":t ¢'1 o Ii' - ~ o ¢ eedback of Correcting :~librmafion ~u I osteditmg 
h~ a Machhte Translath~n 8ys{;em 
Fujio NIStIIDA, Shinobu TAKAMATSU, Tadaaki TANI and Tsunehisa DOI 
Department of Electrical Enqineering~ 
Faculty of Engineering~ University of Osaka Prefecture 
M ozu, Sakai~ Osaka~ 591~ Japan 
Abstract 
This paper presents an attempt to construct a feed- 
back systenr PECOF which improves a Japauese- 
English Machine Translation system by feedback of 
correcting information given by posteditors. PECOF 
analyzes the error-correcting information by using an 
English-Japmmse Machine Translation system which 
works in the reverse direction to the original MT sys~ 
tern, corot)ares the intermediate expressions of the 
corrected patterns with those of the erroneous parts of 
the originai MT output at every transfer stage and 
idenLifies the responsible parts of the original Japanese- 
English M'P system. Then PECOF corrects the 
irrelevant parts of tt,e database or adds error correcting 
patterns to a document of postediting to ask users for 
further exmninations for corrections. 
L Introduction 
Iv_ recent years, studies of machine translation have 
becn rapidly developed and tend to be put to practical 
use in various specific fields. (0 However, it is expected 
that the output sentences from machine translation 
systems need post-editing, more or less, over a long 
years for practical use as seen in the report on the prae- 
eical experience of the Systrau Machine Translation sys- 
tem and the posteditors' experience. 0)-0) 
As can be easily seen, feedback of correcting informa- 
tion given by posteditors to the original MT system 
undoubtedly will bring a remarkable improvement of 
the translation proficiency to the system itselfi IIow- 
ever, it does not seem that adequate discussion about 
feedback of the information for improving of the MT 
system has been done so far. One of the main reasons 
will be that it is difficult to identify the part to be 
corrected in the MT system only by using brief correct- 
ins intbrmation. 
This paper presents an attempt to identify the 
responsible parts of a machine translation system in the 
case of Japanese to English translation. The part to be 
corrected in the MT system is identified by applying an 
English-Japanese Machine Translation system to the 
postedited output in the reverse direction to the orlgi- 
nal translation. Tire English-aapauese MT system is 
assumed here to be capable of mmlyzing and 
comprehending the postedited output at least by using 
the fnndamentM and the general linguistic knowledge of 
the ~arget language. Associated with the assumption, 
476 
many of erroneous patterns arise due to lack of infbrm~-- 
tion about specific usages of words rather tha.u h~(:k of 
general linguistic information° Accordingly, such kind of 
correcting information given by posteditors can be 
understood from the general linguistic information. 
PEOOF(abbreviation of a PostEditing COrrecting 
inforrnation Feedback system) analyzes the correcting 
information by using the English-aapaxtese MT ~ystem 
and tries to perform feedback of it to tim MT systems. 
In the following sections~ the basic idea and the con.. 
struction of PECOF axe described in some details mid 
correcting of typical error patterns is illustrated with 
some examples. 
2. The principle of systein construction 
MTo 
Word Structure Transfer Transfer 
So~ 8Io ~ Tlol ............. ~' T!r°2 "-~To 
Sr~.. ~/'I'r Sir • ~ Tlrl <- ........... Tit2 
Word Structure Treamfer Transfer 
MTr 
Fig.l Machine translation sy,ztems MTo and MTr 
Let us assume that the MT system is constructed 
based on a trm, sfer system. As shown in Fig.l, the orl- 
ginal machine translation system MTo parses a block of 
source language sentences So, constructs the intermedi- 
ate expression or form SIo,transforms it to the tin-get 
language intermediate fornr TIol and Tlo2 by word and 
structure transfer and finally generates a block of the 
tat'get sentences To. 
