e 
O 
O 
O 
O 
O 
O 
O 
O 
0 
O 
O @ 
O 
O 
0 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
0 
O 
O 
O @ 
O 
O 
O 
O @ 
O 
O 
O 
Approachesto Japanese zero pronouns: 
Centering and relevance 
Tomoko Matsui 
International Christian University 
Division of Languages 
• 3-10-20sawa 
Mitaka, Tokyo, JAPAN 181 
matsui@icu.ac.jp 
Abstract 
Accessibility of candidate referents is a crucial 
factor in successful reference assignment and 
this intuition has been adequately 
accommodated and developed in centering 
theory. By contrast, another equally important 
factor involved in reference assignment, 
namely, accessibility of contextual 
assumptions, seems to have not been 
addressed fully as a part of a computational 
theory of reference. In the field of pragmatics, 
however, the role of context has always been 
one of the central issues and the recent 
approach to context selection based on the 
notion of relevance (Sperber & Wilson 
1986/1995) seems to be currently the most 
promising. In this paper, I will review and 
assess the recent centering approach to the 
interpretation of Japanese zero pronouns 
(Walker et al. 1994) as a case study, and 
suggest that relevance theory can provide one 
way of complementing it. In order •to 
incorporate a needed mechanism of contextual 
selection into a model of reference assignment, 
model builders must take into account the 
heater's ability to anticipate certain cognitive 
effects that may be achieved by upcoming 
utterances. It is my hope that this paper will 
lead to a fruitful discussion of finding possible 
ways of using Sperber & Wilson-'s notion of 
relevance in practical applications. 
1. Introduction 
When a system has to assign a referent to a 
referring expression, it is almost always the 
case that there is more than one candidate 
referents and one has to resort to some Way of 
eliminating the wrong candidates in order to 
choose the right one. 
There are two main ways of doing this. 
First, candidate referents can be ranked on the 
basis of accessibility (Erlm & Gundel 1987; 
Fretheim and Gundal 1996; C~mbacher and 
Hargreaves 1988, Grosz et al. 1995; Sidner 
1983a, b; Walker et al. 1998). Alternatively, 
they can be checked against the accessibility 
of contextual assumptions (F'mcher-Kiefex 
1993; Kintsch 1988, Magliano et al. 1993; 
McKoon & Ratcliff 1992; Sanford & Garrod 
1981; Sharkey & Sharkey 1987; Singer 1993, 
1995). 
In either case, once a candidate is singled 
out, the acceptability of the referent needs to 
be tested against some pragmatic criteria. 
Currently, we seem to have at least three 
distinct criteria available: 
i. Truth-based criterion- i.e. whether the 
overall, interpretation is likely to be 
factually plausible (Clark 1977; Clark & 
Haviland 1977; Erku & Gundel 1987; 
Sanford & Gan~ 1981; Sidner 1983a); 
iL Coherence-based criterion -i.e. whether 
the overall interpretation is likely to be 
coherent (Asher & Lascaries 1993; Grosz 
et al. 1995, 1998; Hobbs 1979; Lascarides 
& Asher 1993; Sanders et aL 1992; Walker 
et aL 1994, 1998); 
Hi. Relevance-based criterion - i.e. whether 
the overall interpretation is likely to be 
optimally relevant (Matsui 1993, 1995, 
1998; Wilson 1992; Wilson & Matsui 
1998). 
11 
o 
In this paper, I will focus on one version of 
coherence-based criterion, namely, centering 
theory. One of the goals of centering theory is 
to sort out the various mechanisms used to 
maintain discourse coherence, and the use of 
various referring expressions is regarded as 
one such mechanism. Among the various 
hypotheses put forward by centering theory, 
what concerns us most is the following: that 
'each utterance \[except the initial utterance\] in 
a coherent discourse segment contains a single 
semantic entity - the backward-looking center 
\[or Cb\] - that provides a link to the previous 
utterance, and an ordered set of entities - the 
forward-looking centers \[or Cf\] - that offer 
potential links to the next utterance'(Gordon et 
al. 1993:311). There are two rules to provide 
constraints on choosing centers, which are 
shown in (1): 
( l)Rule 1: If any element of Cf (Un) is 
realised by a pronoun in Un+l then the 
Cb (Un+l) must he realised by a 
pronoun also. 
Rule 2: Sequences of continuation are 
preferred over sequences of retaining; 
and sequences of retaining are to be 
preferred over sequences of shifting. 
