisa 1982: 243-250. 
Stegentritt, Erwin 1984 Automatische Subjektidentifizierung in 
franzdsisehen Infinitivs~itzen des Typs FAIRE + Infinitiv. In Figge, 
Udo L., Ed., Romanistik und Datenverarbeitung 1981. Akten der 10. 
Sektion des Deutschen Romanistentages, Regensberg 1981. Sprachwis- 
senschaft - Computerlinguistik 7. AQ, Dudweiler, W. Germany: 
22-38. 
Vauquois, Bernard 1975 La traduction automatique gt Grenoble. Dunod, 
Paris, France. 
Woods, William A. 1980 Cascaded ATN Grammars. AJCL 6(1): 
1-12. 
