INDEPENDENT TRANSFER USING GRAPH UNIFICATION 
Lauri Carlson 
Maria Vilkuna 
Department of General Linguistics 
University of Helsinki 
Hallituskatu 11 
00100 Helsinki, Finland 
Icarlson@finuh.bitnet, viikuna@finuh.bitnet 
Abstract 
We present a MT system that applies graph unification 
in transfer from English to Finnish. The work described 
below is an outgrowth of a multilingual MT project 
initiated by the IBM in 1987 with the aim of studying 
multilingual translation using a common English lan- 
guage parser. 
The transfer system presented here is independent of the 
parsing and generation modules. Any source language 
parser can be used whose output can be expressed in a 
directed graph form. The transfer system is responsible 
for generating target language phrase structure. Target 
language word order and morphology are left to the 
generation modules. 
The transfer system is lexically based. Transfer rules, 
presented in the form of bilingual graphs, are declarative 
statements of symmetric transfer relationships between 
words, phrases or constructions in the two intertranslat- 
able languages. 
Transfer is structure driven in that the transfer algorithm 
traverses the source language graph, nondeterministically 
trying to apply the relevant transfer rules in the lexicon. 
Each successful transfer yields a bilingual graph, whose 
target hmguage half is extracted and subjected to lineari- 
zation and morphological generation. 
The main f(ycus of attention in our project is the devel- 
opment of the lexicon subsystem. The lexicon system 
consisk~ of separate transfer and monolingual lexicons 
and a common lexicon of language independent defini- 
tions. 
Keywords: unification, machine translation, transfer, 
bilingual lexicon 
I. Unification based transfer 
Our approach is more transfer oriented than some other 
unification based approaches to MT (e.g., Beaven and 
Whitelock 1988). However, we argue, use of graph uni- 
fication blurs the distinction between transfer and inter- 
lingua. 
A feature structure representing a phrase will contain 
information at several levels of linguistic analysis ranging 
from lexical identity to logical argument structure. Trans- 
fer rules can express bilingual correspondences at any 
level of abstraction as well as across different levels of 
structure. (Cf. Kaplan & al. 1989.) A transfer rule in our 
sense can consist of an arbitrary pairing of lexical entries, 
a complex correspondence across structures (e.g., 
"change" of grmnmatical construction including part of 
speech assignments), or a straightforward identification 
of arguments in logical form. 
When the translation relation is best stated in language 
independent (semantic) terms, transfer is trivial. Then 
monolingual lexicons, analysis and generation modules 
will do most of the work. Thus, to what extent a given rule 
has the character of a genuine transfer rule will depend 
on the degree of similarity of the languages under trans- 
lation in the relevant respect. For instance, languages with 
similar tense systems can allow a straightforward identi- 
fication of low level tense distinctions. Low level transfer 
simplifies the tasks of analysis and generation and allows 
tighter control of the translation relation. In particular, 
transfer idioms (multiword equivalences) can be stated 
directly without a detour through more abstract repre- 
sentations. In this sense, unification based transfer fills 
out the space separating interlingua and transfer. 
2. Parsing 
Unlike approaches such as Kaplan & al (1989), which 
produce bilingual descriptions in the course of parsing 
source language text, transfer in our system has a com- 
pleted parse as a starting point. Currently, this parse is 
produced by a general-purpose parser, PEG of IBM 
T.J.Watson Research Center (Jensen 1986), which is not 
unification-based. However, its output is close enough to 
a directed graph to 'allow conversion into the form re- 
quired by the transfer system using a simple conversion 
interface. 
It appears to us that this decoupling of parsing from 
transfer is a safe move. Knowledge of the target language 
is not likely to influence ~)arsing of the source language 
in any significant fashion . 
60 1 
3. lhe transfer system 
Our Iransfer system consists of two modules. A decla- 
rative module defines translation correst)ondenees of in- 
dividual phrases, structures and features. The informa- 
tion is given in bilingual (or multilingual) transfer diction- 
aries. 
An algorithmic modtde actually builds the correspond- 
ence structure out of the source language f-structure and 
the transfer dictionaries. This component ensures that all 
necessary alternatives are considered and the relevant 
information is incorporated into a correct location in the 
correspondence structure. 
We discuss these two modules in turn. 