In order to identify the kind of corrections given in ~ 
ta~rget lax~guage expression, PECOF needs ~z ~imple t',y'a- 
tactic mid lexical mmlyzer of the target language ~d 
least. If a part of the target sentences is corrected trod 
yet Tlo2 remains unchanged except ibr some syntactic 
term expressions, a syntactic rule corresponding to ~omc 
specific word usages of the tat'get language is checked~ 
the inappropriate paact of syntactic rules is identified 
and modified according to the correctioil given by po~. 
editor~. 
If the correction spread~ all over the par~ of the oraL. 
pat sentences or is nmde by entire replacement~ a p~xs.. 
i~g system of the target language is needed. Further° 
~nox'e, if t.ome wrong trmisfers axe made over several 
transt~r stages, an MT system which works in the 
reverse db:ection is much required to identify them. It is 
called the reverse MT system briefly mid denoted with 
MTr. 
In ~imilar to the original MT system, the reverse MT 
system p~rses the corrected target language expression 
':Cr, const,:ucts the intermediate form Tit2 and TIrl, 
~ransfortu,q Tit1 to the intermediate form Sir of the 
source lmlguage and generates the source language 
expressiox~ Sr which should be almost the same as the 
original :~ouree language expression So. The order of 
word trmmfer and structure transfer in the reverse MT 
systeru is reverse to that of the original MT system. 
The relation between both the MT systems is illustrated 
in Fig.1. The ease system, the semantic category sys- 
tem and *;he intermediate form in the reverse MT sys- 
tem are the same as those of the original MT system. 
The intermediate form in both the systems consists of 
several p,firs of a case label attd a term having the 
semantic .~nd the syntactic category name in option. 
P~COF analyzes post-edited output by using the 
reverse ~\[T system and searches for the first transfer 
stage in which the intermediate form of the corrected 
output obtained by the reverse MT system differs from 
that of the original MT system. 
After identifying the word block to be corrected mid 
£he kind of correction PECOF tries to correct the 
corresponding part of the word dictionary and the 
transfer xules. Various methods from documentation to 
automatic correction can be considered. From a practi- 
cM point of view, it will be efficient to give various error 
parmesans through the intermediate forms and 
correspo,dingly to provide the correcting procedure of 
the database of the original MT .~y.ql.em, If unknown 
error patterns occur, PECOF only classifies the pat- 
terns to ask the posteditors about the correcting 
method. 
AKer processing the wrong parts in the current 
transfer ~tage, PECOF updates the intermediate forms 
of the following stages of the original MT system output 
based on the corrections performed in the current 
stage. When some discrepancies still remain between 
~:he interlnediate forms of the original output and those 
o| the postedited output , PECOF applies the same 
correcting procedures to the following transfer stages 
repeatedly. 
The typical error patterns are classified into three 
classes. They are related to syntactic structure of tar- 
get language, structure transfer and word transfer. In 
section 4, some types of the error patterns mid the 
corresponding correcting procedures arc described. 
$o System construction and correcting information 
feedback 
In tMs section, the construction and the function of 
P~COF as well as those of both the MT systems are 
dezc:cib¢~'d in some details° 
~0:~.o Ovt:rvlew of the MT sysiems 
iFigul'e 2 shows a schematic construction of our MT 
systems and the database. The systems belong to a 
kinds of trmlsfcr systems. (s) The body part of the pro- 
cessing, MAPTRAN, is divided into parsing, transfer 
mid generation. Three kinds of dictionaries are imple- 
mented for word transfer, rewriting rules and structure 
transfer. 
MAPTRAN 
parallel bottom-up parsing 
word and structure transfer 
target sentence generation 
database 
a word-transfer dictionary 
a rewrltlng-rule dictionary 
a structure transfer dictionary 
a semantic category table 
Fig.2 Construction of the MT systems and tile datable 
Both the MT system from Japanese to English and 
that of the reverse direction use the same processing 
system MAPTRAN. (7) It is constructed on a hierarchi- 
cal module basis and can be expanded into some com- 
puter languages such as C and LISP. (s) It parses the 
input sentences in a parallel bottom-up manner. 
The word transfer dictionary used here is con- 
structed by combining an original word transfer diction- 
axy with a source language word dictionary for analysis 
and a target language word diction~y for generation. 
