Corpus-Centered Computation  
 
Eiichiro SUMITA 
ATR Spoken Language Translation Research Laboratories 
2-2 Hikaridai, Seika, Souraku 
Kyoto 619-0288, JAPAN 
http://www.atr.co.jp/slt 
eiichiro.sumita@atr.co.jp 
 
Abstract 
To achieve translation technology that 
is adequate for speech-to-speech 
translation (S2S), this paper introduces 
a new attempt named Corpus-Centered 
Computation, (abbreviated to C
3
 and 
pronounced c-cube). As opposed to 
conventional approaches adopted by 
machine translation systems for 
written language, C
3
 places corpora at 
the center of the technology. For 
example, translation knowledge is 
extracted from corpora, translation 
quality is gauged by referring to 
corpora and the corpora themselves are 
normalized by paraphrasing or 
filtering. High-quality translation has 
been demonstrated in the domain of 
travel conversation, and the prospects 
of this approach are promising due to 
the benefits of synergistic effects. 
1 
                                                     
Introduction 
Text-based MT systems are not suitable for 
speech-to-speech translation (S2S) partly 
because they have not been designed to cope with 
the deviations from conventional grammar that 
characterize spoken language input and partly 
because they have been designed to be as general 
as possible to cover as many domains as possible. 
Consequently, the translation quality is not good
1
 
enough for S2S purposes. Furthermore, since 
such systems have been constructed by human 
experts, the development of machine translation 
 
2 
                                                     
1
 For our travel domain, a famous translation system 
on the WEB between Japanese and English produced 
a good translation for only about 10~20% of our test 
sentences. 
systems and porting them to different domains 
are expensive and snail-paced processes. 
This paper introduces a new attempt 
named Corpus-Centered Computation, 
(abbreviated to C
3
 and pronounced c-cube).  C
3
 
places corpora at the center of the technology, 
where, for example, translation knowledge is 
extracted from corpora, translation quality is 
gauged by referring to corpora, and the corpora 
themselves are normalized by paraphrasing or 
filtering.  
C
3
 has demonstrated its ability to provide 
high-quality translation. The construction is done 
by machine, allowing quick and low-cost 
development. 
Section 2 introduces the corpus we are 
currently dealing with, Section 3 briefly explains 
our three corpus-based machine translation 
systems, Section 4 demonstrates the first round 
of competition between the three systems on the 
same corpus, Section 5 touches on the automatic 
selection of the best translation, Section 6 
introduces a combination of corpus-based 
processes, such as translation and paraphrasing, 
Section 7 discusses the implications of our 
approach, and finally Section 8 concludes the 
paper.  
The Corpus 
We are aiming at the development of a S2S 
system to be used in place of a phrasebook by 
foreign tourists. Table 1 shows the English part 
of some sample translation pairs from our 
Japanese and English corpus. 
 Table 2 compares our corpus with two 
other spoken language corpora developed by 
ATR (Takezawa, T. et al., 2002) and Verbmobil 
(Ney, H. et al., 2000).
2
 Our corpus has the shortest 
 
2
 J, E, and G stand for Japanese, English, and German. 
                                            Association for Computational Linguistics.
                             Algorithms and Systems, Philadelphia, July 2002, pp. 1-8.
                          Proceedings of the Workshop on Speech-to-Speech Translation:
average sentence length. On the other hand, it is 
rich in topics and thus has the largest vocabulary 
and volumes.  
Table 1. Sample English sentences in our corpus 
I want to buy a roll of film. 
I’d like to reserve a table for eight. 
Do you have some tea?  
I’d like to return the car. 
You need to cross the bridge to go there. 
My friend was hit by a car and badly injured.
 
