PRI~;DICTING NOUN~PIIRASE SURFACh; I~'ORMS USING Q~ONTEXTUAL \[NFORMA'PION 
Takayuki YAMAOKA: Hitoshi IIDA~ and Hidekazu AILITA~ 
ATR Interpreting Telephony Research Laboratories, Souraku-guu, Kyoto, JAPAN 
tMitsubishi Electric Corporation, Amagasaki, tlyogo, JAPAN 
Abstract 
We propose a context-sensitive method to predict noun- 
phrases ill the next utterance of a telephone inquiry di- 
alogue. First, information about tile utterance type and 
the discourse entities of the next utterance is graspcd us- 
ing a dialogue interpretation model. Second, a domain- 
dependent knowledge base for noun-phrase usage is de- 
veloped, focusing on the dialogue situations in context. 
Finally, we propose a strategy to make a set of the appro- 
priate expressions in the next utterance, using the infor- 
mation and the knowledge base. This set of expressions 
is used to select the correct candidate from tim speech 
recognition output. This paper exantines some of the pro- 
cesses creating sets of polite expressions, deicti6 expres- 
sions, and compound ilOUll pllra-~es, which ~tre COlnlUOll 
ill telephone inquiry dialogue. 
1 Introduction 
A high-quality spoken-language processing system 
must use knowledge of dialogue and spoken-language. 
Using dialogue knowledge facilitates understanding 
and predicting utterances in context. Using spoken- 
language knowledge, that is knowledge about how tile 
speaker expresses what he/she wants to say, makes it 
possible for the system to recognize and generate the 
more complex expressions that are nornmlly ased ill 
our daily dialogues. 
To make language processing in the whole spoken- 
langnage processing system more efficient, it. is vital 
how to select the correct speech recognition output in 
the speech-language interface. The use of discourse- 
level knowledge is an effective way to do this\[6\]\[11\]. 
For example, MINDS\[6\] applied dialogue-level knowl- 
edge, particularly for propositional contents, to pre- 
dict the expected utterance form for the speech recog- 
nition, ilowever, although MINDS showed good re- 
suits, several llroblems remain before it can lie made 
into a complete spoken-language processing system: 
1. how to construct the dialogue structure for the 
given domain, 
2. how to treat predictive concepts regarding not 
only the propositional contents but also the 
speaker's intention, 
3. and, how to etmose a set of surface forms tbat 
the speaker might utter about the predicted con- 
cept. 
Also, MINDS was concerned with a system to par- 
ticipate in human-machine dialogue. On the other 
band, we want to monitor a human-human dialogue. 
We proposed a dialogue understanding model\[7\], and 
a context-sensitive method to predict abstract infor- 
mation allout both tile intentional and propositional 
contents of tile next utterance\[11\]. These are our an- 
swers to the above problems 1 and 2. 
From tile point of view of human behavior, a po- 
tential approacb to selecting the appropriate sur- 
face erpression forms (SEFs) is using spoken-lauguage 
knowledge. In general, when we are talking about a 
concept X, there are many possible surface expres- 
sious and forms to represent X. From a psychological 
(or psyeholinguistic) point of view, Clark\[3\] pointed 
out five abstract factors which should be considered 
ill ,asking what linguistic devices should speakers use ?. 
These are: knowledge of the listener, the coopeva- 
live principle, the reality principle, the social con- 
text, and the linguistic devices available. In the 
eomputatioual linguistics area, Appelt\[1\] has devel- 
oped a framework to generate a sentence in a context- 
sensitive way, based on speech act tbeories. Unfortu- 
nately, however, there also remains, as he described 
ms a future study, the problem of choosing a lexically 
appropriate SEF from among candidates in a social 
conlexl. 