The data structure of the dictionary can be semi- 
automatically transformed to an appropriate form 
corrcspondlng to the change of the programming 
language. 
3.2. Designation of correction 
In order to designate the location to be corrected, a 
number is attached to each word in the output sere 
tences. A word and a word sequence ca~l be designated 
by a number or by a pair of the first word number ul 
and the last word number n2 like nl-n2. Replacement 
of words is designated as follows: 
replace ~ ~2-~ by " ~t new word sequence " (1) 
Insertion and deletion can be indicated in a similar 
manner. 
Movement of a word group ranging from the word 
nmnber nl to n2 to the front of the word of the nmnber 
n3 is designated as follows: 
move nl-n2 to n3. (2) 
For a given correction, it will be not sometimes easy 
for PECOF to identify the key item to be corrected in 
the dictionary when the original MT system does not 
have the sufficient linguistic knowledge and has a lot of 
possible reasons for correction. In such cases, it will be 
effective that PECOF is informed of the key informa- 
tion of correction by posteditors. 
However, detailed descriptions are laborious for post- 
editors and hard for PECOF to comprehend. One way 
to solve this problem is to indicate one or two words 
that conflict with the words to be corrected in the out- 
put sentences or phrases. Besides, it is sometimes desir- 
able to add some words that stand for the kind of 
correction. The designation is given in a form where 
the reason is added to correction in option. For exam.- 
pie, the reasoning for (1) is written as follows: where it 
(nl or nl-n2) conflicts with n3 in terms of ~RI" In the 
above the underlined parts are words to be given by a 
posteditor. R1 stands for a kind of correction like 
477 
TENSE conflict and SEMantic CATegory conflict. 
Similarly, the posteditor can add a syntactic symbol 
such as a part of speech a technical term a compound 
and an idiom to a word group involved in the given new 
word sequence if necessary. 
3.3. Correction by PECOF 
Figure 3 shows the main functions of PECOF. 
(I) Analysis of the corrected parts 
by using the reverse MT system 
PECOF (2) Identification of the part to be corrected 
in the database 
(3) Correction of the part of database 
or documentation of the corrected patterns 
which cannot be completely identified 
Fig.3 The main functions of PECOF 
3.3.1. Installation 
In order to keep lexical information of the words 
appearing in the source language sentences till the end 
of correcting, PECOF needs some record type database. 
The current dictionaries are constructed there by copy- 
ing the parts of the dictionary in the file and modifica- 
tion of the dictionary are performed on the database. 
The reverse MT system needs the fundamental 
rewriting rules to be implemented based on the case 
grammar to parse and comprehend the target language 
• °utput modified by posteditors. Most of parsing sys- 
tems based on a case grammar will be available with 
some modification. 
3.3.2. Performances 
First, PECOF makes the reverse MT system to 
analyze the postedited output and construct the inter- 
mediate forms. When the postedited output involves 
some words which are not contained in the word dic- 
tionary and also in the syntactic information given by 
the postcditor, PECOF tries to identify the syntactic 
information and the equivalents ef the unknown words 
by referring to the information of the related words in 
both the target and the source language expressions or 
by asking the posteditor about them later. 
After constructing the intermediate form, PECOF 
compares it with that of the original MT system. If 
there are some differences between them, PECOF 
makes the reverse MT system to further transfer the 
intermediate form in the reverse direction. If the reverse 
MT system has sufficient capability of translation, it 
will be able to yield almost the same intermediate form 
as the original MT system at a certain transfer stage 
though the same form might be able to be obtained only 
in the given source language sentential expressions. 
Furthermore, if the original MT system can parse and 
normalize the source language sentences correctly, both 
the intermediate forms coincides with each other by the 
end of the word transfer stage of the reverse MT system 
at the latest. 
If the same intermediate form is obtained, PECOF 
stops the transfer by the reverse MT system and begins 
to backtrack. Then PECOF tries to remove the differ- 
ence between the next-stage intermediate form of the 
original MT system and that of the reverse MT system. 
