Constraints and Defaults on Zero Pronouns in Japanese 
Instruction Manuals 
Tatsunori Mori, Mamoru Matsuo, Hiroshi Nakagawa 
Division of Electrical and Computer Engineering 
Yokohama National University 
79-5 Tokiwadai, Hodogaya-ku, Yokohama 240, JAPAN 
{ mori,ma moru }@forest.d nj.ynu.ac.jp, na kagawa@na kla b.dnj.ynu.ac.jp 
Abstract 
In this paper, we propose a method for 
anaphora resolution of zero subjects in 
Japanese manual sentences based on both 
the nature of language expressions and the 
ontology of ordinary instruction manuals. 
In instruction manuals written in Japanese, 
zero subjects often introduce ambiguity 
into sentences. In order to resolve them, 
we consider the property of several types of 
expressions including some forms of verbal 
phrases and some conjunctives of clauses, 
and so on. As the result, we have a set 
of constraints and defaults for zero subject 
resolution. We examine the precision of the 
constraints and defaults with real manual 
sentences, and we have the result that they 
make a good estimate with precision of over 
80%. 
1 Introduction 
From simple electrical appliances to complex com- 
puter systems, almost all machines are accompa- 
nied by instruction manuals. Since recently there 
are many machines whose operating procedures are 
complicated, we have much more trouble in many 
cases including translating their manuals into other 
languages, maintaining consistency between the de- 
scription in manuals and the actual behavior of the 
machines. To solve these problems, we have to have 
a computer assisted system for processing manual 
sentences. In processing instruction manuals written 
in Japanese, however, it is problematic that almost 
all subjects are omitted. They are called "zero sub- 
jects." For example, machine translation systems 
have to supply appropriate subjects to translate sen- 
tences. Therefore, we have focused on anaphora 
resolution of zero subjects in Japanese manual sen- 
tences. Mori et al.(Mori and Nakagawa, 1996) show 
that properties of Japanese conditionals can be used 
to resolve them. In this paper, we propose new con- 
straints and defaults based on properties of linguis- 
tic expressions, which are useful to estimate omitted 
subjects in addition to the constraints and defaults 
proposed by Mori et al. 
A large number of researchers have come to grip 
with the method of understanding some types of text 
including instruction manuals(Abe et al., 1988; No- 
mura, 1992; Eugenio, 1992). One of the most im- 
portant matters of concern in these types of sys- 
tem is how we can resolve ambiguities in seman- 
tic representations and fill underspecified parts of 
them. Generally speaking, almost all systems de- 
scribed above take the following scheme. Firstly, 
each sentence in a text is translated into a semantic 
representation. In this process, the system uses only 
non-defeasible syntactic and semantic constraints. 
This way of analysis is known as the Nondefeasibility Thesis(Kameyama, 
1995). Secondly, all of undeter- 
mined parts of the semantic representation are filled 
or settled by some kind of inferences based on the 
domain knowledge. 
This type of method, which uses a large amount of 
domain knowledge, seems to be dominant from the 
viewpoint of disambiguation. Moreover it scarcely 
depends on the language in use because the way of 
disambiguation is based on the inference with a cer- 
tain knowledge base. On the other hand, in order to 
use this method, we have to prepare the amount of 
knowledge being large enough to cope with various 
types of described objects. Unfortunately, so far we 
have not had such a commonsense knowledge base. 
One of the ways to get rid of this situation is to 
adopt some knowledge which is independent of any 
particular domain. As such a kind of knowledge, we 
pay attention to pragmatic constraints, which have 
not been used sufficiently in the former methods. 
We expect that owing to pragmatic constraints the 
ambiguity in manual sentences would be resolved to 
some extent not in the process of inference but in 
the process of the translation of manual sentences 
into semantic representations. 
