Analysts Grammar or Japanese tn the Nu-ProJect 
- A Procedural Approach to Analysts Grammar - 
Jun-tcht TSUJII. Jun-tcht NAKANURA and Nakoto NAGAO 
Department of Electrical Engineering 
Kyoto University 
Kyoto. JAPAN 
Abstract 
Analysts grammar of Japanese tn the Mu-proJect 
ts presented, It is emphasized that rules 
expressing constraints on stngle linguistic 
structures and rules for selecting the most 
preferable readtngs are completely different In 
nature, and that rules for selecting preferale 
readings should be utilized tn analysts grammars of 
practical HT systems. It ts also clatmed that 
procedural control ts essential tn integrating such 
rules tnto a unified grammar. Some sample rules 
are gtven to make the points of discussion clear 
and concrete. 
1. Introduction 
The Hu-ProJect ts a Japanese nattonal project 
supported by grants from the Special Coordination 
Funds for Promoting Science & Technology of 
STA(Sctence and Technology Agency). whlch atms to 
develop Japanese-English and English-Japanese 
machine translation systems. Ve currently restrict 
the domain of translation to abstracts of 
scientific and technological papers. The systems 
are based on the transfer approach\[;\], and consist 
of three phases: analysts, transfer and generation. 
In thts paper, we focus on the analysts grammar of 
Japanese tn the Japanese-English system. The 
grammar has been developed by using GRADE which ts 
a programming language specially designed for thts 
project\[2\]. The grammar now consists of about 900 
GRADE rules. The experiments so far show that the 
grammar works very well and ts comprehensive enough 
to treat various linguistic phenomena tn abstracts. 
In thts paper we wtll discuss some of the basic 
design principles of the grammar together wtth its 
detatled construction. Some examples of grammar 
rules and analysts results wtll be shown to make 
the points of our discussion clear and concrete. 
2. Procedural Grammar 
There has been a prominent tendency tn recent 
computational linguistics to re-evaluate CFG and 
use tt dtrectly or augment tt to analyze 
sentences\[3.4.5\]. In these systems(frameworks), 
CFG rules Independently describe constraints on 
stngle linguistic structures, and a universal rule 
application mechanism automatically produces a set 
of posstble structures which satisfy the given 
constraints. It ts well-known, however, that such 
sets of posstble structures often become 
unmanageably large. 
Because two separate rules such as 
NP ..... • NP PREP-P 
VP ..... • VP PREP-P 
are usually prepared tn CFG grammars tn order to 
analyze noun and verb phrases modifted by 
prepositional phrases. CFG grammars provide two 
syntactic analyses for 
She was given flowers by her uncle. 
Furthermore. the ambiguity of the sentence ts 
doubled by the lexlcal ambiguity of "by". which can 
be read as etther a locattve or an agenttve 
preposition. Since the two syntactic structures 
are recognized by compZetely independent ru\]es and 
the semantic interpretations of "by" are given by 
independent processes tn the \]ater stages. It ts 
difficult to compare these four readings during the 
anaZysts to gtve a preference to one of these four 
readings. 
A rule such as 
"If a sentence ts passlve and there ts a 
"by"-prepostttonal phrase, tt ts often the case 
that the prepositional phrase ftlls the deep 
agenttve case. (try thts ana\]ysts first)" 
seems reasonable and quite useful for choosing the 
most preferable interpretation, but tt cannot be 
expressed by refining the ordinary CFG rules. Thts 
ktnd of ru\]e ts quite different In nature from a 
CFG ru\]e. It ts not a rule of constraint on a 
stng\]e \]tngutsttc structure(in fact. the above four 
readings are a\]l \]tngulsttcal\]y posstb\]e), but tt 
ts a "heuristic" ru\]e concerned with preference of 
readings, which compares several alternative 
analysts paths and chooses the most feastble one. 
Human translaters (or humans tn general) have many 
267 
such preference rules based on vartous sorts of cue 
such as morphological forms of words, collocations 
of words, text styles, word semantics, etc. These 
heuristic rules are quite useful not only for 
increasing efficiency but also for preventing 
proliferation of analysts results. As Wllks\[6\] 
potnted out, we cannot use semanttc Information as 
constraints on stngle linguistic structures, but 
Just as preference cues to choose the most feastble 
Interpretations among linguistically posstble 
Interpretations. We clatm that many sorts of 
preference cues other than semanttc ones exist tn 
real texts whtch cannot be captured by CFG rules. 
