V~TJ~CY AND MT: RECENT DEVELOPMENTS IN THE METAL SYSTEM 
Rudi Gebruers 
Siemens-METAL project 
Katholieke Universiteit Leuven 
Maria Theresiastraat 21 
B-3000 LEUVEN 
BELGIUM 
ABSTRACT 
This paper describes a valency model, 
developed within the Belgian METAL project, 
aimed at enhancing the modularity and 
multilinguality of the METAL system. The 
introduction provides background, section 1 
discusses the existing valency framework, and 
section 2 presents the alternative model. 
The final section deals with some results and 
problems with this model. 
0. Introduction 
The task of MT is to map between equivalent 
linguistic objects. One of the central design 
questions in MT is that of the best method to 
decompose the translation relation. The ideal 
would be to have a system that produces a 
(natural) language-independent representation 
from a source language (SL) text, which could 
then be synthesized in any target language 
(TL). However, this ideal not being feasible 
for real-world texts, it has become customary 
to adopt a model where a transfer module, 
specific to one language pair, defines a 
mapping between language-dependent structural 
representations. In principle it should be 
possible to design a 'transfer' model in such 
a way that the analysis module for mapping 
surface strings onto structural 
representations and the synthesis module for 
mapping structural representations onto 
surface strings remain the same, regardless of 
the TL and SL, respectively. The advantage of 
this 'multilingual' design is that existing 
modules will not be seriously affected by the 
addition of a new language to the system. A 
still more attractive, but also more 
ambitious, design would be one in which the 
same grammar can be used for both parsing and 
generating, and the same translation rules for 
translating between two languages in either 
direction (see Jin and Simmons, 1986 for an 
example of a 'symmetric' translation system). 
Whereas early MT systems blended the rules 
of grammar and the analysis procedure for 
efficiency reasons, it has also become 
customary, given current system optimization 
techniques, to make a clear separation between 
programming logic and data on the one hand, 
and linguistic logic and data on the other. 
This separation is convenient for the division 
of labour between the linguist and the 
programmer, and it enables the former to 
revise and complete his rule systems without 
the latter having to constantly change his 
programs. 
The METAL automatic translation system tries 
to be multilingual in the above sense. More- 
over, it makes an attempt at separating soft- 
ware and lingware (= linguistic knowledge 
written in a specialised user language). In 
the following, I will show how the adoption of 
a new kind of valency framework, developed at 
the K.U.Leuven in the course of the last two 
years, enhances the multilinguality and 
modularity even further. For the sake of 
clarity, I will first review the relevant 
aspects of the valency framework in the 
current METAL system. 
I. Valency in the METAL system 
Since the main claim to be advanced in this 
paper bears on the relation between a valency 
framework and the general design of an MT 
system, I will first say a few words on the 
METAL system architecture. I will then review 
and comment upon the valency framework in this 
system. 
i.I. The METAL architecturo 
In METAL the translation process proceeds in 
three phases. During the analysis phase an 
input sentence is mapped onto one or more 
interpretations. Each interpretation is 
represented as a flattened phrase structure, 
consisting of a predicate node followed by one 
or more arguments (and zero or more 
modifiers). Anaphoric links are resolved 
during the integration phase. The resulting 
analysis trees are not intended to be 
language-independent representations, but are 
passed to a bilingual transfer phase. During 
transfer, analysis trees are structurally and 
lexically modified according to TL 
specifications. The output sentence is the 
string of terminal nodes of this transformed 
tree. 
168 
The METAL system accommodates two kinds of 
transfer/generation approaches. Most transfer 
instructions are paired one-to-one with the 
grammar rules used to perform the SL analysis. 
However, provisions have also been made to 
complement this "direct transfer' approach 
with an independent transfer grammar (see 
Root, 1985). The latter approach is becoming 
more and more important in METAL because it 
greatly enhances the modularity of the system 
(viz. with an eye on using it for several 
different language pairs). 
1.2o The valency framework 
It is well-known that the dependency relations 
between a verb and its arguments can influence 
greatly the lexical and structural transfer of 
both, as well as the structural transfer of 
the clause as a whole. Though the dependency 
relations themselves may be language- 
independent, their encoding varies from one 
language to another, and, within one language, 
from one verb to another. It is therefore 
essential to know for each verb what its 
dependents should look like. This topic 
being central in Valency Grammar, it is not 
surprising that many MT systems (e.g. 