We do not commit ourselves to the domain specific 
knowledge, but use some ontological knowledge of 
ordinary manuals. For example, the correspondence 
of objects in the manual sentences to the objects 
in linguistic constraints, namely linguistic roles like 
the speaker, the hearer, and so on. Note that the 
ontology in this paper does not refer to all of the 
objects in the world described by manuals, like a 
certain part of machine. Aiming at independence 
from the domain knowledge of objects, we adopt one 
of general ontologies which is applicable to almost all 
manuals. 
Now we have to define the term 'SUBJECT' we 
used in this paper. Since our final goal is the deter- 
mination of "the main participant" which is omit- 
ted, both of the term 'subject' and the term 'agent' 
are not suitable for referring to the omitted objects. 
For example, in a sentence in passive voice, the sub- 
ject corresponds to not the agent(namely the main 
participant), but the patient. Moreover, there are 
several types of sentences whose subjects are main 
participants even if they are not agents, like the de- 
scription of states, attributes and so on. Therefore, 
we use the term 'SUBJECT' to denote the main par- 
ticipant of the sentence, namely ether the agent or 
the surface subject(in the case where the agent is 
not defined). 
2 Zero pronouns in manual 
sentences 
Let's consider the following Japanese sentence, 
which shows a certain instruction. 
(1) ¢a kono-botan-o osu -to, 
Ca-NOM this-button-Acc push -TO, 
¢b der -are -mas -u. 
Cb-NOM go out -can -POL -NONPAST. 
If ¢a push(es) this button, then ¢b can go out. 
Native speakers of Japanese have the following intu- 
itive interpretation for (1) without any special con- 
text. 
(2) ¢~ = ¢b = the hearer (= the user) 
Here, 'TO' is a Japanese conjunctive particle which 
represents a causal relation, and 'ARE' shows ability 
or permission. The symbol ¢ denotes a zero pro- 
noun. 
On the other hand, the following sentence, which 
does not have the suffix 'ARE', has a different inter- 
pretation. 
(3) ¢~ kono-botan-o osu -to, 
¢c-NOM this-button-AcC push -TO, 
¢d de -mas -u. 
Cd-NOM come out -POL -NONPAST. 1 
If ¢c push(es) this button, then ¢d will come 
out. 
The zero pronoun ¢d refers to not the hearer(the 
user) but the machine, even though ¢¢ refers to the 
1The English translation of 'DERU' in (3) is different 
from the translation in (1). It is due to the difference 
of the viewpoint between Japanese and English. The 
difference has no effect on the selection of zero pronoun's 
referent. 
user as well as (1). Note that when only the ma- 
trix clause of (3) is used as shown in (4), ¢~ can be 
interpreted as either the hearer or the machine 2. 
(4) ¢e de -mas-u. 
Ce-NOM go out -POL -NONPAST. 
¢~ will go out. 
These examples show that the expressions TO and 
ARE impose some constraints on the referents of StJB- 
JECIS of the sentences. As described so far, there 
are many cases that linguistic expressions give us key 
information to resolve some type of ambiguity like 
the anaphora of zero pronouns. In the rest of this 
paper, we will show several pragmatic constraints, 
which can account for the interpretations of zero 
subjects including the cases described above. 
Dohsaka(Dohsaka, 1994) proposes a similar ap- 
proach, in which several pragmatic constraints are 
used to determine referents of zero pronouns. For ex- 
ample, honorific expressions and the speaker's point 
of view are used in his approach. While his approach 
treats dialogue, our targets are manual sentences. 
Nakaiwa et.al.(Nakaiwa and Shirai, 1996) also 
propose the method which is based on semantic and 
pragmatic constraints. Although they report that 
their method estimates over 90% of zero subjects 
correctly, there are several difficulties including the 
fact that the test corpus is identical with the corpus 
from which the pragmatic constraints are extracted, 
and the fact that there are so many rules(46 rules to 
estimate 175 sentences). 
