Recycling Lingware in a Multilingual MT System 
Steffen Leo Hansen 
Manny Rayner 
David Carter 
Ivan Bretan 
Robert Eklund 
Mats Wir~n 
Sabine Kirchmeier-Andersen 
Christina Philp 
Finn Sorensen 
Hanne Erdman Thomsen 
SRI International 
Suite 23, Millers Yard 
Cambridge CB2 1RQ 
United Kingdom 
manny@cam.sri.com 
dmc~cam.sri.com 
Telia Research AB 
Spoken Language Processing 
S-136 80 Haninge 
Sweden 
Ivan.P.Bretan@telia.se 
Robert.H.Eklund@telia.se 
Mats.G.Wiren@telia.se 
Handelsh0jskolen i K0benhavn 
Institut for Datalingvistik 
Dalgas Have 15 
DK-2000 Frederiksberg 
Denmark 
slh.id@cbs.dk 
Abstract 
We describe two methods relevant to multi- 
lingual machine translation systems, which 
can be used to port linguistic data (gram- 
mars, lexicons and transfer rules) be- 
tween systems used for processing related 
languages. The methods are fully im- 
plemented within the Spoken Language 
Translator system, and were used to create 
versions of the system for two new language 
pairs using only a month of expert effort. 
1 Introduction 
The basic idea of this paper is simple and uncon- 
troversial. All natural languages are in some sense 
similar (some are obviously very similar), so software 
written to process one language ought to some ex- 
tent to be applicable to other languages. If the lan- 
guages L1 and 52 are similar enough, then it should 
be easier to recycle software applicable to L, than 
to rewrite it from scratch for L2. 
This paper describes two related approaches in 
this general direction, which have been success- 
fully applied within the Spoken Language Transla- 
tor (SLT) project (Rayner and Carter, 1997). The 
first is the most obvious: we start with a function- 
ing grammar and lexicon for L1, and port it to the 
similar language L2. This is not, of course, a novel 
idea, but we think that we have refined it in a num- 
ber of ways. In particular, we show that it is prac- 
tically feasible in the case of sufficiently close lan- 
guages to generalize an existing grammar for L1 to 
cover both L1 and L2 (i.e. produce a single gram- 
mar which through the setting of a single param- 
eter becomes valid for either language). We also 
describe a method which makes it possible to port 
the language-dependent lexicon for L1 so as to max- 
imize sharing of data between the systems for the 
two languages. 
The second idea is specifically related to trans- 
lation. Suppose we have already developed sets of 
transfer rules for the two language-pairs L1 --+ L2 
and L2 ~ L3. We describe a method which allows 
us to compose the two sets of rules off-line to create 
a new set for the pair L1 --+ L3. 
Both methods might be said to operate according 
to the principle memorably described by Mary Pop- 
pins as "Well begun is half done". They do not solve 
either problem completely, but automatically take 
care of most of the drudgery before any human has 
to become involved. In each case, the initial result is 
a machine-written set of linguistic data (lexicon en- 
tries and transfer rules) which is not quite adequate 
as it stands; a system expert can however clean it up 
into satisfactory shape in a small fraction of the time 
that would have been required to write the relevant 
rules and lexicon entries from scratch. 
The practical experiments we describe have been 
carried out using versions of the SLT system involv- 
ing the languages English, French, Swedish and Dan- 
ish. Initial results are extremely promising. In par- 
ticular, we were able to combine both methods to 
create fairly credible Swedish-to-French and English- 
to-Danish spoken language translation systems I us- 
~In fact we do not currently use a Danish speech syn- 
thesizer, but it would be straightforward to incorporate 
55 
ing only a few person-weeks of expert effort. 
The rest of the paper is structured as follows. Sec- 
tion 2 gives a very brief overview of the relevant as- 
peers of the SLT system. Section 3 describes the 
methods we have developed for porting linguistic 
descriptions between closely related languages. Sec- 
tion 4 summarizes the transfer composition method. 
Section 5 describes preliminary experiments. 
2 An overview of the SLT system 
The SLT system has been described in detail else- 
where (most recently (Rayner and Bouillon, 1995; 
Rayner and Carter, 1997)), so this section will only 
provide the minimum information necessary to un- 
derstand the main body of the paper. 