We will show tn thts paper that. by utilizing 
vartous sorts of preference cues. our analysts 
grammar of Japanese can work almost 
determtntsttcally to gtve the most preferable 
Interpretation as the ftrst output, wtthout any 
extensive semanttc processing (note that even 
"semant|c" processing cannot dtsambtguate the above 
sentence. The four readings are semantically 
possible. It requtres deep understanding of 
contexts or situations, whtch we cannot expect tn a 
practical MT system). 
In order to Integrate heuristic rules based on 
var|ous levels of cues tnto a untfted analysts 
grammar, we have developed a programming langauage. 
GRADE. GRADE provtdes us wtth the following 
facilities. 
Expllctt Control of Rule Appl|cattons : 
Heuristic rules can be ordered according to thetr 
strength(See 4-2). 
- Nulttple Relatton Representation : Vartous 
levels of Informer|on Including morphological. 
syntactic, semantic, logtcal etc. are expressed tn 
a s|ngle annotated tree and can be manipulated at 
any ttme durtng the analysts. Thts ts requtred not 
only because many heuristic rules are based on 
heterogeneous levels of cues. but also because the 
analysts grammar should perform semantic/logical 
Interpretation of sentences at the same ttme and 
the rules for these phases should be wrttten tn the 
same framework as syntactic analysis rules (See 
4-2. 4-4). 
- Lextcon Drtven Processing : We can wrtte 
heuristic rules spectftc to a stngle or a 11mtted 
number of words such as rules concerned wtth 
collocations among words. These rules are strong 
tn the sense that they almost always succeed. They 
are stored tn the lextcon and tnvoked at 
appropriate times durtng the analysts wtthout 
decreasing efficiency (See 4-1). 
- Expltct% Definition of Analysts Strategies : 
The whole analysts phase can be dtvtded into steps. 
Thts makes the whole grammar efficient, natural and 
easy %o read. Furthermore. strategic consideration 
plays an essential role tn preventing undesirable 
interpretations from betng generated (See 4-3). 
3 Organization of Grammar 
In thts sectton, we will give the organization 
of the grammar necessary for understanding the 
discuss|on |n the follow|ng sections. The matn 
components of the grammar are as follows. 
(1) Post-Morphological Analysts 
(2) Determination of Scopes 
(3) Analysts of Stmple Noun Phrases 
(4) Analysts of Stmple Sentences 
(5) Analysts of Embedded Sentences (Relative 
Clauses) 
(6) Analysts of Relationships of SentenCes 
(7) Analysts of Outer Cases 
(8) Contextual Processing (Processing of Omttted 
case elements. Interpretation of 'Ha' . etc.) 
(9) Reduction of Structures for Transfer Phase 
Each component conststs of from 60 to 120 
GRADE rules. 
47 morpho-syntacttc categories are provtded 
for Japanese analysts, each of whtch has tts own 
lextcal description format. 12.000 lextcal entrtes 
have already been prepared according to the 
formats. In thts classification. Japanese nouns 
are categorized |nto 8 sub-classes according to 
thetr morpho-syntacttc behavtour, and 53 semanttc 
markers are used to characterize thetr semanttc 
behaviour. Each verb has a set of case frame 
descriptions (CFD) whtch correspond to different 
usages of the verb. A CFD g|ves mapping rules 
between surface case markers (SCN - postpostttonal 
case particles are used as SCN's tn Japanese) and 
thetr deep case interpretations (DCZ 33 deep 
cases are used). DC! of an SCM often depends on 
verbs so that the mapping rules are given %o CFD's 
of Individual verbs. A CFO also gtves a normal 
collocation between the verb and 
SCM's(postpositonal case particles). Oetatled 
lextcal descriptions are gtven and discussed tn 
another paper\[7\]. 
The analysts results are dependency trees 
whtch show the semanttc relationships among tnput 
words. 
4. Typtcal Steps of Analysts Grammar 
In the following, we w111 take some sample 
rules to Illustrate our points of discussion. 
4-; Relative Clauses 
Relative clause constructions in Japanese 
express several different relationships between 
modifying clauses (relative clauses) and thelr 
antecedents. Some relattve clause constructions 
268 
cannot be translated as relative clauses tn 
Engltsh. Me classified Japanese relattve clauses 
Into the followtn 9 four types, according to the 
relationships between clauses and their 
antecedents. 