TAUM-AVIATION, SUSY, GETA, ARIANA-78, EUROTRA) 
incorporate valency notions (see Somers, 
1986). One essential notion borrowed from 
valency theory is that of 'valency frame', 
i.e. a pattern listing all the complements 
allowed and/or required by a verb, together 
with associated morpho-syntactic and/or 
syntactic features. 
Since German-English is the furthest 
developed language pair of METAL at present, I 
will now discuss what the valency frames look 
like for German. In the METAL system the 
German valency frames mainly include 
morpho-syntactic information (syntactic 
(sub) category and/or surface case) with 
respect to non-subject arguments. For 
instance, the pattern (A-X (CP TH)) signals 
that a German verb carrying it may take, 
besides a nominative subject, an accusative 
reflexive pronoun (A-X) plus a complement 
phrase introduced by dass (CP TH). The 
optionality of arguments is not signalled in 
the frame itself, but in a separate feature on 
the verbal predicate. 
In analysis, predicate-argument structures, 
resulting from a flattening and rearrangement 
of constituent structures, are passed to a 
case frame processor. The latter attempts to 
'use up' the available sentence constituents 
by matching them against the argument 
specifications in the valency frame(s) 
specified for the verbal predicate. When it 
finds a constituent that matches an argument 
specification in the frame, it updates this 
constituent with a grammatical role (SUBJECT, 
(IN)DIRECT OBJECT, etc.) according to some 
general implication relation holding between 
grammatical roles and case markings (or some 
other sort of coding). For a clause to be 
well-formed, at least one of its verb's frames 
must have a 'filler" for each of its argument 
'slots'. If a frame is found to be applicable 
in more than one way, preference is given to 
one application on the basis of word order 
criteria and/or semantic properties of the 
subject argument. Eventually, the case frame 
processor returns either an analysis tree that 
has been updated with grammatical role 
information, or it discards the input 
sentence. 
During the transfer phase, morpho-syntactic 
information of the sort present in valency 
frames may be used, both in tests and in 
transformations associated with lexical 
transfer entries, to attune SL argument 
specifications to the TL. In addition, 
transfer entries may contain further 
semantico-syntactic restrictions on argument 
positions, which may help in choosing the 
right translation for a verb. The grammatical 
roles are used to convert the canonical 
ordering of the translations of the verb and 
its arguments into the appropriate TL 
ordering, as well as to generate the 
appropriate forms of the TL constituents. 
1.3. Discussion 
Before discussing the new, extended valency 
framework, we will briefly point out how the 
existing system does not yet exploit the 
potential of a valency grammar to the full. 
The only valency frames referred to by the 
case framing package in the course of 
translation are SL valency frames. There is no 
procedure for mapping SL frames onto TL 
frames, and the information provided in TL 
frames is not used when TL strings are 
generated. The underlying assumption seems to 
be that argument structures are more or less 
the same across languages. Any discrepancies 
with respect to the argument structure are 
resolved by means of a small set of 
transformations specified in the relevant 
lexical transfer entries. Any discrepancies 
with respect to the expression of the argument 
structure (e.g case-marked vs. unmarked NPs) 
are handled in the relevant grammar rules. 
The assumption that argument structures are 
more or less the same across languages is also 
reflected in the status of the canonical 
clause representations employed in the 
METAL system. The latter are considered to 
be some sort of interlingual structures from 
which TL surface strings are to be generated 
directly (cf. the direct transfer approach). 
However, the general philosophy in the 
METAL system has been to start off with a 
rather 'shallow' level of analysis, rather 
than a 'deep representation' of some sort (see 
\[SLOCUM 83\]). Thus, there seems to be a 
conflict between the reluctance to work with a 
more semantically oriented analysis and the 
desire to have an interlingua. This conflict 
may have been negligible for the 
German-English system, because these languages 
are 'cut' along very similar patterns. 
Nevertheless, even these two languages display 
subtle differences as to the way they 'model' 
169 
extralinguistic reality. For instance, 
'helping' is a real-world relationship 
involving two entities, A and B. In English, 
this relationship is construed as an action of 
A which affects B (A helps B is similar to A 
hits B); in German, it is modelled as if A 
transferred something to B (A \[nom\] hilft B 
\[dat\] is similar to A \[nom\] gibt B \[dat\] C 
\[acc\]). Similar differences may be expected to 
increase as languages more divergent than 
English and German are to be handled. If, for 
some reason or other, it is not feasible or 
desirable to reduce language-specific models 
of some real-world relationship to a language- 
independent case frame, there is nothing but 
to state translation equivalences between 
clause structures in terms of equivalences 
between language-dependent argument 
structures. (For similar views, see Alam, 
1986; Kudo and Nomura, 1986; Van der Korst, 
1987.) 