Table 2. Comparison with other bilingual spoken 
language corpora 
 ATR/ 
dialogue 
Our corpus Verbmobil
#Sent. 
16,110 204,108 58,332
#Word 
(J) 231,267 (J) 1,689,449 (G) 519,523
 
(E) 181,263 (E) 1,235,747 (E) 549,921
Voc. 
(J) 4,895 (J) 19,640 (G) 7,940
 
(E) 4,032 (E) 15,374 (E) 4,673
Length 
(J) 14.3 (J) 8.3  (G) 8.9 
 
(E) 11.3 (E) 6.1 (E) 9.4
 
Three Corpus-based Machine 
Translation Systems 
3 
3.1 
Research on corpus-based translation is a 
growing trend and has become indispensable to 
the MT industry.  
There are two main strategies used in 
corpus-based translation: 
1. Example-Based Machine Translation 
(EBMT): EBMT uses the corpus directly. 
EBMT retrieves the translation examples that 
are best matched to an input expression and 
adjusts the examples to obtain the translation 
(Nagao, 1984; Somers, 1999).  
2. Statistical Machine Translation (SMT): SMT 
learns models for translation from corpora 
and dictionaries and searches at run-time for 
the best translation according to the models 
(Brown et al., 1993; Knight, 1997; Ney et al., 
2000). 
 
We have developed two EBMT systems and one 
SMT system.  
An EBMT, D
3 
 
Sumita (2001) proposed D
3
 (Dp-match Driven 
transDucer).  The characteristics of D
3
 are 
different from previous EBMT approaches: a) 
Most EBMT proposals assume syntactic parsing 
and bilingual tree-banks, but D
3
 does not; b) 
Most EBMT proposals divide the translation 
process in two, i.e. learning of translation 
patterns in advance and application of the 
translation patterns, but D
3
 generates translation 
patterns on the fly according to the input and the 
retrieved translation examples as needed.  
As shown in Figure 1, our language 
resources are [i] a bilingual corpus, in which 
sentences are aligned beforehand; [ii] a bilingual 
dictionary, which is used for generating target 
sentences; and [iii] thesauri of both languages, 
which are used for incorporating the semantic 
distance between words into the distance 
between word sequences. Furthermore, [ii] and 
[iii] are also used for word alignment. 
 
Generate 
Select 
Substitute 
 
Retrieve Aligned
Bilingual 
Corpus
[i]
Sentence
[ii] 
Bilingual 
Dictionary
[iii]  
Thesauri
Input 
sentence 
Target 
sentence 
Figure 1. Configuration 
Suppose we are translating a Japanese sentence 
into English. Let’s review the process with a 
simple sample below. The Japanese input (1-j) is 
translated into English (1-e) by utilizing (2-e), 
whose source (2-j) is similar to (1-j). The 
common parts are unchanged, and the different 
portions are substituted by consulting a bilingual 
dictionary. 
 
;;;A Japanese input 
(1-j) iro/ga/ki/ni/iri/masen  
     
;;; Japanese part of an example in corpus [i]  
(2-j) dezain/ga/ki/ni/iri/masen 
;;; English part of an example in corpus [i]  
(2-e) I do not like the design.  
 
;;; the English output  
(1-e) I do not like the color. 
 
We retrieve the most similar source sentence of 
examples from a bilingual corpus. For this, we 
use DP-matching, which tells us the distance 
between word sequences, dist while giving us the 
matched portions between the input and the 
example. According to equation [1], dist, is 
calculated. The counts of the Insertion (I), 
Deletion (D), and substitution operations are 
summed. Then, this total is normalized by the 
sum of the lengths of the source and example 
sequences. According to equation [2], 
substitution considers the semantic distance, 
SEMDIST, between two substituted words.  
 
[1] 
exampleinput
LL
SEMDISTDI
dist
g43
g43g43
g61
g229
2
 
 
[2] 
N
K
SEMDIST g61  
 
Figure 2 illustrates the SEMDIST calculation 
between “potato” and “beef.” The most specific 
common abstraction of the two words is 
“ingredients,” as shown in boldface. The 
SEMDIST is K divided by N in the figure. K is 
the level of the most specific common abstraction 
of two words, and N is the height of the 
thesaurus. 
 
 
fruit
apple carrot potato beef chicken orange
vegetable meat 
ingredients
food
TOP
K
N
most specific common abstraction 
 
Figure 2. SEMDIST calculated using thesaurus 
 
Let us show the latest performance of D
3
. 
The translation speed is sufficiently fast in that 
the average translation time is 0.04 seconds per 
sentence with the 200 K corpus shown in Section 
2.      The translation quality is so high that the 
method can achieve a TOEIC score
3
 of 750. This 
is equivalent to the average score of a Japanese 
businessperson in overseas department of 
Japanese corporations. 
                                                      