3.1. The transfer lexicon 
A leading idea of the lexicon system is the separation of 
four different lexicons as follows: 
DGLEX 
c c T: X ~E r 
PFLEX 
DGLEX is a lexicon of general linguistic definitions of 
terms. There are two monolingtml lexicons, ELEX and 
FLEX, and a bilingual u'ansfer lexicon, TFLEX. The 
monolingual lexicons depend on DGLEX, and TFLEX 
can refer to the other three. No further dependencies are 
allowed. This increases the independence between tile 
component lexicons and makes them reusable for multi- 
lingual translation. 
The descriptions in both monolingual lexicons are kept 
independent of one gmother and linguistically motivated. 
Complex and ad hoc statements belong in TFLEX; it 
cannot be expected that all bilingual intertranslatability 
relations should follow linguistic generalizations. Corre- 
spondingly, we may distinguish two kinds of multi-word 
expressions. Language-internal idioms (e.g., keep tabs in 
English) are given in the monolingual lexicons, whereas 
the other type, which might be called "transfer idioms", 
are referred m at the level of tnmsfer entries only (e.g., 
have access to, which translates into one Finnish verb). 
3.2. The specification language 
The linguistic description language has two levels, an 
internal representation in terms of attribute value graphs, 
and a delinition language consisting of templates abbre- 
viating such graphs. As examples of the latter, conskler 
the simple entries below. 
(i) (d~scuss v slmpleobj-e) 
(2) (keskustella v sJmpJeob\]-ela) 
(3) (d\]scuss (e (@ e::discnss)) 
(f (@ f: :keskusteila) ) 
tra) 
\[E:\[I,EX:Big 
CAT:VERB 
SUBJ:#3\[E: \[LEX:I\[T 
CAT:PRON 
SEM:#2\[ANIM:F 
HUM:f'\]\] 
E:\[LEX:SE 
CAT:PRON 
CASE : EI,A 
SEM:#2\]\] 
VCOMP:#4\[E: \[LEX:DISCUSS 
CAT:VERB 
SUBJ:#3 
PREI):\[ARGI: 
#5\[E:\[LEX:*NONE*\] 
V:\[LEX:*NONE* 
SEM:\[HUM:T\]\]\] 
ARG2:#3 
ARG3:*NONE *\] 
VFORM:PASTPART 
VOICE:PASS\] 
F:#I0\[LEX:KESKUSTELLA 
CAT:VERB 
THEMA:#3 
SUBJ : # 5 
OBL:#3 
PILED : \[ARGI : #5 
ARG2:#3 
ARG3:*NONE*\] 
VFORM:FINITE 
VOICE:PASS\]\] 
PRED:\[ARGI:#4 
ARG2:*NONE* 
ARG3:*NONE*\] 
VFORM:FIN!TE 
VOICE:PASS\] 
F:#!0\] 
Fig. 1: Simplified TFS of "it was discussed" (next page) 
The entries are from ELEX, FLEX, and TFLEX, respec- 
tively, and together they specify the transfer relation 
between English discus,; ~d its Finnish equivalent kes- 
kustella. (The transfer entry is shown expanded into graph 
form in fig. 4.) 
The graph formalism we use is a standard attribute value 
unification formalism except for the use of cyclic graphs. 
The graph specification language extends the template 
language used in D-PATR in the following respects: 
• Compile-time disjunction is included 
• Parametric templates are included 
3.3. Transfer feature structures (TFS) 
The transfer relation between source and target lan- 
guage feature structures could be represented in different 
ways. Separate feature structures could be set up for the 
source language and the target language, and all explicit 
transfer relation between these two structures could be 
defined (Kaplan & al. 1989). in our system, there is only 
one larger transfer feature structure (TFS) which includes 
both feature structures and specifies the explicit transfer 
relation for intertranslatable phrases of source and target 
languages. 
