Machine Translation of Sentences with Fixed Expressions 
Naoto Katoh 1 Teruaki Aizawa 
NHK Science and Technical Research Laboratories 
1-10-11, Kinuta, Setagaya-ku, 
Tokyo 157, Japan 
{ katonao, aizaw a } @ strl.nhk, or.jp 
Abstract 
This paper presents a practical machine translation system 
based on sentence types for economic news stories. 
Conventional English-to-Japanese machine translation 
(MT) systems which are rule-based approaches, are diffi- 
cult to translate certain types of Associated Press (AP) 
wire service news stories, such as economics and sports, 
because these topics include many fixed expressions (such 
as compound words or collocations) which are difficult to 
be processed by conventional syntactic analysis and/or 
word selection methods. 
The proposed MT system, an economic-news stories 
machine translation system (ENTS), can translate eco- 
nomic news sentences with fixed expressions. The system 
consists of three processes, to handle different types of 
sentences, fixed type, economics-specific type and general 
type. This paper focuses mainly on the translation method 
for fixed-type sentences, which is a kind of example-based 
approach. In this translation method, fixed sentence trans- 
lation (STRA) data plays a key role. The STRA data is a 
set of bilingual templates, which is built automatically 
from fixed English sentences and their Japanese transla- 
tion equivalents. The fixed English sentences are extracted 
automatically from the AP corpus. 
A series of experiments to evaluate ENTS using eco- 
nomic news in the AP news stories showed the translation 
accuracy was about 50 % higher than with our conven- 
tional rule-based MT method. 
1 Introduction 
We are developing an English-to-Japanese machine trans- 
lation (MT) system to produce real-time rough translations 
for Associated Press (AP) wire service news stories. With 
some news topics, troubles with fixed expressions lows 
the translation accuracy of the MT system. Economic 
news stories in particular are difficult to translate by con- 
ventional rule-based methods, because they contain many 
fixed expressions sharing two major characteristics: 
cl) The fixed expressions produce economics-specific 
syntactic structure. 
c2) Equivalents of the fixed expressions require Japa- 
nese economic jargons. 
These characteristics respectively cause two major bot- 
tlenecks for the conventional rule-based MT system: 
bl) General-purpose grammatical rules are not suffi- 
cient to yield correct analysis of economic news stories. 
(Simple addition of grammatical rules increases syntactic 
ambiguities.) 
b2) It is difficult to select the appropriate Japanese 
words for the translation. 
Actually, these problems reduce the translation accu- 
racy of our rule-based MT system to only 20%, which is 
too low for practical use. 
This paper presents a new English-to-Japanese MT sys- 
tem for economic news stories, which is called ENTS 
(Economic News stories machine Translation System), to 
process fixed expressions effectively. ENTS consists of 
three sequential processes (as shown in Fig. 1), based on 
the three basic types of economic news sentence. Process 1 
is a kind of example-based approach, while Processes 2 
and 3 are rule-based ones that differ in grammatical rules. 
This paper focuses mainly on Process 1, which is com- 
posed of fixed sentence translation, compound word trans- 
lation, fixed sentence translation data production and fixed 
sentence extraction, Fixed sentence translation data 
(STRA data), which is a kind of bilingual template, plays 
a key role in the fixed sentence translation. The STRA data 
is built automatically from fixed English sentences ex- 
tracted from a large corpus and their corresponding Japa- 
nese translations. 
input 
I Process 2 
Process 3 
Figure 1 
STRA : Fixed sentence translation CTRA : Compound word translation 
DTRA : Data production for STRA EXTRA : Fixed sentence extraction 
............................. .J 
An overview of ENTS 
1Currently he is working at ATR Interpreting Telecom- 
munications Research Laboratories. 
28 
Recently, several example-based MTs were proposed 
for processing fixed expressions \[Nagao841\[Sumita91 \]. 
Furuse proposed a cooperative method using tightly 
woven combination of example- and rule-based ap- 
proaches \[Furuse92\]. In contrast to their approach, we use 
the two methods independently. Therefore, the translation 
accuracy of our example-based method is guaranteed to be 
100%. 
Creating an example-based MT requires bilingual trans- 
lation data. Kaji proposed acquiring the bilingual transla- 
tion data from bilingual texts \[Kaji92\]. However that 
would require a complete syntactic analysis of bilingual 
texts. Our method is more robust, because it requires only 
a partial analysis. 
