SYNTACTIC CONSTRAINTS ON RELATIVIZATION 
IN JAPANESE 
SIRAI, HIDETOSI 
School of Computer and Cognitive Sciences 
Chukyo University 
101 Tokodute, Kaizu-cho, Toyota, Aichi, Japan 470-03 
TOMIOKA, YUTAKA 
Faculty of Engineering 
University of Tokyo 
7-3-1 Hongo, Bunkyo-ku, Tokyo, Japan 113 
Abstract 
This paper discusses the formalization of relative 
clauses in Japanese based on JPSG framework. We 
characterize them as adjuncts to nouns, and formal- 
ize them in terms of constraints among grammatical 
features. Furthermore, we claim that there is a con- 
straint on the number of slash elements and show the 
supporting facts. 
1 Introduction 
This paper discusses the formalization of relative 
clauses in Japanese based on JPSG (Japanese Phrase 
Structure Grammar)\[I, 2\], which is a constraint- 
based grammar formalism like tlPSG(Head-driven 
Phrase Structure Grammar)\[7, 8\]. We have worked 
on JPSG with Prof. Gunji, and also have developed 
a parser based on an efficient mechanism for dealing 
with constraints\[3, 4, 10\] to show that JPSG is effec- 
tive even for the computational processing of natural 
language. 
In the next section, we briefly introduce JPSG the- 
ory. Following a simple characterization of relative 
clauses in Japanese language in section 3, we disscuss 
the variety of acceptability in secton 4, and describe 
its formalization in terms of constraints among the 
grammatical features in section 5. And in section 6 
we will claim that there is a constraint on tile number 
of slash elements and show tile supporting facts. 
2 Basics of JPSG 
This section describes a brief introduction to JPSG, 
which is a grammar formalism originally for the 
Japanese language. As with IIPSG, JPSG is feature 
based and constraint based grammar. 
2.1 Features 
Grammatical categories are represented as sets of fea- 
tures. We list the features used in this paper. 
(1) Features used in this paper 
pos (part of speech) same ms in IIPSG. 
gr (grammatical relation) takes either subj, obj, or 
iobj as the value. 
subcat (subcategorization) designates the set of cat- 
egories (complements) that a particular category 
(head) requires. Though we have to distingush 
two types of complements (is., agglutinated or 
not) in Japanese, for simplicity, we assume that 
subeat designates tile both types of comple- 
ments. 
dep (dependent) designates the category that a par- 
ticular category (adjunct) modifies. 
core roughly corresponds to CAT feature in tlPSG 
\[8\]. Tim value is a set of features including sub- 
cat feature and so-called head features such as 
pos, gr and dep. 
slash designates a syntactic gap within the gram- 
matical category involved. 
sere (semantics) designates the semantic representa- 
tion of the grammatical category involved. 
In this paper, categories are designated by a left 
square bracket ("\[') followed by an indefinite num- 
ber of feature specifications (a feature name followed 
by its value) separated by commas(",") followed by 
a right square bracket ("1"). When the value is null 
or not relevant, the entire feature specification can be 
omitted. The sharing structure is indicated by vari- 
ables such as X, Y,..., which is distinguished by an 
initial capital letter as in programming language Pro- 
log. Finally, a category of the form \[core {pos c, 
... } ..... sere s\] is often abbreviated as el...\]:s. 
Because grammatical relations play no role in or- 
dering complements in Japanese, we assume only a 
binary-branching phrase structure schema as seeu in 
the next subsection. 
2.2 The Phrase Structure Schema and Gram- 
matlcal Principles 
We are assuming three basic phrase structure schema 
for Japanese: complementation, adjunclion, and co- 
ACTES DE COLING.92, NANTES. 23-28 AO~'r 1992 1 1 0 0 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
ordination: 
(2) a. G'omplementation M--~ C If 
b. Adjure(ion M --* A tl 
c. Coordination M ~ tit II2 
where M stands for an arbitrary mother category, C 
a complement, lI a tread category, and A an adjunct. 