Although there is little doubt that the 
framing facilities provided in the system work 
quite well and yield very good results for 
translations from German into English, we have 
tried to improve the framing module beyond 
this language-pair. One should also bear in 
mind that, with a less well-structured MT 
system than METAL, we could never have 
developed a more language-independent valency 
mechanism in such an easy and straightforward 
way. 
2+ An alternative valency fzamawozk 
2.1. The architectuze 
The general philosophy behind the development 
of the Leuven valency framework has been to 
maintain an essentially syntax-driven MT 
system, while enhancing the latter's 
modularity in view of extensions to other 
language pairs. This required reconsidering 
not only the relation between lingware and 
software, but also the general architecture 
behind the system. 
With respect to the general translation theory 
behind the METAL system, enhancing the 
modularity boils down to increasing the 
relative independence of the analysis, 
transfer, and synthesis modules. More 
specifically, we assume that 
(a) an analysis module must provide 
representations which are useful 
starting-points for translation into 
multiple TLs; 
(b) major parts of a synthesis module must be 
independent of the SL under consideration 
so that they can also be used for 
translation from other SLs; 
(c) mappings between SL and TL representations 
must be defined in terms of a minimum of, 
preferably, lexically governed transfor- 
mations. 
Though we are still far away from the ideal 
transfer-based MT system, we believe that the 
alternative valency framework may be an 
important step in the right direction. 
2.2. The linguistic fundamentals of the 
alternative valency framework 
The basic assumption is that simple clauses 
have a predicational structure and that 
(partial) equivalences between SL and TL 
clauses can be defined in terms of (partial) 
equivalences between SL and TL predications. 
The structural centre of a predication is a 
lexical predicate with which one or more 
valency frames are associated. Each valency 
frame is a sequence of typed argument slots to 
be filled with appropriate terms, i.e. 
sentence constituents of the appropriate 
types. Sentence constituents which cannot be 
related to any of a predicate's argument slots 
should be 'legal satellites', i.e. legal 
circumstantial modifiers, of the predication 
as a whole. 
The structure of valency frames is language- 
independent, and can be defined as follows: 
<frame> : :- " (" <slot>+ \[ "OPT" <slot>+\] ") = 
<slot> ::-- "(" <slot_label> <key>+ ")" 
<slot label> ::- one of a set of user-definable atoms, 
startinq with a "S'-sign 
<key> ::- <codepointer> I <feature-value-pair> 
<code_/3ointer> ::- one of a set of user-definable atoms 
<feature-value-pair> ::_ .(m <featname> <feat_val>+ ")" 
<feat_name> ::- one of a set of user-definable atoms 
<feat val> : :- one of a set of user-definable atoms 
The number of argument slots for a given 
frame is primarily determined on the basis of 
formal, language-specific criteria. Thus 
tests to distinguish behween arguments (args) 
and satellites (sats) include, besides the 
elimination test, paraphrase tests (cf. the do 
so and und zwar tests, for English and German, 
respectively), as well as distributional and 
substitutional criteria (e.g. sats are freer 
to move than args, whereas elements of 
pronominal paradigms substitute more easily 
for args than for sats). Whenever those tests 
are not decisive with respect to the status of 
a sentence constituent, the latter is assumed 
to be an arg, since, for transfer, it is 
arguable that it is easier to operate on args 
than on sats. 
Argument slots may be obligatory or 
optional. Optional slots, which need not be 
present, but are always semantically implied, 
are set apart from obligatory ones by means of 
the symbol OPT. In fact, OPT is a means to 
collapse frames whose obligatory slots are 
identical and whose optional slots are not 
mutually exclusive. Thus, a frame containing 
n optional slots is an abbreviation for 2expn 
different frames. 