The first rule states that the most highly 
ranked element of the forward-looking center 
of a previous utterance is the backward- 
looking center of the current utterance, and 
must be realised by a pronoun if any element 
of the Cf of the previous utterance is realised 
by a pronoun in the current utterance. The 
following example from Gordon et ai.(1993). 
shown here in (2). illustrates this nde: 
(2)\[l\]Susan gave Betsy a pet hamster. 
Cf={Susan, Betsy, hamster I ) 
\[2\]She reminded her such hamsters were 
quite shy.. 
Cb=Susa,; Cf={St~m. Betsy. hamsters} 
\[3\]She asked Betsy whether she liked 
the gift. 
Cb~usan; Cf=-\[Susan. Betsy, 
gift=hamster I } 
\[3 "\]Susan asked her whether she liked 
the gift. 
Here, the first utterance has no Cb because it is 
12 
the initial sen'tence of a discourse. Its Cf 
includes the referents off'Susan' and 'Betsy' 
and the semantic interpretation of 'a hamster', 
ranked in that order. The second utterance has 
Susan as the Cb and a Cf with Susan as its 
most highly ranked element. The third 
utterance preserves the Cb and prominent Cf 
from the previous utterance, therefore it 
pronominalises the Cb successfully. By 
contrast, utterance \[3 1, in which Susan is 
realised by a name and Betsy is realised by a 
pronoun, leads to stylistic infelicity. According 
to Gordon et al., this is due to violation of 
Rule 1 mentioned above. 
The ranking of forward-looking centers is 
generally based on the discourse salience of 
each candidate entity. According to Grosz et 
aL (1995), although an ultimate criteria for 
deciding the ranking has not been worked out 
yet, there are evidences to support the idea that 
grammatical role such as SUBJECT, OBJECT, 
etc., can affect the Cf ranking. Thus, I will 
simply assume here the following preference 
in ranking forward-looking center shown in 
(3), as suggested by Grosz et al. (ibid.): 
(3)SUBJECT>OBJECT>OTHERS 
The highest ranked member of the set of 
forward-looking centers is called the 
'preferred center; or Up" As mentioned 
above, Cp is regarded as the most likely 
candidate for Cb in the following utterance. 
Mother important claim made by centering 
theory is that discourse segments are more 
coherent if they share the same Cb. On the 
basis of this idea, different degrees of 
coherence are proposed. For example, Walker 
et al. adopt the following 4 types of transition 
between discourse segments, each 
corresponding to different degree of 
coherence, using the notion of Cb and Cp;. 
namely. 'continue. 'retain; 'smooth-shift" 
and "rough-shift'. These are shown in (4). 
When two utterances, say Ui-I and U share the 
same Cb, and the same entity is also the 
highest-ranked Cf, i.e. Cp, in Ui-l, the 
transition from Ui-I to Ui is called 'continue: 
When Ui-1 and Ui share the same Cb, but the 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 @ 
0 @ 
0 
0 
0 
0 
0 @ 
0 @ 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
@ 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
0 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
same entity is not the highest-ranked Cf in Ui, 
the transition is called 'retain '. When Ui-1 and 
Ui do not share the same Cb, there are two 
possibilities: if Cb in Ui is the same as Cp in 
Ui, the transition is 'smooth-shift'; if Cb in Ui 
is not the same as Cp in Ui, it is 'rough-shift '. 
It is claimed that when the hearer has to 
choose one from several possible 
interpretations, the one based on the most 
coherent transition should be chosen. The 4 
transition states are ordered in the following 
way according to their preference: 
(4)CONTINUF~RETAIN>SMOOTH- 
SHIFT>ROUGH-SHIFF 
The second of the two rules in (1) is about this 
ordered preference of transition states.' 
2. Walker et al. (1994) and Japanese zero 
pronoun resolution 
Walker et ai. (1994) propose the following 
ranking order of forward-looking centers to 
deal with Japanese: 
(5)(Grammatical or Zero) TOPIC > 
EMPATHY > SUBJECT • OBJECT2 • 
OBJECT> OTHERS 
As you can see, they add two new grammatical 
roles, namely, topic and empathy, to the list of 
factors affecting pronoun resolution. Since the 
notion of topic will become important in the 
discussion which follows, I will briefly 
describe the Japanese topic marker below. For 
empathy-loaded verbs, please refer to Kuno 
1987 and Kuno & Kaburaki 1977. 
In Japanese, both in written and spoken 
discourse, lqPs which can be recovered from 
context are often omitted. The omitted NPs are 
often called 'zero pronouns'. It is widely 
agreed that missing NPs in JapaneSe behave 
like pronouns in other languages such as 
There were only three aTtn:idon statea in t&e origml 
\[o~ulmion by Gron et ~ (1983, /99S), namely, 
CONTINUE, RETAIN and SHIFZ. Fhe distinction 
between SMOOTH-SHIFT and ROUGII-SHIFF w~r 
\[u~ proposed in Brenn~ (1987). 