Section 2 describes some relevant features of economic 
news stories. In Section 3, we present an overview of 
ENTS. The following sections describe the fixed sentence 
translation method in ENTS, and the results of experi- 
ments using ENTS for AP economic news stories. 
2 Features of economic news sentences 
The AP delivers about 350 wire-service news stories a 
day, of which about 50 are concerned with economics. 
Each news story has its own title related to the contents. 
Because the titles on economic news stories are fixed, such 
stories can be selected easily. Most sentences in these 
economic news stories have fixed expressions comprised 
of compound words and/or collocations. 
Example of fixed expressions 
el) compound words 
"5 cents", "17.76 dollars per kilo", 
and "The U.S. dollar" 
e2) collocations 
"Malaysian tin closed at", "The U.S. dollar opened", 
and "as share prices rose" 
Based on the fixed expressions, the sentences in eco- 
nomic news stories, called economic sentences, are classi- 
fied into three types: 
Type I : Fixed sentences 
Example 1 
1-1) "In Kuala Lumpur, Malaysian tin closed at 17.76 
dollars per kilo, up 5 cents." 
1-2) "In Kuala Lumpur, Malaysian tin closed at 16.83 
dollars per kilo, up 19 cents." 
1-3) "In Kuala Lumpur, Malaysian tin closed at 16.40 
dollars per kilo, down 8 cents." 
Type II : Economics-specific sentences 
Example 2 
2-1) "The U.S. dollar opened slightly higher against the 
Japanese yen Tuesday morning in Tokyo, while share 
prices inched up." 
2-2) "The U.S. dollar drifted lower against the Japanese 
yen Wednesday morning, while share prices on the Tokyo 
Stock Exchange rose sharply." 
2-3) "The U.S. dollar opened higher against the Japanese 
yen in Tokyo Thursday, as share prices rose in early trad- 
ing." 
Type III : General sentences 
Example 3 
3-1 ) "Kagawa added, however, that the market still antici- 
pates a rising dollar." 
3-2) "Shigeru Sato, an analyst with Sanyo Securities, said 
the index fell some 65 points at one point in the afternoon, 
but last-minute arbitrage buying pulled it back up." 
3-3) "But Tobo said the market's basic sentiment remained 
bearish because of a lack of incentives to focus on." 
(1) Type I 
The sentences in Type I contain fixed expressions in 
which the words change a little form day to day. The parts 
of speech of the translation equivalents of these fixed ex- 
pressions are nouns in Japanese. For example, the transla- 
tion equivalents of compound words like "17.76 dollars 
per kilo" are nouns, as are those of "up" and "down"in 1-1, 
2, 3. The verb in a Type I sentence, such as "close" in ex- 
amples 1-1, 2, 3, is fixed. 
(2) Type II 
Although each sentence in Type II has a unique style with 
fixed expressions, there is a greater variety of fixed ex- 
pressions than that in Type I sentences. For example: 
"opened" and "drifted" 
or "slightly higher", "lower" and "higher" 
The parts of speech of their translation equivalents of 
these fixed expressions are verb or adjective in Japanese. 
Therefore, their translation equivalences require a produc- 
tion method of their inflections in Japanese generation 
process of MT. 
(3) Type III 
The Type IlI sentences have no features that make MT 
appropriate. Most of the general sentences are dealers' 
comments. 
3 Outline of ENTS 
ENTS consists of three translation methods corresponding 
to the types of economic sentences. ENTS processing fol- 
lows the flow in Fig. 2. 
Figure 2 ENTS flow chart 
(1) Process 1 
Process 1 translates fixed sentences (Type I) using bilin- 
gual templates that directly handle fixed expressions. 
(2) Process 2 
Process 2 translates sentences of Type II using a conven- 
tional rule-based approach with grammatical rules tuned 
to economic sentences obtained from two data worth of 
AP stories \[Aizawa93\]. The grammatical rules are built 
reflecting features of fixed expressions. These economics- 
specific grammatical rules total about 500, which is 1/5 of 
29 
the number of rules for general sentences. Therefore, there 
are few ambiguities in syntactic structure. 
(3) Process 3 
Process 3 translates those sentences not processed by 
Process 1 or Process 2. It is a rule-based MT with general- 
purpose grammatical rules. 