(1) Type 1 : Gaps In Cases 
One of the case elements of the relattve 
clause ts deleted and the antecedent fills the gap. 
(2) Type 2 : Gaps In Case Elements 
The antecedent modifies a case element tn the 
clause. That ts. a gap exists tn a noun phrase tn 
the clause. 
(3) Type 3 : Apposition 
The clause describes the content of the 
antecedent as the Engltsh "that"-clause tn 'the 
tdea that the earth ts round'. 
(4) Type 4 : Partlal Apposltlon 
The antecedent and the clause are related by 
certain semantic/pragmatic relationships. The 
relative clause of thts type doesn't have any gaps. 
This type cannot be translated dtrectly lnto 
English relative clauses. Me have to Interpolate 
In English appropriate phrases or clauses whtch are 
Implicit tn Japanese. tn order to express the 
semantic/pragmatic relationships between the 
antecedents and relative clauses explicitly. In 
other words, gaps extst tn the Interpolated phrases 
or clauses. 
Because the above four types of relattve 
clauses have the same surface forms fn Japanese 
......... (verb) (noun). 
RelattvefClause Antecedent 
careful processing ts requtred to d|sttngutsh them 
(note that the "antecedents' -modified nouns- ape 
located after the relat|ve clauses tn Japanese). A 
sophisticated analysis procedure has already been 
developed, which fully ut|ltzes vartous levels of 
heuristic cues as follows. 
(Rule 1) There are a 11mtted number of nouns whtch 
are often used as antecedents of Type 3 clauses. 
(Rule 2) Vhen nouns with certa|n semanttc markers 
appear tn the relattve clauses and those nouns are 
followed by one of spectflc postpostttonal case 
part4cles, there ts a htgh possibility that the 
relattve clauses are Type 2. In the following 
example, the word "SHORISOKUDO"(processtn 9 speed) 
has the semanttc marker AO (attribute). 
\[ex-1\] \[Type 2\] 
"SHORZSOKUDO" "GA" 
(processing speed) (case 
particle: 
subject 
I case) 
RelattvetClause 
"HAYA\[" "KEISANK\[" 
(htgh) I (computer) I /t 
Antecedent 
-->(English Translation) 
A computer whose processing speed ts htgh 
(Rule 3) Nouns such as "MOKUTEKZ"(puPpose). 
"GEN ZN"(reason), "SHUDAN"(method) etc. express 
deep case relationships by themselves, and. when 
these nouns appear as antecedents. |t is often the 
case that they ft11 the gaps of the corresponding 
deep cases tn the relattve clauses. 
\[ex-2\] \[Type 1\] 
"KONO" "SOUCHI" "O" "TSUKAT" "TA" "MOKUTEK\[" (th,s)l(dev,c.  (c.. ICpurpos.) 
|part,cle:h /,ormat,ve: I J 
I / °bJect l / pest) l 
/case) ~ / 
RelattvetClause Antecedent 
--> (English Translation) 
The purpose for wh|ch (someone) used thts devtce 
The purpose of ustn9 thts devtce 
(Rule 4) There ts a 11mtted number of nouns whtch 
are often used as antecedents In Type 4 relattve 
clauses. Each of such nouns requtres a specific 
phrase or clause to be Interpolated tn Engltsh. 
\[ex-3\] \[Type 4\] 
"KONO" "SOUCHI" "0" "TSUKAT"-- "TA" "KEKKA" (th,s),(devlce)/~case e.~. (to use)/~tense ~'...(;esult) 
...l fformat,ve:h J 
1 ,object , Ipast) I 1 
\[ I case) l 
Rel at tve ~ Clause Antecedent 
--> (Engllsh Translation) 
The result which was obtatned by ustng thts dev|ce 
In the above example, the clause "the result whtch 
someone obtatned (the result : gap)" ts onmitted tn 
Japanese. whtch relates the antecedent 
"KEKKA"(result) and the relattve clause "KONO 
SOUCHI 0 TSUKAT_TA"(someone used thts devtce). 