478 
More precisely, PECOF identifies the irrelevant part of 
the intermediate form of the original MT system by 
comparing it with that of the corrected results given by 
the reverse MT system and corrects the data and the 
applied conditions in the database according to the 
procedures determined from the difference patterns. 
If the reasoning of corrections is given in a form of 
the conflicting words and the associated information as 
mentioned in section 3.2, PECOF examines the data to 
be corrected mid the irrelevant applied conditions by 
referring to the syntactic and semantic attributes of the 
conflicting words, and corrects the data and the 
applied condition to be more relevant by refining unify- 
ing or replacing the old conditions. 
Some databases such as word dictionaries omits 
detailed items like the applied conditions of a word if 
they are generally held. These deficit items are impli- 
citly designated by a general condition table impleo 
mented for each category of words. In such cases, if a 
specific equivalent is designated together with the rea- 
soning by a posteditor, the applied conditions of the 
equivalent derived from the reasoning are written in the 
corresponding part of the record of the current word 
dictionary directly or through a pointer. 
If the correction in postediting lacks detailed infor- 
mation about wrong translation and confident reasons 
necessary for correcting the database, PECOF arranges 
the related parts of corrections of the corresponding ori- 
ginal target and source expressions, classifies them by 
some attributes of the error patterns and adds them to 
a document of error patterns. Then PECOF urges the 
users to formulate the correcting procedures of the part 
of database corresponding to the error patterns. 
4. Miscellaneous correction information in posteditlng 
4.1. Syntactic structure correction 
Every part of the translated output is required to 
meet syntactic patterns of the target language even by 
modifying and complementing the given source language 
sentences. The occurrence of some syntactic errors and 
their corrections in the target language expressions can 
be detected when some parts of the target language 
expressions are corrected though the intermediate forms 
are the same as those of the corrected expressions 
except for some syntactic term expressions° 
4.1ol. Word expressions 
Let us describe the rewriting rules of a noun phrase 
as follows: 
RR:ADJP (Rewriting Rule: ADJective Phrases) 
<NP(nl{v}(PRED:., OBJ:n2,..))> 
::=<NPI(nl)><PREP><NP(n2)> 
<NP(nl(PRED:v, KI:.,K2:n2,...))> 
:: = <NPI(nl) ><INF(PRED:v,KI:nl,K2.:n2,...) > 
1 
2 (3) 
The above expressions are useful for transformation 
between sentences and the intermediate expressions in 
parsing and generation of sentences, 
In the rewriting rule expressions, ni(i-l,2) and v are 
a noun term and a verb term respectively, nl{v} means 
that nl is a noun term derived from a verb term. 
<:NP(t)> and <INF(t)> denote the non-ternfinal sym~ 
bolz of ~z noun phrase and an infinitive phrase 
correspondirq~ to a term expression o1' an intermediate 
form t. The symbol * denotes the term prefixed to a 
frame which includes and modifies the symbol *. In this 
ease, it stands for nl{v}. PRED and OB3 denote a 
PREDicate case label and art OBject ease label respec- 
tively. K1 and K2 stand for some case labels. The term 
ekpresslons of the left side hand of the rewriting rule 
(3)d describe that nl{v} is modified by n2 which 
depends on a predicative noun nl as the objective 
term. 
In usual cases, the preposition used for modifying a 
noun by a noun is the preposition "of" and "of" is taken 
for the dcfidt value of the preposition. If a specific 
preposition "prep" is indicated to the noun term nl{vt} 
by a posteditor, PECOF records "prep" in the preposi- 
I;ion item of the noun word nl of the word dictionary 
together with the applied rewriting rule as follows: 
MODifier:< PREP >,<PlZEP>:prep,It~ADJP1 (4) 
~Bxample 1 
(a)..inquiry %f *-- into' the question.. 
(b)..discussion 'of +- on' the question.. 
where'a ~- b' means the replacement of %' by 'b'. 
Corresponding to error corrections in postediting in 
the above, the recorded items of "prep" in the word dic- 
tionary are "into" and "on" in a form of (4) respec- 
tively. 