4 A translation method of fixed sentence 
In our translation method, STRA (a fixed Sentence 
TRAnslation method), the bilingual templates in which 
translation equivalents of the fixed expressions are repre- 
sented as variables are created using STRA data. That 
data is built automatically by DTRA (a Data production 
method for STRA) from fixed English sentences and their 
corresponding Japanese translations. The fixed English 
sentences are extracted automatically from a corpus by 
EXTRA (a fixed sentence EXTRAction method). CTRA 
(a Compound word TRAnslation method) plays a main 
role in STRA and DTRA. 
Fig. 3 visually summarizes the translation system. 
.......................... ,I i° ....................... 
::~-::':':iii 
STRA : Fixed sentence translation 
CTRA : Compound word translation 
DTRA : Data production for STRA 
................................. EXTRA : Fixed sentence extraction 
Figure 3 Fixed-sentence translation method 
4.1 Compound word translation (CTRA) 
The compound word translation module (CTRA) trans- 
lates compound words in fixed expressions \[Katoh91\]. In 
STRA and DTRA, CTRA is the main processing unit, 
while it is used in one step of analysis in Processes 2 and 3. 
In our MT system used in Processes 2 and 3, the CTRA 
step occurs between morphological and syntactic analyses 
as shown in Fig. 4. 
ICTR 
~ ~-SYni-aciic" v a ou~tpta i . i ~ Generation analys~s. 
~__ _ L _I 
dat~ 
Figure 4 Our rule-based MT system with CTRA 
CTRA extracts fixed expressions and defines their ap- 
propriate translation equivalents, the parts of speech and 
the semantic markers. For example, fixed expressions in 
example 1-1 are processed as: 
idiomatic translation part of semantic expression equivalents speech marker 
"17.76 dollars per kilo .... le~ ~ 17.76 ~')v" noun unit expression 
"5 cents" "5 ~ > I," noun unit expression 
In CTRA, English analysis is done by CHART parser 
based on CFG rules which represent fixed expressions. 
On the other hand, Japanese generation is not based on a 
rule-based method, but conducted by substituting the 
translation equivalents of the English words for variables 
in Japanese templates. Fig. 5 shows examples of CFG 
rules and their corresponding Japanese templates. Both 
these CFG rules and Japanese templates are named as 
CTRA data. 
English fixed expressions Japanese templates 
1: S--> UNTEXP r#1#\] 
2: UNTEXP--> UNTEXP PER UNIT \[- 1 #3##1#J 
3: --> NUMEXP UNIT \[#1##2#J 
4: UNIT --> "dollar", "cents", r b')l/.\] , I-~ 2/b / , 
"kilo", "yen", etc \[~ ~ J , \[\['qJ ... 
5: PER--> "per", "a" rJ 
6: NUMEXP--> "1", "12", etc \[-lJ , \[12J "" 
7: CMA--> "," \['J 
8: UPDW --> "up", "down" r T .:~ -TOj , F ¢¢ ey 2/,, j 
9: CITY --> "Kuala Lumpur", I-~' 7 ~ )1/2/"7"-- )l/J , r.~.j -.. 
"Tokyo", etc. 
where part of the #i# denotes the translation equivalent of the "~ 
ith symbol in the fight-hand of CFG rule, and \[-(null)_\] 
means the rule has no corresponding translation equivalent. ) 
Figure 5 Sample CTRA data 
4.2 Fixed sentence translation (STRA) 
The fixed sentence translation module (STRA) is an ex- 
panded CTRA with added CTRA data (named as STRA 
data) for translating not only fixed expressions but also 
fixed sentences. STRA data is produced automatically, as 
described in next section. 
An example of STRA data used to translate example 1- 
1 in Section 2, are shown in Fig. 6. 
English fixed expressions Japanese templates 
h S--> PATI CITY CMA PAT2 UNTEXP CMA UPDW UNTEXP 
\[#20~7 I..-,-- ~,'70-9'--J~l~, #8##7#©#5#"~f3"~/zJ 
2:PAT1-->"In" \[\] 
3: CITY --> "Kuala Lumpur" \[ ~ 7" "~ )P U 7*-- )I/J 
4: CMA --> "," \[J 
5:PAT2--> "Malaysian tin closed at" \[J 
6: UPDW --> "up" \[7 7 "7*\] 
7: UNTEXP -->"17.76 dollarsper kilo" \[ 1 ~ t'/17.76 I¢)l/J 
8: --> "5 cents" 15 q~ 7./b J 
(UNTEXP is obtained by CTRA) 
Figure 6 STRA data for example 1-1 in Section 2 
30 
At the top of the CFG rules in Fig. 6 is an English tem- 
plate of example 1-1, and its corresponding translation 
with variables is a Japanese template. The CFG rules are 
based not on English grammar but on an English sentence 
pattern, although they represent the word order of a fixed 
sentence. For example, "Malaysian tin closed at", which is 
arranged in one phrase, cannot usually be represented as 
one grammatical category according to English grammar. 