Each category is construed as complex symbols, or 
features, with internal structures. The above struc- 
tures are uniquely characterized by the features. For 
example, the head in the complements(ion structure 
should have subeat feature one of whose value is uni- 
tied with the complement, C. 
Furthermore, we assume the following grammarti- 
cal principles, which are applied to every structures: 
Head Feature Principle: same as in III)SG. We 
assume that pos, dep, and other several features 
are head features. 
Subcat Feature Principle: In the complementa- 
tion structure shcema (2b), the value of subeat 
of the mother category unifies with that obtained 
by subtracting the left daughter category from 
the value of subcat of the head category. In 
the other strucrues, the value of subcat of the 
mother category unifies with that one of the head 
category. 
Binding Feature Principle: The value of a bin(l- 
ing fcaturc of the nlother category unifies with 
the ration of its value of the left daughter cat- 
egory and its value of the head category minus 
the category bound of this local branching t. We 
assume that slash anti some other features are 
binding features. 
Slash feature aud Binding Feature Principle play 
a crucial role in describing so-called long distance de- 
pendency phenomena. We will discuss them in detail 
in the following sections. 
3 Relative Clauses in Japanese 
Traditionally the relative clause in Japanese has been 
considered as a kind of so-called ren-1ai syusyoku setu 
'the clause for modifying noun'. Because it is char- 
aeterized as an embedded sentence for modifying the 
head noun in the structure \[s ...\]N, it can be consid- 
ered ~.s an adjunct to the head noun. 
In this paper, we adopt this idea. Thus, the struc- 
ture of relative clause-head noun construction can 
be described as shown in Figure 1, where the left 
daughter category is a relative clause, and the right 
daughter category is a head noun. The fact that the 
left daughter category has dep feature whose value 
l For example, when the left daughter and the head daughter 
have { A } and { B } as their slash feature value respectively, 
the value of slash of the mother will he { A, B } in most cases. 
llowever, in cane that A can be unifiable with B, the result can 
be { A ). Furthermore, in case that A (or B I is bound in the 
local brnaaching stnmture (as in Figure 4), the result will be { 
n}(o~{A }). 
\[ core tl \] 
where II = {pos n} 
Figure 1: Relative Clause as Adjunct 
i1\] 
is unifial)le with the right daughter category, charac- 
terizes the left daughter (is. relative clause) as all 
adjunct. 
tlowever, we have to explain where this dep fea- 
ture comes from. In Japanse, a relative clause ha.~ 
a tense form at its end. Thus we assume that tense 
form is categorized as a verb which takes a verb cat- 
egory as its coinplement, l"urthermore, we assume 
that for every tense form (e.g. -ta and -ru), tbere are 
two different lexieal entries With the same phouetic 
form, which we call the preaomiual tense form and 
the assertive tense form 2. Tile former type is as- 
sumed to be the ending form of relative clauses, and 
has a non-mall value in dep feature. This value is 
bound with the head noun of the relative clause a 
4 Types of Relative Clauses 
Relative clauses can be classilled into the following 
three types in terms of their structural characteristics 
(where ¢ marks a gap): 
1. SS: Simple Sentence 
The relative clanse consists of a simple sentence, 
and the gap plays a role in that sentence, e.g. 
(3) \[ Naomi-ga 5 yore -da\] hon 
NOM ACC read PAST book 
'tile book which Naomi read' 
2. ES: Embedded Sentence 
The relative clause includes an embedded sen- 
tence in which the gap plays a role, e.g. 
(4) \[Naomi-ga \[Taro-ga ¢ yom-da\] 
NOM NOM ACC read PAST 
-to sinzitei -ru \] hou 
CMPL believe PRES book 
'the book which Naomi believes Taro read' 
3. RC: Relative Clause 
The relative clause inchnles a relative chmse in 
which the gap plays a role, e.g. 