Argument slots are not in themselves 
labelled semantically, though they do tie up 
with semantic relations (deep cases) as all 
valency relations are ultimately semantically 
motivated. (See Helbig and Schenkel (1973) for 
a discussion of the relation between logical, 
semantic, and syntactic valency.) Instead, a 
slot label is taken to signal that certain 
rules or regularities apply to all the args 
carrying that label. Our basic principles for 
labelling slots are the following: 
170 
(a) Args labelled $0 are 'deep subjects'; 
typical surface reflexes in Dutch and 
French are "nominative case', position to 
the left of the finite verb in unmarked 
declaratives, ability to become the 
aqentive phrase in passives (under certain 
conditions); 
(b) Args labelled $I are 'deep objects'; 
typical surface reflexes in Dutch and 
French are "accusative case', position 
strongly tied to the main verb, ability to 
become the 'surface subject' in passives 
(under certain conditions); 
(c) Args labelled $2 are 'indirect objects'; 
typical surface reflexes in Dutch and 
French are indirect object prepositions 
(aan vs° a), alternation between PP\[aan 
vs. a\] and NP\[non-clitical and lexical vs. 
clitical and pronominal\]; 
(d) Args labelled $3 are 'oblique objects'; a 
typical surface reflex in Dutch and French 
is that these args can be replaced by 
adverbial constructions; 
(e) Args labelled $4 are "prepositional 
objects'; in both Dutch and French these 
args are PPs, with strongly governed, 
idiosyncratic prepositions; 
(f) Arqs labelled $5 are "subjective 
complements'; these arqs are attributes of 
the subject with bivalent verbs (e.g° 
zijn/etre 'be'); 
(g) Args labelled $6 are "objective 
complements'; these args are attributes of 
the direct object with trivalent verbs 
(e.g. noemen/appeler 'call'). 
It is important to note that those principles 
are rules of thumb, rather than clear-cut 
definitions in terms of necessary and 
sufficient conditions. However, far from being 
arbitrary, the inventory of arg labels should 
be justified both language-internally and 
cross-linguistically. Language-internally, 
this means that one has to come up with a 
number of indications that a given arg label 
allows for significant language-specific 
generalizations. Cross-linguistically, this 
means that, in assigning the same label to 
slots in different languages, it must be 
possible to reveal a reasonable degree of 
overlap in the behaviour of fillers for the 
slots in the respective languages. Of course, 
we do not pretend that our list of arg labels 
is in any way exhaustive and we grant that it 
may have to be adjusted in the light of 
further research. 
Apart from a slot label, an argument slot 
contains a number of 'keys' which refer to 
procedures, frame tests and frame 
constructors, to be called during analysis and 
synthesis, respectively. 
Frame tests consist of morpho-syntactic and, 
possibly, semantic conditions which 
constituents must satisfy in order to become 
potential slot fillers. The actual contents of 
the morpho-syntactic constraints is 
co-determined by such parameters as the slot 
label and the clause's Mood and Voice values. 
Regarding the use of semantic selection 
restrictions, a fairly pragmatic course has 
been pursued. That is, we started off with 
a rather limited inventory of semantic 
features (\[~ person(alized)\], \[~ abstract\], 
etc.) which we think to be consistently 
applicable and flexible enough to be extended 
and/or changed when the need arises. Apart 
from semantic selection restrictions on filler 
constituents, it is possible to include 
(canonical) lexical forms in a slot 
specification list. These may refer to the 
form required in either the 'relator' (e.g. 
the preposition in a PP or the conjunction in 
a subordinate clause) or the 'head' of filler 
constituents. The latter functionality has 
been provided in order to handle more or less 
idiomatic NPs (e.g. een keuze in een keuze 
doen, "make a choice') which are still 
sensitive to regular syntactic operations 
(e.g. passivization). 
Frame constructors consist of instructions 
according to which the system should generate 
the appropriate surface form required for 
fillers of the slots from which those 
constructors are called. Again, the actual 
content of the instructions is co-determined 
by such parameters as the slot label and the 
clause's Mood and Voice values, as well as by 
lexical information provided in the slot. 
It is important to note two things. First, 
it is actually the same codes that are used 
for both tests (Analysis) and constructors 
(Synthesis). The system knows from the 
translation phase whether it has to interpret 
the code as a test or as a constructor. 
Second, both frame tests and frame 
constructors are stored in separate files, in 
order to enhance the modularity of the 
lingware. However, they are written in the 
same format as ordinary grammar rules, and a 
special interface for loading, inspecting and 
editing these procedures has been developed, 
so that they can be easily accessed and 
updated by the linguist. 