English. Japanese zero pronouns should be 
distinguished from missing NPs in 'pro-drop' 
languages such as Italian, since in 'pro-drop' 
languages, information to recover missing NPs 
is morphologically encoded elsewhere in the 
sentence, e.g. in the form of verb inflection, 
whereas Japanese lacks such an overt 
encoding. 
Japanese zero pronouns are used equally 
often as inter-sentential discourse anaphors 
and intra-sentential anaphors. Here, I will 
concentrate on their use as discourse anaphors, 
where the role of context and adequate 
pragmatic criterion is crucial. For those who 
are interested in intra-sentential use of zero 
pronouns, please refer to e.g. Hasegawa 1985 
and Kameyama 1986. 
In Japanese, there are several postpositional 
particles. The one we are interested in here is 
'wa; which is often called a ~opic marker: 
As the name suggests, 'via' is typically used to 
construct a grammatical topic of a sentence, 
which is characterised as an entity whose 
existence is presupposed. The function of 'wa" 
might become clearer when it is compared 
with another particle 'ga; which marks a hrP 
in the subject position, which typically 
conveys a new information. Compare (6a) and 
(6b): 
(6) a. John ga hana o kaua. 
SUB flowers ACe bought 
• 'John bought flowers.' 
b. John wa hana o katta. 
TOP flowers A~ bought 
"John bought flowers.' 
(6a) and (6b) share the same propositional 
content. However, only (6a) is acceptable as 
an answer to the question such as Who 
bought flowers?" while only (6b) is acceptable 
as an answer to the question such as What did 
John buy?'. There are various suggestions 
about how to characterise functions of the two 
particles (see e.g. Shibatani 1990; Tanaka 
1991), and although it is an interesting 
question on its own, it shouldn't concern us 
here. This is because Walker et al. are only 
interested in the surface form of 'lqP+wa; 
13 
which is automatically given the highest 
accessibility ranking in their framework. 
Walker et al. claim that a topic NP is more 
likely to be realised as a zero pronoun in the 
subsequent discourse than any other NPs due 
to its highest degree of accessibility. 
Moreover, in Walker et al. 's framework, topic 
NP is given two further advantages: they 
suggest (a) that a NP marked by %va' becomes 
the backward-looking center even at the onset 
of the discourse; and (b) that once topic NP is 
realised as the backward-looking center ,as 
long as it continues to be realised as a zero 
pronoun in subsequent discourse, it could 
continue to be the backward-looking center. 
The second advantage given to the topic NP is 
called 'zero topic assignment', which is 
defined as in (7): 
('/)Zero Topic Assignment (ZTA) 
(optional) 
When a zero in Ui+l representsan entity 
that was Cb (Ui), and when no other 
CONTHqUE transition is available, that 
zero may be interpreted as the ZERO 
TOPIC of Ui+l 
Walker et al. demonstrate that their 
framework, including an optional rule of zero 
topic assignment, can successfully explain the 
preference in the interpretation of the last 
sentence in (8): 
(8)\[1\] Hanako wa siken o oete, 
TOP exam ACC finish 
kyooahitu ni modorim~hlta. 
classroom to returned 
1-1anako returned to the classroom, 
finishing her exams' 
Cb=Hanako; Cff\[Hanako, exam} 
\[2\] 0 ~n o rotta ni shimaimashito. 
book Ace locker in took-away 
'(She) put her books in the locker' 
Cb=Hanako; Cff-{Hanako, book} 
\[3\]/tsumo no yooni Michiko ga 
always like SUB 
0 monda/no tok/kata o 
questionssolve-way ^CO 
14 
setumeishimashita. 
explained 
'Michiko, as usual, explained (to her) 
how to answer questions.' 
Cb=Hanako; 
Cfl={Hanako, Michiko ...} ZTA 
\[col, rrn,lt~\] 
Cf2={ Michiko, Hanako .... } 
\[RETAIN\] 
\[4\] 0 0 oh/ru ni sasoimashita 
lunch to invited 
'(She) invited (her) to lunch' 
Cbl= Hanako; 
Cfl={Hanako, lunch, Michiko} 
\[CONTINUE\]from \[3\]-Cfl 
Cb2=Michiko 
Cf2={Michiko, lunch, Hanako} 
\[S-SHIFYJfrom \[3\]-Cf2 
According to the questionnaire carried out by 
Walker et al., the preferred interpretation of 
\[4\] is that Hanako invited Michiko to lunch. 