This paper describes a context-sensitive framework 
for selectiug all SEF for noun-phrases(NPs). This 
method is sensitive to botb tbe utterance situation 
and the history of the dialogue. To do this, first, we 
analyze the relations between concepts and SEFs, and 
between applicable situations and contexts, using a 
corpus of Japanese inquiry dialogues. Then, we make 
a domain-dependent knowledge source for NP usage, 
and define rules driven by applicable conditions to 
determine a set of possible SEFs in the knowledge 
base. Finally, we give exanaples of the SEI" selection, 
especially for polite expressions, deictifi expressions, 
and compound NPs, which are common in our target 
domain, and describe a simple experiment to evaluate 
using the ATR dialogue database. The result show 
tllat tile method can choose tbe contextually correct 
expression from the speech recognition output can- 
didates, and can be used in tile generation module 
of a spoken-language processing system to generate 
and determine all appropriate expression under tile 
dialogue situation. 
Throughout this paper, all examples are in 
Japanese and written in italic. English translations 
follow in parentheses. NP denotes a noun phrase, and 
SEF denotes a surface expression form. SEFs are en- 
closed ill double quotation marks and concepts are 
enclosed in single quotation marks. 
2 Dialogue Interpretation and Predicting 
the Next Utterance 
The next. utterance call be predicted after under- 
standing the previous utterances, because predicted 
information must be affected by tile dialogue struc- 
AcrEs DE COLING-92, NANTES, 23-28 Aour 1992 1 1 5 2 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 
lure. This section brielly describes tile model for in- 
terpreting a dialogue\[7\] and tile ulethod of pre(lictng 
the next uttcrance\[ll\]\[12\]. 
In tile lnodel, an utterance is represented, by a 
predicate form. An typical Japanese sentence, "(;0- 
juusyo we ouegai-shi-masu." (May 1 have your ad 
dress?), uttered by the secretariat in a inquiry dia- 
logue, is shown below: 
(ASK-VALUE u q (address q) (IS (address q) ?va\].)) 
where constant s denotes the secretariat, q the 
questioner, address tile concel)t of an a.ddress, and 
the variable, ?val, is the value for the address of q. 
The dialogue interpretation model h;us \[our types 
of plan and can interpret input utterances as the di- 
alogue proceeds, using au extended plain inference 
mechanism\[7\]. Thus, a dialogue structure can be con- 
structed. 
lit order to provide contextual ieforlnation about 
discourse entities we use typed variable notation\[2\] to 
describe a discourse entity in a plan schema, l';ach 
type in this notation corresponds to a particular con- 
cept node ill the domaiu-del)eudent NP knowledge 
base (described in Section a.3) '\['he following de- 
scription is an example of a Domailt Pla~J to send 
something: 
(Domain-Plan: SERD-SOMETHIN~ 
(HEADER (SEND ?a:person ?r:person ?s:object)) 
(PRECONDITION (KNON ?a ?d:destination)) 
(EFFEECT (HAS ?r ?s)) 
(CORSTRRINT (BELDNG ?d ?r))) 
The state of understanding is managed nsieg two 
pushdown stacks. The mlderstanding llst stores 
completed plans as the current understanding state, 
and tile goal list maintains incomplete plans ~Ls pos- 
sibilities and expectations tbr fltture goals. By rc 
ferring to tile goal list., the next utl.erance can I)e 
predicted on an abstract, level as the dialogue pro- 
ceeds, using the two generalized rules: e×imctation 
mid prefere.nee\[12\]. 
Predicted utterances are tel)resented in the slttnc 
style as intmt utterances. As a result,, we can predict 
two types of infornlatiou, one about the COlllllltlUiCa- 
live act, types and the other about discourse entities 
in the propositional contents (or in the topic slot,) (t\[" 
the ilext uttcrarlce, lnforrnatioll abotlt a discourse 
entity lnay appear in the forlu of ;ill particular ex 
pression if it is ill a prvvious utterance that can be 
related to the. current atterauce. Othe.rwise infornl~ 
ties will be in the tbrm of a type represcuting a par- 
ticular concept in tile related domain plan. We call 
such information conteztual information in tile t~sk 
of selecting ttle constituents of tile next utterance. 