The STRA data is flexible in its ability to translate fixed 
sentences. For example, the STRA data shown in Fig. 6 
and the CTRA data in Fig. 5 can translate : 
1-4) "In Tokyo, Malaysian tin closed at 1941 yen per kilo, 
down 19 yen." 
into Japanese: 
1941H'd" O' ~" fz J 
because, the fixed expressions in the examples are 
matched: 
Kuala Lumpur <---> Tokyo 
17.76 dollars per kilo <---> 1941 yen per kilo 
up <---> down 
5 cents <---> 19 yen 
To make STRA data flexible, the words used in fixed 
expressions, such as "Kuala Lumpur", "Tokyo", "cents" 
and "yen", should be registered in CTRA data. These 
words are selected by hand, referring to frequently appear- 
ing fixed expressions collected from corpora. 
4.3 Data production for STRA (DTRA) 
A data production module for STRA (DTRA) builds 
STRA data automatically from English fixed sentences 
and their Japanese equivalent sentences. In DTRA, CFG 
rules are constructed by transforming English fixed sen- 
tences, and Japanese templates are made by replacing 
fixed expressions in their Japanese equivalent sentences 
with variables. DTRA's algorithm is as follows: 
\[DTRA's Algorithm\] 
STEP0 
Translate a fixed sentence wl...wn into Japanese by 
hand. 
STEP 1 
CTRA makes candidate variables for a bilingual tem- 
plate. 
STEP2 
Define weights for the candidates by the algorithm 
shown in Fig. 7. 
STEP3 
DP selects the optimal set of candidates. 
STEP 4 
Make CFG rule. 
STEPS 
Make Japanese template. 
for i :=0ton-1 do 
for j := i+l to n do 
if (there is a non-active edge including w i... wj.) 
& (there is an equivalent in the Japanese sentence.) 
then 
weight(i, j) = 3 ji 
else then 
weight(i, j) = 0 
Figure 7 Algorithm for calculating relative 
weights of positions 
In STEP 0, a fixed sentence wl...wn is translated into 
Japanese by hand. STEP 1 collects candidates for vari- 
ables in the Japanese sentence. Actually, the fixed sen- 
tence is analyzed by CTRA, and various fixed expressions 
are extracted as symbols (pre-terminal or terminal sym- 
bols) used in non-active edges. STEP 2 is to calculate the 
weights of the symbols by the algorithm shown in Fig. 7 to 
select an optimal set of fixed expressions. If the translation 
equivalent of a symbol exist in the Japanese equivalent 
sentence, its weight is defined according to the number of 
words in edges, otherwise it is zero. STEP 3 selects an 
optimal set of edges by calculating the maximum in sums 
of the weights between positions 0 and n by Dynamic 
Programming (DP). STEP 4 produces pre-terminal sym- 
bols for the word sequences not selected in STEP 3, and 
lines up the symbols in order of their appearance to make 
CFG rules. In STEP 5, each translation equivalent of the 
edges in the optimal set is replaced with a variable in the 
Japanese equivalent sentence to make Japanese templates. 
DTRA is illustrated by processing sentence 1-1. 
STEP 0 translates the sentence into Japanese by hand: 
7 "y 7*O') 1 ~ t~ 17.76 F')I,"C O'~?t:J 
The non-active edges obtained by CTRA in STEP 1 are 
shown in Fig. 8. 
CMA CMA---I r-- UPDW tnCrEXp // 
CITY ~ I UNTEXP II~i._~Xp I Ifl 
a Kuala Lumpur, Malaysian tin closed at 17.76 dollars per kilo, up 5 cents" 
I II II I I UI I 
NUMEXPUNIT PERUNIT #UNIT 
NIJNEXP 
Figure 8 Non-active edges in sentence 1-1 
(by CTRA) 
STEP 2 calculates the weights of the non-active edges 
as shown in Fig. 9. For example, the weight of "Kuala 
Lumpur" is 9. 
