EXPLOITING REFERENCE INTERACTION IN RESOLVING TEMPORAL REFERENCE 
Kohji DOHSAKA 
NTT Basic Research Laboratories 
3-1 Morinosato-Wakamiya, Ai:ugi-shi, Kanagawa, 2/1:3-01 Japan 
dohsaka~_atom.ntt.jp 
ABSTRACT 
q'his paper provides an account of the role that the 
interaction between nominal and temporal reference 
plays in resolving temporal reference. Exploiting this 
interaction in resolving temporal reference clarifies 
how the process of resolving nominal reference in- 
teracts with the process of resolving temporal refer- 
ence, and how a restricted set of worht knowledge con- 
tributes to resolving temporal reference. 
1 INTRODUCTION 
B.esolving reference or anaphora is of great interest 
in computational linguistics and formM semantics. 
Research on reference began with the development 
of models to account for nominal reference bronght 
about by the usage of nominal expressions such as 
definite noun phrases and pronominals (Grosz, Joshi 
and Weinstein 1983; Kameyama 1986). Recently a 
number of researchers have indicated that temporal 
expressions such as tense morphemes and temporal 
adverbials can be regarded as referring expressions 
and have proposed models for temporal reference res- 
olution (Hinrichs 1986; Hwang and Schubert 1992; 
Kameyama, Passonnean and Poesio 1993; Lascarides 
1.992; Partee 1984; Song and Cohen 1991; Webber 
1988). 
Sentences in a dialogue describe eventualities 1 and 
introduce them into the context. The time of an even- 
tuality described by a sentence is interpreted as tem- 
porally related to the times of other eventualities that 
were introduced into the context earlier in the dia- 
logue. Temporal expressions are regarded as referring 
expressions since they convey information about the 
time of an eventuality, which is interpreted relative to 
the times of other eventualities in the context. Re- 
solving temporal reference means determining what 
temporal relationships exist between an eventuality 
described by a sentence of a dialogue and eventuali- 
ties that have been introduced into the context ear- 
lier in the dialogue. 2 Resolving temporal reference 
is requisite to many tasks, such as designing a natu- 
ral language interface to a planning system in which 
temporal information is crucial (Crouch and Pullman 
tan event.uality is the general term for an event, process or 
state, due to Bach (1986). 
~The chief concern here is temporal reference to intersen- 
tential context. Thus this paper does not address the probiem 
of determining what temporal relationships an eventuality de- 
scribed by a clause of a sent, ence has with eventualities described 
by other clauses of the same sentence. This problem is covered 
by Brent (1990), Itwang et al. (1992), and Reichenbaeh (1947). 
1993). 
To understand the notion of temporal reference, 
consider the following dialogue, (dl). a Assume tha~ 
dialogue participants, Bill and John, mutually know 
that John received an e-mail message from Mary. 
Dialogue (dl) 
(sl) Bill: Mary karano mail wa mot-teirn-yone? 
from mail ToP have~PRESENT 
You have the mail from Mary, don't your 
(s2) John: Keshi-ta-kedo. 
delete-PAST 
I deleted it. 
(s3) Bilh Hontoni! 
really 
Really\] 
(s4) John: Sochira niwa tensoushi-ta-hazuda-yo. 
you DAT forward~PhsT-sure. 
I'm sure that l forwarded it to you. 
(s5) John: Sochira no mail box ni at-ta-desho? 
you GFN mail boz" in be-PAST. 
II was in your mail bo~', wasn't it?. 
Sentence (s2) describes an eventnality of deleting, 
Ed, and sentence (s4) describes an eventuality of for- 
warding, F/. Both eventualities are past ones, since 
the matrix verbs, "tensousnru (forward)" and "kesu 
(delete)", 4 describe momentary acts and are accom- 
panied with the past-tense morpheme "ta". Although 
the tense morphemes of sentences convey information 
about the times of eventualities described by the sen- 
tences, the context also imposes restrictions on the 
times of the eventualities. In dialogue (dl), the time 
of tQ described by (s4) is interpreted as relative to 
the time of \]'d~ in the context: i.e. the time of E/ is 
before the time of 1,2d. in this sense, the time of 1~ is 
referred to in uttering sentence (s4). 
Existing models of temporal reference account for 
the parMlel between temporal and nominal reference. 
Itowever, as I will state in see. 2, existing models 
fail to explain the interaction between temporal ref- 
erence and nominal reference (reference interaction 
for short), i{eference interaction is a phenomenon in 
which tile process of resolving nominal reference in- 
ahl extfibiting a .Japanese dialogue, English words such as 
"Mary" mtd "mail" are used only for easy comprehensibility, hi 
addlt, ion, sentence-final forms such as "yone", "kedo", "yo", and 
"desho" indicate mental states of file speaker that are/mrelated 
to the subject of this paper. 
4 "tensoushi-" and "keshi-" in file dialogue are inflections of 
the verbs "tensousm'u" and "kesu" respectively. 