(5) \[\[¢ ¢ kak -ta\]l ..... ga 
NOM ACC write PAST book NOM 
yoku ure 4a\] sakka 
well sell I'AST writer 
'the writer whose book sold well' 
2There are several l'ea.qon~ wily we Call /L~StUlle there are two 
different lexical el(tries for tense with the same phonetic form 
{91, 3The assertive tense form is assumed to be the ending form 
of assertive clauses, and has a null value in dep feature. 
AcrEs DE COLING-92, NAN'IT!S, 23-28 AO~I' 1992 1 l 0 1 PROC. OF COLING-92, NAI~rES, AUG. 23-28, 1992 
Table 1: Acceptance of Relative Clause 
Grammatical Pole SS ES RC 
Subject o o o 
Object o o ? 
Adjunct 1 o o ? 
Adjunct 2 * * * 
Table 2: Simplification of Judgement 
Grammatical Role SS ES RC 
Subject o o o 
Object o o o 
Adjunct 1 o o o 
Adj.unct 2 * * * 
As lnoue\[5, 6\] pointed out, the acceptability of the 
relative clauses varies depending on their types shown 
above and what roles are played by tbe gaps. 
The grammatical roles other than subject and ob- 
ject can be classified into the following two types de- 
pending on the acceptability of relative clauses sbown 
in Tablel, where, o, ? and * represent our subjects' 
judgements are acceptable, marginal, and unaccept- 
able, respectively. : 
Adjunct 1 : This includes 'locative'(marked by hi, 
we and de), 'goal'(marked by ni and he) and 
'source'(marked by kara). 
Adjunct 2 : This includes 'instrumental'(marked 
by de), 'reason'(marked by kara) and 'compar- 
ative'(marked by yori). 
In order to simplify the formalization, we will re- 
gard the marginal cases as acceptable, and the re- 
sult is shown in Table 2. This simplification, how- 
ever, may be too permissive, because usually in the 
RC clauses are acceptable only when tile gap's role is 
subject. 
5 Formalization 
In this section we will formalize the syntactic struc- 
ture of the relative clause in Japanese. First we will 
formalize the SS type relative clause. Then we formal- 
ize the distribution of the slash feature. And finally, 
we examine the other eases, that is, those in which 
the gap is derived in either an embedded sentence or 
a relative clause. 
5.1 Simple Relative Clause 
As discussed in section 3, a relative clause is regarded 
as an adjunct. Thus, the gap should be bound with 
tile dep feature value of the prenominal tense cate- 
gory. 
As shown in the last section, gaps are not restricted 
to subjects or objects. Optional Adjunct 1 categories 
('locative','goal' and 'source') can be a gap. Thus, for 
simplicity, we here assume that Adjunct 1 type roles 
are also designated in the value of subcat feature, 
core subcat } 
dov {hi I:X} 
\[ /' .... }l core subcat { \[core el} 
dep {n\[ \]:X} 
whereC= { pos v } 
subeat {VII:X} oY 
Figure 2: Binding of Dep with Subcat 
Furthermore, we assume that a gap is directly 
bound with dop feature, instead of being bomtd with 
an element of slash which is ill turn bound with dep. 
We call t~his hypothesis the Direct Binding Hypothe- 
sis (that is, subcat element can be directly bound 
with dep). We will discuss tile correctness of this 
assumption in section 6. 
This relationship is expressed ill tile lexical entry 
for pronominal tense category, eg. ~ta, -ru, and -i. 
The structure is shown ill Figure 2, where tile head 
daughter is a pronominal tense category. For exam- 
pie, kak 'write' is assumed to take bottt subject and 
object complements and have the following lexical en- 
try as follows: 
(6) v\[eore {subcat {p\[gr subj\]:X, p\[gr obj\]:Y}}\] 
Then, a relative clause kak-ta 'wrote' corresponds to 
the following categories: 
(7) a. v\[eore {subcat { }, dep { n\[g .... bjl }}\] 
b. v\[eore {subcat { }, dep { n\[gr obj\] }}\] 
Thus for kak.ta X (where X is a noun), there are at 
least two possibilities in the interpretation of the role 
of X, that is, either subject or object. 