The TFS contains extra levels ofa|tributes for the soume 
and target lar~guagc. Intertranslalable phrases form sub- 
descriptions which have two altributes, one for each 
language. The values of these attributes are always trims- 
2 61 
\[F:#10\[LEX:KESKUSTELLA 
CAT:VERB 
THEMA:#3\[F:\[LEX:SE 
CAT:PRON 
CASE:ELA 
SEM:#2\[ANIM:F 
HUM:F\]\]\] 
SUBJ:#5\[F:\[LEX:*NONE* 
SEMi\[HUM:T\]\]\] 
OBL:#3 
PRED: \[ARGI:#5 
ARG2:#3 
ARG3:*NONE*\] 
VFORM:FINITE 
VOICE:PASS\]\] 
Fig. 2: Simplified Finnish FS of 'It was discussed' 
lafions of each other, and they may share values of com- 
mon features and especially component phrases which, in 
turn, are translations of each other. 
An example of a ~anslation relation expressed in one 
feature structure is given in fig. 1. This structure contains 
the feature descriptions of both the English and Finnish 
sentences and coreferential links that bind the corre- 
sponding units together. 
Monolingual feature representations can be read off the 
bilingual one by omitting all attribt, te-value pairs where 
\[E:\[TENSE:#I\] 
F:\[TENSE:#1\]\] 
Fig. 3: Simple tense transfer rule 
\[E:\[LEX:DISCUSS 
CAT:VERB 
SUBJ:#2\[E:\[DUMMY:F\]\] 
OBJ:#3\[F:\[CASE:ELA\]\] 
PRED: \[ARGI:#2 
ARG2:#3 
ARG3:*NONE*\] 
F:\[LEX:KESKUSTELLA 
CA'f:VE\[<S 
SUBJ:#2 
OBL:#3 
PRED: \[ARGI:#2 
ARG2:#3 
ARG3:*NONE*\]\]\] 
Fig. 4: Partial transfer rule for "discuss" 
\[E:\[LEX:BE 
SUBJ:#2 
VCOMP:#5\[E:\[SUBJ:#2 
BY-PASS:F 
VFORM:PASTPART 
VOICE:PASS\] 
F:#1\[THEMA:#2 
SUBJ: \[F:\[LEX:*NONE* 
SEM:\[HUM:T\]\]\] 
VOICE:PASS 
NOMOBJ:T\] \] 
Fig. 5: Simplified transfer rule for agentless passive 
tile attribute is the name of tile other language. The 
Finnish language subgraph of the previous example is 
given in fig. 2. 
3.4. Transfer rules 
A transfer rule in this approach is formally just another 
transfer feature structure, similar to the bilingual struc- 
ture. It is a partial specification of an acceptable inter- 
translatability relation. The rule is applied to a TFS by 
unifying it with a specified node in the "ITS. The transfer 
process consists simply of adding of further information 
into a partially described instance of the transfer relation. 
There is no formal distinction between lexical and gram- 
matical transfer rules. Examples of different types of rule 
are given in figures 3-5. 
Some aspects of our linguistic description will be briefly 
described. In monolingual lexicons, shills in grammatical 
function like the English active and passive are described 
as different Iinkings of arguments to grammatical func- 
tions, in this case, the subject and the object function. 
In transfer of complement-taking elements, we can then 
for the most part rely on the simple rule "equate argu- 
ments", which resulLs in correct bilingual corresponden- 
ces given the language-particular linkings. For example, 
the verb disc~s (fig. 4) rakes as its second argument a 
direct object in English but an oblique complement in 
Finnish, but this language-particular informatkm need not 
be recapitulated in the transfer entry. 
There are also translation equivalents whose arguments 
do not match, and these receive slightly more complex 
transfer rules where argument equations are expressed 
separately. 
Graph unification descriptions are particularly simple 
and effective where the relevant structures consist of 
predicates u~king a restricted number of unique argument 
types, such as subject, object, or sentential complement. 
Adjuncts, which may have multiple instantiations for 
each head, need a different treatment. Each of the adjuncts 
has a unique modifiend (modif = the modified word), 
#1\[E: 'T ..... o - t ~z,X. mXAM~ LE 
CAT : NOUN 
ADJT:#2\[E: \[LEX:ADDITIONAL 
CAT : ADJ 
PRED: \[ARGI:#1 
ARG2 : *NONE': 
ARG3 : *NONE*\] 
ADJT : 
\[E: \[CAT :ADV 
MODIF:#2\]\] 
MODIF:#I\] 
F: \[LEX:LISA 
CAT : NOUN 
ADJT : \[F : *NONE*\] 
MOD IF : # i 
NUM:SG\] \] 
NUM:PL 
PERS : 3 
F: \[LEX : ES IMERKKI 
CAT:NOUN\] \] 
Fig. 6: A cyclic TFS 
62 3 
which it may share with other adjuncts. We allow adjuncts 
to point back to the modifiend so as to let transfer rules 
refer to properties of the rnodifiend. This means that a 
TFS can be a cyclic graph. This is illustrated in fig. 6. 
