Lexical Information for Determining Japanese 
Unbounded Dependency 
Shin-ichiro KAMEI, Kazunori MURAKI and Shin'ichi DOI 
Information Teclmology Research Laboratories, NEC Corporation 
4-1-1 Miyazaki, Miyamae-ku, Kawasaki 216, JAPAN 
{kamei, k-muraki, doi}@hum, cl. nec. co. jp 
Abstract 
This paper presents a practical method for a global 
structure analyzing Mgorithm of Japanese long sen- 
tences with lexical information, a method which we 
call Lexical Discourse Grammar (LDG). This method 
assumes that .Japanese function words, such as con- 
junctive particles (postpositions) located at the end of 
each clause, have modality and suggest global struc- 
tures of Japanese long sentences in cooperation with 
modality within predicates or auxiliary verbs. LDG 
classifies the encapsulating powers of function words 
into six levels, and modality in predicates into four 
types. LDG presumes tile inter-clausal dependency 
within Japanese sentences prior to syntactic and se- 
mantic analyses, by utilizing the differences of the en- 
capsulating powers each Japanese function word has, 
and by utilizing modification preference between func- 
tion words and predicates that reflects consistency of 
modality in them. In order to confirm the encapsula- 
tion power of Japanese function words, we analyzed the 
speech utterances of a male announcer and found the 
correlation between a particle's encapsulating power 
and the pause length inserted after the clause with a 
conjunctive particle. 
1 Introduction 
When analyzing long sentences with two or more pred- 
icates (i.e. compound and complex sentences), it is dif- 
ficult to grasp the proper structure of sentences hav- 
ing a large nmnber of possible dependency (modifier- 
modifee relation) structures. This difficulty is more 
marked in Japanese than in English, since there are 
more syntactically ambiguous structures in Japanese. 
Tile Japanese language has few syntactic indicators for 
dividing sentences into phrases or clauses, unlike En- 
glish with its relative pronouns and subordinate con- 
junctions. One of the most critical features of Japanese 
is that the difference between a phra~se and a clause is 
not cleat'. Even subjects or other obligatory elements 
of clauses are omitted very often when they aye indi- 
cated by contexts. In addition, the Japanese language 
does not have any parts of speech to clearly indicate 
either the beginning or end of a phrase or a clause. 
Another critical feature is that the Japanese language 
is an almost pure Head-final language, i.e., predicates 
and function words to signify the sentence structure ap- 
pear at the end of the clause or sentence. This means 
that it is syntactically possible for all phrases or clauses 
that can modit) predicates to modify all other phrases 
or clauses that appear in the latter part of long sen- 
tences. 
These syntactic characteristics of the Japanese lan- 
guage make it difficult to determine the dependency 
(modification) structure of hmg sentences. Simple 
parsing of Japanese long sentences inevitably produces 
a huge number of possible modification structures. A 
conventional bottom-up parsing method can reduce 
ambiguity in modification by local information in tim 
surface structure. However, this inclines toward an 
improper output, since the locally highest likelihood is 
sometimes low on the whole. 
To overcome this problem, several methods to pre- 
dict the global structure of long sentences have been 
proposed. One is a top-down parsing method by 
matching the input sentence and the domain-specific 
patterns (Furuse et al., 1992). Improvements made by 
other researchers enabled this method to parse irregu- 
lar, incomplete and multiplex patterns, by describing 
the domain-dependent patterns in the form of gram- 
mar (Doi, Muraki, et M., 1993). 
Another method employs global structure presump- 
tion to divide a sentence into clauses by utilizing 
general lexical information. It predicts the sentence 
structure prior to syntactic analysis only by utilizing 
domain-independent lexical information such as con- 
junctive particles, parallel expressions, theme transi- 
tion, etc. (Mizuno et al., 1990; Kurohashi et al., 1992). 