As you can see, in fact, there are two possible 
ways of ranking forward-looking center in \[3\], 
and two possible ways of deciding both the 
backward-looking center and the ranking of 
forward-looking center in \[4\]. In their analysis 
of (8), the preference in the interpretation in 
\[4\] is explained by zero topic assignment in 
\[3\] and preference on 'continue' transition in 
\[4\]. 
In summary, Walker et aL's account of 
Japanese zero pronoun is based on two 
independent preference mechanisms: the first 
one is the forward center farting, and the 
second is the ordered transition states. 
However, as Walker et al. themselves point 
out, each transition state between discourse 
segments is determined by the ordering of 
forward looking centers, the predictions of the 
theory tend to depend more largely on the 
forward center ranking. 
3. Problems with Walker et al. 
Now I would like to discuss some problems 
With Walker et al. "s framework. Needless to 
say. it has great advantages, such as relative 
ease of computational imPlementation. 
Moreover, I agree that accessibility of 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 @ 
0 
0 
0 
0 
0 
0 
0 @ 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
discourse entities plays an important role in 
reference assignment, and their forward center 
ranking is an adequate enough approximation 
of accessibility of discourse entities in 
different grammatical categories in Japanese. 
However, as I mentioned before, there are 
cases whose interpretation process cannot 
poss!bly be explained by the accessibility 
factor alone. Accounts of reference assignment 
which are largely based on accessibility of 
discourse entities tend to exhibit their 
weaknesses when they face cases which 
require some pragmatic inferences, and 
Walker et al. is not an exception here. I will 
illustrate two problems they need to solve 
below. 
3.1, Multiple topics 
One of the most obvious shortcomings of 
Walker et al. 's approach is that it cannot 
handle situations where there is more than one 
topic in a sentence. In their framework, in 
order to identify zero pronouns, the backward- 
looking center has to be identified first. The 
backward-looking center, in turn, is 
determined by the way forw~d-looking 
centers are ranked. Therefore, the most 
powerful mechanism in their framework is the 
forward-looking center ranking shown in (5) 
above. However, notice that it is only useful if 
there is no more than one entity in each 
category in a sentence. 
Let us concentrate on the category of topic 
here. In Walker et al. 's framework, the topic 
marker 'wa' is given a special status: the topic 
marker 'wa' is so powerful that the topic NP 
becomes the most highly ranked forward- 
looking center even at the onset of a discourse; 
in addition, once a topic NIP is realised as the 
backward-looking center, as long as it 
continues to be realised as zero pronoun in 
subsequent discourse, it could continue to be 
the backward-looking center. This status of the 
topic NP rightly allows the possibility of 
multiple topics as in (9), which is very often 
seen in Japanese discourse: 
(9)\[1\] Mary to Jane war shinyuu ck 
and TOP best friends are 
'Mary and Jane are each other's best. 
friend.' 
\[2\]Senshuu no Doyoubi, Mary wa kaze 
last week GE~ Saturday TOP cold 
o hiite nete-ita 
^CC had lying-was 
Last Saturday, Mary had a cold and 
was lying on the bed.' 
\[3\]ltsumo no youni, Jane wa 
always GEN as TOP 
ohiru goro denwa shita. 
noon around telephone did 
'As always, Jane phoned (Mary) around 
noon. 
\[4\]0 0 eiga ni so, sou tsurnori datta. 
film to invite planning to was 
'(She) was planning to invite (her) to a 
film.' 
However, cases of multiple topics cause a 
serious problem to Walker et al. Here, the 
preferred interpretation of \[4\] is that Jane was 
planning to invite Mary to a film. Walker et al. 
might explain this preference by saying that 
this is because 'Jane'is the backward-looking 
center in \[4\]. According to their framework, 
however, the alternative interpretation, 
namely, that Mary was planning to invite Jane 
to a film, is equally accessible, since 1Vfary" 
could continue to be the backward-looking 
center in \[4\]. The problem is that in their 
current framework, Walker et al. do not 
provide any mechanism to choose one 
interpretation and discard the other. 