4. ~,~neroHon 
Since complex aspects of the transfer mapping are 
handled by the parser and the transfer system, generation 
in our model remains simple. It involves a recursive sort 
of the lexical entxies of the target language and the gener- 
ation of morphologically inflected forms from sets of 
morphological features. 
The linearization component uses a set of unification 
based LP rules operating on information in the final 
Finnish feature structure. Discourse-related information 
relevant tot linearization is included in the feature struc- 
ture. 
For Finnish subjectless clause types, we use a transfer 
ntle thai requires equation of rite English subject with the 
Finnish discourse function THEMA. Depending on 
clause type, any one of the Finnish arguments may appear 
as a TI-iEMA (e.g., "about it one-must discuss"; see fig. 
7). The linearization rule then places the THEMA before 
the finite verb, preserving, in effect, the characteristic 
information structure of the English sentence. 
our experience. In conclusion, we survey the properties 
of graph unification that have proved valuable. 
o Recursive structure of qTS: No limit to the complex~ 
ity of an entry. Multiword entries on a par with one 
word entries. 
Uniformity: Linguistic infommtion at different le~ 
vels represented in a uniform way. No dichotomy of 
lexical and structural transfer. 
Unification: Structure changing correspondences 
can be expressed through coindexing. 
Subsumption: Inheritance of definitions allows 
making generalisations across entries and lexicons. 
, Partial infornmtion: No requirement of complete~ 
ness of linguistc descriptions for transfer to work. 
Disjunctions eliminated by underspecification. No 
need to make translation related sense distinctions 
in monolingual lexicons. 
. Monotonicity: Entries remain valid when lexicon is 
extended and enriched. Enables incremental refine- 
ment of individual entries and grammatical corre~ 
spondences. 
• Commutativity and associativity: Entries remain 
valid when entries or sense definitions are rear- 
ranged or regrouped. 
IF: \[LEX:TAYTYA 'must' 
CAT : VERB 
THEMA:#3 IF: \[LEX:SE ' it' 
CAT : PRON 
CASE :ELA\] \] 
VCOMP : # 9 \[F : #i0 \[LEX : KESKUSTELLA 
' discuss' 
CAT : VERB 
THEMA : #3 
SUBJ:#b\[F: \[LEX:*NONE* 
SEN: \[I!UM:T\]\]\] 
OBL : # 3 
VFORM : INFI 
VOICF, :ACT\] \] 
VFORM:FINITE 
VOICE :ACT\] \] 
Notes 
1 Since unification-based transfer is monotonic, the assump- 
tion of completeness of input is not essential for us. Nothing in 
principle rules out incremental transfer during parsing. 
Acknowledgements 
This research has been supported by IBM Finland. We thank 
Kimmo Koskenniemi for his insights at the system pl,'mning 
stage and to Krister Linddn for discussions on this paper. 
Fig. 7: A Finnish impersonal. Thema percolated from 
VCOMP 
Morphological generation involves production of Fin- 
nish inflected word forms from morphological tags ob- 
tained from the Finnish feature strnctrue using Kosken- 
niemi's two-level morphological processor. 
5. Conclusion 
"\['he choice of unification as a descriptive tool in develo-- 
ping the transfer lexicon system has been productive in 

References 

Beaven, J.L. - Whitelock, P. 1988: Machine Translation 
Using Isomorphic UCGs. Proceedings of COLING 
'88, Budapest. 

Jensen, K. 1986: PEG 1986: A Broad-coverage Compu- 
tational Syntax of English. Technical Report, IBM 
T.J. Watson Research Center. 

Kaplan, R. - Netter, K. o Wedekind, J. - Zaenen, A. 1989: 
Translation by Structural Correspondendes. Proceed- 
ings of the Fourth Conference of the European Chap- 
ter of ACL, Manchester. 