Lexical Discourse Grammar (LDG) is one of the 
approaches with which a global structure of a long 
sentence is presumed by focusing on function words 
(Kamei et al, 1986; Doi et al., 1991). LDG assumes 
that Japanese function words, such as conjunctive par- 
tides (postpositions) located at the end of each clause, 
convey modMity, or propositional attitude, and suggest 
global structures of Japanese long sentences in cooper- 
ation with modality in predicates, especially within 
310 
auxiliary verbs. LDG can presume the inter-clausal de- 
i)endency within Japanese sentences prior to syntactic 
and semantic alLalyses by utilizing tilt; difDrences of 
the encapsulating powers each ,\]apaiLese function word 
has, and by utilizing modification preference-between 
function words and predicates that reflects consistency 
(if lnodality reading or propositional attitude inte, rpre- 
tation. 
LDG is effective in reducing the syntactic ambigu- 
ities, and it him alre~My been applied to a machine 
translation system. Ilowever, it has not claritied the 
level of tile encapsulation powers of Japanese flint:- 
tion words or tile relation between modality and level. 
IIence we refined the concept of LDG, lmrticularly tile 
conjunction level of function words, and explain the 
outline of the refilled LDG in tilts paper. First, we 
present the encapsulation power of Japanese function 
words, which are classified into six levels. Second, we 
state moditication l)refi3renee of ,}apanese conjunctiw'. 
particles that reflects modality within them. Filmlly, 
we present evidence of tile h;vels of aaplnLese func- 
tion words. We think that tile h;vels of clauses pro- 
duce prosodic infbrmation, especially tlw location and 
h;ngth of pauses, which are influenced by tile sentence 
51obal struchn'e. We atnalyzed the speech utterances 
of a professional new,~ announcer (male) and fonnd a 
correla,tion between a particle's encapsulating power 
and tile pause hmgt;h inserted ai'ter dm clause with a 
conjunctiw; i~ar rich;. 
2 Lexical Discourse Grammar 
2.1 Levels of Conjunctive Particles in 
Japanese 
\]Jr ,Jap;~n(;se eOlnl)lex or compound SelLtellCes~ Sll})()r- 
dinate clauses have several dependency levels relative 
to the main clause. Conjunctive particles, which are 
located at dm end of clauses and which link them, are 
classified according to the elemeni;s that the clause 
can contain, or to the correh~tion between clauses. 
Se, e tl~e fbllowing examples with conjunctive pard- 
cles ",,ode"(© "¢) ~t,,(t "nagara"( ~,~ 7>" 6) (* is added to 
meaningless sentences). 
m') 
 ,. oao6She)-.e.,(s ECT) 
t,~suke(Help)--ta(PAST) -node(Because) 
seikou (succeed) -shit a(PAST). 
Ite succeeded because she helped him. 
K,.re(HJ--w, ( OeIC) 
kaet (Rot.m) -t~ 
He retunmd while lie was talking. 
3)* ~ ~,to~ i~r, " b*~ ~ ~')i~}o ?:,, 
ka lojo(She)-ga(Sm3.U C'r) 
hanashi(Ta110-,,~gara(While ) 
 a(mSa'). 
He returned while she was talking. 
{ 
warai(Smile)-nagara(While) 
tazune(Ask)- t~(PAST)- node(Becavse) } 
wa tad(1)--wa (TOPIC) 
k ot~e( Answe~9-ta(PAST ). 
i answered as lie asked while tie was smiling. 
{  4SU13JEC ') 
famine(Ask) t~(l'AST)- node(ibca,,se) } 
wa,'ai(S,*li 4,,ag, r,(Wm,O 
watasi(.l) w,ffTOPKO kotae(Answel9 ta(l"AST). 
I answered while i was smiling as lie asked. 
A clatuse with the conjunctive p~Lrticle to express 
~reason' "node"too'el can contaii, ~ subjective noun 
phrase and an auxiliary verb of past tense "t~d'(7:), 
while a clanse with the particle indicating attendant 
action %agara"('&~/d 6) cannot, as shown in 1) 3). 