As for the identifying method available in gen- 
eral discourses, the centering theory(Brennan et al., 
1987; Walker et al., 1990) and the property shar- 
ing theory(Kameyama, 1988) are proposed. The im- 
portant feature of these theories is the fact that it 
is independent of' the type of discourse. However, 
according to our experimental result, it seems that 
these kinds of theory do not estimate zero subjects in 
high precision for manual sentences 3. The linguistic 
constraints specific to expressions are more accurate 
than theirs if the constraints are applicable. 
3 Hypothesis and general ontology 
of manuals 
Kaiho et al.(Kaiho et al., 1987) explain the basic 
function of instruction manuals as follows: 
* A manual is an interface between humans and 
machines based on language information. 
2It seems to be more natural that ¢¢ is interpreted as 
the hearer. 
3The result precision of centering theory is 60% to 
70% in our experiment. One reason why the precision is 
not so good is that the structure of texts in (Japanese) 
manuals is slightly different from the ordinary discourses 
structure. 
• Since the essential function of manuals is to pro- 
vide users with information to make the ma- 
chine operate properly, the existence of users 
should be considered at all times. 
• Manuals should appropriately provide informa- 
tion which is required by users. 
On the other hand, the following tendency are 
pointed out in many linguistic literatures. 
• The readers have the same point of view as the 
writer. 
• Generally, the first candidate of the point of 
view is the nominative. 
According to these considerations, we make the 
following hypothesis: 
Hypothesis 1 (Manuals easy to understand) 
• All descriptions are written from the viewpoint 
of users. Therefore, in general, subjects in man- 
ual sentences tend to be users. 
• Things users know, tend to be omitted for read- 
ability unless they are needed. Therefore, the 
subject of the sentence whose agent is a user 
tends to be omitted. 
• On the other hand, things which readers do 
not known, like reactions of operations, prompt 
from machines and so on, tend to be specified 
explicitly. 
As the parts of ontology, we should consider, at 
least, two types of information: the properties of 
the objects in manuals and the discourse situation 
that is characterized by linguistic roles like a writer 
and a reader. 
Constraint 1 (Objects) 
User has intention. 
Manufacturer has intention. 
Machine has no intention. 
Constraint 2 (Discourse Situation) 
Speaker(Writer) = Manufacturer 
Hearer(Reader) = User 
From these constraints of the ontology, we can ob- 
tain the constraint of persons as follows. 
Constraint 3 (Persons) 
First Person = Manufacturer 
Second Person = User 
Third Person = Machine 
In the rest of this paper, we will propose several 
constraints and defaults based on the property of 
linguistic expression under the hypothesis and the 
constraints described above. Then, we will exam- 
ine them with test examples from several manual 
sentences. Note that the constraints and defaults 
we propose here are derived not from some spe- 
cific manuals but from our linguistic consideration 
for each of linguistic expressions. Therefore, we do 
not adopt strict validation method like 'cross valida- 
tions', which is used in machine learning, to examine 
them. However, in order to confirm the validity of 
our constraints and defaults, we have checked them 
out with 24 manuals from various areas. Although 
we cannot explain all of our defaults and constraints 
here because of shortage of space, we will briefly 
show the table of our all defaults and constraints in 
Section 6. 
4 Constraints and defaults based on 
the type of verbs 
4.1 Request form 
The speaker uses the sentences in the request form 
or the solicitation form to prompt hearers to do the 
action described by the sentence. Therefore, 
Constraint 4 (SUBJECT of sentence in the re- 
quest form) 
A SUBJECT of a sentence in either the request form 
or the solicitation form is the hearer. 
The combination of this constraint and Constraint 
3 (Persons) shows that the SUBJECT is the user in 
such a case. In example manuals, there are 123 sen- 
tences in the request form and all of them satisfy 
Constraint 4. 
4.2 Modality expressions 
Manual sentences may have a kind of modality ex- 
pressing permission, ability, obligation, and so on. 