The language-processing (translation) part of the 
system is supplied with N-best sentence hypotheses 
by the system's recognizer, and itself uses a hybrid 
architecture, which combines rules and trainable sta- 
tistical models. To summarize the argument from 
(Rayner and Bouillon, 1995), there are good reasons 
for requiring both these components to be present. 
Rules are useful for capturing many kinds of regular 
linguistic facts that are independent of any partic- 
ular domain of application, prime examples being 
grammatical agreement and question-formation. In 
contrast, there are other types of phenomena which 
intuitively are more naturally conceptualized as id- 
iosyncratic and domain-dependent: the most obvi- 
otis examples here are word-choice problems. 
The system uses two translation mechanisms, ap- 
plied bottom-up in parallel (Rayner and Carter, 
1997). The primary, rule-based translation mecha- 
nism performs transfer at the level of Quasi-Logical 
Form (QLF), a type of predicate/argument style 
notation (Alshawi et al., 1991). The source- and 
target-language grammars provide a declarative def- 
inition of a many-to-many mapping between sur- 
face form and QLF. The grammars are domain- 
independent, and can be compiled to run efficiently 
either in the direction surface form ~ QLF (anal- 
ysis) or QLF --+ surface form (generation). In 
transfer, unification-based rules are used to define 
a space of possible candidate translations; domain- 
dependent, statistically trained preferences then se- 
lect the most preferred candidate translation. This 
division of effort has the important consequence of 
allowing the transfer rules to be fairly simple, since 
much of the complexity is "factored out" into the 
trained preferences. 
In order to deal with the brittleness inherent in 
az~y rule-based system, a second, much less sophisti- 
one. 
cated translation method is also used, which simply 
maps surface phrases from the source language into 
possible target-language counterparts. We refer to 
the backup method as "word-to-word" (WW) trans- 
lation. The two methods are combined, roughly 
speaking, by using rule-based QLF transfer to trans- 
late as much as possible, filling in any gaps with ap- 
plications of the WW rules. 
The parts of the system of central interest here 
are the rule-based components, in particular the 
morphologies, grammars, lexica and transfer rules. 
Morphologies are written in a variant of two-level 
morphology (Carter, 1995), and grammars in a 
unification-based formalism (Alshawi (ed), 1992). 
The lexicon for each language is divided into three 
main parts: 
Domain-independent function (closed class) 
word entries are written directly in terms of def- 
initions of suitable feature-value assignments, 
and can from a software-engineering standpoint 
be regarded as part of the grammar. 
A collection of language-dependent but domain- 
independent macros define the feature-value 
assignments needed for each type of regular 
content-word, e.g. "count noun", "transitive 
verb" and so on. These macros are called 
paradigm macros. 
Content word entries, which in general may be 
domain-dependent, are defined in terms of these 
lexical macros. An entry of this kind contains 
the following information: the name of the rel- 
evant macro, the base surface form of the word, 
the associated logical-form (QLF) constant, and 
if necessary a pointer to the correct inflectional 
type (conjugation or declension). 
Structurally, transfer rules have much in common 
with lexicon entries. (Bear in mind that conven- 
tional bilingual and monolingual dictionaries have 
similar structures too). A small set of domain- 
independent transfer rules encode cross-linguistic 
divergences significant enough that they need to 
be represented at the QLF level: these rules may 
contain arbitrary pieces of QLF form. The ma- 
jority of the transfer rules, however, are "atomic- 
atomic" : they associate a logical-form constant from 
the source language with one or more logical-form 
constants from the target language. Transfer rules 
of this type have a close connection with the macro- 
defined monolingual content-word lexicon, and may 
also be domain-dependent. 
55 
3 Porting grammars and lexica 
between closely related languages 
The original version of the Core Language Engine 
had a single language description for English, writ- 
ten by hand from scratch (Pulman, 1992; Rayner, 
1994). Subsequently, language descriptions have 
been developed for Swedish (Gamb~ck and Rayner, 
1992), French and Spanish (Rayner, Carter and 
Bouillon, 1995). In each of these cases, the new lan- 
guage description was created by manually editing 
the relevant files for the closest existing language. 