269 
A set of lextcal rules ts defined for 
"KEKKA"(resulL). which basically works as follows : 
tt examines first whether the deep object case has 
already been filled by a noun phrase tn the 
relattve clause. If so, the relattve clause ts 
taken as type 4 and an appropriate phrase ts 
Interpolated as tn \[ex-3\]. If not, the relattve 
clause ts taken as type 1 as tn the following 
example where the noun *KEKKA" (result) ftlls the 
gap of object case tn the relattve clause. 
\[ex-4\] \[Type 1\] 
"KONO" "JIKKEN • / •GA". "TSUKAT• J"TA" l "KEKKA" 
(thts)J(expertment)//(case~(to use)~(tense (r~ult) rParticle~ iformsttve:\]l 
IsubJect I I past)| I \[ _ll case) l / I 
Relattve Clause Antecedent 
-->(English Translation) 
The result whtch thts experiment used 
Such lextcal rules are Invoked at the beginning of 
the relattve clause analysts by a rule tn the math 
flow of processing. The noun "KEKKA • (result) is 
given a mark as a lexlcal property which Indicates 
the noun has special rules to be Invoked when tt 
appears as an antecedent of a relatlve clause. A11 
the nouns which requlre speclal treatments In the 
relative clause analysts are given the same marker. 
The rule tn the matn flow only checks thts mark and 
Invokes the lextcal rules defined tn the lextcon. 
(Rule 5) Only the cases marked by postpostttonal 
case particles 'GA'. 'WO" and 'NI" can be deleted 
tn Type 1 relattve clauses, when the antecedents 
are ordtnary nouns. Gaps tn Type 1 relative clauses 
can have other surface case marks, only when the 
antecedents are spectal nouns such as described tn 
Rule (3). 
4-2 ConJuncted Noun Phrases 
ConJuncted noun phrases often appear in 
abstracts of scientific and technological papers. 
It ts Important to analyze them correctly. 
especially to determine scopes of conjunctions 
correctly, because they often lead to proliferation 
of analysis results. The particle "TO" plays 
almost the same role as the Engllsh "and" to 
conjunct noun phrases. There are several heuristic 
rules based on various levels of information to 
determine the scopes. 
<Scope Decision Rules of ConJuncted Noun Phrases 
by Partlcle 'TO'> 
(Rule 1) Stnce parttcle "TO" ts also used as a case 
particle, tf It appears tn the position: 
Noun 'TO" verb Noun, 
Noun 'TO' adjective Noun. 
there are two posstble Interpretations. one tn 
whlch "TO" Is a case parttcle and "noun TO 
adjective(verb)' forms a relattve clause that 
modifies the second noun. and the other one tn 
which "TO" ts a conjunctive particle to form a 
conJuncted noun phrase. However. it ts very 11kely 
that the parttcle 'TO' ts not 8 conjunctive 
parttcle but a post-positional case particle, if 
the adjective (verb) ts one of adjectives (verbs) 
which requtre case elements wtth surface case mark 
"TO' and there are no extra words between "TO • end 
the adjective (verb). In the following example. 
"KOTONARU(to be different)" ts an adjective which 
ts often collocated wtth a noun phrase followed by 
case particle "TO". 
\[ex-5\] 
YOSOKU-CHI "TO" KOTONARU ATAI 
(predicted value) (to be different) (value) 
\[dominant interpretation\] 
IYOSOKU-CHI "TO" KOTONARU ATIAI 
relattve~clause ant/cedent 
• the value which ts different from the 
predicted value 
\[less domtnant Interpretation\] 
YOSOKU-CHI "TO" KOTONARU ATAI 
Me N~ I I 
conJuncte~ noun phrase 
= the predicted value and the different value 
(Rule 2) If two "TO* particles appear tn the 
position: 
Noun-1 'TO' . ......... Noun-2 'TO' 'NO" NOUN-3 
the right boundary of the scope of the conJuctton 
ts almost always Noun-2. The second 'TO" plays a 
role of a delimiter which deltmtts the right 
boundary of the conjunction. Thts 'TO" tS 
optional, but tn real texts one often places tt to 
make the scope unambiguous, especially when the 
second conjunct IS a long noun phrase and the scope 
is highly ambiguous without tt. Because the second 
'TO' can be Interpreted as a case parttcle (not as 
a delimiter of the conjunction) and 'NO' following 
a case parttcle turns the preceding phrase to a 
270 
modlfter of s noun. on Interpretation tn whtch 
"NOUN-2 TO NO" ts taken as o modtrter of NOUN-3 and 
NOUN-3 ts token as the hood noun of the second 
conJunt ts also linguistically possible. However, 
In most cases, when two 'TO" particles appear tn 
the above position, the second "TO' Is Just a 
delimiter of the scope(see \[ex-6\]). 