31 
81 
9 
0 f"-~ 3 O--1 2 3--405o6-£7 ° 231331 
In ...... 17.76 ...... kilo ...... cents 
Figure 9. Weights of non-active edges 
in sentence 1-1 
STEP 3 has DP select the maximum in sum of the 
weights between edge 0 and edge 16. In Fig. 9, the maxi- 
mum is 108 and the optimal set of edges is selected as 
{"Kuala Lumpur", ",", "17.76 dollars per kilo", ",", "up", 
"5 cents"}. 
In STEP 4, the word sequences not selected in STEP 3 
are given pre-terminal symbols automatically: 
"In" PAT1 
"Malaysian tin closed at" PAT2 
and setting these symbols in a line, the CFG rule is: 
S --> PATI CITY CMA PAT2 UNTEXP CMA UPDW UNTEXP 
( 1 2 3 4 5 6 7 8 ) 
The variables are defined as: 
#2#= rBg~., j 
#5# = \[ 1 ~ ~ 17.76 b")l.,J 
#7#= I-7 "~ 7"3 
#8# = 1-5-e :./b J 
STEP 5 replaces their translation equivalents of the se- 
lected edges in the Japanese sentence with variables: 
\[#2#~7 1t -- -5" Z a)~-'~, #8##7#¢)#5#'eU'~ 
tcJ 
4.4 Fixed sentence extraction (EXTRA) 
A method of extracting fixed sentences (EXTRA) collects 
fixed sentences for DTRA from a corpus using the fixed 
pattern ratio (FPR) defined below. 
The first step in EXTRA is to extract the fixed-word 
sequences which appear in a corpus most frequently, ig- 
noting differences of days of the week (e.g., Monday and 
Tuesday) and digits (e.g., 123 and 1000). The fixed-word 
sequences are not only compound words such as "\[DIGIT\] 
dollars per kilo"(where \[DIGIT\] denotes digits) and "on 
condition of anonymity", but also some parts of fixed ex- 
pressions, such as "said in a" and "condition of anonym- 
ity". The fixed-word sequences are called "fixed patterns" 
and the compiled fixed patterns are called "fixed pattern 
data". 
Using fixed patterns, FPR is defined as follows: 
sum of words in fixed sequences of a sentence 
FPR ........................................................... 
the total number of words in a sentence 
Example 
4-1) The NYSE's composite index rose 0.39 to 196.61. 
4-2) The NYSE's composite index edged up 0.33 to 
186.51. 
FPD1) "The NYSE's composite index rose \[DIGITI to 
\[DIGITI" (8 words), 
FPD2) "The NYSE's composite index"(4 words), 
FPD3) "\[DIGITI to IDIGITI"(3 words) 
FPD1, FPD2 and FPD3 are assumed to be in fixed pattern 
data. Thus, 
The FPR of example 4-1 = 8/8 =1.0, because 4-1 itself is 
FPD1. 
The FPR of example 4-2 = (4+3)/9 = 0.78, because 4-2 
includes FPD2 and FPD3. 
Fixed sentences are defined as those with FWR values 
above a certain threshold. EXTRA analyze each sentence 
in a corpus and extracts the sentences with sufficiently 
high FPR as fixed sentences. 
EXTRA has three parameters: 
P1) range of fixed patterns 
P2) frequency of fixed patterns 
P3) threshold of FPR 
5 Experiments 
5.1 Extracting fixed sentences 
The parameters for EXTRA were selected as: 
P1) 3 to 6 words in fixed patterns 
P2) more than 10 times 
P3) 0.8 
To satisfy conditions P1 and P2, about 92,000 fixed pat- 
terns were collected from AP wire-service news stories 
from a two-year period, which include about 1.6 million 
sentences. Using these fixed patterns, about 21,000 fixed 
sentences were extracted under the condition P3. The ex- 
periment was not limited to economic news stories. Ex- 
amples of the extracted results are shown in Appendix. 
Since most of the sentences are economic ones with many 
idiomatic expressions, EXTRA would be a good method 
enough to extract fixed sentences. 
5.2 Production of STRA data 
The 388 most frequently occurring economics-related 
fixed sentences were manually sampled from the 21,000 
fixed sentences. After manually translating them into 
Japanese, STRA data was produced by DTRA. 