1138 
teracts with the process of resolving temporal refer- 
ence, when an utterance involves temporal and nomi- 
nal reference at the same time. l)ue to as1 insutticient 
account of the reference interaction, existing models 
cannot show how the process of resolving nominal ref- 
erence Mfects the process of resolving temporal refer- 
ence. 
The chief concern of this paper is to describe the 
role that the reference interaction plays in resolving 
telnporal reference and demonstrate that the reference 
interaction serves to clarity how a restricted set of 
world knowledge contributes to the resolution process. 
I focus on sentences with past-tense morphemes in 
Japanese dialogues. Previous work also used past- 
tense sentences as a touchstone to show the validity 
of a model (IIinrichs 1986; Kameyama ct al. 1993; 
Lascarides 1992; Partee 1984; Webber 1988). As I 
will state in sec. 2.2, the retk'xence interaction is not a 
domestic plleuolnenon in Japanese dialogues. 
In sec. 2, I argue that existing models do not de-. 
count for the reference interaction, that the reference 
interaction plays an important role in an account of 
how temporal reference is resolved, and that exploit- 
ing the reference interaction clarifies how a rest, rioted 
set of world knowledge serves to resolve tempered ref- 
erence. In sec. 3, a tYamework is presented, within 
which the reference interaction is exploited in resolv- 
ing temporal reference. In sec. 4, 1 demonstrate how 
f, he framework works. In sec. 5, \[ smnmarize the 
claims of this paper and describe fnture, work. 
2 PROBLEMS WITH EXISTING MODELS 
2.1 Approaches in existing models 
In existing models, when the t, ime, 7', of an eventalality 
described by a sentence of a dialogue is given, the 
problem of resolving temporal reference is divided into 
(i) that of identifying the time, 71/, of an eventuality 
that has been introduced into the context, and (ii) 
that of determining what temporal relationship hohls 
between 7' and 7):. 
The notion of temporal Ibcus is nsed to solve proh.- 
lem (i). ','he notion was introduced by Webl,er (~988), 
Dialogue participants pay attention to the time of an 
eventuality that is in the temporal focus. '\['he time, 51', 
of an eventuality described by the cnrrent sentence is 
interpreted according to the time, Ttf, in the temporal 
focus. Existing models that apply to intersentential 
context use the notion of temporal focus (llwang el el, 
1992; Song ct el. 19911. Kameyama et el. (1993) re- 
fined the. notion of temporal focus and proposed the 
notion of temporal centering. In formal semantics, 
models of temporal anaphora based on discourse rep- 
resentation theory have been proposed (ltinrichs 7198{3; 
Partee 1984). 'Fhese models concentrate on the tense 
interpretation of adjacent sentences its a narrative dis- 
course. Thus, the time of ;m ew;ntuality described l)5' 
a sentence immediately preceding tile current sentence 
can be regarded as being in the temporal {hens. 
There ~r0. two ~tpproaches to prol)lcm (ii): (ii-a) 
av approach based on linguistic cuen such as tense 
morphelnes, the aspectual class of verbs, and the de- 
scriptive order of eventualities in a dialogne; (ii-b) an 
approach based oil worht knowledge of the causal re- 
lationships between eveutnalities. 
With regard Co approach (ii-a), the following de- 
fault interpretation rules (ilinrichs 1986; Kameyama 
et el. 1993; bascarides 1992; Partce 1984; Webb0r 
19881 have been used. '~ 
'De.fault into.rpretation rules) 
Provided that an eventuality, t¢, is described by the 
current sentence, the time of an eventuality, let\], is 
in the temporal focus, and both/',' and \[¢t/ are past. 
eventualities, then 
(R1) The time of E is after the time of/¢tf if both 
I¢ and l';tf are non-st, drive, 
(1{2) The time of/¢ contains the time of ICe\] if \]¢ is 
stative and l','tf is non-stative. 
These rules are t;ermed default interpretation rules 
since they are utilized when world knowledge of 
causality is not available. 
\[n approach (ii-b), the temporal relationship be- 
tween l,\] and .l¢tf is determined according to world 
knowledge of the causal relationships between eventu- 
alities. Lascarides (1.9!t2) presented a model based on 
a theory of defeasihle inference for integrating world 
knowledge of causality with the process of determining 
the temporal and causal relationships between even 
tualities mentioned in a discourse. This approach was 
also adopted in a model proposed by Kameyama ct el. 
(1993). In these models, approaches (ii-~) and (ii-b) 
are unified within a single framework. Let me call 
the interpretation based on approach (ii-a) the (le- 
fanlt interpretation, and the interpretation based on 
approach (ii-b) the knowledge-based interpretation. 
The knowledge~based interpretation is preferred over 
the default interpretation when they disagree. 
l,et me explain how existing models work by using 
(tialogue (dl) ~ as, example. 1 focus on how (s4) i,, 
interpreted, assuming that the time of /~'a is in the. 
temporal focus, l)eNnlt interpretation rule (hi.l) says 
that the Lime of \]Q must be after the time of lea sittce 
both eventualities are not>stative, past eventualities. 