\[ex-6\] 
YOSOKU-CHI TO JIKKEN DE.NO JISSOKU-CHI TO 60 SA 
(predtctedl'~expertment~'~case'~(octual valu~ I 
value) J ~orttcle~ (dtt'ference) 
t pl°c°) \] 
\[dominant Interpretation\] 
YOSOKU-CHI TO J\[KKEN DE 60 O\[$$OKU-CH\] TO NO SA 
NP NP 
1 I 
ConJuncted HP 
I 
NP 
• the difference between the predicted value 
and the actual value tn the experiment 
\[less domtnant tnterpnetattons\] 
(A) 
YOSOKU-CHI TO JIKKEN DE NO JISSOKU-CHI TO NO $A 
NP NP I I 
ConJuncted NP 
- the difference wtth the actual value tn the 
predicted value and the experiment 
(e) YOS~KU-CH\] 
.p ~p 
l I 
ConJun~ted NP 
TO J\[KKEN DE NO JZSSOKU-CH\[ TO NO SA 
"l "" I 
• the predicted value and the difference wtth 
the actual value tn the experiment 
(Rule 3) If a spectal noun whtch ts often 
collocated wtth conjunctive noun phrases appear tn 
the position: 
Noun-1 'TO' . ....... Noun-2 "NO'<spectal-noun>, 
the rtght boundary of the conjunction ts almost 
always Noun-2. Such spectal nouns are marked tn 
the lextcon. \[n the following example. "KANKEI" ts 
such a spectal noun. 
\[ex-7\] 
JISSOKU-CHI~O" 
(actual value) I 
RIRON-DE E-TA YOSOKU-CHI. NO, KANKE\[__ 
1(theory \]( ( to~( prod tcted~ (l:e lot ton~ 
" Iobtatn)l value) // shtp)J 
II 
spectal noun 
\[dominant Interpretation\] 
JISSOKU-CH! "TO" . ...... YOSOKU-CH\[ NO KANKEI L._;___I 
(relative antecedent 
clsuse)l J 
NP ~P I I 
con\]u~cted NP 
• the relationship between the actual value 
and the predicted value obtatned by the 
theory 
\[less domtnant Interpretations\] 
(A) 
JIS$OKU-CHI "TO" R\]ROH-DE ...YO$OKU-CH\[ NO KANKE! 
NP I I 
conJun~ted NF 
I 
relattvetclouse antecedent 
• the relationship of the predicted value whtch 
was obtatned by the actual value and the theory 
(e) 
JX$SOKU-CH! "TO" . ......... YO$OKU-CHX NO KANKEX 
~P NP I 
I conJuncted NP 
• the actual value and the relationship of 
the predicted value whtch was obtatned by 
the theory 
(Rule 4) Zn 
Noun-1 'TO' . ..... Noun-2, 
tf Noun-1 and Noun-2 are the same nouns, the rtght 
boundary of the conjunction ts almost always 
Noun-2. 
(Rule 5) In 
Noun-! 'TO' . ...... Noun-2. 
tf Noun-! and Noun-2 are not exactly the some but 
nouns wtth the same morphemes, the rtght boundary 
271 
ts often Noun-2. In \[ex-7\] above, both of the heed 
nouns of the conJuncts. JISSOKU°CHI(actual value) 
and YOSOKU-CH\[(predtcted value), have the same 
morpheme "CH\[" (whtch meams "value"). Thus, thts 
rule can correctly determine the scope, even tf the 
spectal word "KANKE1"(relattonshtp) does not extst. 
(Rule 6) If some spectal words (11ke 'SONO" 
'SORE-NO' etc. whtch roughly correspond to 'the'. 
'1iS' tn Engllsh) appear tn the position: 
Phrases whtchlNoun-1 "TO' <spectal word> Noun-2. 
modtfy noun 
phrases 
the modifiers preceding Noun-1 modtfy only Noun*l 
but not the whole conJuncted noun phrase. 
(Rule 7) \[n 
...... Noun-1 'TO' . ........... Noun-2. 
tf Noun-1 and flour-2 belong to the same spectftc 
semanttc categories, 11Le actton nouns, abstract 
nouns etc, the rtght boundary ts often Noun-2. 
