An Integrated Architecture for Example-Based Machine Translation 
Alexander Franz, Keiko Horiguchi, Lei Duan, 
Doris Ecker, Eugene Koontz, and Kazami Uchida 
Spoken Language Technology, Sony US Research Labs 
3300 Zanker P, oad 
San Jose, CA 95134, USA 
{ amf, kei ko, l ei, do ri s,eko ontz, kuchid a } @ sit. sel. so ny.com 
Abstract 
This paper describes a machine translation 
architecture that integrates the use el' 
examples for flexible, idiomatic translations 
with the use o1' linguistic rules for broad 
coverage and grammatical acctu'acy. We 
have implemented a prototype for 
English-to-Japanese translation, and our 
ewfluation shows lhat tile system has good 
translation quality, and only requires 
reasonable computational resources. 
1 Introduction 
Machine translation by analogy lo pairs of 
corresponding expressions in the source and 
target s, or "example-based transhtlion", 
was firs! proposed by (Nagao 1984). Recent 
work in the example-based I'ramework inchldes 
memory-based translation (Sate & Nagao 1990), 
similarity-driven translation (Watanabe 1992), 
transl'cr-driven nlachine translation (Furusc & 
Iida 1996), and patten>based machine 
translation (Watanabe & Takeda 1998). 
The example-based approach promises easy 
translation knowledge acquisition, more flexible 
transfer than brittle rule-based approaches, and 
idiomatic translations. At the same time, the use 
o1' linguistic rules offers a number of important 
benel'its. Detailed linguistic analysis can allow 
an example-based machine translation system to 
handle a wide variety of input, since rules can be 
used to factor out all linguistic wMations that do 
not influence tile exampled)ased transfer. 
Rule-based  generation from detailed 
linguistic representations can lead to higher 
grammatical output quality. Finally, a modular 
system architecture that uses 
domain-independent linguistic regularities in 
separate linguistic modules allows extending the 
system to much broader domains. The 
HARMONY architecture for lkybrid Analogical 
aud rule-based njachine translation of naturally_ 
occurring colloquial hmguagc combines the 
adwmtages of both these approaches. 
2 The Travel Domain 
Our prototype implementation o1' the HARMONY 
architecture was designed to cover the "travel 
domain". This is composed of words, phrases, 
expressions, and sentences related to 
international travel, similar to what is covered 
by typical travel phrase books. 
Two principles guided our detailed definition of 
the translation domain. FirsL the translation 
domain should not be limited to a narrow 
sub-domain, such as appointment scheduling or 
hotel rcscrwttions. Second, the expressions 
considered in the domain should reflect the fact 
that people quickly adapt to limitations in 
human-machiue or machine-mediated 
communication by simplifying the input. For 
example, (Sugaya et al. 1999) found that the 
average length el' actual human utterances in a 
hotel resmwation task using speech translation 
was only 6.1 words, much shorter than some o1' 
the data that has been used in previous work on 
speech translation. 
'File current vocabulary o1' 7,500 woMs is 
divided into a group o1' general words, a number 
of extensiblc word groups (such as names el' 
food items or diseases), and a number of 
area-specific woM groups (such as names of 
cities or tourist destinations).The travel domain 
is divided into eight "situations": A general 
situation (including everyday conversation); 
transportation; accommodation; sightseeing; 
1031 
shopping; wining, dining, and nightlife; banking 
and postal; and doctor and pharmacy. 
We created a corpus for this dolnain, and 
divided it into a development set o1' 7,000 
expressions, and a separate, unseen test set of 
5,000 expressions. The development set is 
used for creation and refinelnent of the 
translation knowledge sources, and the test set is 
only used for evaluations. (Each evaluation 
uses a new, random 500-word sample from the 
5,000 word test set.) 
The corpus was balanced to illustrate the widest 
possible variety of types o1' words, phrases, 
syntactic structures, semantic patterns, and 
pragmatic functions. The average length el' the 
expressions in the corpus is 6.5 words. Some 
examples from the development corpus are 
shown below. Even though this dolnain might 
seem rather limited, it still contains inany 
challenges for machine translation. 
• Can I have your last name, please ? 
• Is this the bus for Shinagawa station ? 
• 1 would like to make a reservation for two 
people for eight nights. 
• Can you tell us where we can see some 
Buddhist temples ? 
• Most supermarkets sell liquor. 
• Can you recommend a good Chinese 
restaurant in this area ? 
• I'd like to change 500 Dollars' in traveller's 
checks into Yen. 
• Are there any English..speaking doctolw at 
the hospital? 
3 NLP Infrastructure 
The prototype implementation is coustructed out 
of components that are based on a powerful 
infrastructure for natural  processing 
and  engineering. The three inain 
aspects of this infrastructure are the Grammar 
Programming Language (GPL), the GPL 
compiler, and the GPL runtime environment. 
3.1 The Grammar Programming Language 