Ilowevcr, the relevant interpretation is that F,/ tem- 
porally precedes l','a. Thus, the default interpretation 
does not coincide with the relevant interpretation. In 
this case, existing models resort to world knowledge 
of the ransal relationships between £'a and Ej. ltow. 
cw'.r, existing models do not clarify what kind of world 
knowledge is use.d (Kameyama et el. 1993; I,ascarides 
1992). In this sense, the world knowledge used in ex- 
isting models is unrestricted. 1 agree with such an 
approach its that various causal relationships are in-- 
volved ill the process of resolving reference. IIoweveL 
it is desirable to find a restricted set of world knowl- 
edge and avoid resorting directly to the entire set of 
~Note that the time of a stative even~uallty is usually nc)t 
regarded as being in the t.emporal focus. 
1139 
world knowledge of causality since such knowledge is 
enormous. 
Consequently, existing models exploit the notion 
of temporal focus or temporal center, which serves 
only to show that temporal reference is accounted for 
in the same way as nominal reference. However, exist- 
ing models do not explain how the process of resolv- 
ing nominal reference affects the process of resolving 
temporal reference and assume that the entire set of 
knowledge of causality can be used. 
2.2 Reference interaction 
In this section, I argue that the notion of reference 
interaction provides an account of the temporal refer- 
ence in (s4), which existing approaches ignore. More- 
over, a restricted set of world knowledge can be used 
to explain the reference interaction. 
Sentence (sl) introduces an individual 6, M, which 
is an e-mail message. Sentence (s4) involves nominal 
reference to M: tile object of "tensousnrn (forward)" 
is zero-pronominalized and refers to M. Moreover, 
sentence (s4) involves the interaction between nomi- 
nal and temporal reference. That is to say, what the 
object being forwarded is affects what the time of for- 
warding is. I will explain this below. 
When someone forwards something to someone 
else, the time of forwarding must be during the time 
when the object being forwarded exists. In other 
words, the time of forwarding depends on what the 
object being forwarded is. Moreover, when some- 
one deletes something, the time when the object be- 
ing deleted exists must be before the time of delet- 
ing. This kind of world knowledge can be regarded as 
knowledge of temporal relationships between eventu- 
alities and the existential status of individuals. Judg- 
ing fi:om such knowledge, the eventuality of forward- 
ing, EI, cannot temporally follows the eventuality 
of deleting, Ea. This is because both the object of 
"kesu (delete)" and the object of "tensousuru (for- 
ward)"refer to M, so the time of t~'f must be during 
the time when M exists and the time when M exists 
rnust be before the time of Ed, The interpretation 
consistent with such knowledge is that /~Sf temporally 
precedes Ea. Consequently, identifying the referent of 
the zero-pronominalized object of/Q serves to resolve 
temporal reference, and the knowledge of temporal re- 
lationships between eventualities and the existential 
status of individuals can be used to explain the refer- 
ence interaction. 7 
Moreover, knowledge of the temporal relationships 
between eventualities and the existential status of in- 
dividnals can be regarded as a restricted set of world 
knowledge of causality. It is restricted because given 
an eventuality, E, we have only to allow for the in- 
dividuals that constitute E and do not have to allow 
°flere, "individual" is used as a term for a single person or 
thing. 
7In this paper, only nominal reference to individuals is al- 
lowed for. 
for all eventualities that can be causally related to 
E. Exploiting the reference interaction clarifies how 
such a restricted set of world knowledge contributes 
to resolving temporal reference. When such restricted 
world knowledge is sufficient to resolve temporal ref- 
erence, immediate recourse to the entire set of world 
knowledge of causality call be avoided. Note that such 
knowledge is not selected arbitrarily. It is the knowl- 
edge that is necessary for explaining the reference in- 
teraction. 
Next I review two existing proposals that are re- 
lated to the argument above. First, Hwang el al. 
(1992) argued that resolving temporal reference re- 
quires plausible inference that can interact with var- 
ious processes such as resolution of anaphora, intro- 
duction of new individuals and identification of spa- 
tial and temporal frames. They also argue that the 
plausible inference has to rely on world knowledge 
such as that one normally would not buy broken 
things (IIwang e¢ al. 1992: p.239). Their argument 
agrees with the approach presented here. They did 
not, however, present a concrete model to support the 
argument. 
Second, Webber and Baldwin (1992) discussed 
the integration of two independent mechanisms for 
context-change by entity introduction and by event 
simulation. The idea of integrating these context- 
change mechanisms and that of exploiting the refer- 
ence interaction in resolving reference share the view 
that the relationships between eventualities and the 
existential status of entities or individuals serve in in- 
terpreting referring expressions. They, however, fo- 
cused on interpreting nominal reference made by the 
use of definite nouns, rather than on the problem of re- 
solving temporal reference by exploiting the reference 
interaction, which is the chief concern of this paper. 