There is another problem concerning 
multiplicity of entities with equal degree of 
accessibility. In Japanese, post-positional 
particles such as ~va' and 'ga" shouldn't be 
used more than once in a sentence. Thus, you 
wouldn't come across sentences with two 
overtly marked topics or subjects. However, it 
is possible to have more than one NP with 
more or less equal salience in one sentence, for 
example, when two nouns; the first being a 
modifier and the second being the head noun, 
form a NP. A noun modifier is followed by a 
particle 'no', the Genitive Case particle. Some 
examples of NPs which contain noun 
modifiers are shown in (10): 
15 
(lO)a. Mary no tomodachi 
GEN friend 
qVIary's friend' 
b. Niwa no ki 
garden GEN tree 
'A tree in the garden" 
c. Tegami no henji 
letter GEN reply 
'A reply to the letter' 
Here, the first noun in each NP is the modifier. 
The most typical relation exhibited between 
two nouns combined by the particle 'no' is the 
'possessive' relation, as in (lOa). However, the 
use of 'no' is by no means restricted to that 
relation, as illustrated in (lOb) and (lOc). Now 
consider (11), which include a NP with this 
structure: 
( i I)(A memo written by a man, and 
addressed to his wife) 
\[l\]Kooto no botanga toreta. 
', coat GEN button SUB came off 
~3ne of the buttons of (my) coat has 
off' 
\[2\]0 0 sagashi.temo, 0 mitsukaranakatta 
search-although was not found 
'(I) tried to find (i0, but failed.' 
\[3\]gyou jyuu ni 0 0 0 sagashite 
today within at find 
tsukete-hos/u¥. 
fix -want 
'(1) want (you) to find and fix (it) 
today." 
\[4\]Ash/~ 0 0 h/tauyou Ja, 
tomorrow need 
Tomorrow, (1) will need (it)' 
( 12)\[1\]Sakuba~ John wa fitensha no kagi 
last night TOP bicycle GEN key 
o kake-wasweta. 
ACe lock-forget 
l.ast night, John forgot" to use the key to 
lock his bicycle' 
\[2\]gesa, too, 0 na/amatteita. 
this morning already gone-has 
This morning, (it) has already gone' 
In (11), 'one of the buttons of my coat' is 
introduced in \[1\] as the subject of the sentence. 
What is important here is that the NP 
introduces two conceptual entities, namely, a 
'coat' and a ~outton', which are equally 
accessible. The question is whether Walker et 
al. 's framework can handle cases like (11). For 
\[2\], the preferred interpretation is that the 
speaker wants his wife to find and fix the 
button today. Walker et al. 's system would 
successfully predict that the button is the 
backward-center+ for \[2\] and \[3\]. However, for 
\[4\], the preferred interpretation is that the 
speaker will need his coat tomorrow, rather 
than the button. I do not see how Walker et al. 
can explain this. Example (12) causes exactly 
the same problem for them. Walker et al. 
would predict that the preferred interpretation 
should be that the key has gone. This 
obviously is the wrong prediction. The point I 
would like to make here is that there are many 
cases in Japanese discourse where there is 
more than one roughly equally salient 
discourse entity in a sentence which 
subsequently become equally strong candidate 
referents for zero pronouns, and some 
mechanism of choosing the right one is 
needed. 
3.2. Ordered preference of transition states 
Now I would like to move on to a different 
kind of problem, which concerns their ordered 
preference of transition states shown in (4). 
Walker et al., as well as centering theorists in 
general, assume that when there are more than 
one possible overall interpretation available, 
the one which exhibits 'continue' transition is 
prefene& This is based, on the assumption that 
maximally coherent segments are those that 
require less processing effort and the heater 
will prefer an interpretation which requires 
less processing effort. However, it is not 
difficult to think of examples which go against 
their assumption. For example, look at (13) 
and (14): 
(13)Mary \[l\] KinouO Peterto denwa de 
yesterday with telephone by 
j/tan o kahnn/n dek/m? 
16 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 @ 
0 
0 
0 @ 
0 @ 
0 
0 
0 
0 
0 
0 @ 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
O 
O 
O 
O 
O 
O 
O 
O @ 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O @ 
O 
O 
0 
O 
O 
O 
O 
0 @ 
O 
O 
O 
time ACC confirm could 
'could you confirm the time with Peter by 
phone yesterday?' 
John \[2\] lya, 0 dekinakatta. 
no couldn't 
~10, (I) couldn't' 
John \[3\] 0 ie ni inakattn 
home at present-not-was 
aL (I) was not at home. \[CONTINUE\] 
b. (He) was not at home \[SHIFI'\] 
(I 4)\[ ~ \]John wa jou~hi ni mama ga 
TOP boss with head SUB 
agaranai. 
hold-not 
'John Cannot hold up his head before his boss 
(i.e. John cannot help feeling ashamed of 
himself in front of his boss). 
\[2\]ltsumo 0 kaisha ni 0 saki ni kuru. 
always office to earlier come 
'(He) always comes to the office earlier 
(than him).' 
a. (John) always comes to the office earlier 
(than his boss). 
b. (John's boss) always comes to the office 
than (John). 