(Rule 8) \[n most conJuncted noun phrases, the 
structures of conJuncts are well-balanced. 
Therefore, tf a relattve clause precedes the first 
conjunct and the length of the second conjunct (the 
number of words between 'TO" and Noun-2) ts short 
11ke 
\[Relative Clause\] Noun-1 'TO" . ....... Noun-2 
the relattve clause modtftes both conJuncts, that 
ts. the antecedent of the relattve clause ts the 
whole conJuncted phrase. 
These heuristic rules are based on different 
levels of Information (some are based on surface 
lexlcal Items. some are based on morphemes of 
words, some on semanttc |nformatton) and may lead 
to different decisions about scopes. However. we 
can distinguish strong heuristic rules (t.e. rules 
whtch almost always give correct scopes when they 
are applled) from others. In fact. there extsts 
some ordertng of heuristic rules according to thetr 
strength. Rules (1). (2). (3), (4) and (6). for 
example, almost always succeed, and rules like (7) 
and (8) often lead to wrong decisions. Rules 11ke 
(7) and (8) should be treated as default rules 
whtch are applted only when the other stronger 
rules cannot dectde the scopes. We can deftne tn 
GRADE an arbitrary ordertng of rule applications. 
Thts capability of contro114ng the sequences of 
rule applications ts essential tn Integrating 
heuristic rules based on heterogeneous levels of 
Information tnto a untried set of rules. 
Note that most of these rules cannot be 
naturally expressed by ordtnary CFG rules. Rule 
(2). for example, ts a rule whtch blocks the 
application of the ordtnary CFG rule such as 
NP ---> NP <case-particle> NO N 
when the <case-particle> ts 'TO' and a conjunctive 
parttcle 'TO' precedes thts sequence of words. 
4-3 Determination of Scopes 
Scopes of conJuncted noun phrases often 
overlap wtth scopes of relattve clauses, whtch 
males the problem of scope determination more 
complicated. For the surface sequence of phrases 
11ke 
NP-1 'TO' NP-2 <case-particle> ..... <verb> NP-3 
there are two passable 
scopes of conJuncted noun 
clause 11ke 
relationships between the 
phrase and the relattve 
(1) NP-1 'TO" NP-2 <case-particle> .... <verb> NP-3 
I J 
conJ~ncted 
noun phrase I 
Relattv~ Clause 
I 
Antecedent 
I t 
NP 
(2)NP-2 'TO' NP-2 <case-particle> ..... <verb> NP-3 
I Relattve ~ Clause Antecedent J I 
N,P 
ConJuncted* Noun Phrase 
Thts ambiguity together with genutne ambtgu|ttes tn 
scopes of conJuncted noun phrases tn 4-2 produces 
combinatorial Interpretations tn CFG grammars, most 
of whtch are linguistically posstble but 
practically unth|nkable. It Is not only 
Inefficient but also almost Impossible to compare 
such an enormous number of linguistically posstble 
structures after they have been generated. In our 
analys|s grammar, a set of scope dectston rules are 
applted in the early stages of processing tn order 
to block the generation of combinatorial 
Interpretations. \]n fact. the structure (2) tn 
whtch a relsttve clause extsts wtthtn the scope of 
• conJuncted noun phrase is relatively rare tn real 
texts, especially when the relattve clause ts 
rather long. Such constructions wtth long relattve 
clauses are a ktnd or garden path sentence. 
Therefore. unless strong heuristic rules like (2). 
(3) and (4) tn 4-2 suggest the structure (2). the 
structure (1) ts adopted as the ftrst chotce (Note 
that, tn \[ex-7\] tn 4-2, the strong heuristic 
rule\[rule (3)\] suggests the structure (2)). Stnce 
272 
the result of such a decision ts explicitly 
expressed tn the tree: 
SCOPE-OF-CONUN~CTI~ 
and the grammar rules in the later stages of 
processing work on thts structure, the other 
interpretations of scopes will not be tried unless 
the ftrst choice fatls at e later stage for some 
reason or alternative interpretations are 
explicitly requested by a human operator. Note 
that a structure llke 
NP-1 'TO' . ..... <verb> NP-2 ....... <verb> NP-3 r\[ 
relettve~clause 8!tecedent 
I 
relattve ~clause antecedent I I 
I 
conJunct~d noun phrase 
which ts linguistically posstble but extremely rare 
tn real texts, is naturally blocked. 