Finally, it is easy to see that the reference interac- 
tion also occurs in an English dialogue. In an English 
counterpart of (dl), the pronoun "it" is used the same 
way as a Japanese zero pronoun is used. l,ikewise, 
the restricted knowledge stated above is common to 
Japanese and English. 
3 A FRAMEWORK 
A framework is presented, within which the reference 
interaction is exploited in resolving temporal refer- 
ence. First, in sec. 3.1, the descriptive device is shown. 
Next, in sec. 3.2, the process of resolving temporal ref- 
erence is described. 
3.1 A descriptive device 
In a diMogue, sentences convey information about tim 
speakers' mental attitudes toward eventualities. This 
framework does not concern the mental attitudes, but 
focuses on the eventualities described in the sentences. 
An eventuality is written as a variant of the Davidso- 
nian representation (l)avidson 1980; Vlach 1993). For 
example, the informational content of the eventuality 
of forwarding described by (s4) is written as follows. 
1140 
(rl) 7'ypc( .c \] l , .t,'orward)& 
Agcnt( *e \]l, *agcnt~ )&leccp( *e \] 1, *rccpl )& 
Objert(*c\]l, *objl)&'l'imc(*c\] 1, ,t\] z)& 
NonStativc( *c \]l )& ll carer( *rccpl )& 
Tl?.el( Bc f orc, *Q, , *t,v l )& 
T t?.el( *rcl z , *t :,, *t \] .... 1 )& 
SpecchT'irne( .t,v~ )&:/'cmpFocus(*t\] ..... 1). 
Symbols starting with a cN)ital letter represent 
constants and symbols starting witl, the ctmracter '*' 
represent varlahles. Symbol '~;' is an of)erator \[or con- 
junetkm. 
Type(L',T) means that eventuality l,; is classified 
as a type T. t{.epresentation (rl) specifies the. even- 
tuality, *e/l, that *agcntl tbrwards *objt to *rcepl 
at time *t\],. The agem, and tire object of forward- 
ing are represented as variables, since they are zero- 
pronominalized. Llcarcr(l') means that P is the 
hearer. In the above, the recipient of forwarding is 
identified with the hearer since it is specified by tire 
pronoun "sochira", designating the hearer. In addi- 
tion, ewmtuality *c:1 is non-st~d.ive. 
SpecchTimc(T) means that time 7' is the speech 
time, Tcmp\]Coeus(7 ') means that time T is in the 
temporal focus, and 7'Rel(l{, 7't,7:e) means that tem- 
poral relationship lg holds between time ~I'\] and 7). 
In this framework, temporal relationships are repre- 
sented based on temporal logic proposed by Allen 
(1983) and times are treated as temporal intervals. In 
the above, time *tSl is before sl)eech time *GpJ since 
*Ill is specified by the past-tense morpheme "ta". 
Time *tlo~.,t,t is in the temporal focus. Temporal re- 
lationship betweeen *Ill attd *l\]o~u,l is represented 
as a w~riable, *roll. Resolving temporal rel)rence 
means determining a relevant tentporal relationship 
betweeen *tit and *t\]ocu.,1. 
Knowledge of the temporal relationship between 
an ewmtuality and the existential status of individu- 
als is represented as a triplet < \]~;R, {i\[~.,...},Tlg >, 
where 1,;1~ is a represent,ttion for the eventuality, 
{\[R,...} is a set of representations of the existen-. 
tial status of the individuals, and Tt~ is the teulpo- 
ral relationship between them. This framework con- 
cerns the existential status of individuals such that 
an individual exists at a certain space-time location 
of the physical world. For example, knowledge about 
an eventuality of forwarding is written as follows. 
(r2) <:l~3lQ,{ll~ja, \[IQb}, '\]'\[{f> , where 
l,;le: do: 7'gpc(.e\], Forward)&, 
Agcnt(*c:, *agcnt)& 
ltccp(*cj, *recp)& 
Objcct(*c\], *obj)& 
7'ime( *c : , *t : ), 
iiQa <le=r 7'ypc(*Cobj, t;'xi~t)&, 
Object(*eo~j, *obj),~.'. 
Loe( *eobj , *lobj )& 
Owner( *lobj , .agent)& 
Time( *covj , *toni), 
1 l~\]b d~=r !l'ypc(*c~.ovy, Exist)& 
Objeet( *e¢o~)v, *copy)& 
CopyO f(*copy, *obj)& 
Loc( *C~o~,y, *l~ovy )& 
Owncr( *l~ol, y, *','eep)& 
"l~imc( *e~op,j , *t~ot,,j ), 
T I~,\] de=f !I'll.el( During, *t j, *tobj)~ 
7' l?.cl( Bc f orc, *t j, *teovy ). 
Loe(*x, *1) means that individual *x exists at lo.. 
cation *l. in this dialogue domain, the location where 
an e-mail message exists is a mail box. Owner(,l, .p) 
means that person *p owns location */. I assulne here 
that the owner of a mail box is uniquely identified. 