Sentence 4) in comparison with 5) shows that a clause 
with "node"((/)(?) can subordinate at clause, with "na 
g~md'()7')~6), but the reverse is impossible. In these 
two selt~,011(;eS, brackets { } show subordinaU: clauses. 
Consequently, "nagara"()'/2 6) is ranked at a lower 
h,vel than "nodr"( (O -0"). 
in I,DG, conjunction levels of clauses are divided 
into six classiiic~tions according to the elements the 
clause clm contain, a,s listed in Table 1. These levels 
construct ~t hierarchy, i.e., at lower level clause caltlctofi 
subordinate a higher level one. The levels also repre- 
sent the encapsulating powers of each Japanese. fltIlC- 
tion words located at tile end of the clauses. Besides 
em0unctiw' partMes, Japanese conjunction nouns (u: 
relative nouns are also classiiied avd assigned it level. 
ltere, Japanese conjunction ILtltllLS, such as "toki" (lt~/: 
: :when) are nouns that can often be nsed just like COL> 
juncdve l~articles when alley are attached at the end 
of clause. Japanese relative notln8, such as "lnae" (\]ii'j 
::before) are another type of nouns that play roles sim- 
il;~r to those of conjunctions in English when they are 
moditied by predicative phrases or clauses. 
2.2 Modality in Conjunctive Particles 
and Modification Preferences 
The conjunction lewis we introduced above reduce the 
syntactic ambiguities of long sentences. However, in 
order to select the most reliable struchtre of sentences, 
we use another important discourse feature tile con- 
junctive pt~rticles have, i.e., modMity. 
LDG assumes Japanese function words have moda\]- 
ity or ~proposidonal attitudes' ~md suggest global 
,~tructures of Japanese hmg sentences in cooperation 
with modality within ~mxiliary verbs. Wc mssllIne du~t 
the same kind of modality in a conjunctive particle anti 
311 
Table 1 Conjunction Level in LDG 
levels 
LEV.0 
LEV.1 
LEV.2 
LEV.3 
LEV.4 
LEV.5 
Definition of level Example of function words 
can contain every element 
cannot contain 
sentential particles 
cannot contain 
conclusive modal 
cannot contain 
probable modal 
"to"( ~ =that(quotation)) 
"k~ra"(~ 6 =because), "node"(69 ~ =because), 
"keredomo"( ~:~ ~ E" % =but) 
"nara"( t~¢ 6 =if), "hoka"(~N =besides), "to"( a =when), 
"baai"(t~@ =in case of), "toki"("# =when) 
"mae"(~ =before), "ato"(~ a =after) 
cannot contain "totomoni"( ~ & ~ K- =as), "tame"(£: ~5 =because) 
tense expressions "todoujini"( ~ ~1~# ~- =at the same time) 
cannot contain "nagara"( re ~" 6 =while), "tsu ts u"(o o =while), 
particle "ga"(;O ~') "kotonaku"( t ~ t¢ < =without) 
higher 
t 
lower 
a predicate (or an auxiliary verb) correspond to each 
other. From the parsing viewpoint, this suggests that 
each conjunctive particle has modification preference 
with certain predicates or auxiliary verbs. 
From the viewpoint of modality, there are four pred- 
icate types in Japanese; (1) Auxiliary verbs of the first- 
type modality (conjecture etc.), (2) Auxiliary verbs of 
the second-type modality (necessity etc.), (3) Copula, 
and (4) Plain (present and past tense) forms of Verbs. 
Here, first-type modality includes conjecture, such as 
"darou"(f2"7~ ")) which corresponds to 'may,' 'can,' 
'maybe,' and 'possibly' in English auxiliary verbs, ad- 
verbs, and adjectives. Second-type modality includes 
necessity, such as "nakereba-naranai" (re t~ fc t:ft3 6 
to¢ u,) and "ta-hou-ga-yoi"(t: t! ") ~3: u~) which corre- 
spond to 'have to' or 'must,' and 'had better,' 'should,' 
or 'preferably' in English. The Japanese Copula "da" 
(?Z) or "desu" (~¢'Y) means definition or speaker's 
judgment with confidence. Phdn forms of verbs are 
the present or past tense forms of verbs without any 
modal auxiliary verbs. These forms do not have any 
modal morpheme, but when they which appear at the 
end of the sentence and are followed by a period they 
CAN convey modality, that is, attitudes or intentions 
of the subject or speaker. Plain forms of verbs in a 
relative clause which modify a nominal phrase do not 
have such modMity. 