(The Swedish and French systems are modified ver- 
sions of the original English one; the Spanish system 
is modified from the French one). There are however 
some serious drawbacks to this approach. Firstly, 
it requires a considerable quantity of expert effort; 
secondly, there is no mechanism for keeping the re- 
suiting grammars in step with each other. Changes 
are often made to one grammar and not percolated 
to the other ones until concrete problems show up in 
test suites or demos. The net result is that the var- 
ious grammars tend to drift steadily further apart. 
When we recently decided to create a language de- 
scription for Danish, we thought it would be inter- 
esting to experiment with a more principled method- 
ology, which explicitly attempts to address the prob- 
lems mentioned above. The conditions appeared 
ideal: we were porting from Swedish, Swedish and 
Danish being an extremely closely related language 
pair. The basic principles we have attempted to ob- 
serve are the following: 
• Whenever feasible, we have tried to arrange 
things so that the linguistic descriptions for the 
two languages consist of shared files. In partic- 
ular, the grammar rules files for the two lan- 
guages are shared. When required, rules or 
parts of rules specific to one language are placed 
inside macros whose expansion depends on the 
identity of the current language, so that the 
rule expands when loaded to an appropriate 
language-specific version. 
• When files cannot easily be shared iin particu- 
lar, for the content-word lexica), we define the 
file for the new language in terms of declara- 
tions listing the explicit differences against the 
corresponding file for the old language. We have 
attempted to make the structure of these dec- 
larations as simple as possible, so that they can 
be written by linguists who lack prior familiar- 
ity with the system and its notation. 
Although we are uncertain how much generality to 
claim for the results (Swedish and Danish, as already 
noted, are exceptionally close), we found them en- 
couraging. Four of the 175 existing Swedish gram- 
mar rules turned out to be inapplicable to Danish, 
and two had to be replaced by corresponding Danish 
rules. Five more rules had to be parameterized by 
language-specific macros. Some of the morphology 
rules needed to be rewritten, but this only required 
about two days of effort from a system specialist 
working together with a Danish linguist. The most 
significant piece of work, which we will now describe 
in more detail, concerned the lexicon. 
Our original intuition here was that the function- 
word lexicon and the paradigm macros (cf Section 2) 
would be essentially the same between the two lan- 
guages, except that the surface forms of function 
words would vary. To put it slightly differently, we 
anticipated that it would make sense as a first ap- 
proximation to say that there was a one-to-one cor- 
respondence between Swedish and Danish function- 
words, and that their QLF representations could be 
left identical. This assumption does indeed appear 
to be borne out by the facts. The only complica- 
tion we have come across so far concerns definite 
determiners: the feature-value assignments between 
the two languages need to differ slightly in order to 
handle the different rules in Swedish and Danish for 
determiner/noun agreement. This was handled, as 
with the grammar rules, by introduction of a suit- 
able call to a language-specific macro. 
With regard to content words, the situation is 
somewhat different. Since word choice in transla- 
tion is frequently determined both by collocational 
and by semantic considerations, it does not make as 
much sense to insist on one-to-one correspondences 
and identical semantic representations. We conse- 
quently decided that content-words would have a 
language-dependent QLF representation, so as to 
make it possible to use our normal strategy of letting 
the Swedish-to-Danish translation rules in general be 
many-to-many, with collocational preferences filter- 
ing the space of possible transfers. 
The remarks above motivate the concrete lexicon- 
porting strategy which we now sketch. All work 
was carried out by Danish linguists who had a good 
knowledge of computational linguistics and Swedish, 
but no previous exposure to the system. The start- 
ing point was to write a set of word-to-word trans- 
lation rules (cf Section 2), which for each Swedish 
surface lexical item defined a set of possible Danish 
translations. The left-hand side of each WW rule 
specified a Swedish surface word-form and an asso- 
ciated grammatical category (verb, noun, etc), and 
the right-hand side a possible Danish translation. 
An initial "blank" version of the rules was created 
57 
automatically by machine analysis of a corpus; the 
left-hand side of the rule was filled in correctly, and 
a set of examples taken from the corpus was listed 
above. The linguist only needed to fill in the right- 
hand side appropriately with reference to the exam- 
ples supplied. 