5.2 Distribution of Slash 
As sentence (5) shows, there is a so-called long- 
distance dependency phenomenon in Japanese. Tiffs 
phenomenon is described by using the slash feature 
as in tlPSG. We will characterize the distribution of 
slash in tile following three sections. 
5.2.1 SLASII INTRODUCTION 
Lexical entries with non-null value in slash feature 
are introduced by tile following lexical rnle: 4 
Subeat-Slash Lexical Rule: 
\[subcat { p\[\]:X } US\] 
\[subeat S, slash { n\[l:X }\] 
Tbe number of slash elements introduced by this rule 
can be at most onc, a fact which we explaiu more 
precisely in the next section. 
4There are some re~trlctions on the application of tiffs rule. 
For exmnple, an agghttinated complement emmet become a 
slash value. 
ACRES DE COLING-92, NANTES, 23-28 Aour 1992 I 1 0 2 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
slash 
v dep 
kak -ta 
write PAST 
For example, kak 'write' will have tile following lex- 
ieal entries other than (6) by this lexcial rule: 
(S) a. v\[core {subcat {p\[gr subj\]:X}}, 
slash {n\[gr obj\]:Y}\] 
b. v\[eore {uubcat {p\[gr objl:Y}}, 
slash {u\[gr subj\]:X}\] 
5.2.2 SLASH INnEaITANCE 
Each slash inheritance is constrained by means of 
Binding Feature Priuciple. lIere, again, we assume 
that the number of slash elements of each category 
carl be at most one. 
In the RC type relative clause, the gap is inherited 
from the innermost embedded clause to the outermost 
relative clause shown in Figure :L 
v\[slash X\] 
p\[slash X\] v\[subcat {p\[ \]}\] 
n\[slash X\] p\[subcat {n\[ \]} 
{n\[\]} \] / NOM sell well 
holt 
book 
Figure 3: Exanrple of Slash Inheritance 
5.2.3 SLASH BINDING 
A long-distance inherited gap can be bound with tile 
dep of tile prenominal tense category, and finally 
bound with the head nmm of the relative clause as 
shown in Figure 4 (where only core features of the 
mother and the head are shown). For example, if 
the left daughter category corresponds to kak-ta hou- 
ga yoku ure 'hook which ¢ wrote sells well'), anti the 
right daughter (i.e. the bead) to the prcnomiaal tense 
form -to, then the mother, kak-ta hon-ga yoku ure-ta, 
will he described as follows: 
v\[core {dep {n\[\] }}\] 
It should be noted that we do not necessarily need 
a specific phrase structure rule for the slash binding 
with tile dep \[9\], though it may be uatural that we 
assume to have such a phrase structure rule as shown 
in Figure 4. 
5.3 Complex Relative Clause 
In both ES and RC type relative clauses, the Bind- 
ing Feature Principle specifies that a slash element 
which is introduced by the Subcat-Slash Lexical Rule 
is iuherited and finally bound with the dep value of 
the prenominal tense category in the ontermost rel- 
ative clause. Figure 5 shows an example of such a 
structure. 
\[.os \] 
subeat } 
dep { X } 
slash { X } subcat \[core C\] 
dep { X } 
Figure 4: Slasil Binding 
n\[\] 
v\[ .... {aep {x}}\] nil 
v\[slash {X}\] v\[ .... { dep {X} }\] hito 
kak-ta hon-ga yoku ure -ta 
where X = n\[\] 
Figure 5: Complex l~lative Clause 
6 Constraints oil tile Number of Slash 
Elenaents 
As the consequence of tile constraints on slash shown 
above, an Adjunct 2 type role cannot construct ally 
acceptable rclatlvization, because it cannot introduce 
all clement of subcat ill ally lexical entries. 
As discussed ill section 3, we assume tile relative 
clause of Japanese language is an adjunct, Thus ill 
the relativc clause, tile value of tile dep feature is 
bound with the gap. And according to tile Direct 
Binding Hypothesis which we proposed ill section 5, 
a subcat element call he directly bound with the dep. 
Ilere, we will discuss why this hypothesis is correct. 