/it the ahove, \[l~\]a specifies the existenti~fi status 
of *obj, whM1 is the object being forwarded, and says 
that *obj exists at time *to~j and at location *lobd, 
which the agent of forwarding owns. lleZb specifies 
the existential status of *copy, which is a copy of *obj 
aud is generated by forwarding *obj. The object thai; 
the recipient of forwarding receives is ,lot identified 
with *obj becmlse of domestic constraints concerning 
an e-mail system, lJgIi, says that *copy exists at time 
*tcopu and at lo('ation *le~py, which the recipient of 
forwarding owns. 3'l¢.j says that the time: of forward- 
ing, *t j, must be during the time, *to,j, when .obj 
exists, and that the time of forwarding, *t: nmst be 
before the time, *tcopy, when *copy exists. 
l(nowledge about an eventuality of deletiug is writ- 
ten as follows. 
(r3) <l';l{d,{\[ll.d}, 'Pl{d>, where 
L'tQ d:-f 7'ypc(*ed, Delete)& 
A:lcnt(*ed, *agent)& 
Object(*cd, *obj)& 
7'imc(*Cd, *td), 
def Itgd = Typc(*Cobj, \]'\]xi.~t)& 
Object( *co~j, *oh j)& 
Loe( *eobj , *lobj )&. 
Owncr( *lobj, *agent)& 
7'imc( *Cobj , *tobj ), 
(\['led d(~t' ,\[,\[~el(Mect, *td, j, *td). 
l lgd says that the object being deleted, *oh j, exists 
at time *ld and at h)cation *loci owned by the agent 
of deleting, *agent. 7'lQ says that the time when 
,obj exists must be immediately before the time of 
deleting. 
3.2 Resolving temporal reference 
This framework assumes that a representation of an 
eventuality described by a sentence is given, in the 
representation, pronominalized individuals an(l inde- 
terminate teml)oral relationships are represented as 
variables such as *agent1 ~md *rOll in (rl). When the 
representation of the eventuality described by the cur- 
rent sentence is given, representations of the existen- 
tial status of individuals and temporal relationships 
between the eventuality amt the existential status of 
individuals are derived by using knowledge such gm 
1141 
(r2) and (r3). These representations are interpreted 
within the context as described below. In the interpre- 
tation process, appropriate constants are substituted 
for variables. After the interpretation process, the 
representations are introduced into the context. Thus, 
the context includes representations of eventnalities, 
the existential status of individuals, and temporal re- 
lationships among the eventualities and the existential 
status of individuals that have been mentioned in a di- 
alogue. In addition, it includes representations that 
show who is the speaker~ who is the hearer, what is 
the speech time, and what is the time in the temporal 
focus. 
Let RSet be a set of representations of an eventu- 
ality, E, described by a sentence, the existential status 
of individuals mentioned in E, and temporal relation- 
ships between them. The interpretation process is as 
follows. 
(I1) A representation, /~, in 12set is unified with a 
representation, R~, in the context, if possible. 
Through unification, constants in R~ are substi- 
tuted for variables in/C 
(i2) An indeterminate temporal relationship between 
the time of eventuality E and the time in the 
temporal focus is identified with the relevant tem- 
poral relation according to default interpretation 
rules and transitive and reflexive laws governing 
temporal relationships. 
(I3) Constants are generated and substituted for vari- 
ables that cannot be identified in steps (I1) and 
(I2) of this process. 
(I4) Representations in Rset are added to the context. 
Some nominal reference is resolved in (I1), al- 
though this paper does not go into how nominal ref- 
erence is resolved. 
In (I2), temporal reference is resolved. The inter- 
pretation by transitive and reflexive laws governing 
temporal relationships is preferred over the interpre- 
tation by default interpretation rules. 
This framework uses the defanlt interpretation 
rules (R1) and (R2), which are used in existing mod- 
els. The default interpretation rules (R1) and (R2) 
are represented as the following theorems, (r4) and 
(r5) respectively, s 
(r4) Time(*ei, *ti)&NonStative(*e~)& 
Time( *et\] , *tt\] )&NonStative( *etf )& 
TempVocus(.t~/) D TRel(After, .t~, .its). 
(1"5) Tirrte(*ei, *ti)gStative(*ei)& 
Tirne(*eti , *hi )&NonStative( *etj )& 
TempFocus(*tt\]) D TRel(Contains, *t~, *ttf ). 
In the above, R1 D R2 means that RI implies R> 
The transitive and reflexive laws governing tem- 
poral relationships are also represented as theorems. 
For brevity's sake, I will not present all the laws. 
Allen (1983) presents a exhanstive list of transitive 
SHare i ignore conditions where the eventualities at stake 
are past ones. 
laws governing temporal relationships. The following 
theorems are sufficient for dealing with dialogue (dl). 
(r6) T~el(1)u,'i~q, **, *z)~7','el(M~< *~, *V) 
Trel(Befove, *x, *y). 