3 NP Identification Model 
3.1 Chauge to NP Linguistic Expressions 
hi general, when we are talking about a concept X, 
there are many possible surface expressions and fornls 
to represent X. In llarticular, Japanese ha.s several 
possible SEFs for a given X, one from ttle Chinese 
reading and another based on tile original Japanese 
language (e.g. "oka~sakf' and "atesakf' for 'destina- 
tion' in Fig. 1). In addition, there are particular phe- 
nomena of expression variations depending hi)on the 
particular dialogue. For exmnple, if a speaker is ut- 
tering his/her own address for tile concept 'address', 
he/she will use "3uusyo"(\[my\] address), e.g. "Juusyo- 
wa Oosaka-sht desu."(My address is in Osaka~city.). 
On the other hand, if he/she is uttering tile other par- 
ticipant's address, he/she will use ".qo.juusyo"(\[your\] 
address (polite fornl)), e.g. "Go-juusyo-wo onegai- 
shi-masu."(Your address, please?) These facts lead 
rlS to ilnplenmltt knowledge SOllrces elf s/Idl vari~. 
lions(we call them changes) in a computational l)rc~ 
cessilr g systenl. 
Only by liltcriug using any intra-sentential knowl- 
edge sources, several candidates may remain ~m syn- 
tactically and semantically correct senteuces. For 
example, "9ojuu-shichi'(fifl.y-seven) sounds like "yo- 
3uusyo', and the sentence "Gojuu-shichi-wo onegai- 
shi-masu."(Fifty-seveu, i)lea-ne.) is not only well- 
formed but also correct in a particular context. It is 
possible to select the correct candidate by referring 
to both the context and the situation of the ongoing 
dialogue. Ewm so, to pick the surface form, we nmst 
kllow wily tile speaker haq used ~t given expression to 
represel~t a COIIgellt, 
If we can determine llow these NPs ehallge, ~,lld 
what effect they bawl then we can choose the speech 
recognition candidates more accurately. 
3.2 Analysis of NP Changes 
In order to analyze NI' changes in a dialogue we. in- 
spected 50 dialogues in a corpus. As a result of the 
analysis, NP changes arc categorized into three main 
classes: 1) Change by lexical e.ohesion: (tiffs class 
corresponds to reiteration\[5\]), 2) Change by dif- 
fere.nt viewpoints: (described in detail in the next 
paragraph), and 3) Change by misrecognition. 
There are two aspects of viewpoint, which are the 
standpoinl of the agent and the node of the concept. 
in an inquiry dialogue, the standpoints of the agents 
are. always different. Thus, this class has only two 
stlb cla~ssl!s : 
2(a) iioint keeping: both agents see the s;uue 
t~odc of a concept, and this subclass is divided 
according to the SEI.' : 
i. ditferent expression- 
e.g. "walashf' and "Yama0ka-san". 
il. additimt of prefix 
e.g. "lUUSyO" and "9o-juusyo". 
iii. COml)lex a mixture of 2(a)i and 2(a)ii, 
2(I)) shifting: the viewpoint of one of the agents 
shifts from thc node of a conccpt to a node of 
a related concept, and this is divided into: 
i. shortening 
e.g. "Kokusai-kaigf'(lntenrational Confer- 
euee) and "kaiqf' (tire conference). 
it. uniting- 
e.g. "ryousyuu-syo-to saNka-touroku-syo"(a 
receipt and an application form) and '"2- 
syurnt-uo syorltf'(two types of forms). 
ACRES DE COLlNG-92, NA~rzs, 23-28 lot's 1992 1 1 5 3 PRec. OF COLING-92, NA~ri~s, Auo. 23-28, 1992 
I 
node I ~-.,v..~.~. ....... - 
Figure h Example of NP knowledge base 
iii. specification- 
e.g. "niNzuu"(nnmber of people) and "saNka- 
niNzuu" (number of participants). 