The Gralnmar Programming Language (GPL) is 
an imperative programming  for 
feature-structure-based rewrite grammars. GPL 
is a l'ormalism that allows the direct expression 
of linguistic algorithms l'or parsing, transfer, and 
generation. Some ideas in GPL can be traced 
back to Tolnita's pseudo-unification fornmlisln 
(Tomita 1988), and to Lexical-Functional 
Grammar (Dalrymple et al. 1995). GPL 
includes variables, simple and complex tests, 
and various manipulation operators. GPL also 
includes control flow statements including 
if-then-else, switch, iteration over 
sub-feature-structures, and other features. An 
example of a silnplified GPL rule for English 
generation is shown in Figure 1. 
Wll SENT --) NP YN_SENT { 
!exist\[Sin VP SUBJ WH\]; 
local-variable WIt_VP = \[$m VP\]; 
local-wlriable WH PHP, ASE; 
$WH PHRASE = find-subfstruct in $WHVP 
where (?exist\[$x WH\]); 
$d I = \[$WH_PHRASE SLOT-VALUE\]; 
\[$WH PHRASE SLOT-VALUE TRACE\] = '+'; 
$d2 = $m; 
/ 
Figure I Example of a GPLGeneration P, ule 
3.2 The GPL Compiler 
GPL grammars are compiled into C code by the 
GPL compiler. The GPL compiler was created 
using the Unix tools lex and yacc (Levine et al. 
1990). For each rewrite rule, the GPL compiler 
creates a main action function, which carries out 
most of the tests and manipulatious specified by 
the GPL statements. 
The GPL compiler handles disiunctive feature 
structures in an efficient manner by keeping 
track of sub-feature-structure references within 
each GPL rule, and by generating an expansion 
function that is called once before the action 
function. The coinpiler also tracks variable 
references, and generates and tracks separate test 
functions for nested test expressions. 
3.3 The GPL Run-time Environment 
The result of compiling a GPL grainmar is an 
encapsulated object that can be accessed via a 
public interface function. This interface fnnction 
serves as the link between the compiled GPL 
grammars, and the various 
-independent and domain-independent 
software engines for parsing, transfer, generation, 
and others. This is illustrated in Figure 2. 
1032 
GI)I+ Grammar J Feature Slructure \] 
l)cfinitions 
.,...---- 
GI)L Compiler \] 
1 
hllcrface Function 
Aclion leunctions 
l£xpansion Funclions Test Ftmctions 
Compiled GI)L Grammar 
Software Engine 
(parsers, Iransfers, 
gellcrators, ...) 
I"7 
Feature Structtlre Library 
P, el)resentaiion Testing code 
Manipulalion code Memory Management 
) Figure 2 GI L Run-time Environmenl 
The compiled GPI. grammars use the feature 
structure library, which provides services for 
efficiently representing, testing, manipulating, 
and managing memory for feature structures. A 
special-purpose inemory manager maintains 
separate stacks of memory pages for each object 
size. This scheme allows garbage colleclion 
that is so fast that it can be performed after every 
aUempted GPL rule execution. In our 
experiments with Japanese and English parsing, 
we found that l)el'-rule garbage collection 
reduced the overall read/write memory 
requirements by as much as a factor of four to 
six. 
4 Source Language Analysis 
Translation is divided into the steps ot' analysis, 
transfer, and generation. Sourcc-hmguage 
analysis is illustrated in Figure 3. 
English analysis begins with tokenization and 
morphological analysis, which creates a lattice 
that contains lexical feature structures. I)uring 
multi-word matching, expressions from tile 
multi-word lexicon (such as White House or take 
on) are detected in the word lattice, and new arcs 
with the appropriate lexical feature structures are 
added. 
English lnpul 
Morphological Analysis 
I lmtlice with Single-word 
gexical l:cature ,qlructures ~, 
Mulli-word Matching 
~.~ LaHice with Single and Multi-word texical t"eature 
Structures 
I gexical P, educlion Anlbiguity 
P.educed l.exical Feature 
N \[I'UCI.LII'e Lattice 
V I'm'sing 
Senlei|lial Fealure Slruclure 
Thesaurus 
I Compiled 
English 
l'arsing Gialnnlar 
Figure 3 Architecture el'the Analysis Module 
Lexical ambiguity reduction reduces the nulnber 
el' arcs in the word lattice. This module carries 
out part-of-speech tagging over the lattice, and 
reduces the lattice to those lexical feature 
structures that arc part of the number of best 
paths that represents tile best speed/accuracy 
trade-off (currently two). This calculation is 
based on the usual lexical and contextual bigram 
probabilities that were estimated from a training 
corpus, but it also takes into account manual 
costs that can be added to lexicon entries, or to 
individual part-of-speech bigrams. 
The resulting reduced lattice with lcxical 
single-word and multi-word feature structures is 
parsed using tilt GLR parsing algorithm 
extended to lattice input (Tomita 1986). The 
English parsing grammar consists of 540 GPL 
rules. The output is a sentential feature structure 
that represents the input to the transfer 
component. 
1033 
5 Transfer 
Transfer I¥om the source- sentential 
feature structure to the target- sentential 
feature structure is accomplished with a hybrid 
rule-based and example-based method. This is 
illustrated in Figure 4. 
I English Scntcntial ) 
Feature Structure 