(r7) Trel(t3efore, *x, *y) D Trel(After, .y, .x). 
(rS) T~el(D,m.g, **, ,y) ~ :C~l(Co.~,i..~, w, *~). 
The interpretation by the transitive and rellexive 
laws governing temporal relationships can be regarded 
as a kind of of knowledge-based interpretation as de- 
scribed in sec. 2.1, although only a restricted set of 
world knowledge is used in this framework. As demon- 
strated in see. 4, the interaction between references to 
individuals and times plays an important role in re- 
solving temporal reference according to the transitive 
and reflexive laws governing temporal ,:elationships. 
4 EXAMPLES 
This section demonstrates how the framework works 
by using sentence (s4) and (s5) as examples. First, 
consider the interpretation of (s4) under the context 
established by (s2). Sentence (s2) introduces an even- 
tuality of deleting, Ed, into the context. The eventu- 
ality is represented as follows. 
(r9) Tyl)e( \]2,'d, Delete)& 
Agent( Ed, J ohn )&Objeet( lS d, M )& 
Time( Ed, Td)&NonStative(13d)& 
Be fore( IQ, 7;p). 
in (rg), M represents an individual, which is an 
e-mail message, Td represents the time of Ed, and ?\['~p 
represents the time when sentence (s2) is uttered. 9 
By using knowledge (r3), the existential status, 
Era, of M and the temporal relationship between Ed 
and Em are derived and introduced into the context. 
They are written as follows. 
(rl0) Type( E.,, Exist)& 
Object( E,u, M)&Loc( t'Jm, Lm)& 
Owner( L .... J ohn )& Time( t'~'m , 5/;,~). 
(,.11) :t'Rd( M ~et, 7;,~, T~) 
Tm represents the time when the e-mail message 
M exists and L,~ represents the location of M. 
The context also includes the following represen- 
tation when (s4) is interpreted. 
(r12) Speaker( John)&llearer(13ill). 
(r 13) Spa e chTime ( T,p )& 7'e mp F ocus'( Td). 
Namely, I assume that the time of deleting~ Td, is 
in the temporal focus when (s4) is interpreted. 
Now, let me explain how sentence (s4) is inter- 
preted under the above context. Sentence (s4) de- 
scribes an eventuality of forwarding, *ell , which is 
written ~ representation (rl) described in sac. 3.1. 
Likewise, by using knowledge @2), the following rep- 
resentations are derived. 
(rl.4)Type(*eobj~, l~Txist)& 
Object(*eobjl, *objl)&Loc(*eobji, *lobjJ)& 
Owner(*lobjl, *agentl)&Time (*e0bj 1, *toU 1). 
9I asstune that all sentences of a dialogue are uttered in the 
same temporM interval, 7'sp. 
7142 
(rlS)Type(*e~ol~yl, I';xist)& 
Objeet( *c~ovvl , *copy1)& 
CopyO f ( *copy, , *obj~ )& 
Loc(*e~o1,,al, *&ovv\])& 
Owner(*l~o~,ji , *recpl )& 
Timc( *e~ovvl, *t~ov,j1). 
(rl 6) ~l '\[ee~(\]Olt,'iTtq, *t\]l, ,let objl ). 
(r17) T Rel( Be fore, *t f ~ , *t~o~vz ). 
In (rl), *agentx, *rc.cpl and *objl represent re- 
spectively the pronominalized agent, recipient, and 
object of fbrwarding. Variable *reep~ is identified 
with Bill by ,rallying \[learer(*recp~) in (rl) with 
(r12) in the coutext. 
Representation (r14) is unified with (rl0), ~° and 
then *ayent~ is identified with John and *obj\] is iden- 
tiffed with M. Likewise, *eobj ~, *lobj~, and *to~j~ are 
identified respectively with Era, Lm and {g,,. 
Consequently, nominal reference in sentence (s4) 
is resolved. As stated below, identifying *obji with 
3//serves to resolve temporal reference. 
By using (r13), variables, *t,~)~ and *tf ..... 1 are 
identitied respectively with T.~ v and 7'd. 
The following temporal relationship is deriw;d 
fi'om (r16) bee~mse *objl is identitied with M and then 
*tobjl is identified with "/;,~. 
(r18)Tltel( I)uring, ,ti~, 7;,). 
The tbllowing temporal relationship is derived 
t.'1"o111 (rl) because *t\]o~,.,~ is identitied with 7'd. 
(r l 9 ) T \[eel( *rel , , *t \], , Td ). 
Resolving temporal reference here means deter- 
mining temporal relationship *tel, in (r19). 
By default interi)retation rule (r4), the tbllowing 
representation is derived from (r19). 
(r20) 7'ted(After, *t \] ~ , :l a ). 
On the other hand, by applying theorem (r6) to 
(r19) and then using (r18)and (r\]l), the following 
represent~tion is derived. 
(r21) :l'~.d Utefo," <*t\]\], "~'~). 