3.3 Domain-dependent Knowledge 
Configuration: q-'he domain-dependent knowledge 
base consists of a network of nodes and links. Basic 
nodes are divided into three types: a concept node 
representing a particular thing or concept retained in 
human memory, a lexleal node representing a par- 
tieular word or phrase used when expressing some- 
thing, and a variable node representing a particu- 
lar value corresponding to a valuable concept, which 
can have a specified value. A variable node can be in- 
stantiated by executiug tbe effect of a completcd plan 
(usually by GET-VALUE-UNIT ill Inleractiou plan\[11\]), 
so that it can have a particular SEF as the valnc of the 
node. For example, "Yamaoka" could be the value of 
a variable node corresponding to a concept node of 
'name' in a sentence like "My name is Yamaoka ". 
The following types of links are defined: is-a 
link, representing a superordinate/subordiuate rela- 
tion between two concept nodes, part-of link, rep- 
resenting a whole/part relation between two concept 
nodes, causal link, representing a causal relation 
between two concept nodes, prag llnk, representing 
a pragmatic relation to connect a particular concept 
node to a lexical node representing the tyllical SEFfor 
the concept, value link, representing an instance 
value relation between a particular valuable concept 
node and a variable node which has been bound to 
the SEF of its value, and eq llnk, representing the 
same meaning between two lexical nodes. 
Extension of eq link: In order to make the knowl- 
edge base sensitive to the changes considered in Sec- 
tion 3.2, tile eq link is extended. This lets us to 
add applicable conditions to eq links as sub-types of 
tile link. Applicable conditions are defined based ou 
classes of the categorization in 3.2. For example, if 
one lexical node is a polite SEF of another, the two 
lexical nodes can be emmected with an eq-if-polite 
llnk, e.g. "juusyo" and "go-juusyo"(see Fig. 1). 
4 Selnctlon Strategy 
Ill the dialogue, a speaker chooses an expression ac- 
cording to tile situation, the preceding context, and 
his/her beliefs. Assuming that tile system has recog- 
nized such conditions, we can efficiently choose the 
correct speech recognition candidate by searching the 
SEFs that are appropriate under the conditions. 
4.1 Rules of Applicable Conditions 
\[terc, two terms are defined for explanation: 
seed: if the predicted contextual information is 
bolmd to a particular SEF, then the seed of tile 
contextual information is tile SEI", otherwise the 
seed is the value of the lexieal node linked by tbe 
l)rag llnk to the concept node corresponding to 
the contextual information, 
preferable set: a set of SEFs derived from a seed by 
an applicable rnlc, whicb then takes first priority 
for selecting the candidate. 
The basic rule for making a preferable set is: col- 
lect tile SEFs by following the eq link from the sced. 
Because in this paper we are focusing on dialogue 
situations and contcxts ratber than the speaker's be- 
liefs, we only cover rules regarding changes by dif- 
ferent viewpoints. 
For a predicted contextual information I, consid- 
ering the dialogue situations ill Class 2(a): 
I. if I is in the territory of information of the other 
agent, then make a preferable set by following 
the eq-if-polite link from the seed, 
additionally, considering the preceding context: 
2. if \[ has an antecedent which denotes the status 
of the other agent, i.e., there is an instantiated 
variallle node corresponding to I , then replace 
the seed with the antecedent, i.e., the SEF of 
I.he variable node, and make a preferable sct by 
following the eq-if-pollte llnk from the seed. 
Considering the contexts ill Class 2(b): 
3. if 1 is a compound noun, (it's obviously the an- 
tecedent) then shift the seed to the concept one- 
level up t and make a prefcrablc set using the 
basic rule, 
4. if l includes two or more concepts or SEFs and 
there is a concept node which is the upper node 
of both of these concepts, then shift the seed to 
the upper concept 
and make a preferable set using the basic rule 2, 
I Precisely, shifting a seed to a concept metals an operation 
to replace the seed with the SI~F of the lexical node followed 
by the prag llnk from the concept node. 
In this case an auxilisa'y word is usually added. 