As an illustration of the above, let us 
consider the following (incomplete) list of 
valency frames which occur as elements in a 
value list to the feature ARGS for the verb 
faire: 
( 
(($0 N1 PI) ($I N1 P0) 
OPT ($2 N1 P1 (PREP pour))) :'make" 
(($0 NI)($5 A)) :'look" 
(($0 NI)($3 MEA)) :'do" 
) 
What is said here is that faire can show up in 
(at least) three different frames. The first 
one contains two necessary and one optional 
slot. The first slot requires a nominal filler 
with selection restriction \[+personal(ized)\], 
the second one a nominal filler with selection 
restriction \[-personal\], and the optional one 
a prepositional complement introduced by pour 
and having selection restriction 
\[+personal(ized)\]. The second frame contains 
two obligatory slots, the first of which 
requiring a nominal filler and the second one 
an adjectival complement. Finally, the third 
frame requires a nominal filler and a measure 
constituent. The respective frames are 
illustrated in the following sentences: 
171 
(1) Je ($0) fais ce Jouet ($I) pour mon ami ($2) 
"I make this toy for my friend" 
(2) Ella ($0) fair vieille ($5) 
"She looks old" 
(3) Carte volture ($0} fair 100 km/h ($3) 
• This oar does i00 km/h, 
An example of a (French) frame test is given 
in fig. i. It is invoked by the key PI, when 
called for the slot labelled $2 (as in the 
first frame of faire). Comments explaining the 
test instructions are given in italics. Note 
how different contextual restrictions have 
been assembled in one test procedure. 
In fig. 2, we give an example of one 
(French) frame constructor which is called by 
the key N1 from a slot labelled $2. 
PI-$2 
(SONS ($2SON - (INT ? $2 N1)) 
single OUt nodes which pa~sed the $2-NI test 
DO 
for each of these nodes, 
(COND ((INT $2SON TY HUM HI) 
if it has semantic type \[+human\] 
or \[+human intervention\], 
then succeed 
(PUT ($2 PI))) 
((INT $2SON RP REL) 
if it contains a relative pronoun, 
then succeed 
(PUT ($2 PI) (TR HUM HI))) 
((RET $2SON KP) 
if it confains a pronoun, 
then succeed 
(PUT ($2 PI))) 
(T 
else, unmmrk the node as a candidate 
for the $2 slot 
(RMV $2SON $2)))) 
fig.l: frame test 
NI-$2 
(AND (RET 0 PREP) 
if there is a PREP feature on the father node 
(this feature has been retrieved fr~ a 
verbal entry) 
(OR (XFM (&:l (--:2 (NP:3 NIL (INT 3 ROL $2)) ---:4)) 
and if thera is a NP son marked $2, 
then create a TL-PREP node in front of 
it and make both dependent on a 
new PP node marked $2 
(&:l (--:2 
(PP:5 ((PREP:6 NIL (TRF 1 PREP)) 
(NP:3 NIL (RMV ROL))) 
(PUT (ROL $2))) 
--:4))) 
(XFM (&:l (--:2 
(PP:3 (pREP:4 &:5) (INT 3 ROL $2)) 
--:6)) 
or if there is a PP node marked $2, 
then translate its PREP node 
(&:l (--:2 
(PP:3 ((PREP:7 NIL (TRF i PREP)) &:5)) 
--:6))))) 
fig. 2 : frame constructor 
2.3. The valency procedure 
The valency procedure is composed of three 
subprocedures (see fig. 3). Two of them use 
purely language-specific material, while the 
third one has to establish a link between 
material from two different languages. What 
is important to know, however, is that we hold 
the overall organization of all three 
subprocedures to be completely language- 
independent° As a consequence, it should be 
easy to plug language and language-pair 
specific information into these procedures, 
without the latter having to be adapted for 
each new language pair. We think this 
modularity is a substantial improvement as 
compared with the valency procedure 
incorporated in the LRC-METAL system. 
canonical clause structure 
CORE SOFTWARE 
frame matching 
frame selection 
tree updating 
P 
m IAm 
LINGWARE 
SL lexical entries 
frame tests 
canonical clause structure 
with SL role distribution and identification 
of the matching SL frame 
.... 