LDG assumes that each conjunctive particle has a 
preference in modifying predicates or auxiliary verbs 
with consistent modMity. There are six levels of modal- 
ity in conjunctive particles, and there are four types of 
modality in predicates, as mentioned above. A sub- 
ordinate clause with modality modifies a consistent 
modality predicate type. The following figure illus- 
trates the modality consistency between particles and 
auxiliary verbs in Japanese sentences. 
{Clause + particle} 
Modality 
{Clause + auxiliary verb} 
Modality 
Take the Japanese conjunction noun "toki" (1~ = 
when, if) for example. This word corresponds to ei- 
ther the English conjunction 'when' with neutral read- 
ing or 'if' with conjecture modality. When the word 
"toki" is used as the 'if' reading, this word modifies 
a clause in which the modality is expressed. In most 
cases, auxiliary verbs such as "darou" (?Z,5 ") = may, 
maybe) or "ta-hou-ga-yoi"(t~ It 5 7~3: ~ = had better, 
should, preferably) express the modality of the modifee 
clause. The Japanese language has some words that 
indicate or emphasize the fact that the word "toki" is 
being used as the "if" reading. One of them is the ad- 
verb "moshi" (~o L) that indicates a supposition read- 
ing is applicable. This adverb is never used by itself 
and always modifies conjunctive forms such as "toki," 
"nara," "to," and so on, and selects or emphasizes the 
supposition reading of the conjunctive forms. Another 
such word is the particle "w£' (~:t) , which is usually 
used as a topic marker for a sentence. When "wa" is 
attached to "toki," that is, in the form of "toki-wa," 
the supposition reading is enhanced. This tendency 
is strengthened by the use of comma after the phrase 
"toki-wa." The phrase "toki-w£' tends to be used to 
modify phrases with auxiliary verbs of modality. 
When this phrase with modality modifies a plain 
form of a verb with a period at the end of the sentence, 
the readers recognize that the plain form of the verb 
contains a kind of modality, such as the subject's or 
speaker's intention. In other words, modality informa- 
tion of the subordinate clauses is attached to the plain 
form of the main verb. The following figure illustrates 
this interpreting mechanism. 
312 
\[ Input Sentence Reader 
\[ Morphological Analysis 
\[ Discourse Structure Reference 
\[ Discourse Structure Assumption 
<=> ~ictionary I 
Discourse Form Preservation \[ 
Discourse Structure Analysis 
\[ Synta.ctic & Sem~mtic Analysis 
Figure 1 Analysis based on LDG 
(Modality}-- 
1 
{Clause + particle} {Clause (Phdn form).} 
Modality 
In contrast, when a subordinate clause does not have 
modality explicitly and modifies a clause with modal- 
ity, the readers interpret the subordinate clause as that 
with a kind of modality such as conjecture. The fol- 
lowing figure illustrates this situation. 
? (Modality) 
{Clause + particle} 
1 
{Clause + auxiliary verb} 
Modality 
The modality coincidence described in this section 
is the base for analyzing Japanese long sentences. The 
.Japanese language has few syntactic indicators to show 
the segments of sentences, but is rich in semantic indi- 
cators which suggest sentence structure. The sem,~n- 
tic indicators are the modalities that a wide range of 
parts of speech have. Conjunctive particles, adverbs, 
and even plain forms of verbs can have modMity in the 
Japanese btnguage. The modality structure is the key 
to comprehending Japanese long sentences. 