While most of CFG rules in the STRA data include vari- 
ables, a few do not, such as for "Gold prices were mixed." 
The STRA data produced was as simple as: 
S --> PAT225 r#1 #\] 
PAT225 --> "Gold prices were mixed" 
The total number of symbols given in STEP 4, such as 
PAT1 and PAT2, are approximately 230. 
5.3 Experiment for ENTS 
A series of experiments was conducted using the STRA 
data discussed in Section 5.2 to evaluate the accuracy of 
ENTS. 
Table 1 and 2 show each process's volume and transla- 
tion accuracy, respectively for two data sets: Datal in- 
cludes 193 economic sentences used to tune to the CFG 
rules of Process 2, and data2 includes 167 sentences which 
were not used in the tuning. 
32 
datal 
data2 
Table 1 Processing volume (%) 
Process 1 
29.1 
31.2 
Process 2 
61.3 
57.5 
Process 3 Total 
9.6 100 
11.3 100 
\[Uratani91\] Uratani, N., Katoh, N. and Aizawa T.: "Ex- 
traction of Fixed Patterns from AP Economic News" Proc. 
of 42nd Annual Convention of IPSJ, 6E-4 (1991 ) ; in Japa- 
nese. 
Table 2 Translation accuracy (%) 
data I 
data2 
Process 1 
1130 
100 
Process 2 
70.1 
58.7 
Process 3 Total 
10.2 73.0 
22.2 66.3 
About 30% of each data set is translated in Process 1 
and its translation accuracy is 100% for both cases. The 
translation accuracy of Process 2 for data2 is so high as for 
datal, although Process 2 is not tuned to data2. The overall 
translation accuracy increases from about 20% with our 
conventional MT system to about 70%. 
6 Conclusion 
We described our new machine translation system 
(ENTS), which is an economics-specific MT system for 
processing fixed expressions. We focused mainly on the 
method of translating fixed sentences in ENTS. The re- 
sults of experiments show that translation accuracy in- 
creases from 20% with our conventional MT to 70% with 
ENTS. We conclude ENTS will be effective for translat- 
ing AP economic news stories into Japanese. 
The processing rate in Process 1 will be improved by 
increasing CFG rules in CTRA data and by collecting 
more fixed sentences and their translation equivalents in 
DTRA. Moreover we intend to apply the translation 
method to sports and general news stories. 
Appendix 
(FPR: extracted sentence) 
1.000: He did not elaborate. 
1.000: No injuries were reported. 
1.000; The U.S. dollar opened at 159.97 yen on the Tokyo 
foreign exchange market Monday, up from last Friday's 
close of 157.65 yen. 
1.000; The Federal Reserve Board's index measuring the 
value of the dollar against 10 other currencies weighted on 
the basis of trade was 97.46 Tuesday, off 0.74 points or 
0.74 percent from Monday's 98.20. 
0.970: The average price for strict low middling 1 1-16 
inch spot cotton declined 99 points to 78.64 cents a pound 
Wednesday for the seven markets, according to the New 
York Cotton Exchange. 
0.952: The Nikkei Stock Average closed at 25,194.10, 
down 48.30 points, or 0.19 percent on the Tokyo Stock 
Exchange Wednesday. 

References 

Aizawa, T., Katoh, N. and Kamata, M.: "Tuning of a Machine Translation System to Wire-Service 
Economic News", Proc. of PACLING-93, pp. 304-308 
(1993). 

Furuse, O. and Iida, H.: "Cooperation between 
Transfer and Analysis in Example-Based Framework", 
Proc. of COLING-92, pp. 645-651 (1992). 

Kaji, H. and Morimoto, Y.: "Learning Transla- 
tion Templates from Bilingual Text", Proc. of COLING-92, pp. 672-678 (1992). 

Katoh, N., Uratani, N. and Aizawa, T.: "Processing Proper Nouns in machine Translation for English 
News", Proc. of International Conference on Current Issues in Computational Linguistics, pp. 431-439 (1991). 

Nagao, M.: "A Framework of a Mechanical 
Translation between Japanese and English by Analogy 
Principle", in Elithorn, A. and R. Bemerji (eds.) Artificial 
and Human Intelligence, North-Holland, pp. 173-180 
(1984). 

Sumita, E. and lida, H.: "Experiments and 
Prospects of Example-Based Machine Translation", Proc. 
of ACL-91, pp. 185-192 (1991). 