AcrEs DE COLING-92. NA,'CrE.s, 23-28 AOt~r 1992 1 I 5 4 PRec. ot: COLING-92, NANTES. AUG. 23-28. 1992 
In daily dialogue, speakers apply combinations of 
the above rules and other rules, but in this study we 
arc concentrating on simpler cases. 
4.2 Seleetion Algorithm 
Our ultimate goal is to select the correct speech 
recognition candidate from the predicted contextual 
information. An algorithm to do this is roughly de 
fined by following the three steps: 
1. provide contextual information, 
2. make a prefcrable set from i by the rules, 
3. compare speech recognition outputs with 2, 
and if all equivalent is found 
then pick it ms the ai)propriate candidate, 
else goto 2. 
Steps 1 and 2 above are I)acktracking points. For 
details of Step l, see \[1I\],\[12\]. l,'urthcr large-scale 
experiments may deternfinc hcuristically how many 
times Step 2 should be iterated. 
5 Examples and Evaluation 
hi this section, we examine some polite expressions 
aud compound NPs that are common in telel)holm 
inquiry dialogues. 
5.1 Polite expressions 
An examl)le of the process for detecting the appropri 
ate SEF given a polite expression, is shown through 
the following subdialogue, focusing on discourse en- 
tities. 
(ul) Q: TouT~oku-youshi-wo okutte-kudasai. 
(Please send me a registration from.) 
(u2) S: Go-juusyo-wo onegai-shi-masu. 
(May \[ haw~ your address?) 
(u3) Q: Juusyo-wa Osaka.shi ...... desu. 
(My address is Osaka-city ....... ) 
where agent Q is the questioner and S is tfie secre- 
tariat. 
This example can be recognized in the send- 
somthing domain plan (m Section 2). First ul is 
recognized and understood a.s an utterance which ill- 
troduces the domain plan. Then, fi)r the next utter 
ance by S (u2), since the system does not hohl the 
statement that S knows where to send a form, e.g. 
the wdue of Q's address, an utterance requesting the 
value of the destination is tirst predicted, and con- 
textual information about the 'destination' concept 
can be provided (Step 1). 
Next, due to the constraint in tile plan, Rule 1 
is applied to the contextual information. 'fhen, the 
preferable set of SEFs is derived by the rule(Step 2). 
Although the tirst Step 3 fails because "go- 
juusyo" is not the exact polite form of the first seed 
"okurisakf'(destination), the second time it, picks 
"yo-juusyo" ms the appropriate SEF because one of 
tim lower concepts, 'address', can be the next seed. 
On the other band, when processing 113, the set of 
polite tbrms for the 'address' is not preferred. 
Table 1: Result of l)ataba.~e Inspection 
Comm Act Type ~ SF, F Type Number 
\[)FFFR-ACTION SP -- polite 25 
normal 4 
Hit m)rmM 22 
REQUEST-ACTION SP normal 2 
}IR polite 4 
normal 3 
CDNFI tlII-AC'I'I fin HR polite 3 
normal 5 
The communicative act type in the first column:is ~-e 
type of the utterance "to send". The speaker in the sec 
ond colulnn is the speaker of the target SEF, with SP in- 
dicating the speaker in the first COblntn, and HR, hearer. 
Evaluati(m: We evaluated this tnethod by inspect- 
lug SEFs for 'destinatiou' in the ATR diak)gue 
database\[4\]. The target corpus, whose topic is "Con 
fi~rence registration", has 85 conversations, 195fi ut- 
terance units, and 3085 sentences. Moreover, the tar. 
get expressions are restricted to those uttered in a 
segment of the send-somet.hing domain plan. The 
evaluaLion was done in the tollowing way: 
1. H.etrieve sentences which have the verb "okumg' 
(to Selld) or syllonylllOllS w~.rbs as tile lllaill verb 
of the sentelice (lfll sentences). 
Then, output the utterance refit together with 
the next utterance unit ( 161 pairs). 
2. Pick tim i)airs in which there is a expression 
about 'destination' (43 pairs). 