The next step was to use the word-to-word rules to 
induce a Danish lexicon. As a first approximation, 
we assumed that the possible grammatical (syntac- 
tic/semantic) categories of the word on the right- 
hand side of a WW rule would be the same as those 
of the word on its left-hand side. (Note that in gen- 
eral a word will have more than one lexical entry). 
Thus lexicon entries could be copied across from 
Swedish to Danish with appropriate modifications. 
In the case of function-words, the entry is copied 
across with only the surface form changed. For 
content-words, the porting routines query the lin- 
guist for the additional information needed to trans- 
form each specific item as follows. 
If the left-hand (Swedish) word belongs to a lexi- 
cal category subject to morphological inflection, the 
linguist is asked for the root form of the right-hand 
(Danish) word and its inflectional pattern. If the 
inflectional pattern is marked as wholly or partly ir- 
regular (e.g. with strong verbs), the linguist is also 
queried for the values of the relevant irregular in- 
flections. All requests for lexical information are 
output in a single file at the end of the run, for- 
matted for easy editing. This makes it possible for 
the linguist to process large numbers of information 
requests quickly and efficiently, and feed the revised 
declarations back into the porting process in an it- 
erative fashion. 
one particularly attractive aspect of the scheme 
is that transfer rules are automatically generated as 
a byproduct of the porting process. Grammar rules 
and function-words are regarded as interlingual; thus 
for each QLF constant C involved in the definition 
of a grammar rule or a function-word definition, the 
system adds a transfer rule which maps C into itself. 
Content-words are not interlingual. However, since 
each target lexical entry L is created from a source 
counterpart L', it is trivial to create simultaneously 
a transfer rule which maps the source QLF constant 
associated with L' into the target QLF constant as- 
sociated with L. 
4 Transfer composition 
The previous sections have hopefully conveyed some 
of the flavour of our translation framework, which 
conceptually can be thought of as half-way between 
transfer and interlingua. We would if possible like 
to move closer to the interlingual end; however, the 
problems touched on above mean that we do not see 
this as being a realistic short-term possibility. Mean- 
while, we are stuck with the problem that dogs all 
multilingual transfer-based systems: the number of 
sets of transfer rules required increases quadratically 
in the number of system languages. Even three lan- 
guages are enough to make the problem non-trivial. 
In a recent paper (Rayner et al, 1996), we de- 
scribed a novel approach to the problem which we 
have implemented within the SLT system. Exploit- 
ing the declarative nature of our transfer formalism, 
we compose (off-line) existing sets of rules for the 
language pairs L1 --+ L2 and L2 ~ L3, to create 
a new set of rules for L1 ~ L3. It is clear that 
this can be done for rules which map atomic con- 
stants into atomic constants. What is less obvious 
is that complex rules, recursively defined in terms 
of translation of their sub-constituents, can also be 
composed. The method used is based on program- 
transformation ideas taken from logic programming, 
and is described in detail in the earlier paper. Simple 
methods, described in the same paper, can also be 
used to compose an approximate transfer preference 
model for the new language-pair. 
The rule composition algorithm is not complete; 
we strongly suspect that, because of recursion ef- 
fects, the problem of finding a complete set of com- 
posed transfer rules is undecidable. But in practice, 
the set of composed rules produced is good enough 
that it can be improved quickly to an acceptable 
level of performance. Our methodology for perform- 
ing this task makes use of rationally constructed, 
balanced domain corpora to focus the effort on fre- 
quently occurring problems (Rayner, Carter and 
Bouillon, 1995). It involves making declarations to 
reduce the overgeneration of composed rules; adding 
hand-coded rules to fill coverage holes; and adjust- 
ing preferences. The details reported in (Rayner et 
al, 1996). 
5 Experiments 
We will now present results for concrete experi- 
ments, where we applied the methods described 
above so as to rapidly construct translation systems 
for two new language pairs. All of the translation 
modules involved operate within the same Air Travel 
Inquiry (ATIS; (Hemphill et ai., 1990)) domain as 
other versions of SLT, using a vocabulary of about 
1 500 source-language stem entries, and have been 
integrated into the main SLT system to produce ver- 
sions which can perform credible translation of spo- 
ken Swedish into French and spoken English into 
Danish respectively. 