\[Compiled l \[Tra,,srer / I \[Database I 
1 1.inguistic ~ Examl)le 
! I Tra,,.<er ~ Matching\[ 
;iiiiii!iiiiiiiiiiii ................... ...... 
Figure 4 Architecture of the Transfer Module 
The input feature structure is passed to the 
linguistic transfer procedure. This consists of a 
rule-rewriting software engine that executes the 
compiled English-to-Japanese transfer grammar. 
The transfer grammar consists of 140 GPL rules, 
and its job is to specify linguistic constraints on 
examples, combine multiple examples, transfer 
informatiou that is beyond the scope of the 
example database, and perl'orm various other 
transformations. The overall effect is to broaden 
the linguistic coverage, and to raise the 
grammatical accuracy far beyond the level of a 
traditional example-based transfer procedure. 
The linguistic transfer procedure operates on the 
input feature structure in a recursive manner, 
and it invokes the example matching procedure 
to find the best translation example for various 
parts of the input. The example matching 
procedure retrieves the best translation examples 
from the example database, which contains 
14,000 example pairs ranging from individual 
words to entire sentences. In an ofl'-line step, the 
example pairs are parsed, disambiguated, and 
indexed for corresponding constituents using a 
Treebanking tool. 
At each invocation of the example matching 
procedure, linguistic constraints fl'om the 
transfer grammar are used to limit the search 
space to appropriate examples. In an ol'l'-line 
step, these constraints are pre-compiled into a 
complex index that allows a preliminary fast 
match. Examples that survive the fast match are 
matched and aligned with the input feature 
structure (or sub-feature-structure, during 
recursive invocations) using the thesaurus to 
calculate word similarity, and using various 
other constraints and costs for inserting, deleting, 
or altering slots and features. Rather than rely on 
the exact distance in the thesaurus to calculate 
lexical similarity, we use a scheme that is based 
on the information content of thesaurus nodes, 
similar to (Resnik 1995). 
6 Target- Generation 
The Japanese target- feature structure 
l'orms the input to the generation module, which 
is summarized in Figure 5 below. This module 
also consists o1' a rule-rewriting software engine, 
executing the compiled GPL Japanese 
generation grammar, which consists ol' 200 GPL 
rules. The generator uses the Japanese lexicon to 
create the Japanese target- expression. 
Feature Slrtlclure Compiled Generation 
Granll/lar 
Generation 
Module 
Expression 
Figure 5 Architecture of~.he Generation Module 
7 Evaluation and Conclusions 
We evaluated the trauslation system using a 
random 500-expression sample from the unseen 
test set (see Section 2 above). The translatious 
were manually assigned to one of the following 
categories o1' translatiou quality: 
Failure. Complete translation failure, due to 
lack of coverage of a rule-based component. 
1034 
Wrong. A translation that is COlnpletely wrong, 
or that has major errors in an important part, 
such as in the main clause. 
Major Problem. A translation that has a 
missing, extra, or incorrect constituent, such as a 
subject, object, or adjectival/prepositional 
predicate. 
Minor Problem. A translation that l-ms a 
missing, extra, or incorrect minor part, such as 
an intensifier, tense, aspect, temporal or locative 
adjunct, adverb, adjective or other prenominal 
modifier, prepositional phrase, verb conjugation 
form, adjective form, or required word or 
constituent order. 
Stylistic Problem. Slylistic problems include 
awkward but tolerable word order, incorrect 
Japanese particles, incorrect idioms, and silnilar. 
Flawless. A translation that does not exhibit any 
of the above problems is considered flawless. 
The results of the ewtluation are shown in Table 
1 below. Overall, 84% of the translations 
convey the meaning in an acceptable manner. 
We also ewtluated the computational resource 
requirements of the system. On a Pentium II1 
running at 500 MHz, the average translation 
speed was 0.44 seconds. The memory 
requirements are summarized in Table 2 below. 
Flawless 60% 
Stylistic Problem 9% 
Minor F'roblcm 14% 
Acceptable with OOV 1% 
Major ProMem 9% 
Wrong Translation 5% 
Translation IVailure 3% 
Table 1 Translalion Quality 
6MB Read-only Memory lbr Code and Data 
Read-only Melnory for l)iclionary, 
Examples, Fast Match Index, etc. 
Read/Write Memory for Feature Slructures 
Read/Write Memory fo," Software Engines 
Table 2 Memory l,Iequirements 
23MB 
14MB 
4MB 
Our plans for further work include extending the 
size of the input w)cabulary, and developing 
mechanisms for closer integration with speech 
recognition and speech synthesis components for 
speech-to-speech translation. We are also 
working on the Japanese-to-English translation 
direction, and we plan to report results on this in 
the future. 
Acknowledgements 
Our thanks go to Robert Bowen, Benjamin 
Hartwell, Chigusa Inaba, Kaori Shibatani, 
Hirono Stonelake, and Kazue Watanabe for their 
 engineering elTorts, and to Edward tto 
for user interface and application development. 