CORE SOFTWARE ~ LINGWARE 
frame mapping 
F transfer entries 
slot mapping 
canonical clause structure 
with TL role distribution and indication of which 
TL frame corresponds to the SL frame 
TL lexical entries 
filler adaptation FI 
frame constructors 
canonical clause structure 
with TL role distributic,~ and 
updated valency-bound sentence constituents 
fig. 3 : valency procedure 
(FRA, FRX, FRG are the drive functions called from 
within grammar rules during analysis, transfer, and 
synthesis, respectively) 
2.3.1. During analysis, the valency 
procedure is invoked from within grammar rules 
for building clausal structures of the 
following format: 
<clausal category> 
ARGI PRED ARG2 .'.. ARGn 
<ARGS> 
Given a verb with a set of valency frames and 
a set of sentence constituents, this procedure 
has to make sure that one of these frames is 
realized in the sentence at issue, and, if so, 
171 
it has to make clear how that frame is 
realized. It will take a frame to be realized 
in a tree, if and only if each of the frame's 
non-optional slots, and possibly, one or more 
of its optional slots, is matched by at least 
one sentence constituent. If the system 
finds a matching frame, it will ultimately 
return a tree structure in which each of the 
valency-bound constituents is marked for a 
slot and whose root node contains a reference 
to the matching frame. 
Now, the general idea is to have the 
procedure look for the most ambitious frame 
matching the tree structure, as well as for 
the best realization of this frame in the 
structure, while avoiding superfluous 
processing as much as possible. This implies 
that two kinds of preference mechanisms had to 
be introduced in the valency procedure: one 
to choose the best candidate from a set of 
potential fillers for a given slot (instead of 
always choosing the first constituent matching 
the specifications of that slot), and one to 
choose the most ambitious frame from a set of 
successful frames (instead of always choosing 
the first frame that matches a given analysis 
tree). 
The first preference mechanism is 
implemented in the following way. When 
checking whether a frame matches the tree, the 
valency procedure collects for each slot ($i 
keyl...keyN) all sentence constituents which 
pass all the frame tests associated with the 
keys in that slot. Furthermore, in the action 
part of frame tests, each potential filler 
gets a number indicating the probability that 
this constituent will be taken as the ultimate 
filler for a slot. As the linguist can easily 
alter this number, he has significant control 
over the assignment of constituents to slots. 
The actual assignment procedure is fairly 
economical and runs as follows. Whenever 
there is only one potential filler for a slot 
(which may be either an obligatory or an 
optional one), this constituent loses its 
marking as a candidate filler for other slots, 
if it was one. Furthermore, it is marked as 
the only remaining filler for the slot that is 
being matched. A side-effect of this 
marking may be that one of the other slots 
will now have only one candidate filler. In 
that case, the latter constituent will be 
marked as the actual filler for that slot and 
lose its marking as a potential filler for 
still other slots, if it was one. This may 
again cause the number of potential fillers 
for yet another slot to be reduced to one, in 
which case the above marking (and unmarking) 
procedure starts over again. If, eventually, 
there is more than one potential filler for a 
slot, the procedure takes the leftmost 
candidate which received the highest prefer- 
ence value for the slot in the frame tests. 
The second preference mechanism is fairly 
economical as well. The general assumption is 
that, in order to find the most 'ambitious' 
frame, one should look for the frame which has 
the largest number of slots realized. In 
order to avoid superfluous processing, we let 
the more complex frames (i.e. the ones with 
the larger number of slots, whether obligatory 
or optional) precede the less complex ones in 
the ARGS value of a verb. Consequently, the 
system comes across the former before it sees 
the latter. Whenever two or more alternative 
frames happen to have the same number of 
slots, the lexicographer has to determine 
(e.g., on the basis of frequency of 
occurrence) which frame to try first. Given 
that the system has found a matching frame, it 
will only explore an alternative frame if the 
number of (optional and non-optional) slots 
contained in the alternative frame outnumbers 
the number of slots found to be realized 
during the matching of the first frame. An 
alternative frame will be preferred only when 
it has more slots realized than the previous 
matching frame. 
2.3.2. During transfer, since we take 
argument structures to be language-specific 
entities, the valency procedure has to 
accomplish two tasks. First, it has to 
determine which TL frame corresponds with the 
SL frame that has been found to be applicable 
to the analysis tree. Secondly, it has to 
specify which slots in the TL frame correspond 
with which slots in the SL frame. It performs 
those tasks in the following way. 