68 
Swe --+ Fre Eng -+ Swe 
FLflly acceptable 29.4% 56.5% 
Unnatural style 16.3% 7.75% 
Minor syntactic errors 15.2% 11.75% 
Major syntactic errors 2.0% 4.75% 
Partial translation 7.0% 8.75% 
Nonsense 22.9% 5.0% 
Bad translation 7.0% ! 4.0% 
No translation 0.2% 1.5% 
Table 1: Translation results for Swedish --+ French 
and English --+ Swedish on unseen speech data 
5.1 Swedish --+ English ~ French 
This section describes an exercise which involved us- 
ing transfer composition to construct a Swedish 
French translation system by composing Swedish 
English and English ~ French versions of the sys- 
tem. The total expert effort was about two person- 
weeks. We start by summarizing results, and then 
sketch the main points of the manual work needed 
to adjust the composed rule-sets. 
We used a corpus of 442 previously unseen spoken 
utterances, and processed the N-best lists output for 
them by the speech recognizer. The results are as 
given in Table 1; for comparison, we also give the 
results for English --+ Swedish, the language pair to 
which we have devoted the most effort (and which 
does not involve any transfer composition). 
Thus almost 30% (top row) of the translations 
produced were completely acceptable, with another 
30% or so (rows 2-3) having only minor problems, 
giving a total of 60% that would probably be ac- 
ceptable in practical use. A further 9% (rows 4-5) 
contained major errors but also some correct infor- 
mation, while nearly all the remaining 30% (bot- 
tom 3 rows) were clearly unacceptable, consisting 
either of nonsense or of a translation that made some 
sense but was wrong. The reasons for these 30% of 
outright failures, compared to only about 10% for 
English --~ Swedish, are firstly, that recognizer per- 
formance is slightly less good for Swedish than for 
English, owing to less training data being available; 
second, that Swedish and French differ more than 
English and Swedish do; thirdly, that transfer rules 
for both the component pairs (Swedish ~ English 
and English --+ French) have had much less work 
devoted to them than English --+ Swedish; and last 
but not least, of course, that transfer composition is 
being used. 
When cleaning up the automatically composed 
Swedish -+ French rule-set, the task on which we 
spent most effort was that of limiting overgeneration 
of composed transfer rules. The second most impor~ 
tant task was manual improvement of the Composed 
transfer preference model. The methods used are 
described in more detail in (Rayner et al, 1996). 
5.2 English --+ Swedish --+ Danish 
This section briefly describes a second series of ex- 
periments, in which we converted an English --~ 
Swedish system into an English --+ Danish system 
using the methods described earlier. The total in- 
vestment of system expert effort was again around 
two person-weeks. 
About half the effort was used to port the Swedish 
language description to Danish, employing the meth- 
ods of Section 3. After this, we carried out two 
rounds of testing and bug-fixing on the Swedish --~ 
Danish translation task. For this, we used a Swedish 
representative corpus, containing 331 sentences rep- 
resenting 9 385 words from the original Swedish cor- 
pus. These tests uncovered a number of new prob- 
lems resulting from previously unnoted divergences 
between the Swedish and Danish grammars. About 
half the problems disappeared after the addition of 
20 or so small hand-coded adjustments to the mor- 
phology, function-word lexicon, transfer rules and 
transfer preferences. 
After the second round of bug-fixing, 95% of the 
Swedish sentences received a Danish translation, and 
79% a fully acceptable translation. (When measur- 
ing results on representative corpora, we count cov- 
erage in terms of "weighted scores". The weight as- 
signed to sentence is proportional to the number of 
words it represents in the original corpus: that is, 
its length in words times the number of sentences it 
represents). Most of the translation errors that did 
occur were minor ones. Finally, we composed the 
English ~ Swedish and Swedish --+ Danish rules to 
create a English -+ Danish rule-set, and used this, 
after a day's editing by an expert, to test English --+ 
Danish translation using a representative text corpus 
(we will present results for unseen speech input at 
the workshop). Our results, using the same scheme 
as above, were as given in Table 2. 