Furthermore, we will show that if we adopt the Direct 
Binding Hypothesis, the number of slash elements can 
be at most one. 
If we don't accept this hypothesis, the only alter- 
native is to assume that the gap is bound only with 
an element of slash. That is, if an element of subcat 
cannot be bound directly with the value of the dell 
the gap should he initially ill the value of slash, and 
should then be bound with the elenlent of dep. We 
will call this hypothesis the Indirect B~nding Hypoth- 
esis (that is, subcat elenmnt can be indirectly bound 
with the dep through a lexical rule). 
First consider the following seutence: 
(9) \[\[ kak -ta\] I ........ yuppan-si ota\] 
write PAST book ACC publish PAST 
syuppansya-ga yuumei-ni nar 
publisher NOM famous GOAL become 
-ta \] gakusya 
PAST scholar 
'the scholar such that the publisher that 
published his book became famous' 
AC1V.S DECOLING-92, NANTES, 23~28 Aot)r 1992 1 1 0 3 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
This is acceptable to everyone. If we accept the In- 
direel Binding Hypothesis, the number of slash ele- 
ments would be at least two, because in the above, 
every relative clause other than outermost one has at 
least two slash elements. 
However, consider the following strnctures (wbere 
i and j are either 1 or 2, and ¢1 and ¢2 are assumed 
to be bound with NI and N2, respectively): 
(10) a. Double Relative Clause 
\[s \[s-"\[s'"¢i"'\]N'"\[s'"¢,-"\]N'"\]N .... \]N1 
b. Double Gap in one Relative Clause 
\[s \[s \[s...~i..-~b,... IN... \]N2... \]N1 
c. Embedded Sentence in Relative Clause 
\[s \[s... ¢ .... \[s, .,¢~ ... \]-t .... \]N .... \]N~ 
The following are example sentences with the above 
structures: 
(ll) a. Double Relative Clause (10.a) 
\[\[\[¢, kak -ta\]e -ga \[¢2 mac-hi 
paint PAST picture NOM before 
tot -ta \] syasin-ni nitei -ta\] 
take PAST photo DAT resemble PAST 
syasin-k% -ga sin -da \] gaka 1 
photographer NOM die PAST painter 
'the painter such that the photograpber died 
whose photo taken before resembled the 
picture that he painted' 
b. Double Gap in one Relative Clause (10.b) 
\[\[\[¢1¢2~bz kaka -se -ta \] gakusy% -ga 
write make PAST scholar NOM 
sin -da\] hons -ga yuumei -hi 
die PAST book NOM famous DAT 
nar -ta \] hensyusya 1 
become PAST editor 
'the editor who made the sctmlar who died 
write a book that became famous' 
c. Embedded Sentence in Relative Clause (19.c) 
\[\[¢1\[¢2 syuwai-si -ta\] -to kak -ta \] 
take-bribe PAST CMPL write PAST 
seizlka2 -ga zisatu-si -ta\] kisya 1 
politician NOM kill-himself PAST journalist 
'the journalist who wrote a report that the 
politician, who killed himself, took a bribe' 
Though the sentence (ll.c) is better than the oth- 
ers, these are all unacceptable. Taking Indirect Bind- 
ing Hypothesis, the number of slash elements for the 
above is 3, 3, and 2, respectively. It follows that in 
terms of tbe number of slash elements, we cannot ex- 
plain the difference of the acceptability of the above 
structures. 
Ilowever, we accept the Direct Binding Hypothe- 
sis, sentence (9) needs only one slash element, and 
sentences (11) all need at least two. Thus the Direct 
Binding Hypothesis is better than the Indirect Bind- 
ing Hypothesis. And adopting this hypothesis leads 
to the concusion that the number of slash elements 
should be at most one. 
7 Concluding Remarks 
We have described a grammatical formalization of 
the relative clause in Japanese. Based oil tile JPSG 
framework, this formalization is characterized in 
terms of constraints, that is, relationships among tile 
grammatical features in a phrase structure and in- 
formation in tile lexicai entries. In this paper we 
proposed the Direct Binding Hypothesis, and showed 
that the number of slash elements should be at most 
one in Japanese. 