3.2 
3
 The TOEIC
 
(Test of English for International 
Communcation) test is an English language 
proficiency test for people whose native language is 
not English (http://www.chauncey.com/). The Total 
score ranges from 10 to 990. ATR has developed a 
method to measure the TOEIC score of a machine 
translation system. (Sugaya et al., 2000) 
In brief, D
3
 uses DP-matching, which 
features the semantic distance between words. D
3
 
has demonstrated good quality and short 
turnaround in travel conversations such as those 
in a phrasebook. 
EBMT and SMT based on Hierarchical 
Phrase Alignment (HPA) 
Here we intoroduce Hierarchical Phrase 
Alignment (HPA) and its application to EBMT 
and SMT. 
3.2.1 Hierarchical Phrase Alignment (HPA) 
This subsection introduces a new phrase 
alignment approach (Imamura, 2001) called 
Hierarchical Phrase Alignment (HPA). 
 
 
I have arrivejus in New York
ni desu
NewYork
tatsui bakari 
VP 
VP 
(3)
(3)
VP
AUXVP
S
VP
AUXVP 
S
(4)
(5)
(6)
(4)
(5)
(6)
NP
VMP
NP
(1)
(2)
(1)
(2)
VMP
 
Figure 3. Bilingual trees and alignment 
 
First, two sentences are tagged and parsed 
independently. This operation obtains two 
syntactic trees. Next, words are linked by the 
word alignment program. Then, HPA retrieves 
equivalent phrases that satisfy two conditions: 1) 
words in the pair correspond with no deficiency 
and no excess; 2) the phrases are of the same 
syntactic category. 
Let’s look at a sample pair of a Japanese  
                                                                                
tree and the corresponding English tree (Figure 
3). The retrieval of equivalent phrases is done in 
a bottom-up fashion. First, the syntactic node pair 
that consists of only the ‘New York’ and 
‘NewYork’ link, having the same syntactic 
category, is retrieved. Then, NP(1) and VMP(2) 
are found. Next, the syntactic node pair that 
consists of only the ‘arrived’ and ‘tsui’ link, 
having the same syntactic category, is retrieved. 
Then, VP(3) is found. Finally, the syntactic node 
pairs that include two word links having the same 
syntactic category are retrieved. Then VP(4), 
AUXVP(5), and S(6) are found. Accordingly, six 
equivalent phrases are hierarchically extracted. 
3.2.2 EBMT based on HPA 
Imamura (2002) proposed an application of HPA 
in EBMT called HPA-based translation (HPAT). 
HPAed bilingual trees include all information 
necessary to automatically generate transfer 
patterns. Translation is done according to transfer 
patterns using the TDMT engine (Sumita et al., 
1999), our previous EBMT system. First, the 
source part of transfer patterns are utilized, and 
source structure is obtained. Second, structural 
changes are performed by mapping source 
patterns to target patterns.  Finally, lexical items 
are inserted by referring to a bilingual dictionary, 
and then a conventional generation is performed. 
HPAT achieved about 70% accuracy. 
3.2.3 SMT based on HPA 
Statistical machine translation (SMT) represents 
a translation process as a noisy channel model 
that consists of a source-channel model and a 
language model of the target language. 
The translation model is based on 
word-for-word translation and limited to allow 
only one channel source word to be aligned from 
a channel target word. Although phrasal 
correspondence is implicitly implemented in 
some translation models by means of distortion, 
careful parameter training is required. 
In addition, the training procedure relies 
on the EM algorithm, which can converge to an 
optimal solution but does not assure the global 
maximum parameter assignment. Furthermore, 
the numbers of parameters represent the 
translation models, so that easily suffered from 
the over-fitting problem. In order to overcome 
these problems, simpler models, such as 
word-for-word translation models (Brown et al., 
1993) or HMM models (Och et al., 2000), have 
been introduced to determine the initial 
parameters and to bootstrap the training. 
We have proposed two methods to 
overcome the above problems by using HPA. (1) 
The first method converts the hierarchically 
aligned phrasal texts into a pair of sequences of 
chunks of words, treating the word-for-word 
translation model as a chunk-for-chunk 
translation model. (2) The second method 
computes the parameters for the translation 
model from the computed phrase alignments and 
uses the parameters as a starting point for training 
iterations. 
The experimental results on 
Japanese-to-English translation indicated that the 
model trained from the parameters derived from 
the HPA could improve the quality of translation 
(Watanabe et al., 2002a).  
4 
4.1 
4.2 
Competition between the Three 
MTs on Same Corpus 
Competition Conditions 
We used the corpus shown in Section 2, which is 
a collection of Japanese sentences and their 
English translations, typically found in 
phrasebooks for foreign tourists. We lemmatized 
and POS-tagged both the Japanese and English 
sentences. A quality evaluation was done for the 
test set consisting of 510 sentences selected 
randomly from the above corpus, and the 
remaining sentences were used for learning and 
verification.  
We also used a bilingual dictionary 
previously developed for TDMT. The size of the 
dictionary is 24,658 words. We used thesauri 
whose hierarchies are based on the Kadokawa 
Ruigo-shin-jiten (Ohno and Hamanishi, 1984). 
The size of the Japanese thesaurus is 21,608 and 
that of the English thesaurus is 11,359. 
Results 
SMT has been applied to language pairs of 
similar European languages. We implemented 
SMT for translation between Japanese and 
English, which are dissimilar in many points 
such as word order. Table 3 shows the accuracy 
of our SMT system. The four ranks are defined as 
follows (Sumita et al., 1999): (A) Perfect: no 
problems in both information and grammar; (B) 
Fair: easy-to-understand, with either some 
unimportant information missing or flawed 
grammar; (C) Acceptable: broken, but 
understandable with effort; (D) Nonsense: 
important information has been translated 
incorrectly. It worked in both J-to-E and E-to-J 
directions
4
 in spite of the negative opinions 
previously expressed.  
Table 3. SMT worked for J and E 
Rank(s) A A+B A+B+C 
SMT(JE) 25% 46% 64% 
SMT(EJ) 41% 48% 57% 
 