The correcting information of the other kinds of syn- 
tactic errors can be fed back to the original MT system 
in a similar way. 
4.L2° lhewrltlng rule eons~ructlon 
An MT system sometimes lacks some rewriting rules. 
For example, technical papers oll.en devise and use a 
concise phra~e expression instead of a long complicated 
expression under ~lle condition that no ambiguity is 
brought. 
Let u,~ consider the following rewriting rules of a 
noun phrase: 
<NP(nl (PRED:v,KI:*,K2:n2,..)) > 
:: = <Nl?l(n 1)><RLC(PRED:v,KI:*,K2:n2,...) > (5.1) 
<NPI(n(OBJ:*,DET:det)) > 
:: = <DET(det)> <NP2(n)> (5.2) 
where RLC and DET stand for RELative Clause and 
DETerminer respectively. 
When an interinediate form 
car(OBJ:*(OBJ:*,DET:a),PRED:be drive-en, 
AG:motor) (6.1) 
is given~ a relative clause 
" a car that is driven by motor" (6.2) 
is generated by the rewriting rule (5.1) aud others. 
Now ~ suppose that the relative clause is replaced by a 
phrase 
"a motor driven tax" (7.1) 
in postediti~,g. PECOF infers the rewriting rule which 
wewrites (6.1) to the above corrected phrase and adds to 
(SA) an alternative rewriting rule 
:: = <DET(det)> <NP(n2)><PastParticiple(v)> 
<NP(nl)> (7.2) 
Some words can be used only by a specific syntactic 
rule which belongs to a general syntactic rule. The sen- 
tences that include these words can be parsed and inter- 
preted by the general rules which are not conditioned 
by various syntactic patterns in details. On the con- 
trary, the generation of the sentences needs the infor- 
mation of the specific usages of the words. For example, 
the verb "doubt" conventionally takes the OBject term 
through WHETHER-CLAUSE in an affirmative sen- 
tence and THAT-CLAUSE in a negative sentence. The 
information is needed to generate the target sentence 
from the intermediate form TIo2 and is given by some 
rewriting rules or by complementing the rewriting rules 
hy means of adding the necessary information to the 
word dictionary as shown in Expr.(4). 
4°2. Transfer rule ¢orrectlon 
4.2.1. Structure transfer 
If the intermediate form TIol is the same as that of 
the postedited output Tlrl though TIo2 is different 
from TIr2, PECOF tries to search for some structure 
transfer rules to meet the corrections. Structure transfer 
is needed so that natural and conventional target 
language expressions can be generated corresponding to 
given source language expressions. 
As well known, Japanese tends to avoid the use of 
non-animate subjects and also interpret an event as a 
change of a state due to some causes rather than action 
on some objects by an agent. A general structure 
transfer rule between state chauge and action is given 
as ibllows: 
STR:CRl(Structure Transfer Rule :Cause Result 1) 
(PRED:vl~ OBJ:nl, CAUSE:n2) 
-,(PRED:v2, AG:n2, OBJ:nl) (8.1) 
where v2 and vl form a pair of verbs of 'cmtse and 
result' or 'action and state change'. The typical con- 
crete instances of them are "cause to do" and "do", 
"lead" and"reach", "show" aald "be seen" mid others. 
The above fundamental rule (8.1) is implemented in 
the database of the original MT system as well as the 
reverse MT system in the structure transfer database. 
The specific information of cause verb "v2" for a given 
result verb "vl" is sometimes recorded in the word dic- 
tionary together with the name of the applied transfer 
rule as follows: 
CAUSE-V:v2 STR:CR1 (8.2) 
When PECOF finds that an original output is pos- 
,edited by the structure transfer of this type and by 
using a cause verb v2, PECOF records v2 in the vl item 
of the word dictionary according to the form (8.2). 