First, the mechanism looks up all the 
transfers for the SL verbal predicate. Each 
of the verbal transfer entries describes a 
transition between the SL verb with one of its 
frames, and an equivalent TL verb with one of 
its frames. It does so in terms of (a) condi- 
tions on the transition and (b) (possibly 
partial) mappings between SL slots and TL 
slots. The condition part of a verbal transfer 
entry may be empty or take any of the follow- 
ing forms, for disambiguation w.r.t, the TL: 
(a) a test on the presence (absence) of some 
slot filler in (from) the SL tree; 
(b) a test on the presence (absence) of 
certain lexical or grammatical information 
on some slot filler in (from) the SL tree; 
(c) a test on the presence (absence) of some 
feature on (from) the root node of the SL 
tree. 
As for the mappings between SL slots and TL 
slots, three possibilities have been catered 
for so far: 
(a) SL slot maps onto TL slot. In this case, 
we only state contrastive information 
which is strictly necessary to effect an 
appropriate mapping between SL slots and 
TL slots. That is, in General, 
equivalences between distinct slot labels 
will suffice, though additional 
information may be provided for 
disambiguation. 
(b) SL slot without TL counterpart. Again, 
only minimal information needs to be 
specified in order to identify the SL slot 
whose filler must be removed from the tree 
structure. 
(c) TL slot without SL counterpart. Here, the 
coder should be able to describe the 
internal structure of a TL constituent to 
be created at clause level in the tree 
structure. At the moment, the 
functionality provided is limited to the 
creation of new TL nodes without internal 
structure. 
173 
Having retrieved all the transfers for the 
SL verb, the system reduces the potential 
transfer ambiguity of a SL verb in two steps. 
It first discards any transfers for which not 
all conditions are fulfilled. Afterwards, it 
checks which of the remaining transfer entries 
(there should be at least one) provides a 
frame equivalence whose 'left-hand side" can 
be linked to the frame realized in the SL 
tree. Once this equivalence has been found, 
the TL frame matched by the 'right-hand side' 
of the frame equivalence will be substituted 
for the SL frame referenced by a feature on 
the root node of the tree. At this stage, 
nodes can be pruned from or added to the tree 
structure, if lexical instructions tell the 
system to do so. 
Finally, after translation of the verbal 
predicate, the system exploits the 
equivalences between SL slots and TL slots in 
order to determine how the translations of the 
sentence-level constituents fit into the slots 
of the TL frame. 
The above matching procedure complicates the 
lexicographer's task considerably (see section 
3. for how we try to remedy this situation). 
This is because it requires that verbal 
transfer entries be written in such a way that 
each of them provides 
(a) a link with a SL verb and exactly one of 
its frames; 
(b) a link with a TL verb and exactly one of 
its frames; 
(c) sufficient information concerning the slot 
equivalences holding between these two 
frames. 
On the other hand, it has the advantage of 
giving the dictionary writer significant 
control over verbal and clausal transfer. 
2.3.3. The task of the valency procedure 
during synthesis, as we conceived of it, 
consists in guiding the generation of the 
appropriate surface form of valency-bound 
sentence constituents. Furthermore, the 
synthesis component contains constituent 
ordering procedures which may refer to slot 
labels in order to rearrange the canonical 
clause structure into the appropriate TL 
ordering. 
First, the valency procedure retrieves the 
TL-frame from the root of the tree. For each 
slot contained in this frame, it then checks 
whether it contains any lexical information 
specific to the TL verb and the frame that are 
at issue (e.g., the third (optional) slot in 
the first frame of faire has to be filled by a 
prepositional phrase introduced by pour) and 
makes this information available for further 
processing. Next, it calls all of the frame 
constructors associated with the keys in the 
slot. During those calls the relevant 
sentence constituents will be modified and 
updated according to the instructions that the 
linguist specified in the constructors. 
Eventually, the valency procedure should 
return a tree all of whose valency-bound 
constituents contain sufficient information so 
that morphological rules and linearization 
rules can generate the appropriate TL forms 
and constituent ordering (the latter rules may 
occasionally alter the constituents' forms). 
Again, the linguist has significant control 
over the generation process as he can easily 
specify and update the contents of both frame 
constructors and linearization rules. 