We implemented two EBMT systems, D
3
 
and HPAT, using the same corpus. D
3
 and HPAT 
surpassed SMT in the travel conversation task 
(Tables 3 and 4). 
Table 4. EBMTs on the same corpus 
Rank(s) A A+B A+B+C 
D
3
(JE) 47% 66% 77% 
HPAT(EJ) 50% 61% 71% 
 
Finally, it became clear that word-based 
SMT, a revival of the direct method of the ’50s, is 
suitable for pairs of European languages but not 
for Japanese and English. This is because 
word-based SMT cannot capture the major 
differences such as word order between Japanese 
and English.  
Several organizations (Yamada et al., 
2001; Alshawi et al.,  2000) are pursuing 
syntax-based SMT. We plan to join the race. 
Which is suitable for Japanese and English, 
syntax-based SMT or EBMT? 
5 
5.1 
                                                     
Combination of Evaluation and 
Translation 
We are researching automatic evaluation of 
machine translation outputs and multiple 
paradigms for machine translation 
simultaneously. Together, they have synegistic 
effects as explained below. 
Automatic Quality Evaluation Using 
Corpus 
 
5.2 
4
 For this test set, the accuarcy of SMT is at least 
twice as good as that of a famous conventional 
machine translation system on the WEB. 
Translation quality has conventionally been 
evaluated by hand. Likewise, we have evaluated 
the outputs of our translation systems 
subjectively with four ranks from ‘good’ to 
‘bad’: A, B, C, and D (Sumita et al., 1999). 
Such subjective evaluation by ranking, 
however, is taxing on both time and resources. If 
automatic evaluation methods are inexpensive, 
fast, and sufficiently accurate, then such 
automatic evaluation methods would prove 
beneficial. 
Conventional approaches to automatic 
evaluation include methods (Su, 1992; Yasuda et 
al., 2001) that automatically assign one of several 
ranks  to MT output according to a single edit 
distance between an MT output and a correct 
translation example. 
To improve performance, we proposed 
an automatic ranking method that, by using 
multiple edit distances, encodes 
machine-translated sentences with a rank 
assigned by humans into multi-dimensional 
vectors from which a classifier of ranks is learned 
in the form of a decision tree. The proposed 
method assigns a rank to MT output through the 
learned decision tree (Akiba et al., 2001). 
Experimental results show that the 
proposed method is more accurate than the 
single-edit-distance-based ranking methods in 
both closed and open tests. The proposed method 
has the potential to accurately estimate the 
quality of outputs of machine translation 
systems. 
Multiple-engine Machine Translation 
System 
Every researcher has his own way of acquiring 
translation knowledge by generalizing translation 
instances in a corpus. Our approach is no 
exception to this rule. Our MTs are based on 
different paradigms, different development styles, 
and different development periods. This results 
in various behaviors for each input sentence, and 
the translation rank of a given input sentence 
changes system-by-system.  
Table 5. Sample of transaltion variety with quality 
rank 
o-shiharai wa genkin desu ka kurejitto kaado desu ka
[B] Is the payment cash? Or is it the credit card?  
[A] Would you like to pay in cash or with a credit card? 
[C] Could you cash or credit card? 
 