4.2.2. Word transfer 
The word to be corrected in word transfer can be 
identified by using the information about replacement of 
fall words such as verbs adjectives and nouns. More 
precisely, it can be identified from a term of the target 
language intermediate form TIrl modified by a postedi- 
tot. Some patterns to bc corrected are shown in the fol- 
lowing. 
In a ease structure of a language, the semantic 
categories of a governor and the dependants are~ more 
or less, bounded. Furthermore, the conventional 
category boundary set of terms in a frame of the target 
language is not always the same as that of the source 
479 
language and sometimes structure transfer is needed. 
The discrepancy between the semantic category 
boundary of a word and that of the equivalent is 
revealed when the word is linked with some dependants 
or the governor, ttence the equivalent must be chosen 
so that no conflict occurs on the semantic category con- 
ditions of the linked words in the target language. 
Let us assume that a word tl of a source language can 
be represented by an equivalent tll' or t12' of the tar- 
get language in the following condition: 
HEADWOItD: tl 
EQUIV:t Ii ' COND:(KI-Chtl, K2-C21,...), (9.1) 
EQUIV:t 12' COND:(KI-CI:tl, K2-C22,...), (9.2) 
where Ki and Ci (i=1,2,...) denote a case label and a 
semantic category nmne respectively and both t11' and 
t12' are the equivalents of tl under the condition that 
the term tl appears in the context of the case-category 
label of K1-C1 accompanied with a term of case- 
category label K2-C21 or K2-C22. 
Furthermore, suppose that a postedltor replaces the 
equivalent tll' by t13' under a condition that the 
word tl is accompanied with a word t23 of a case- 
category label K2-C23 and C23 is a subcategory name of 
C21. 
Then, the equivalent applied condition (9.1) of the 
word tl of the word dictionary are replaced as follows: 
HEADWORD:t 1 
EQUIV:t 11 ', COND:(K1-CI:t 1,K2-{C21-C23},...), (9.3) 
EQUIV:tl3', COND:(KI-C l:t I,K2-C23,...), (9.4) 
If t13' is the same as t12', (9.4) and (9.2) are uni- 
fied as follows: 
EQUIV:tl2', COND:(K1-Chtl,l¢2-{C22 ~C23},...), (9.5) 
where {Ci±Cj) denotes the union or the difference 
set of the sets expressed by tim category names Ci and 
cj. 
Example 2 
Japanese represents emotional state expressions by 
using adjective verbs. They have the same form 
irrespective of the active or the passive type. On the 
other hand, English has a different expression depend- 
ing on whether the object to be described is active or 
passive. For instance, they are "interesting" versus 
"interested", "enjoyable" versus "enjoy oneself", "excit- 
ing" versus "excited" and others. 
Let us take an example. It is assumed that the main 
part of the Japanese-English word dictionary of a word 
"TAIKUTSU-DE ARU" or " be boring" in English is 
given as follows: 
tIEADWORD :~J~'/?~ ~ (TAIKUTSU-DE ARU) 
Part of Speech:adjective verb, ROLE:PRED-ATTILEMOTION 
EQUIV: be boring (9.6) 
Let us suppose that for a given Japanese sentence 
"WATAKUSHI-WA KARE-NO KOUEN-GA TAIKUTSU-DE 
ATTA" a modified sentence "I was (boring *--bored) in 
his lecture" is given by a posteditor. PECOF identifies 
the wrong expression and refines the corresponding 
items of the word dictionary by referring to the format 
(9) as follows: 
/480 
EQUIV:be boring,COND:(PRED-ATTR:% 
OBJ-THINGS,R.ECIPIENT-HUMAN ), (9°7) 
EQUIV:be bored, COND:(PRED-ATTPa*, 
EXPER-HUMAN,OBJ-THINGS ), (9.8) 
where * stands for the word "TAIKUTSU-DE ARU". 
Example 3 
A Japanese verb "oshieru" has several meanings, 
namely, "teach, tell, show, instruct and so forth". The 
object case categories of "teach" and "tell" of English in 
this context are a school subject like mathematics and 
simple information like a way to a station respectively. 