2.3 Japanese Sentence Structure Pre- 
sumption 
We assume that the modality structure (:an mMnly be 
detected by lexical intbrmation. Based on this assump- 
tion, LDG presumes the sentence structure before syn- 
tactic and semantic analyses on the ba~sis of previously 
collected lexical information that characterizes the lex- 
ical discourse. 
Figure 1 shows the configuration of our Japanese 
long sentence analyzer based on LDG. Input sentences 
are first analyzed morphologically. The part 'Discourse 
Structure Analysis' in Fig. 1 then presumes the sen- 
tence structure, before syntactic and semantic analysis. 
ttere, 'Discourse' means an inner-sentence congruence 
in Japanese long sentences thttt contain two or more 
predicates. 
In order to reduce the huge number of syntactic 
structures of Japanese hmg sentences and give priori- 
ties to each possible structure, the analyzing method 
based on LDG uses global modality structure focusing 
on lexic~tl information. 
First, the Discourse Structure Reference module re- 
duces the number of possible syntactic structures, us- 
ing the level of conjunctive particles described in the 
previous section. After that, the Discourse Structure 
Assumption module gives priorities to each possible 
syntactic structure, using the modification prefl;rence 
based on modality. 
3 An Application of LDG 
3.1 Pause Control with LDG 
The level of conjunctive particles, which indicates the 
structure of the Japanese long sentences, is the most 
important feature of LDG. In this section we apply 
the level to another linguistic phenomenml in order to 
('onfirm the validity of this model. 
The sentence structure influences a wide range of 
linguistic phenomena. One example is prosodic infor- 
mation (Dorffner et al., 1990; Iwata et al., 1990; Kaiki 
et al., 1990; Sakai et al., 1990). If the correct sentence 
structure is acquired for each input sentence, prosodic 
information can be accurately calculated. As yet, 
even the most up-to-date, advanced systems have not 
achieved the analysis in the deep structure, therefore 
sentence structure presumption in the surface struc- 
ture is essentiM for a robust prosodic control system. 
LDG meets this requirement since it presumes the sen- 
tence structure by means of function words occurring 
on the surface (Doi, Kamei, et al., 1993). Hereafter, 
we propose the prosodic control system based on LDG. 
The presumption function ibr sentence structure 
(lexical discourse) by LDG is applied to pre-processing 
313 
Table 2 Pause length data 
Conjunction Levels 
With a comma Without a comma 
Number of cases 
LEV. 1 ( "gW', "ka,'a", %ode",etc. ) 
Number of cases 
(with/without a pause) 
LEV'. 0 ("~o", "~ te",etc.) 0 - 
461.6 11 (11/0) 
LEV:2( "ha', "~o', "tara",ete.) 
LEV.a("a*o",ete.) 
LEV.4( "tame", "temo", %odo',etc. ) 
LEV. 5 ( "tsu tsu ", "naga~'a", "z uni",etc.) 
Verbs in adverbial form 
Verbs in adverbial form + "te"('Q 
Adjectives in advei'bikl form 
Adjectives in adverbial fbrm + "re" 
Predicative auxiliary verb "da"(?2") 
Average pause 
length \[msec\] .. 
1 ( 1/ 0) 
4 (4/0) 
11 (11/ 0) 437.0 
2 ( 2/ 0) 277.5 
5 ( 5/ 0! 421.5 
"4 ( 4/ 0) 293.8 
40 (40/ 0) 468.8 
14 (12/2) 331,8 
3 (3/o) 542.5 
410.0 
603.1 
(with/without a pauae) 
7 ('1/6) 
Avel'age pause 
length \[msee\] 
7.9 
3 ( 3/ 0) 563.3 
13 (11/ 2) 243.8 
1 (0/1) 0.0 
8 ( 6/ 2) 201.6 
5 (2/a) 12o.o 
6 (G/o) 252.1 
30 (15/15) 127.4 
.... i9. (8/i1) 89.2 
6 ( 4/ 2) 56.7 
3 (3/.0) 359.2 
ahead of speech synthesis, in a text-to-speech system. 