6 Conclusions and further directions 
We have demonstrated that it is practically feasible 
in the case of sufficiently close languages to general- 
ize an existing grammar for one language to produce 
a grammar which, through the setting of a single pa- 
rameter, becokes valid for either language. As well 
as providing major efficiency gains over writing a 
grammar for the second language from scratch, this 
69 
Eng ~ Dan 
Fully acceptable 52.5% 
Unnatural style 0.4% 
Minor syntactic errors 24.4% 
Major syntactic errors 0.7% 
Partial translation 0.0% 
Nonsense 0.9% 
Bad translation 10.7% 
No translation 10.3% 
Table 2: Translation results for English --+ Danish 
on representative text data 
technique means that subsequent enhancements to 
the grammar, in those areas where the characteris- 
tics of the two languages are equivalent, will apply 
automatically to both of them. 
We have also described an algorithm for com- 
position of transfer rules. We have demonstrated 
that it can be used to automatically compose non- 
trivial sets of transfer rules containing on the or- 
der of thousands of rules, and shown that by small 
adjustments the performance can be improved to a 
level only slightly inferior to that of a correspond- 
ing set of hand-coded rules. Our experience is that 
the amount of work involved in using these methods 
is only a fraction of that needed to develop similar 
rules from scratch. 
Acknowledgements 
The Danish-related work reported here was funded 
by SRI International and Handelsh0jskolen i 
Kobenhavn. Other work was funded by Telia Re- 
search AB under the SLT-2 project. We would like 
to thank David Milward and Steve Pulman for help- 
ful comments. 

References 
Agn~is, M-S., Alshawi, H., Bretan, I., Carter, D., 
Ceder, K., Collins, M., Crouch, R., Digalakis, V., 
Ekholm, B., Gamb~ck, B., Kaja, J., Karlgren, J., 
Lyberg, B., Price, P., Pulman, S., Rayner, M., 
Samuelsson, C. and Svensson, T. 1994. Spoken 
Language Translator: First Year Report. SRI 
Technical Report CRC-0432 
Alshawi, H., D. Carter, B. Gamb~ck and M. Rayner. 
1991. Transfer through Quasi Logical Form. Proc. 
ACL-91. 
Alshawi, H. (ed.) 1992. The Core Language Engine. 
MIT Press. 
Carter, D. 1995. Rapid Development of Morpho- 
logical Descriptions for Full Language Processing 
Systems. Proceedings of 7th European ACL. Also 
SRI Technical Report CRC-047 
Ganlbgck, B., and M. Rayner. The Swedish Core 
Language Engine. Proc. 3rd Nordic Conference 
on Text Comprehension in Man and Machine, 
LinkSping, Sweden. Also SRI Technical Report 
CRC-025. 
Hemphill, C.T., J.J. Godfrey and G.R. Doddington. 
1990. The ATIS Spoken Language Systems pilot 
corpus. Proc. DARPA Speech and Natural Lan- 
guage Workshop, Hidden Valley, Pa., pp. 96-101. 
Pulman, S. 1992. Unification-Based Syntactic Anal- 
ysis. In (Alshawi (ed), 1992). 
Rayner, M. 1994. English linguistic coverage. In 
(Agn~ et al., 1994). 
Rayner, M. and Bouillon, P. 1995. Hybrid Transfer 
in an English-French Spoken Language Transla- 
tor. Proceedings of IA '95, Montpellier, France. 
Also SRI Technical Report CRC-056. 
Rayner, M., P. Bouillon and D. Carter. 1995. Us- 
ing Corpora to Develop Limited-Domain Speech 
Translation Systems. Proc. Translating and the 
Computer 17. Also SRI Technical Report CRC- 
059. 
Rayner, M., D. Carter and P. Bouillon. 1996. 
Adapting the Core Language Engine to French 
and Spanish. Proc. NLP-IA, Moncton, New 
Brunswick. Also SRI Technical Report CRC-061. 
Rayner, iVl., and D. Carter. 1997 Hybrid language 
processing in the Spoken Language Translator. 
Proc. ICASSP-97, Munich, Germany. Also SRI 
Technical Report CRC-064. 