4-4 Sentence Relationships and Outer Case Analysts 
Corresponding to Engltsh sub-ordinators and 
co-ordinators like 'although'. 'tn order to'. 'and' 
etc.. we have several different syntactic 
constructions as follows. 
(1) .......... (Verb wtthe specific ............ 
Inflection form) I I 
I I 
$1 S2 
(2) ...... (Verb)(a postpostttonal particle) ...... 
! 
S1 S2 
(3) ..... (Verb)(a conjunctive noun) ............. 
! | 
I i 
S1 S2 
(1) roughly corresponds to Engllsh co-ordinate 
constructions, end (2) end (3) to Engltsh 
sub-ordinate constructions. However. the 
correspondence between the forms of Japanese end 
Engltsh sentence connections ts not so 
straightforward. Some postposttional particles tn 
(2). for example, are used to express several 
different semantic relationships between sentences. 
and therefore, should he translated tnto different 
sub-ordtnators in Engltsh according to the semantic 
relationships. The postpostttonal parttcle 'TAME' 
expresses either 'purpose-action" relationships or 
'cause-effect' relationships. In order to 
dtsambtguate the semantic relationships expressed 
by 'TAME'. a set of lextcal rules ts defined in the 
dictionary of "TAME'. The rules are roughly as 
follows. 
(1) If S1 expresses a completed actton or a 
stative assertion, the relationship ts 
"cause-effect'. 
(2) If $1 expresses neither a completed 
event nor e statIve assertion and $2 expresses s 
controllable action, the relationship ts 'purpose- 
action'. 
\[ex-e\] 
(A) $1: TOKYO-NX 
(Tokyo) 
IT- TEITA 
(to go) (aspect 
formative) 
TAME 
52: KAIGI-N! SHUSSEK| DEKINAKA- TA 
(meeting) (to attend) (cennot)(tense format- 
ive : past) 
$1: completed actton 
(the aspect formative "TEITA" means 
completion of an action) 
---> \[cause-effect\] 
- Because I was in Tokyo. I couldn't 
attend the meeting. 
(B) $1: TOKYO-NI IKU 
(Tokyo) (to go) 
TAME 
$2: KAIGI-NI SHUSSEKI DEKINAI 
(meeting) (to attend) (cannot) 
$1: neither a completed action nor 
a stattve assertion 
S2: "whether I can attend the meeting 
or not • ts not controllable. 
---> \[cause-effect\] 
• Because ! go to Tokyo. I cannot attend 
the meeting. 
(C) S1: TOKYO-NI IKU 
(Tokyo) (to go) 
TAME 
S2: KIPPU-O KAT- TA 
(ttcket) (to buy) (tense formative: past) 
$1: neither a completed action nor 
a stative assertion 
S2: volitional action 
---> \[purpose-action\] 
• In order to go to Tokyo. I bought a 
ticket. 
Note that whether S1 expresses a completed 
action or not is determined tn the preceding phases 
273 
by ustng rules whtch uttllze espectual features of 
verbs described tn the dictionary and aspect 
formattves following the verbs (The classification 
of Japanese verbs based on thetr aspectual features 
and related toptcs are discussed tn \[8\]). Ve have 
already wrttten rules (some of whtch are heuristic 
ones) for 57 postpostttonal particles for 
conJucttons of sentences 11ke 'TAME'. 
Postpostttonal particles for cases, whtch 
follow noun phrases and express case relationships, 
are also very ambiguous In the sense that they 
express several different deep cases. Vhtle the 
Interpretation of tnner case elements are dtrectly 
given tn the verb dictionary as the form of mapping 
between surface case part|cles and thetr deep case 
Interpretations. the outer case elements should be 
semantically Interpreted by referring to semanttc 
categories or noun phrases and properties of verbs. 
Lextcal rules for 62 case particles have also been 
Implemented and tested. 
5 Conclusions 
Analysts Grammar of Japanese tn the Mu-proJect 
ts discussed tn thts paper. By Integrating vartous 
levels of heuristic Information, the grammar can 
work very efficiently to produce the most natural 
and preferable readtn 9 as the f|rst output result. 
wtthout any extensive semanttc processtngs. 
The concept of procedural granwars was 
originally proposed by Wtnograd\[9\] and 
Independently persued by other research groups\[lO\]. 