Sentences which have the expressions of ability or 
permission mean not only that it is possible for the 
SUBJECT to do the action, but also that the SUB- 
JECT has his/her own choice of whether to do the 
action or not. Therefore, 
Constraint 5 (SUBJECT of sentence with abil- 
ity expressions) 
A SUBJECT of a sentence with the expressions of 
ability or permission must have his/her intention to 
make a choice about the action described by the sen- 
tence. 
This constraint and Constraint 1 (Objects) show 
that a SUBJECT of a sentence with the expressions 
of ability or permission is a user, because all of the 
actions of manufacturer have been finished when the 
user is reading the manual. In example manuals, 
there are 56 sentences with the ability expressions 
and all of them satisfy Constraint 5. 
4.3 RU form 
In Japanese, simple operation procedures are often 
described as simple sentences with no subjects whose 
verbs are of one of the following types: the RU form, 
the request form or the solicitation form. The RU 
form is the basic form of verbs and it denotes the 
non-past tense. Since the RU form has a neutral 
meaning, it does not impose any restriction on the 
SUBJECT. However, with Hypothesis 1 we expect 
that the zero subject tends to be a user. 
Default 1 (SUBJECT of sentence with a verb in 
the RU form) 
A SUBJECT of a sentence with a verb in the RU form 
is a user. 
In example manuals, there are 214 sentences with a 
verb in the RU form and with no subject, and the 
SUBJECTS of 172 sentences are users. Therefore, the 
precision of the default is about 80.4%. 
4.4 Intransitives 
In almost all cases of machines which come with 
instruction manuals, their actions are initiated by 
some activities of users. The activities are repre- 
sented not by intransitives but by transitives. There- 
fore, we expect that a SUBJECT of a sentence with 
an intransitive tends to be a machine. 
Default 2 (SUBJECT of sentence with an in- 
transitive) 
A SUBJECT of a sentence with an intransitive is a 
machine. 
In example manuals, there are 238 sentences with 
intransitves, and the SUBJECTS of 211 sentences are 
machines. Therefore, the precision of the default is 
about 88.7%. 
4.5 Passives 
The passivization is the transfer of the viewpoint of 
the speaker from the nominative to the objective by 
exchanging their positions. Namely, the passiviza- 
tion is used to bring the objective in the active voice 
to readers' attention, when SUBJECT is not so im- 
portant for readers. Since readers, or users, do not 
have to know what SUBJECT is, it is hard for a SUB- 
JECT of a sentence in passive voice to be a user. 
Default 3 (SUBJECT of passives) 
A SUBJECT of a passive is a machine. 
In example manuals, there are 48 passives and the 
SUBJECTS of 46 sentences are machines. Therefore, 
the precision of the default rule is about 95.8%. 
4.6 Causatives 
Since a causative expresses an event that the SUB- 
JECT of the causative makes someone(or something) 
do some action, the SUBJECT should have some in- 
tention and the initiative in controlling someone's 
action. Since a user has the initiative, we propose 
the following default. 
Default 4 (SUBJECT of causatives) 
A SUBJECT of a causative is a user. 
In example manuals, there are 38 passives and the 
SUBJECTS of 36 sentences are machines. Therefore, 
the precision of the default rule is about 94.7%. 
4.7 Expressions with the suffix -DESU 
Expressions with the suffix -DESU are divided into 
two groups: 
• noun + the suffix of copula 
• Adjective verb 
Each of them expresses that a SUBJECT has some 
property. Since it is unusual to describe user's prop- 
erty in manuals. Therefore, 
Default 5 (SUBJECT of sentence with the suffix 
-DESU) 
A SUBJECT of a sentence with the suffix-DESU is 
a machine. 
In example manuals, there are 25 sentences with the 
expression, and all SUBJECT's of them are machines. 