Filter by the send-something domain plan, and 
those pairs that are not recognized are elimi-. 
nated (32 pairs remain). 
3. Cl~kssify the target expressions (68 expressions) 
into tile othm"s territory (a2 polite and 12 nor- 
real) and the spe.aker's (24 normal). 
The results are shown in the Tal)le 1. Tiffs inspec 
tion shows that ill (an" target, dOiqlaill, I~lle framework 
described m the paper is useful for selecting a sur- 
face expression that is appropriate in the dialogue 
situation. 
Example (Vocative): Consider the subdialogue 
that follows tile above subdialogue: 
(u4) Q: Namae-wa Suzuki-Mayumi-desu. 
(My name is Maymni Suzuki.) 
(u5) £': £'uzuki-Mayumi-sama-desa-ae. 
(Ms. Mayumi Suzuki, correct?) 
After recognizing 114 by the saule interaction plan 
a.s the first example, a variable node corresponding 
to Q's name is instantiated and bound to "Suzuki 
Mayumi". Then, for the next utterance by S (u5), 
we can predict the confirmation utterance including 
the contextual intbrmation about Q's name as a dis- 
course entity. Consequently, we can select the SI';F 
"Suzuki-Mayumi-sama"(polite form) by the contex- 
tual information and the applicable rule 2. 
ACII~S DE COLING-92, NAM'Es, 23-28 hotrr 1992 1 1 5 5 Pace, OF COLING-92, NANTES, AUG. 23-28, 1992 
5.2 Compomld NPs 
Compound NPs can roughly be classified into l)rol)er 
NPs and common NPs. Predicting SEFs from a com- 
mon NP is usually done by shifting the seed to the 
upper level (by Rule 3). For example; 
(u6) S: 7burokn.youshi-wa o-miehi-desyou-ka? 
(Do you have an application form?) 
(uT) Q: Mada-dcsu. 
(Not yet.) 
(u8) Youshi-wo okutte-kudasai. 
(Please send me a form.) 
In this example, u6 instantiates the send 
something domain plan by tile effect chain\[7\]. Then, 
since we know from u7 that the effect (the goal of 
this subdialogue) is not satisfied, we can predict that 
the next utterance hy Q (u8) ruay concern introduc- 
ing the action to send a form, and it. irmludes con 
textual information about 'application form'. In the 
knowledge base, 'form' is tile concept node just above 
'application form'. Consequently, by applicable ruin 
3, we select " YoushF(forn 0 directly. 
On the other l~and, predicting SEFs tbr proper NPs 
requires another rule to creat.e the donlain-dependent 
knowledge base for shortening, llerc, we use the 
dependency relationships within NP\[91 to abbrevi- 
ate a proper compound NP. For exalnph~, applying 
this rule to a proper compound NP "Kyoto-Kokusai- 
Kaigijoff'(Kyoto International Conference Center), 
we get a preferable set of SEFs inelnding "Kyoio- 
Kaigijon"(Kyoto Conference Center) and "Kokusai- 
Katgoou"(International Conference Center), ill addi- 
tion to the basic upper SEF "l(aiyzjou"(confcrence 
center). Consequently, we call select "Kokusai- 
kaigOou" in ul0 in the following subdialogue with 
a take-transportation domain plan; 
(u9) S: Kyoto-koknsai-kaiyljou-ewa basu-ga 
r~you-deki-masu 
(There is a bus that goes to the b:yoto 
hlternational C, onferenee Center.) 
(ul0) Q: Kokusa~-ka~gUou-made ikur'a-desu-ka~ 
(Ilow much is it to the International 
Conference Center?) 
At the moment, we define a short link to con 
noel lexieal nodes created by abbreviation rules to 
the proper compound NP thai instantiates a variable 
node. 
6 Conclusion 
This paper has proposed a context-sensitive method 
of predicting NPs in the next utterance of tele- 
phone inquiry dialogues. Abstract information about 
the constituents of the next utterance can be pre- 
dicted based on the dialogue interpreting model. 