The system MTo chooses the equivalents by examlno 
ing these category conditions. If the postediting results 
suggest that the system needs more precise conditions 
for equivalent selection, PECOF tries to refine the 
category conditions or to add the other case conditions. 
4.3. Idlomatle expression transfer 
An idiomatic expression (including a compound) cow 
sists of several words and has a somewhat definite 
meaning when compared with a single word. Every 
language has various unique idiomatic expressions. It is 
considered that one of the most efficient and reliable 
methods in machine translation is to cover the main 
content to be translated with a frame of an appropriate 
idiomatic expression of the target language. Therefore, 
it will be necessary for a good MT system to have a 
large storage of various idiomatic expressions and to be 
capable of selecting an appropriate target idiomatic 
expression to cover the main part of a given source 
language expressions. PECOF helps gathering the 
idiomatic expressions given by posteditors. 
At first, PECOF identifies an output part modified 
by a posteditor to be an idiomatic expressions. The 
information of identification is given by the postedltor 
or is inferred from a phrase form of the replaced part. 
PECOF selects a keyword among the words involved in 
the idiomatic expression and identifies the correspond° 
ing source language word by referring to the current 
word dictionary mentioned in section 3. Then PECOF 
records the corresponding source language idiomatic 
expressions "al t /~1" at the location labeled with the 
heading of the keyword "t" in the word dictionary so 
that the idiomatic expression can be applied with prior- 
ity in the word transfer by the longest match method. If 
a similar heading idiomatic expression "a2 t f12" which 
has the same target expressions is already involved in 
the location, these heading expressions are unified as 
follows: 
,,{~1,~2} t {~1,~2},, ,-- ,,~1 t ~r,, ,,~2 t ~2', 
The semantic and syntactic information necessary tbr 
the dependants and the governor are added to the 
idiomatic expressions. 
Example 4 
HEADWORD: ~,~.. (JYUTEN) 
e o • • . o 
IDIOM1: 1~_~¢{~<, Ll~'6} 
(NI-JYUTEN-WO{OKU,SHIBORU}) 
(PRED-THINKACT: ~x,~ ~: { ~ <. L l~" 6 }, 
I,O(;.,THINCS:~ t~:) 
put cmph.~si,,~ ox~ 
(PRED-TtHNKACT:put emphasis 
L()C, .THINGS:on-) 
~L ~1~me Rtustrative exar~ples 
Thi:~ ~c,:tion sho.m some illustrative examples of the 
modific~,tion of d{ctiomMez by posteditors ~ correcting 
inlbrmati~t, 'J/hey axe {;o bc tested by the experinmntal 
system PI~\]COF which is uttdcr construction. 
~\]xample ~;~1 
The inpn~ 3apanese sentence So mid tim output To 
of t;he original Mq.' system are given respectively aa lbl.. 
\].OWS .' 
To:The t¢.ol ha.s the ability which removc'~ heat fl:om 
the crit;ieal~cut-rmlge. 
The interr,mdiate form Tio2 of the above is 
TIo2:(Pi~,?'3D:have~I'OSSES,S OR:t~oI(OBJ:*~DET:the), 
( )BJ:~tbility(PRED:remove, 
OBJ:heat, INSTR:*,...)) 
The output Tr corrected by a posteditor is 
'rr:The tool has the ability to remove heat from the 
critical catting re'ca. 
The inte:.:n,ediate Ibrm Tit2 of the above is the same 
as Tio2 e)::cept Nr the technical term "critical cutting 
First~ PFATOI? find~ t, hat the intermediate form Tlol of 
the origimd ouLpnt is the same as that of the postedited 
ou~pu~ T\[:.:I if the word "critical~vuG~ang0" iS l'epl~ced 
by "crith:alocutting-area' ~ according to postedi~or's 
designation. PECOF corrects the word dictionary m~d 
~he intermediate forms of the following tran.st~r stages 
bmsed on ~he replacement and finds that discrepmmies 
remain only in the target surNce expressions. 