Let us look at (13) first. What interests me 
here is the interpretation of the zero pronoun 
in \[3\]. Notice that both \[3a\] and \[3b\] are 
equally acceptable in terms of factual 
plausibility. In such case, Walker et al, would 
predic~ that \[3a\], which exhibits a 'continue' 
transition should be preferred. However, for 
some reason, the preferred interpretation for 
\[3\] is not \[3a\], but \[3b\]. Similarly, in (14), the 
preferred interpretation for \[21 is definitely 
\[2b\]. However, Walker et al. predict that \[2a\] 
should be preferred. Obviously, their 
mechanism based on the accessibility of 
transition states makes wrong predictions. In 
the next section, I will consider why this is the 
case. 
4. A relevance-theoretic solution to 
preblems with Walker et al. 
Sperber & Wilson's relevance theory (Sperber 
& Wilson 1986/95) inherits the Gricean 
assumption that the hearer's goal of verbal 
17 
understanciing is to find an interpretation 
intended by the speaker. However, it differs 
from Gricean approach in two crucial points: it 
does not take the view that we have to follow 
maxims, nor the view that we have to be co- 
operative, to achieve successful 
communication. Sperber & Wilson claim that 
what makes communication achievable at all 
is a fundamental mechanism built in our 
cognitive system, namely, the pursuit of 
relevance. This is expressed as the First, or 
Cognitive, Principle of Relevance: 
(15)Cognitive Principle of Relevance 
• Human Cognition tends to be geared to 
maximisation of relevance. 
the 
.The notion of relevance is defined in terms 
cognitive effects, i.e. some changes in the 
belief system, and processing effort to obtain 
such effects: 
(16)Relevance 
a. The greater the cognitive effects, the greater 
the relevance; 
b. The smaller the effort needed to achieve 
those effects, the greater the relevance. 
Cognitive effects result from the interaction of 
new and old (or contextual) information in one 
of the following three ways: (a) combining 
wi.~..~ an existing assumption to yield 
contextual implications; (b) strengthening an 
existing assumption; (c) contradicting and 
eliminating an existing assumption." Processing 
effort is the mental effort needed to parse the 
utterance, decide what proposition and 
propositional attitude it was intended to 
express, access an appropriate context, and 
work out the contextual effects of the utterance 
in the context. When an utterance has more 
than one possible interpretation, the hearer 
should look for the one which satisfies the 
following conditions of optimal relevance: 
( 17)Optimal relevance 
An utterance is optimally relevant to the heater 
iff: 
a. it is relevant enough to be worth the heater's 
processing effort; 
b. it is the most relevant one compatible with 
the speaker's abilities and preferences. 
The Second, or Communicative Principle of 
Relevance, governs this search process: 
(I 8)Communicative Principle of Relevance 
Every utterance communicates a presumption 
of its own optimal relevance. 
The pursuit of optimally relevant interpretation 
suggests a pattern of comprehension procedure 
the hearer should follow, which can be spelled 
out as (19): 
(19)Relevance-theoretic comprehension 
procedure 
a. consider cognitive effects in their order 
of accessibility (i.e. follow a path of least 
effort); 
b. stop when the expected level of relevance is 
achieved. 
Now let me illustrate how this 
comprehension procedure should work for 
Japanese zero pronoun resolution illustrated in 
(11), (12), (13) and (14). In Matsui (1995, 
1998, also Wilson and Matsui 1998), I have 
developed the idea proposed by Wilson (1992) 
that in addition to the factor of accessibifity of 
candidate referents, there is another important 
factor which affects the bearer's choice of 
referent, namely, accessibifity of contextual 
assumptions. Accessibility of contextual 
assumptions becomes particularly crucial 
when there is more than one roughly equally 
accessible candidate referent. and it is the 
factor which is vital to solve problems with 
Walker et al.. In fact. the importance of 
contextual assumptions in reference resolution 
had been recognised before and various 
proposals were made as to how to retrieve the 
right context: some appeal to situationally 
partitioned knowledge (e.g. Sanford & Oan~ 
1981) and others arc motivated by textual 
coherence (e.g. Hobbs 1979; Asher & 
Lascarides 1993). The account pursued here is 
different from any existing accounts in that it 
18 
claims that the selection of contextual 
assumptions is ordered in terms of both their 
accessibility and likeliness to contribute 
towards cognitive effects of the utterance. In 
other words, in relevance theory, it is assumed 
that these candidate referents are tested in 
parallel, with the one which gives quickest 
access to a context in which the utterance as a 
whole yields an acceptable overall 
interpretation being selected. 