5 Constraints and Defaults based on 
types of Connectives 
5.1 Conditionals 
Japanese has four conditional particles, TO, REBA, 
TARA and NARA, which are attached to the end of 
subordinate clauses as described in (1). The sub- 
ordinate clause and the matrix clause conjoined by 
one of these particles correspond to the antecedent 
and the consequence, respectively. The difference 
of constraints of these expressions are shown in the 
following sentences, which are the variants of the 
sentence (3). 
(5) ¢i kono-botan-o use -ba 
¢i-NOM this-button-Acc push -REBA, 
Cj de -mas -u. 
Cj-NOM come out -POE -NONPAST. 
If ¢i push(es) this button, then Cj will come 
out. 
(6) Ck kono-botan-o osi -tara, 
Ck-NOM this-button-Acc push -TARA, 
el de -mas -u. 
¢/-NOM come out/go out -POL -NONPAST. 
If Ck push(es) this button, then Cz will come 
out/go out. 
(7) Cm kono-botan-o osu -nara, 
Crn-NOM this-button-Acc push -NARA, 
Cn de -mas -u. 
~n-NOM come out/go out -POL -NONPAST. 
If Cm push(es) this button, then Cn will come 
out/go out. 
As well as the sentence (3), for Japanese native 
speakers, the SUBJECT of the matrix clause of (5) 
should be a machine. On the other hand, in the 
case of the sentences (6) and (7), the SUBJECTS of 
the matrix clauses can be either users or machines. 
These phenomena probably due to the nature of each 
conditionals(Masuoka, 1993). Since a causal rela- 
tion, which is shown by TO or REBA, expresses a gen- 
eral rule, the consequence cannot include speaker's 
10 
attitude, like volition and request. Therefore, the 
SUBJECT of the matrix clause should be a machine. 
In contrast, in the case of assumptions, that is TARA 
and NARA, there are no such restrictions on the SUB- 
JECT . 
Based oil these observation, Mort et al. (Mort and 
Nakagawa, 1995; Mort and Nakagawa, 1996) pro- 
pose the defaults of SUBJECTS of sentences with 
these conditionals. Since it depends on the voli- 
tionality of the verb whether a sentence shows a 
speaker's attitude or not, the constraint and defaults 
are described in terms of volitionality of each verb. 
Note that the electronic dictionary IPAL provides 
the information of volitionality for each Japanese 
verb entry(IPA Technology center, 1987). Accord- 
ing to the classification by IPAL, all of Japanese 
verbs are classified into two types, volitional verbs, 
which usually express intentional actions, and non- 
volitional verbs, which express non-intentional ac- 
tions. Although non-volitional verbs only express 
non-volitional actions(non-volitional use), some of 
volitional verbs have not only volitional use but also 
non-volitional use. 
Default 6 (SUBJECT of sentence with TO or REBA) 
The matrix clause does not express user's volitional 
action. Therefore, the SUBJECT of the matrix clause 
is a machine, if the verb of the matrix clause does 
not have the non-volitional use. 
Default 7 (SUBJECT of sentence with TARA or 
NARA) 
The matrix clause expresses only user's volitional 
action. Therefore, the SUBJECT of the matrix clause 
is a user. 
The precision of the default rules of TO,REBA,TARA 
and NARA is 100%, 95.1%, 89.8% and 100%, respec- 
tively. 
5.2 Adverbial conjunctive forms 
Japanese verbs have two major adverbial conjunc- 
tive forms: '-TE form' and 'adverbial form.' Roughly 
speaking, a clause with a verb in one of these forms 
is placed in front of another clause and they con- 
struct a coordinate relation. The following example 
shows the coordination of-TE form. 
(8) ¢o botan-o oshi-te, 
¢o-NOM button-Acc push-TE, 
Cp Cq toridasi -mas -u. 
Cp-NOM Cq-ACC take-out -POL -NONPAST. 
¢o pushes the button and Cp takes out Cq. 