PECOF examines the rewriting rules applied to Tlr2~ 
recognizes that the form to be ~applied to "ability" as a 
modifier is an il/finitive and adds the information to the 
MODifier item of the word dictionary as follows: 
HEADWORD: fl~?\]J (NOURYOKU) 
EQUIV:ability,.., 
MODifier:<INF>, RR:ADJP2 
Bxample 5,2 
The input Japmtese sentence So and the output To 
of the original MT system are given respectively as tbl° 
low~: 
So: = ¢)I,~)~tINJfV4:: • o:, ~c~¢)~Y ~ .. ~--Y- 4 Jto 
ttl')J ~q'4 6 tc Z,. 
To: A phlrMity of outputs d variable duty ratio is 
obtai~,ed with this time division operation. 
The intermediate form TIo2 of the above is 
TIo2:(P\]fOgD:obtMn(VOiCE:p~sive), 
OBJ:oatput(OBJ:*~NUM:plurality, 
ATTR:variable-duty°ratio), 
CA USE:time-divlsiolt ~ 
operation(OBJ:*~l)ET:the)) 
The output Tr corrected by a posteditor is 
Tr:This time division operation provides a plurality of 
outputs of variable duty ratio. 
The internmdiate form Th-2 of the above is 
Th'2:(PRED:provide(VOICE:active), 
OBJ:output( ..... ), 
AG:time-division-operation(....)) 
PECOF compares TIo2 with TIr2~ recognizes that 
TIr2 cau be obtained by applying tim gexleral Structure 
'\]?rausfcr R.ule CR1 to Tit2 with the specification of 
CAUSI~:V being "provide" for RESULT-V of "obtain", 
and then writes down the information in the 
corresponding item in the word dictionary a~s follows: 
IIEADWORD: ~5 (EltU) 
..,, ..... 
EQUIV:obtain, .... 
CAUSE-V:provide 8TR:CR1 
~° Conclusion 
This research is still under the early stage and needs 
a lot of experimental investigation. This paper shows a 
nmthod of modification of the database for a compara- 
tively definite error patterns. There will be left various 
kinds of indefinite error patterns which should be 
characterized chLssified and corrected by some formu- 
lated procedures. IIowever, the basic idea mM system 
presented here will be useful for improving the trmlsla- 
tion quality of the MT system and collecting new techn- 
ical words and idiomatic expressions. Furthermore, if 
both the original and the reverse MT systems have the 
posteditors feedback respectively and cooperate with 
each other, the integrated system will be very powerful 
and efficien£ ¢o improve ~he tra~slatk)n tI*~Mity and 
capability in the bilingual translation. 

Reference

(1)Pigott I.M.(1981) 'The importmme of feedback from 
translators in the development of high-quality 
machine translation.' In:V.Lawson(ed.): Practical 
Experience of Machine Translation: North- 
Holland,pp.61-74. 

(2)Green R.(1981) 'The MR' errors wlficb causes most 
trouble to posteditors.' ibid. pp.101-1.304. 

(3)Bernard Lavorel(1981) 'Experience in English-French 
post-editing.' ibid. pp.105-1.309. 

(4)Rossi F.(1981) 'The impact of posteditors' feedback 
on the quality of MT.' ibid. pp.113-1.317. 

(5)Carbonell J.G. and Tomita M.(1987) 'Knowledge- 
based machine translation.' In:S.Nierenburg(ed.): 
Machine Translation:Cmnhridge Univ.Press,pp.68-89. 

(6)Nishida F. and Takarnatsu S.,(1982) 'Japanese- 
English Translation through Internal Expressions.' 
In: Proc. of COLING 82, pp.271-276. 

(7)Nishida F., Fujita Y. and Takamatsu S.(1986) 'Con- 
struction of a Modular and Portable Machine Trans- 
lation System.' In: Proc. of COLING 86, pp.649-651. 

(8)Nishida F., Fujita Y.and Takmnatsu,S(1987) 'Refine- 
nmnt a~ld Error Detection of Program Specifications 
by a Linking Technique of Library Modules.' 
IPSJ,28, 5,pp.486-498.(written in Japanese) 