As a working hypothesis, let us assume that 
certain contextual assumptions are accessed by 
the hearer after the immediately preceding 
utterance is processed, during and after the 
current utterance is being processed. I have no 
specific claim here concerning what triggers 
the retrieval or the construction of certain 
contextual assumptions, and can go along with 
existing suggestions (e.g. it can be triggered by 
lexical information, or/and by situational 
knowledge). As the second working 
hypothesis, I would like to suggest that after 
having understood an utterance, the hearer 
tends to have, if not always, fairly accurate 
expectation as to what kind of cognitive 
effects he would like to obtain from the next 
utterance.. Relevance theory predicts that when 
an utterance creates in the hearer an 
expectation for a specific cognitive effect to be 
achieved by the next utterance, other things 
being equal, the hearer is more likely to spen d 
his processing effort to find an interpretation 
which can achieve such cognitive effect when 
interpreting the utterance. As a consequence, 
the candidate referent which is not the highest 
in the general accessibility ranking can 
become the most accessible to the hearer if the 
referent is expected to contribute to the 
interpretation hc is looking for. In other words, 
relevance theory predicts the alteration of 
accessibility ranking of the candidate referents 
as a result of the pursuit of certain cognitive 
effects. 
Consider examples (I1) and (12) in the 
context discussed above. In (11), the utterance 
in \[3\] is a request to the wife to find and fix 
the missing button of the speaker's coat before 
tomorrow. Generally, if someone asks you to 
do something by certain time, there should be 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 @ 
0 
0 
0 
0 
0 
0 
0 
0 
0 
P 
0 
0 @ 
0 
O 
O 
O 
O @ 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O @ 
O @ 
O 
O @ 
O 
O 
O 
O 
O 
O 
O 
O 
O @ 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
0 
a good reason for such a time limit, since the 
time limit in turn maY create certain priority. 
Thus, it should be reasonable to assume that 
after hearing the utterance in \[3\], the question 
such as 'why do I have to do it today?' or 
'can't it wait a little while?' occurred in the 
hearer's mind. If so, the utterance in \[4\] can 
readily be interpreted as the reason why he 
made such a request: he wants to wear the coat 
on the following day. The referent is assigned 
automatically during the process of finding 
expected the cognitive effects. Of course, what 
he needs is the coat with the button fixed, and 
this interpretation is only possible with the 
overall interpretation in which the 'coat: 
rather than the 'button" is the referent of the 
zero pronoun. The first utterance in (12), on 
the other hand, seems to create different kind 
of expectation in the hearer. After hearing that 
the bicycle was left unlocked, the most likely 
question occurs to everyone's mind is 'what 
happened to the bicycle?" We all know the 
likely consequence of the unlocked bicycle, 
and the hearer of (12) gets the expected 
contextual effects by interpreting the zero 
pronoun as the ~icycle; rather than the ~ey '. 
In both (13) and (14); it is reasonably 
assumed that the hearer will have a why- 
question in his mind before hearing the final 
utterance. In the case of (13), the only possible 
reason why John could not confirm the time 
with Peter by phone was because Peter was 
not .at home when John phoned, and the zero 
pronoun is resolved .automatically in the 
process of obtaining this interpretation. I will 
Jook at (14) in more detaiL The interpretation 
of (14) might be explained like this. After 
processing the first utterance, certain 
assumptions might become moderately 
accessible (but not necessarily at the conscious 
level) to the hearer:, e.g. various assumptions 
about John and his boss, and more general 
assumptions about l~eing ashamed of; e.g. 
that one must feel unhappy about such 
situation, or that one must have specific 
reasons for such feeling, etc. In this way, 
contextual assumptions might contribute to 
form the hearer's anticipation about the way 
subsequent utterances achieve relevance. 
When the hearer interprets the second 
utterance in (14), further:-assumptions related 
to the event described, such as that 'workers 
are encouraged to come to work early ', or 
~osses like their workers to arrive before 
them' etc., might become highly accessible. 
Such assumptions contribute to the hearer's 
search for the way the utterance could achieve 
cognitive effects. At the final stage of the 
interpretation process, the hearer finds the only 
way in which the second utterance in (14) 
might be intended to achieve relevance in a 
context created by the first - namely, as an 
explanation for why John cannot help feeling 
ashamed of himself in front of his boss - and 
the zero-pronouns are resolved automatically 
in the process. 