Acknowledgments. We are grateful to I)r. Takao 
Gunji, Dr. K6iti Hasida and other members of the 
JPSG working group at ICOT for discussion. And 
we thank Dr. Phillip Morrow for proofreading. 
References 
\[1\] Gunji, T. : Japanese Phrase Structure Gram- 
mar: A Unification-based Approach, D. l~eidel, 
1987. 
\[2\] llarada, Y., Gunji, T., Miyoshi, \]\[1., Sirai, 1I. and 
tlasida, K. : JPSG-A Phrase Structure Gram- 
mar for Japanese, Advances in Software Science 
and Technology, Vol. 1, No. 1, 1989, pp. 1-15. 
\[3\] Hasida, K. : Conditioned Unification for Natural 
Language Processing, Proc. 11th Int. Conf. on 
Cornput. Linguistics, 1986, pp. 85-87. 
\[4\] llasida, K. and Sirai, tl. : Conditioned Unifi- 
cation, CompuL Soflw., Vol. 3, No. 4 (1986), 
pp. 28-38 (in Japanese). 
\[5\] lnoue, K. : Transformational Grammar 
and Japanese language, Taishukan, 1976 (in 
Japanese). 
\[6\] \]noue, K. : Grammar Rules of Japanese lan- 
guage, Taishukan, 1978 (in Japanese). 
\[7\] Pollard, C.J. and Sag, I.A. : h~formation-Based 
Syntax and Semantics, Vol. 1: Fundamentals, 
CSLI Lecture Notes No. 12, Center for tile Study 
of Language and Information, Stanford Univer- 
sity, 1988. 
\[8\] Pollard, C.J. and Sag, I.A. : Information.Based 
Syntax and Semantics, Vol. ~, to be published. 
\[9\] Sirai, II. : Relative Clause and Adnominal 
Clause in Japanese, in preparation. 
\[1O\] Tsuda, H., Hasida, K. and Sirai, lf. : JPSG 
Parser on Constraint Logic Programming, Proc. 
of ~th A CL European Chapter, 1989, pp. 95-102. 
AcrEs DE COLING-92, NAN'rE.S, 23-28 AO(JT 1992 1 1 0 4 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
Abstract in Japanese 
~'5o 
:(E~:~'J~ slash ~ ~k'O~li~fl2K- J: o ~, ~ ~l,~'~'b 
J: < ~416~:cw zo ,,t: 5 I~, U*~-ot2, -~-.d,~ 
J: § ~:~ t~, lION\]© subeat ~-'©ff\[ ~ L-C 
cat ~ ~ t.t ~'~6 ~k w & T 6 o 
ss ~ (~) ~:~5~-c~ 
slash 0~.~: Subcat..Slash a~/~J~l~ll~ & 9, 
9~6o 
c')~a}l~ (_2~)-) ;~g-~%" ~o, ~liO slash 
slash 0")31~tli!l: ~i~q'J., slash r~e)f~(fi: 
IC%~ L'C ~, ES ~-~ R,C ~OtN~ff6 ~ Nt~ g: slash 
Slash ~l, lll\]IC _k t) slash ~N ~: ~ ~, ~c/q, 
I) , subeat -~:5~ dcp !~ti~ a NJ~I~fgJK~J" 
'5o 
6o +. ~ht, 
~-e~'5o (0 ~> (2) ¢~'~¢, ~N~:t~t~ 
60 
:b~. 1 ~l~e) slash ~l~li~ ~,.g,:~I~ & :J" 8 O ~-~j~ L., NE 
(2) tlt~Tt'g ~> 2 tt~l~J_kO slash ~+@,~ & 
¢)~t~, H*~ + ~-5 slash ~ilOll~lt~ 
AcrEs DE COLING-92, NANTES. 23-28 ao£rr 1992 1 1 0 5 PROC. O}: COLING=92, NAN-\] t,,s, AUG. 23-28, 1992 