We show a sample of different English 
translations obtained by three systems for a 
Japanese sentence (Table 5). The brackets show 
the quality rank judged by a human translator.  
Translation systems gain A-ranked 
translations in different subsets of input 
sentences as illustrated in Figure 4. Thus, we 
could obtain a large increase in accuracy by using  
an “ideal” MT, if it were possible to select the 
best one of the three different translations for 
each input sentence. 
 
 
MT1 
MT2 MT3 
 
Figure 4. Subsets of input sentences whose 
translation is A-ranked 
We are investigating methods to utilize 
techniques of automatic evaluation for selector 
(Figure 5).  
 
 
 MT1
 
MT2 
MT3 
Selector 
 
Figure 5. Selector for multi-engine MT 
In our pilot experiment, our selectors (Akiba et 
al., 2002; Yasuda et al., 2002) outperformed not 
only the component systems but also a 
conventional selector using N-gram 
(Callison-Burch et al.,  2001). 
Combination of Paraphrasing 
and Translation 
6 
6.1 
We are automating extraction of paraphrase 
knowledge from a bilingual corpus. In this 
section, we introduce its application to improve 
the performance of corpus-based translation by 
using SMT as a touchstone.  
Extraction of Synonymous Expressions 
We propose an automatic paraphrasing method 
that exploits knowledge from bilingual corpora 
(Shimohata et al., 2002). 
Synonymous expressions are defined as a 
sequence of variant words with surrounding 
common words. The expressions are extracted 
from bilingual corpora by the following 
procedures (Figure 6): 
1. Collect sentences that share the same 
translation in another language. The 
accumulated sentences are defined as 
synonymous sentences. 
2. For all pairs of synonymous sentences, apply 
DP-matching and collect sequences of words, 
synonymous expressions that consist of 
variant words preceded/followed by common 
words. 
3. Remove pairs of synonymous expressions 
with a frequency lower than a given 
threshold. 
4. Cluster the pairs of synonymous expressions 
by transitive relation.  
 
 
<s> how would you like … </s> g243 <s> … wa dou nasai masu </s> 
<s> how long will … </s> g243 <s> dore kurai … </s>  
<s> how much … </s> g243 <s> ikura … </s>  
<s> would you like … </s>
<s> do you like … </s> 
<s> what would you like … </s> 
<s> how do you like … </s> 
<s> would you like …
   |             |     | 
<s> do      you like …  
 
<s>         would you like …  
   |             |       |     | 
<s> what would you like …
<s> would you g243<s> do you   
<s> would       g243<s> what would you 
<s> would you g243<s> how do you   
<s> would you g243<s> how do you g243<s> do you 
Cluster of synonymous expressions 
Synonymous expressions
DP-matching
Synonymous sentences 
Bilingual corpus
g243 <s>…wa ikaga desu ka </s>
 