It can presume the global sentence structure through 
lexical information without any analysis in the deep 
structm'e. It is also possible to consider the pause 
length inserted after each clause in relation to the lex~ 
ical information in LDG. In other words, pauses are 
more fl'equently inserted after the clause of the higher 
conjunction levels than those of the lower levels. Corn 
sequently, the pause length and location can be more 
efficiently controlled with the LDG conjunction levels. 
To develop a text-to-speech conversion system with 
LDG, it is necessary to prepare the LDG conjunction 
level information of a large nmnber of conjunct equiv- 
alents such as conjunctive particles. Statistical data 
should also be collected fi'om human speech and read- 
ing, in regard to the correlations between pause length 
and the LDG conjunction levels. This substantial data 
is added to the lexical information to be used for speech 
synthesis in cooperation with pronunciation and ac- 
cent. 
3.2 Data Analysis 
To confirm the correlation between the conjunction 
level and the pause length, we have analyzed speech 
data spoken by a professional news announcer (male), 
reading newspapers and magazines at a regular speed. 
Wc extracted conjunctive particles and verbs, auxiliary 
verbs and adjectives in adverbial form from the speech 
data, and classified these words by the LDG conjunc- 
tion level. The average pause length for each level was 
calculated for two separate cases; words preceding a 
comma and words without a comma. See Figure 2 and 
Table 2. 
For words without a cormna (marked with white bars 
in Fig. 2), the result shows that the higher the con- 
junction level is, the longer the average pause length is 
(except for LEV.0, which is a particle for quotation). 
This tendency basically does not depend on whether 
or not a comma exists after the words. However, for 
words with a comma (marked with black bars in Fig. 2) 
pause length of Lev. 3 is shorter than that of Lev. 4. 
We suppose that the reason for this phenomenon is 
that a comma adds modality to the words and length- 
ens the pauses, as described in the previous section. 
Taking the comma effect into consideration, we can 
conclude that there is a solid correlation between the 
LDG conjunction level and the pause length. 
LEV.0 ("to"(&) and "tte"(o'C)functioning in a 
similar way as quotation marks), however, requires a 
careful observation. This conjunction level, the highest 
rank, can contain every element, even an independent 
sentence. In this case, the relation between the con- 
junctive particle and its preceding clause is so weak 
that a pause tends to be inserted BEFORE the con- 
junctive particle, not after it. Therefore, in the present 
data, a pause was inserted after the particle in only one 
case out of seven. 
LEV.0 
LEV.I 
LEV.2 
LEV.3 
LEV.4 
LEV.~ 
adv. 
adv, + 're 
.......... I 
II I 
I 
i i 
"--'----1 
I 
I t ...... 0 i00 200 300 400 500 600 
\[msec\] 
r'n: without a comma .,,, : with a comma 
Figure 2 Average Pause Length Diagram 
In Table 1, no level is assigned to two of tile most 
314 
fl'equent groups: verbs and auxiliary verbs in adverbial 
tbrm, and verbs and auxiliary verbs in tile same form 
with a conjunctive particle "te"(-C). These groups 
are difficult to allocate to a single level, as dmy are 
used in expressing many factors such its parataxis, 
cause, means~ attendant circumstances, and because 
they vary semantically and syntactically. However, in 
reference to the pause length data, the adverbial verbs 
ill the former group might fall into LEV.1 or LEV.2, 
while those in the latter group with "te"(-\[) might 
full into LEV.4 or LEV.5. Conventionally, these two 
groups; are often treated as one "adverbial form", al- 
dlough many functional diiferences have been pointed 
out between them. Our data supports the difference 
with respect to the pause length. There are identical 
tendenci~ between two adw;rbial forms of adjectives: 
("-k,F'(~//) and "-kute"(~ a 9=)) and two adverbial 
forms of pseudo adjectives: ("-ni"(~-=-) and "-de"( r)). 