References 

Dalryml)le, M., R.M. Kaplan, J.T. Maxwell lll, and A. 
Zacncn, cds. (1995) l;ormal Issues in 
Lexical-Functional Grammar. CSLI Lecture Notes 
47, Stanford, CA. 

Furuse, O. and H. Iida (1996) "incremental 
Iranslation utilizing constituent-I~oundary patlcrns", 
in Proceedings of COL1NG-96, pages 412-417. 

Lcvine, J.R., T. Mason, and D. Brown (1990), lex & 
yacc (Second Editio,), O'Rcillcy and Associates, 
Sebastopol, CA. 

Nagao, M. (1984) "A framework of a Machine 
Translation between Japanese and English by 
analogy principle", in Artificial and tlumatt 
Intelligence, A. 17Jithorn and R. Baneiji (eds.), 
North Holland, pages 173--180. 

Resnik, P. (1995) "Using information content to 
evalualc semantic similarity in a taxonomy", in 
Proceedings of LICAI-95. 

Sale, S. and M. Nagao (1990) "Toward 
memo,y-based lratlslalioD", ill Proceedings of 
COLING-90, vol. 3, Helsinki, Finhmd, pages 
247--252. 

Sugaya, F., T. Takczawam A. Yokoo, and S. 
Yamamolo (1999) '%rid-to-end ewtlualion in 
ATP,-Matrix", in Proceedings of Fmrospeech-99, 
Bt, dapest, Hungary, pages 2431--2434. 

Tomita, M. (1988) The Generalized LR 
I'ars'eMCompiler (Version 8.1): User's Guide. 
Technical Memorandum CMU-CMT-88-Memo, 
Cenler for Machine Translation, Carnegie Mellon 
University. 

Tonlila, M., "All efficient word lattice parsing 
algoritlun for conlinuous speech recognition", ill 
l'roceedings of ICASSP-86, Tokyo, Japan, pages 
1569-1572. 

Walanabe, H. (1992) "A similarily-driven transfer 
system", in Proceedings of COLING-92, Nantes, 
France, pages 770-776. 

Walanabe, H. and K. Takeda (1998) "A 
pattern-based Machine Translation syslem 
extended by exanll~le-based processing", in 
Proceedings of ACL-COLING-98, pages 
1369-1373. 