Figure 6. Extraction of synonymous expressions 
 
6.2 Corpus Normalization 
After the acquisition of clusters of synonymous 
expressions, normalization is carried out by 
transforming the expressions into major ones, 
selected according to their frequency in the 
corpora. For instance, the cluster obtained 
consists of the expressions ‘<s> would you,’ ‘<s> 
how do you’ and ‘<s> do you.’ Suppose that an 
expression ‘<s> do you’ occurred most 
frequently in a given corpus, an input ‘how do 
you like your coffee’ could be normalized into 
‘do you like your coffee.’ 
6.3 
7 
7.1 
7.2 
8 
SMT on Normalized Corpus 
Statistical approach to machine translation 
demands huge bilingual corpora in good quality 
and broad coverage. However, such an ideal 
corpus is not usually available: one may contain a 
sufficiently large number of samples, for instance, 
derived from web pages with translations, but 
these may not be well-aligned or have low 
translation quality. Others may consist of 
high-quality translations but have a limited 
number of examples. In addition, the variety of 
possible translations makes it even harder to 
estimate parameters for statistical-based machine 
translation. 
We propose a way to overcome these 
problems by creating a corpus that consists of 
normalized expressions, expressions with less 
variety, through automated paraphrasing 
(Watanabe et al., 2002b). As described above, by 
the method of transforming target sentences of a 
given bilingual corpus into a normalized form is 
expected to improve the parameter estimation for 
a statistical machine translation model. The 
normalization method proposed above locally 
replaces word sequences, hence will not affect 
the syntactical coherence. Therefore, 
normalization will not affect the distortion model, 
which accounts for reordering of bilingual texts. 
In addition, reduction of the vocabulary size will 
greatly help improve the parameter estimation 
for lexical models. 
The experimental results on 
Japanese-to-English translation indicated that the 
SMT created on the target normalized sentences 
reduced word-error-rate from 66% to 58%.  
Discussions 
Forecasting from the Obtained 
Performance  
As a component of C
3
, D
3
 has achieved a high 
TOEIC score. We foresee much higher scores for 
C
3
 because it features a multi-engine and selector 
scheme, which is an easy, quick and low-cost 
method of improving total performance, since 
there is no need to investigate the messy 
relationships between resources and processes of 
the component systems by hand.  
Backcasting from the Future S2S System in 
the Real World 
We are aiming to develop technologies for S2S 
that are usable in real-world environments. No 
one knows what the real world will be, but there 
is no doubt that an S2S system should deal with 
variations in length and expressions beyond our 
corpus that explained in Section 2. In other words, 
we divided our “real-world” goal into three 
sub-goals; (1) translation of short and edited 
sentences; (2) translation of long sentences; (3) 
translation of short but non-edited sentences; and 
(4) combining solutions for these sub-goals 
seamlessly. 
 Since we are centering our approaches on 
corpora, we are developing corpora for achieving 
sub-goals (1), (2) and (3) as reported in 
(Takezawa et al., 2002; Sugaya et al., 2002) . 
 For sub-goal (1), we are using a selector 
for multiple engines, for sub-goal (2), we have to 
devise methods to chunk long sentences into 
appropriate translation units, and for sub-goal (3) 
we need a powerful automatic paraphraser.  
Conclusions 
Our attempt called C
3
 places corpora at the center 
of S2S technology. All components of C
3
 are 
corpus-based as shown in the paper. If we have 
sufficient volumes of sentence-aligned bilingual 
corpora, we would be able to build a high-quality 
MT. Corpus-based processes for such tasks as 
translation, evaluation, and paraphrasing have 
synergistic effects. Therefore, we are optimistic 
about the progress of components and their 
integration in C
3
. 
Acknowledgements 
The author’s heartfelt thanks go to Kadokawa-Shoten 
for providing the Ruigo-Shin-Jiten. The research 
reported here was supported in part by a contract with 
the Telecommunications Advancement Organization 
of Japan entitled, "A study of speech dialogue 
translation technology based on a large corpus." 

References 
Akiba, Y., Imamura, K. and Sumita, E.  2001 
Using multiple Edit Distances to automatically 
rank machine translation output, Proc. of 
MT-SUMMIT-VIII 
 
Akiba, Y., Watanabe, T. and Sumita, E.  2002 
Using Language and Translation Models to Select 
the Best among Outputs from Multiple MT 
systems, Proc. of Coling (to appear) 
 
Alshawi, H., Bangalore, S. and Douglas, S. 2000 
Learning Dependency Translation Models as 
Collections of Finite-State Head Transducers, 
Computational Linguistics, 26 (1), pp. 45--60. 
 