According to Teramura(Teramura, 1991), essentially 
these forms of verbs express the coordination and 
cooccurrence of two events. For example, tile most 
plausible interpretation of (8) is that ¢o and Cp are 
identical. Thus it is expected that two SUBJECTS 
of two clause in the coordination are identical or of 
the same type. Especially in manuals, the writer 
does not describe user's actions in the same treat- 
ment as machine's action, because the writer takes 
the viewpoint of users as supposed in Hypothesis 1. 
Therefore, 
Default 8 (Two SUBJECTS of clauses in TE 
form conjunction or adverbial form conjunc- 
tion) 
Two SUBJECTS of two clauses are identical when 
the two clauses are connected by the TE form con- 
junction or the adverbial form conjunction. 
In example manuals, there are 83 sentences with TE 
form conjunction and 75 sentences meets the de- 
fault. Thus the precision of the default for sentences 
with TE form conjunction is about 90.4%. Similarly, 
there are 99 sentences with adverbial form conjunc- 
tion and 98 sentences complies the default. The pre- 
cision of the default for adverbial form conjunction 
is about 99.0%. Moreover, in the majority of the 
cases of TE form conjunctions, SUBJECT is a user 
(85 cases). Therefore we revise the default for ad- 
verbial form conjunction as follows. 
Default 9 (Two SUBJECTS of clauses in TE 
form conjunction) 
Each SUBJECTS of two clauses is a user when the 
clauses are connected by the TE form conjunction. 
6 Results 
The constraints and defaults we proposed here and 
their precision are summarized as Table 1. 
Note that this table shows that each default rule is 
strong enough for anaphora resolution of zero SUB- 
JECT if it is applicable. Therefore, we have to ex- 
amine the total performance of our method, that is, 
have to verify what percentage of zero SUBJECTS in 
manuals are correctly determined with our defaults. 
In order for the verification, we examined the esti- 
mate by our defaults with 9 test example manuals, 
which contain 740 sentences. Table 2 shows the re- 
sult of the verification. Here, the term 'Restricted' 
shows that the candidate of SUBJECT is correctly 
restricted by Default 8. 
As the table shows, our defaults successfully deter- 
mine zero SUBJECTS in manuals with the precision 
of over 80%. It is remarkable that the rate of wrong 
judgment is only 2.5%. Since almost all zero SUB- 
JECTS which cannot be resolved by our defaults are 
still undetermined, it is possible to improve the pre- 
cision by adding new defaults/constraint, combining 
other methods of zero pronoun resolution, and so on. 
7 Conclusion 
In this paper, we proposed a scheme which closely 
depends not on domain knowledge of objects de- 
scribed in manuals but on pragmatic constraints 
which linguistic expressions innately have. This 
method uses only the linguistic constraints and the 
general ontology of the world described by manuals. 
11 
Expression 
Table 1: Constraints and Defaults 
# of occurrences Estimation of SUBJ 
Request 123 User 
56 
385 
225 
Ability 
TO 
REBA 
TARA 59 
NARA 9 
User 
Machine(Matrix) 
Machine(Matrix) 
User(Matrix) 
User(Matrix) 
Simple 214 User Non-past 
Intransitive 238 Machine 
Passive 48 Machine 
Causative 38 User 
'Want to' 4 User 
Copura 25 Machine 
'Automat- ically' 20 Machine 
TE-Conn. 99 User 
Adverbial-Conn. 83 Identical SO BJs 
Precision 
100% 
100% 
100% 
95.1% 
89.8% 
100% 
80.4% 
88.7% 
95.8% 
94.7% 
100% 
100% 
100% 
85.9% 
90.4% 
Table 2: Examination of our method 
Judgment by human 
for zero subjects Correct (Subject is fixed) 
Judgment by our method 
Correct 
(Subject is 'restricted') 
User 692 548 24 
Machine 262 221 2 
Manufacturer 2 0 I 0 
80.6% I 2.7% 
Not 
Wrong applicable 
6 114 
18 21 
0 2 
2.5% I 14.2% 
* Each figure shows the number of sentences. 