4 Concluding Remarks 
We have proposed a practical method for a global 
structure anMyzing algorithm of Japanese long sen- 
tences with lexical information (Lexical Discourse 
Grammar: LDG). This model assumes that Japanese 
conjunctive particles convey modality, and modality 
structure can basically be detected by lexical informa- 
tion. We assign a ~conjunction level' to each conjunc- 
tive particle and reduce the number of possible syntac- 
tic structures of Japanese long sentences. In addition, 
we assume that all conjunctive particles have a mod- 
ification preference according to their modality. This 
preference assigns priorities to the possible structures 
of tile sentences. 
We applied LDG to a prosodic information control 
method in a Japanese text-to-speech conversion system 
to confirm the conjunction level experimentally. This 
method controls pause location and length in speech 
synthesis with the conjunction level in LDG, using only 
lexical information with no need fi~r syntactic analy- 
sis. Even so, it can tune tile pause length more finely 
than methods without sentence structure presumption. 
AnMyzing speech data, we confirmed a correlation be- 
tween the level of a function word and the length of 
a pause inserted ai'~er that word. We are now in the 
process of developing a speech synthesis system with 
this method, by defining the default pause length for 
each conjunction level. In future research, LDG will 
also be applied to other prosodic information (rhythm 
and intonation). 
There can be little doubt that LDG will be more 
etfective when two or nmre conjunct equivalents of dif- 
ferent levels appear in one sentence, since the LDG con- 
junction levels are closely related to the inter-clausal 
dependency. Unfortur~ately there were few such cases 
in the data used in this paper. In future work, we will 
collect such data to proove this hypothesis, in so doing 
will refine our method to improveits ability to analyze 
long Japanese sentences. 
References 
\[1\] S. Doi, K. Muraki and S. Kamei. 1991. Lexical 
Discom'se Grammar and its Application for Decision 
of Global Dependency (II). IEICN Technical R.epor% 
NLC91-29(PRUgl-64). (in Japanese) 
\[2\] S. Doi, K. Muraki, S. Kamei, and K. Yamabana. 
1993. Long Sentence Analysis by Dmnaln-Specific 
Pattern Grammar. Ill Proceedings of EACLg3. 
\[3\] S. Doi, S. Kamei, K. Muraki, Y. Mitome, and 
K. Iwata. 1993. Prosodic Information Control 
by Lexical Discourse Grammar. In Proceedings 
SIG-SLUD of the JSAI, SIG-SLUD-9301-4. (in 
Japanese) 
\[4\] G. Dorffner, E. Buchberger and M. Kommenda. 
1990. Integrating Stress and Intonation into a 
Concept-to-Speech System. In Proceedings of COL- 
ING90, 1.990 
\[5\] O. Furuse and H. Iida. 1992. An Example-Based 
Medlod for Transfer-Driven machine translation. In 
Proceedings of TMI'92, pp. 139-150. 
\[6\] K. Iwata, Y. Mitome and T. Watanabe. 1990. 
Pause Rule for Japanese Text-to-speech Conversion 
Using Pause Insertion Probability. In Proceedings of 
IUSLP, 2, pp.837-840. 
\[7\] N. Kaiki and Y. Sagisaka. 1990. Analysis of Pause 
Duration based on Local Phrase Structure. IEICE 
Technical Report, SP91-130. (in Japanese) 
\[8\] S. Kamei and K. Muraki. 1986. Proposal of Lex- 
ical Discourse Grammar. IEICE Technical Report, 
NLC86-7. (ill Japanese) 
\[9\] S. Kurohashi and M. Nagao. 1992. Dynamic Pro- 
tramming Method for Analyzing Conjunctive Struc- 
tures in Japanese. In Proceedings of COLINGg£, 
pp.170--176. 
\[10\] J. Mizuno and J. Nakagaki. 1990. A Study for 
three Structures of Japanese Sentence. IPSJ SIG 
Notes, 9(I-NL-76 4. (in Japanese) 
\[11\] S. Sakai and K. Muraki. 1990. From Interlin- 
gua to Speech : Generating Prosodic Information 
from Conceptual Representation. In Proceedings of 
ICASSPgO, S6a.10. 
315 