Brown, P., Cocke, J.,  Della Pietra, S. A.,  Della 
Pietra, V. J., Jelinek, F., Laffetry, J. D., Mercer, R. 
L. and Roossin, P. S.  1993 A Statistical Approach 
to Machine Translation, Computational 
Linguistics 16, pp. 79--85 
 
Callison-Burch, C. and Flournoy, S.  2001 A 
Program for Automatically Selecting the Best 
Output from Multiple Machine Translation 
Engines, Proc. of MT-SUMMIT-VIII  
 
Imamura, K. 2001 Hierarchical phrase alignment 
harmonized with parsing, Proc. of NLPRS, pp. 
377--384 
 
Imamura, K. 2002 Application of Translation 
Knowledge Acquired by Hierarchical Phrase 
Alignment, Proc. of TMI 
 
Knight, K. 1997, Automating Knowledge 
Acquisition for Machine Translation, AI 
Magazine, 18 (4), pp. 81--96 
 
Nagao, M. 1984 A Framework of a Mechanical 
Translation between Japanese and English by 
Analogy Principle, in A. Elithorn and R. Banerji 
(eds), Artificial and Human Intelligence, 
Amsterdam: North-Holland, pp. 173--180. 
 
Ney, H., Och, F. J. and Vogel, S. 2000 Statistical 
Translation of Spoken Dialogues in the Vermobil 
System, Proc. of MSC2000, pp. 69--74.  
 
Och et al. 2000. Improved statistical alignment 
models. In Proc. of ACL, pp. 440--447, Hong 
Kong, China, October 
 
Ohno, S. and Hamanishi, M. 1984. 
Ruigo-Shin-Jiten, Kadokawa, Tokyo (in Japanese) 
 
Shimohata, M. and Sumita, E. 2002 Automatic 
paraphrasing based on parallel corpus for 
normalization, Proc. of LREC 
 
Somers, H. 1999 Review Article: Example-based 
Machine Translation, Journal of Machine 
Translation, pp. 113--157 
 
Su, K. -Y, Wu, M. -W. and Chang, J. –S. 1992 A 
new quantitative quality measure for machine 
translation systems, Proc. of Coling, pp. 433--439  
 
Sugaya, F., Takezawa, T., Yokoo, A., Sagisaka, Y, 
and Yamamoto, S. 2000 Evaluation of the 
ATR-MATRIX Speech Translation System with a 
Pair Comparison Method Between the System and 
Humans, Proc. of ICSLP, pp. 1105--1108  
 
Sugaya, F. et al. 2002 Proposal for a very 
large-corpus acquisition method by registering in 
tree-structure form, Proc. of LREC  
 
Sumita, E. 2001 Example-based machine 
translation using DP-matching between word 
sequences, Proc. of DDMT (ACL), pp. 1--8 
 
Sumita, E., Yamada, S., Yamamoto, K., Paul, M., 
Kashioka, H., Ishikawa, K. and Shirai, S.  1999 
Solutions to Problems Inherent in 
Spoken-language Translation: The 
ATR-MATRIX Approach, Proc. of MT Summit 
VII, pp. 229--235 
 
Takezawa, T. et al. 2002 Toward a 
Broad-coverage Bilingual Corpus for Speech 
Translation of Travel Conversations in the Real 
World, Proc. of  LREC 
 
Yamada et al. 2001 A Syntax-Based Statistical 
Translation Model. Proc. of ACL, France 
 
Yasuda, K., Sugaya, F., Takezawa, T., Yamamoto, 
S. and Yanagida, M. 2001 An automatic 
evaluation method of translation quality using 
translation answer candidates queried from a 
parallel corpus, Proc. of MT-SUMMIT-VIII 
 
Yasuda, K., Sugaya, F., Takezawa, T., Yamamoto, 
S. and Yanagida, M. 2002 Automatic Machine 
Translation Selection Scheme to Output the Best 
Result, Proc. of LREC 
 
Watanabe, T., Imamura, K. and Sumita, E. 2002a 
Statistical Machine Translation Based On 
Hierarchical Phrase Alignment, Proc. of TMI 
 
Watanabe, T., Shimohata, M. and Sumita, E. 
2002b Statistical Machine Translation Based On 
Paraphrased Corpora, Proc. of LREC 