12 
We have shown that we can determine the referents 
of zero pronouns to some extent with our linguistic 
constraints and defaults. However, we do not have 
enough knowledge about the following points. They 
are important portions of our future work. 
• Utilization of discourse structure specific to 
manuals. 
• Analysis for the other types of manual sen- 
tences, like definitions. 

References 
Norihiro Abe, Tomohiro Ishikawa, and Tsuji Saburo. 
1988. Generating an assembly procedure from as- 
sembly instruction. Journal of Japanese Society 
for Artificial Intelligence, 3(5):590-598, Septem- 
ber. (in Japanese). 
Susan E. Brennan, Marilyn Walker Friedman, and 
Carl J. Pollard. 1987. A centering approach 
to pronouns. In Proceedings of the 25th Annual 
Meeting of the American Association of Compu- 
tational Linguistics (A CL 87), pages 155-162. 
Kohji Dohsaka. 1994. Identifying the referents of 
Japanese zero pronouns based on pragmatic con- 
dition interpretation. Transactions of Informa- 
tion Processing Society of Japan, 35(5):768-778. 
(in Japanese). 
Barbara Di Eugenio. 1992. Understanding natural 
language instructions: The case of purpose clause. 
In Proceedings of 30th Annual Meeting of ACL, 
pages 120-127. 
IPA Technology center, 1987. The lexicon 
of the Japanese basic verbs for Computers. 
Information-technology Promotion Agency(IPA), 
Japan, March. (in Japanese). 
Hiroyuki Kaiho, Takashi Kato, Keizo Hori, and Et- 
suko Harada. 1987. User/Yomite-no kokoro-o 
tsukamu manual-no kaki-kata (How to write in- 
struction manuals). Kyoritsu Pub., Tokyo. (in 
Japanese). 
Megumi Kameyama. 1988. Japanese zero pronomi- 
nal binding: Where syntax and discourse meet. In 
W. Poser, editor, Japanese Syntax, pages 351-369. 
CSLI,Stanford. 
Megumi Kameyama. 1995. The syntax and se- 
mantics of the Japanese language engine. In 
R. Mazuka and N. Nagai, editors, Japanese Syn- 
tactic Processing. Lawrence Erlbaum Associates. 
Takashi Masuoka, editor. 1993. Nihongo-no Joken- 
Hyogen (Conditional Expressions in Japanese). 
Kurosio Publishers, Tokyo. (in Japanese). 
Tatsunori Mori and Hiroshi Nakagawa. 1995. A 
pragmatic approach to zero pronoun resolution 
in Japanese manual sentences -- the case of 
Japanese conditionals --. In Proceedings of Nat- 
ural Language Processing Pacific Rim Symposium 
'95 (NLPRS'95), pages 296-301, December. 
Tatsunori Mori and Hiroshi Nakagawa. 1996. Zero 
pronouns and conditionals in Japanese instruction 
manuals. In Proceedings of the 16th International 
Conference on Computational Linguistics (COL- 
ING 96), pages 782-787, August. 
Hiromi Nakaiwa and Satoshi Shirai. 1996. 
Anaphora resolution of Japanese zero pronouns 
with deictie reference. In Proceedings of the 16th 
International Conference on Computational Lin- 
guistics (COLING 96), pages 812-817, August. 
Hirosato Nomura. 1992. Linguistic analysis of law 
sentences. In Proceedings of the 6th International 
Symposium: Legal Knowledge and Legal Reason- 
ing Systems, pages 10-18. Legal Expert System 
Association, October. 
Hideo Teramura. 1991. Nihongo-no sintakusn-to 
imi III (Syntax and Semantics of Japanese III). 
Kuroshio shuppan Pub. (in Japanese). 
Marilyn Walker, Masayo Iida, and Sharon Cote. 
1990. Centering in Japanese discourse. In Pro- 
ceedings of COLING 90. 