3. Rem.zlte ~"*d Probl~ 
In the last year, the alternative valency 
framework presented in this paper has been 
applied to the translation of Dutch into 
French, and vice versa. Though the application 
has been limited to "kernel" sentences, i.e. 
simple active and passive declarative 
sentences (possibly containing relative 
clauses of the same type), the results seem 
fairly promising. At the moment, provisions 
are being made to handle non-finite 
valency-bound subclauses. 
In the meantime, small conversion 
experiments have been conducted for the 
language pairs German-English (at K.U.Leuven) 
and German-Spanish (at CDS Barcelona). These 
experiments have shown that, at least for the 
admittedly very limited domain of application, 
the Leuven valency framework works very well. 
Its main advantages seem to be the follow- 
ing: 
(a) it provides the skeleton for a really 
transfer-based MT system, since it clearly 
separates three subprocedures to be 
invoked during Analysis, Transfer, and 
Synthesis, respectively; 
(b) it allows for a neat separation of kernel 
software and application-specific lingware 
and provides user-friendly facilities to 
access and update the latter; 
(c) its methodological underpinning to a 
certain extent allows languages to be 
treated independently of one another. 
Because of these advantages, the Leuven 
valency framework has recently been adopted by 
all sites of the METAL project. 
However, serious problems remain to be 
tackled, with respect to both lexicon coding 
and grammatical parsing. The first kind of 
problems can be traced to the rigidity of the 
frame mapping schema itself. As has been 
pointed out in section 2.3.2., the main 
requirement for frame mapping to be possible 
is that verb lexicons be coded consistently 
across languages. This may indeed be 
profitable in an experimental environment. 
However, it is doubtful whether we can expect 
the average end user to have both source, 
target, and transfer codings in mind at the 
same time and to make sure that no aspect of 
the mapping between frames is overlooked. 
Therefore, we have developed provisional 
relaxations on the rigid schema to the extent 
that at least the mandatory slots of the TL 
frame must have a counterpart in the SL frame. 
Eventually, however, it may turn out to be 
more effective not to have these relaxations 
at run time, but to sort out inconsistencies 
between verbal lexicons at coding time. 
174 
The second kind of problems (among which the 
notorious difficult problem of PP-attachment) 
has to do with a need for still greater 
functionality in the system. In order to 
provide for this functionality, we envisage 
two paths of further research. One path 
concerns how we can make our valency mechanism 
interact with a mechanism to identify 
peripheral constituents, which orbit around 
the verb and its arguments. The other path 
concerns the extensibility of the valency 
framework to nouns and adjectives. 
Preliminary research along both lines has 
revealed that 
there are no objections of principle against 
the valency framework presented in this paper. 
ACKNOWLEDGMENTS 
Thanks are due to all the members of the 
Leuven METAL-project who helped develop the 
valency framework presented in this paper. 
The author would also like to thank Geert 
Adriaens en Herman Caeyers for their helpful 
comments on earlier drafts of this paper. 

REFERENCES 

Alam, Y.S. 1986. 'A lexical-functional 
approach to Japanese for the purpose of 
machine translation.' Computers and 
Translation 1(4): 199-214. 

Bennett, W.S. & J. Slocum 1985. 'The LRC 
machine translation system.' Computational 
Linguistics ii (2-3): 112-121. 

Helbig, G. & W. Schenkel 1973. Woerterbuch 
zur Valenz und Distribution deutscher Verben. 
Leipzig: VEB Verlag Enzyklopaedie. 

Jin, W. & R.F. Simmons 1986. 'Symmetric 
rules for translation of English and Chinese.' 
Com~uters and Translation 1(3): 153-167. 

Kudo, I. & H. Nomura 1986. 'Lexical-functional transfer: a transfer framework in a 
machine translation system based on LFG.' 
Proceedings COLING 86, 112-114. 

Root, R. 1985. 'A two-way approach to 
structural transfer in MT.' Proceedings of 
the 2nd ACL 1985, 70-72. 

Slocum, J. 1983. 'A status report on the 
LRC machine translation system.' Proceedings 
of the ACL-NRL Conference on Applied Natural 
Language Processing, 166-173. 

Somers, H.L. 1986. Valency 
Computational Linguistics. 
Edinburgh University Press. 
and Case in 
Edinburgh: 

Van der Korst, B. 1987. 'Twelve sentences: 
a translation procedure in terms of Functional 
Grammar." Working papers in Functional 
Grammar 19, University of Amsterdam. 