Knowh;dge..based interpretation (r21) is preferred 
over default interpretation (r20). Consequently, the 
relevant temporal relationship between forwarding 
and deleting is determined. 
Finally, new constants IQ, 7:t, l'.'~o~,a, Copy, L~ovy, 
and %.opy are generated for respectively variables *d,tl, 
*t\]l, *eoop,al, *copy\], *l¢opvl, and *t~.opul. Constant 
Copy represents the copy of the e-mail message M, 
generated by forwarding M. l,'inally, the following 
representations are introduced into the context. 
Aw,,t( tc v , Job,,)&. ~eeep(~,.), oiu)& 
Object( IQ , M)&7'ime( IQ , 7 ) )& 
N ond tativc( l,; : )& 
Tlecl(Before, TI, 7a), 
1°Representation (r\]5) cmmot be ratified with (rl0) since 
*rccpi in (r15) is identilicd with Bill. 
(r23) 7'ype( J'&opu, tSxist)& 
Objcet( E~.ovv, Copy)&CopyO f ( Copy, M )& 
l, oe( F~ovy, L~opy)&Owne,'( ~ovy, t~ill)& 
1l 'ime ( l,;~o~,y , 7 '~op~ ). 
(r>l:) :/'t~d(>,..i,,j, '/), :l;,,). (r~5) 
'll t~,el( \]~e Jt(),.~,,, rl :f ) rl :o,)y ) • 
Ne×t, consider the intm:l)retation of sentence (sS), 
which describes the R)llowing eventuality, *e.,:2. 
(r2(~) 'Cypd*e,.~ , Z,:a:i.~t) a~ 
~ aimox( .l~,.2~O,,,,~,.( .l~..~, ,p~,,~o,, )& 
H earer(*person)& 
!l' l{cl( \[fe fore, *tz2, *t ~p~)& 
:\['\[~,el( *rel.e, *tx2, *t \] ...... 2 )~ 
,qpeech.'.l'ime( *t.w'~ )& Templ" ocus( *t f ..... 2). 
Here, I assume that (s5) directly describes the exis- 
tentiM status of an individual *obj.e, which is realized 
as the zero-pronomlnalized subject in the sentence, 
mid that the existential status is construed as a stat, ive 
eventuality. '\['he location, *l,,2, of *obj.e is specilied as 
a mail box that the hearer owns since the location is 
designated hy "sochira no mail 1oox (your mail box)". 
The time, *t.,.~, when *obj.e exists is before the speech 
time since the past-.tense morpheme "ta" is used. 
'\['he referent for *obj.2 is mnbiguous since the refer 
cnt can be either e-mail message M or its copy Copy. 
The relewmt reDrent muss he Copy. As argued below, 
resolving this nominM reibrence is crucial in resolving 
temporal reference. 
I assume that the time of forwarding, 7}, is in the 
temporal focus. Resolving temporal reference here 
means dete.rmining temporal relationship *tel2 t)e~ 
tween *t,2 and ~/}. The relevant interpretation is th{~t 
• t~.~ is after illt. 
Let me explain how the fi'amework interprets (sS). 
First of all, *person in (r26) is identified with Bill, 
since *person nmst be the hearer. '\['hen *%2 in (r26) 
is identitied with E~ovy in (r23). Likewise, *obj.e, *l,~u, 
and *t~.u are identified respectively with Copy, L~ovy, 
and Teovv. Thus tim zero-pronominalized subject is 
appropriately interpreted. As a result, the following 
temporal relationship is derived from (r26) since *t:,,p. 
is identified with "l'~ovy and 7} is in the temporal focus. 
0,27) 7 'Ied (*,'d~, :I'~o,,,~, '.l) ). 
llere, consider how the temporal relationship *tel2 
is determined. By using default interpretation rule 
(rS), *rel.e is identilied with a relatkm, Contains, 
since the eventuality described by(s5) is stative. 
On the other hand, by using (r25) and theo- 
rem (rT), *rel.e is identified with a relation, After. In 
other words, the time of the eventuality deseribe.d by 
(sS) follows the time of forwarding. This knowledge- 
based interpretation is l)referred over the default in-- 
terpretation, and is the relevant interpretation. 
In the above process, identifying the pronominal- 
ized subject, *obj2, of (sS) with Copy is crucial for 
the temporal reference resolution. Assume that *obj~ 
1143 
is identified with the e-mail message M. This case 
happens when *ex2 in (r26) is identified with E,~ in 
(rl0) and *t~2 is identified with 7~. In this case, the 
following temporal relationship is derived from (r26). 
(r28)T~el(.rel~, Tin, T~). 
By using (r24) and theorem (r8), *rel2 is identified 
with a relation, Contains. This interpretation is not 
relevant. 
However, the case that leads to the wrong inter- 
pretation never occurs. When *e~2 is identified with 
Era, the following representation is derived from (r26) 
since *person is identified with Bill. 
(r29) Loc( Em , *l~. 2 )&Owner( *l~ 2 , Bill). 