Furthermore, notice that as a consequence 
of using the notion of accessibility of 
contextual assumptions, an ad hoc system such 
as the ordered preference in transition states in 
Walker et aL's account becomes automatically 
unnecessary. For example, our framework 
could easily accommodate cases such as (13) 
and (14), where the referent of a zero pronoun 
in the current discourse is not the same as that 
of a zero pronoun in the preceding discourse, 
as well as cases where the most accessible 
candidate referent indeed coincides with the 
referent chosen on the basis of the preferred 
overall interpretation. Recall that the ordered 
transition states is created on the basis of the 
assumption that the hearer prefers an 
interpretation which exhibits 'continue" 
relation, because such an interpretation 
requires less processing effort. I suggest that 
this assumption is ultimately wrong for the 
following reason: the preferred interpretation 
by the hearer is the one which provides 
enough cognitive effects worth his processing 
effort, rather than the one which merely 
requires less processing effort. Walker et al. 
would predict that the preferred interpretation 
for the second sentence in (14) is that "John 
always comes to his office earfier than his 
boss.' The question we have to ask here is: 
how could this interpretation possibly achieve 
relevance? I cannot easily see how. Relevance 
theory predicts that such interpretation will 
19 
never be considered when there is an 
alternative interpretation accessible which 
achieves relevance, even if the latter might 
require more processing effort. Notice here 
that the relevance theory shares the view that 
the intended interpretation should be the most 
accessible one for the hearer. However, in the 
framework of relevance theory, it is 
accessibility of contextual assumptions, 
together with the accessibility of contextual 
referents, that determines the overall 
accessibility of an interpretatimL In this way, 
the fact that the preferred interpretation of the 
second sentence in (14) is more accessible 
overall for most of us, than the alternative 
interpretation, is adequately explained in this 
framework. 
5. Summary 
In this paper, I have proposed a way of 
complementing centering theory from a 
relevance-theoretic perspective. I have 
suggested that a model of reference 
assignment which appeals to the expectation 
based on the accessibility of candidate 
referents such as centering theory should 
ultimately accommodate some mechanism of 
contextual selection and pointed out that the 
hearer's expectation of specific contextual 
effects (a type of forward inferences) should 
be taken advantage of in creating such a 
mechanism. Recently, Oberlander (1998) says 
that 'the key lesson from the work on pronoun 
generation and interpretation is that we must 
develop a more sophisticated view of 
'~xpectafion."I hope this paper makes some 
contribution towards that goal. 

References 
Asher, N. & Lascarides, A. 1993. Lexical 
disambiguation in a discourse context. 
Journal of semantics 12:69-108. 
Gordon, P., Grosz, B. & Gilliom, L. 1993. 
Pronouns, names, and the centering of 
attention in discourse. Cognitive Science 
! 7:31 i-347. 
Grosz, B., Joshi, A. & Weinstcin, S. 1995. 
Centering: A framework for modeling the 
local coherence of discourse. 
Computational Linguistics 21 (2):203-224. 
Hobbs, J. 1978. Resolving pronoun references. 
Lingua 44:311-338. 
Matsui, T. 1993. Bridging reference and the 
notions of topic/focus. Lingua 90:49-68. 
Matsui, T. 1995.Bridging and relevance. 
University College London, Phi) thesis. 
Matsui, T. 1998. Pragmatic criteria for 
reference assignment: A relevance-theoretic 
account of the acceptability of bridging. 
Pragmatics and Cognition 6(1/2), 47-97. 
Oberlander, J. 1998. Do the right thing ... but 
expect the unexpected. Computational 
Linguistics 24(3):501-507. 
Sanford, A. & Garrod, S. 1981. Understanding 
Written Language: Explorations in 
Comprehension Beyond Sentence. 
Chichester: Wiley. 
Sidner, C. 1983a. Focusing and discourse. 
Discourse Processes 6:107-130. 
Sped~r, D. & Wilson, D. 198611995. 
Relevance: Communication and Cognition. 
Oxford: Blackwell 
Walker, M.A., Iida, M. & Cotes, S. 1994. 
Japanese Discourse and the Process of 
Centering. Computational Linguistics 
20(2), 193-232. 
Walker, M. A., Joshi, A.K. & Prince, E.F. 
1998. Centering Theory in Discourse. 
Oxford: Clarendon Press. 
Wilson, D. 1992. Reference and relevance. 
UCL Working Papers in Linguistics 4:167- 
191. 
Wilson, D. & Matsui, T. 1998. Recent 
Approaches to bridging: Truth, coherence 
and relevance.UCL Working Papers in 
Linguistics I0:173-200. 