However. thetr clatms have not been well 
appreciated by other researchers (or even by 
themselves). One often argues agatnst procedural 
grammars, saytng that: the linguistic facts 
Wtnograd's grammar captures can also be expressed 
by ATN. and the expressive power of ATN ts 
equivalent wtth that of the augmented CFG. 
Therefore; procedural grammars have no advantages 
over the augmented CFG. They Just make the whole 
grammars complicated and hard to maintain. 
The above argument, however, mtsses an 
Important po|nt and confuses procedural grammar 
wtth the representation of grammars tn the form of 
programs (as Shown tn Vtnograd\[9\]). Ve showed tn 
thts paper that: the rules whtch gtve structural 
constraints on ftnal analysts results and the rules 
whtch choose the most preferable linguistic 
structures (or the rules whtch block "garden path" 
structures) are different tn nature. \[n order to 
Integrate the latter type of rules tn a untfted 
analysts grammar, tt ts essential to control the 
sequence of rule applications explicitly and 
Introduce strategic knowledge tnto grammar 
organizations. Furthermore. Introduction of 
control specifications doesn't necessarily lead to 
the grammar In the form of programs. Our grammar 
wrtttng system GRADE allows us a rule based 
specification of grammar, and the grammar developed 
by ustng GRADE ts easy to maintain. 
Ve also dtscuss the usefulness of lexicon 
driven processing 4n treattng Idiosyncratic 
phenomena tn natural languages. Lax|con drtven 
prcesstng ts extremely useful tn the transfer phase 
of machtne translation systems, because the 
transfer of lextcal ttems (selection of appropriate 
target lextcal ttems) ts htghly dependent on each 
lextcal ttem\[tt\]. 
The current verston of our analysts grammar works 
qutte well on t.O00 sample sentences tn real 
abstracts wtthout any pre-edtttng. 
Acknowledgements 
Appreciations go to the members of the 
Nu-ProJect, especially to the members of the 
Japanese analys4s group \[Mr. E.Sumtta (Japan \[BH). 
Hr. M.gato (Sord Co.). Hr. S.Ten|gucht (Kyosera 
Co.). Hr. A.Kosaka (~EC Co.). Mr. H.Sakamoto (Ok1 
Electr|c Co.), MtSS H.Kume (JCS). Hr. N.\[shtkawa 
(Kyoto Untv.)\] who are engaged tn Implementing the 
comprehensive Japanese analysts grammar, and also 
to Or. 6.Vauquots. Dr. C.Bottet (Grenoble Untv.. 
France) and Dr. P.Sabat|er (CNRS. France) for 
their fnuttful discussions and comments. 
References 
\[t\] S.Vauquots: La Traductton Automat|que 8 
Grenoble, Documents de Linguist|qua Quantitative, 
No. 24, Par|s, Dunod, t975 
\[2\] J.Nakamura et.al.: Granunar Vrtttng System 
(GRADE) of Nu-Machtne Translation Project and tts 
Characteristics, Prec. of COL\[NG 84. t984 
\[3\] J.Slocum: A Status Report on the LRC Nach|ne 
Translation System, Vorktng Paper LRC-82-3. 
Linguistic Research Center, Untv. of Texas, t982 
\[4\] F.Pere|ra et.al.: Oef|ntte Clause GRammars of 
Natural Language Analysts. Artificial Intelligence. 
Vol. 13. 1980 
\[5\] G.Gazdan: Phrase Structure Grammars and Natural 
Languages. Prec. of 8th \[JCA\[. 1983 
\[6\] Y.M|lks: Preference Semantics, tn The Formal 
Semant4cs of Natural Language (ed: E.L.Keenan), 
Cambridge University Press, t975 
\[7\] Y.Sakamoto et.al.: Lextcon Features for 
Japanese Syntactic Analysts In Mu-ProJect-JE, Prec. 
of COLING 84, 1984 
\[8\] J.TsuJ41: The Transfer Phase tn an 
English-Japanese Translation System. Proc. of 
COLING 82. t982 
\[g\] T.Mtnognad: Understanding Natural Language, 
Academic Press, t975 
\[tO\] C.Bottet et.al.: Recent Developments tn 
Russian-French Machtne Translation at Grenoble, 
Linguistics, Vol. 19, tg8t 
\[tt\] M.Nagao. et.al.: Dealing wtth \[ncompleteness 
of L4ngutsttc Knowledge on Language Translation. 
Proc. of COLZNG 84. 1984 
274 