This representation is inconsistent with (rl0) since 
the owner of the location of the e-mail message M 
must be uniquely identified. Thus, *e~.2 must be iden- 
tified with E~opy and then *obj2 must be identified 
with Copy. 
Consequently, identifying appropriately the refer- 
ent of the pronominalized subject of sentence (s5) af- 
fects the process of resolving temporal reference, and 
this reference interaction can be explained by exploit- 
ing knowledge such as (r2) and (r3). 
5 CONCLUSION 
This paper has made a twofold contribution to re- 
search on temporal reference resolution. First, an ac- 
count of the role played by the reference interaction 
in resolving temporal reference is given. As stated 
in see. 2, existing models cannot account for how the 
process of resolving nominal reference affects the pro- 
cess of resolving temporal reference since they do not 
explain the reference interaction. Second, it is shown 
that exploiting the reference interaction clarifies how 
a restricted set of world knowledge (knowledge of the 
temporal relationships between eventualities and the 
existential status of individuals) contributes to the 
resolution process. As stated in see. 2, existing mod- 
els resort to the entire set of causality knowledge. I do 
not entirely deny the validity of the existing approach. 
However, when a restricted set of world knowledge is 
sufficient to resolve temporal reference, immediate re- 
course to the entire set of world knowledge of causality 
can be avoided. A framework is also presented, within 
which temporal reference is resolved by exploiting the 
reference interaction. 
Future work will first extend the framework to 
deal with eventualities specified by various tense mor- 
phemes, whereas this paper focuses on eventualities 
specified by the past-tense morpheme "ta". Next, to 
explore the notion of reference interaction in more de- 
tail, the way resolving temporal reference affects the 
resolution of nominal reference must be investigated, 
whereas this paper concentrates on the converse pro- 
cess. Finally, the coverage of this framework must be 
evaluated. Evaluation will require an exhaustive list- 
ing of possible cases of the existential status of indi- 
viduals that are used to resolve reference. This paper 
accounts for cases where an individual exists at a cer- 
tain space-time location of the physical world. We can 
allow for other kinds of existential status, such as a 
status where an individuM exists in the mental state 
of a dialogue participant. 11 
REFERENCE 
Allen, J. F. (1983). Maintaining knowledge about temporal 
intervals. Communication o\] the ACM, 26, 832-843. 
Bach, E. (1986). The algebra of events. Linguistics and 
Philosophy, 9, 5-16. 
Brent, M. R. (1990). A simplified theory of tense repre- 
sentations and constraints on their composition. In: 
Proc. of the 28th Annual Meeting of the Association 
for Computational Linguistics, 119-126. 
Crouch, R. S. and S. G. Pulmau (1993). Time and modal- 
ity in a natural language interface to a planning sys- 
tem. Artificial Intelligence, 63, 265-304. 
Davidson, D. (1980). Essays on Actions and Events. Ox- 
ford University Press. 
Grosz, B. J., A. K. Joshi and S. Weinstein (1983). Pro- 
viding a unified account of definite noun phrase in dis- 
course. In: Proc. of the 21th Annum Meetings of the 
Association for Computational Linguistics, 44-49. 
Hinrichs, E. (1986). Temporal anaphora in discourses of 
English. Linguistic and Philosophy, 9, 63-82. 
Hirst, G. (199l). Existence assumptions in knowledge rep- 
resentation. Artificial Intelligence, ~9, 199-242. 
Hwang, C. H. and L. K. Schubert (1992). Tense trees as 
the "fine structure" of disconrse. In: Proc. oJ the 30th 
Annual Meeting of the Association \]or Computational 
Linguistics, 232-240. 
Kameyama, M. (1986). A property-sharing constraint in 
centering. In: Proc. of the 24th Annual Meetings of the 
Association for Computational Linguistics, 200-206. 
Kameyama, M., ii. Passonnean and M. Poesio (1993). 
TemporM centering. In: Proc. of the 31th Annual Meet- 
ing of the Association for Computational Linguistics, 
70-77. 
Lascarides, A. (1992). Knowledge, causality, and temporal 
representation. Linguistics, 30, 941-973. 
Partee, B. H. (1984). NominM and temporal anaphora. 
Linguistics and Philosophy, 7, 243-286. 
Reichenbach, H. (1947). Elements of Symbolic Logic. 
Macmillan, New York. 
Song, F. and R. Cohen (1991). Tense interpretation in the 
context of narrative. In: Proc. of the 9th National Con- 
ference on Artificial intelligence, 131-136. 
Vlach, F. (1993). Temporal adverbials, tenses and the per- 
feet. Linguistics and Philosophy, 16, 231-283. 
Webber, B. L. (1988). Tense as discourse anaphor. Com- 
putational Linguistics, 1~, 61-73. 
Webber, B. L. and B. BMdwin (1992). Accommodating 
context change. In: Proc. of the 30th Annual Meeting 
of the Association .for Computational Linguistics, 96- 
103. 
tlVarious kinds of existence are presented by Hirst (1991), 
1144 