Then, domain-dependent knowledge for NP usage 
was developed based on an extended NP identifica- 
tion model. Tile knowledge base is characterized by 
its ability to derive the set of possible surface ex- 
pression forms from the predicted contextual infor- 
nration. We define rules for applicable conditions, 
particularly ill polite Japanese, bmsed on an anal- 
ysis of NP changes. Finally, using the above two 
mechanisms a strategy was proposed for selecting tile 
appropriate surface expression form representing the 
predicted concept in a context-sensitive way. 
In the fllture, we plan to integrate this method with 
a method of predicting expressions of tile speaker's 
intention, tp form a complete system. It is also vital 
to make tile method more powerful, so it can au- 
tomatically construct the domain-dependent knowl- 
edge base from thesauri and/or corpora of tile do- 
main, and call model and recognize various dialogue 
situations. 
Acknowledgements 
Ti~e authors wouhl like to thank Dr. Kurematsu, Presi- 
dent of ATR h, terpreting Telephony Labs., and other col- 
leagues for their encouragement and thought-provoking 
discussio,,. 
II.efi~rnnees 
\[1\] l)oug|~u E. Appelt. Planning English Sentences. 
Studies in Natural Language Processing. Oambridge 
University Dress, 1985. 
\[2\] Eugene Charniak. Motivation analysis, abductive 
unification, and nonmonotonic equMity. Artificial 
Intelligence, 34:275-295, 1988. 
\[3\] llcrbert 11. Clark and Eve V. Clark. Psychology and 
Language -An bltroduction to Psycholinguistics-, 
chapter 6, pages 223-258. llarcourt Brace Jo- 
vanovich, 1977. 
\[4\] Terumasa Ehara, I<entaro Ogura, and Tsuyoshi Mo- 
rimoto. ATR dialogue database. In Proceedings of 
1CSLP'90, pages 1093 1096, November 1990. 
\[5\] M. A. K. llalliday and Ruqaiya Ilasan. Cohesion 
in English, chapter 6, pages 274 292. LONGMAN, 
1976. 
\[6\] Alexander G. llauptmann, Sheryl R. Young, and 
Wayne II. Ward. Using dialog-level kaowledge 
sources to improve speech recognition, in Proceed- 
ings of AAA\['88, pages 729 733, 1988. 
\[7\] ttitoshi Iida, Takayski Yamaoka, and Itidekazu 
Arita. Three typed pragmatics for dialogue struc- 
ture analysis. In Proceedings of COLING'90, pages 
370-372, August 1990. 
\[8\] Akio Kamio. Proximal and Distal \[nforamtion: A 
Theory of Territory of ln\]ormation in En91ish and 
Japauese. PhD thesis, University of Tsukuba, March 
1986. 
\[9\] Masahiro Miyazaki. Automatic segmentation 
nlelhod for conq>ound words nsillg seluantic depen- 
dency relationships between words. Journal oJ lnfof 
marion Processing Society of Japan, 25(6):970-979, 
1984. is Japanese. 
Izuru Nogaito and flitoshi lida. Noun phra.ue identi- 
fication in diMogue and its application. In Proceed- 
ings of ~nd International Conference on Theoretical 
and Methodological Issues in Machine Translation of 
Natu,cd Languages, June 1988. 
Takayuki Yamaoka and Hitoshi lida. A method to 
predict the next utterance using a four-layered plan 
recognitioa model. In Proceedings of ECAI'90, pages 
726 731, Atlgust 1990. 
Takayuki Yamaoka and liitoshi lids. Dialogue ill- 
terl>retation model and its application to next utter- 
anne prediction for spoken langm~ge processing. In 
Proceedings of Eurospeech'91, September 1991. 
\[io\] 
\[nl 
it2\] 
ACRES DE COLING-92, NAi'Z~S, 23-28 Ao~r 1992 1 1 S 6 PRoc. or: COLING-92, N^NrEs, AO~. 23-28. 1992 
