SGS: A SYSTEM FOR MECHANICAL GENERATION OF JAPANESE SENTENCES 
Taisuke Sato 
Electrotechnical Laboratory 
Ibaraki, Japan 
SGS is a compact sentence generation 
system. Inputs are the frames and spe- 
cifications of a sentence. Programs attached to 
context free rules carry out the generation 
task. Output is a surface sentence with an 
associated derivation tree. 
suitable laws for a computer from lingistic 
phenomena. Therefore, this paper first 
describes the overall organization of SGS, 
secondly explains the linguistic structure of 
Japanese with which SGS. tries to deal, and 
lastly gives examples of sentence generation. 
Introduction 
A sentence genration process can be a con- 
sidered to be a process starting from non-linear 
meaning structures and ending in a linear struc- 
ture, i.e. a sentence. Because meaning struc- 
tures reflect the speaker's intension and a 
speaker can easily produce a sentence realizing 
his intension, one tends to think of sentence 
generation as an easy task. 
Famous AI systems including SHRDLU, often 
adopt a fill-in-the-blank method to generate an- 
swering sentences. Efforts are concentrated on 
other tasks, such as sentence understanding, 
planning, deduction, and so on. 
Although study of sentence generation has 
not been receiving much attention, it is valua- 
ble for the following reasons: 
i) to develope a tool which enables a user to 
understand what has been understood by an in- 
telligent system. 
2) to build a machine translation system. 
3) to develop a theory of knowledge 
representation. If some formalism of knowledge 
representation is to be valid, it must be 
readable. In other words, it must be easily 
transformed into sentences. And this readablity 
is checked by means of sentence generation. 
4) to verify correctness of the various lin- 
guistic theories from a computational linguistic 
point of view. 
SGS is an experimental sentence generation 
system, the inputs of which are frames repre- 
senting some meaning. It generates a Japanese 
sentence with the help of a user-supplied 
dictionary and grammar. The generation process 
of the SGS is top-down with backtracking. The 
result is a surface sentence with its derivation 
tree. This system does not generate sentences 
at random but carefully generate one sentence 
obeying the user's control information which is 
given in advance of the generation process. 
In computational study of sentence genera- 
tion, building a system is one facet of the 
study. The other facet is the extraction of 
System Organization 
SGS is written in ETL-LISP and consists of 
about i000 line source statements. To actually 
produce a sentence, it needs three kinds of in- 
put, a dictionary, and a grammar. Accounts are 
given in order. 
Inputs 
A sentence generation corresponds to the 
speaker's speech process. Accordingly, if a 
sentence of good quality is needed, many factors 
of a speaker should be incorporated. However we 
restrict ourselves to treating only the syntac- 
tic and semantic factors. The pragmatic factors 
remained as future problems. 
Conceptually, factors considered here are 
separated into two categories. One is the fac- 
tors governing the intra-sentenceal phenomena, 
which determines the cognitive meaning of a sen- 
tence, and is stated in terms of phrase struc- 
tures, transformation, various features and the 
like. Inputs belonging to this category are the 
frames representing cognitive meaning of the 
sententence to be generated, and the syntactic 
category (e.g. SS ---Simple Sentence) of the 
sentence. In this paper, frames in examples are 
supplied by the Japanese language understanding 
system EXPLUS. 14 
The other cateogry is factors governing 
inter-sentencial phenomena related to "topic and 
comments", such as the distinction between "wa 
and ga" e.t.c.. These factors reflect a 
speaker's views. They can be treated by speci- 
fying the arrangement of noun phrases or the 
surface subject of the sentence. For example, 
if one wants to put emphasis on a certain noun 
whose deep case is THEME, specifying (S-SUBJ = 
THEME) may compel the system to derive a passive 
sentence whose surface subject is the specified 
noun. Therefore, such specifications work as 
conditions on the sentence or control informa- 
tion for the generative process. 
In summary, frames, a syntactic category, 
and conditions on a sentence reflecting a 
speaker's views comprise the inputs of SGS. 
Given these inputs, SGS tries to generate a 
sentence of the specified syntactic category 
from the frames by considering the given 
21-- 
conditions. 
Figure 1 is an example of an input frame. 
This frame represents the fact that HANAKO BUYS 
A BOOK. The REL-TM slot designates relative- 
time relation to other facts. The SF slot des- 
ignates semantic features of the predicate KAW-D 
(BUY). CACT(causal actant), THEME are the deep 
cases of KAW-U. 
((IDEHT = (POOOOOB PROPOSITION)) 
(LIHK 
(THEME (N000807 NON)) 
(CACT (N000008 HANAKO)) 
(REL-TH (P00000S PROPOSITION))) 
(SELF • (a ITU-DOKO with REL-TM) (a MODALITV)) 
(REL-TM , (SORE-WA (PO0000S PROPOSITION) MAE)) 
(SF - %VASPD XVTa) 
(PREDICATE - KAW-U VERB) 
(CACT - (N000808 HAHAKO)) (THEME 
• (N000807 HON))) 
Figure i. Example of Frame. 
Grammar 
Grammar in SGS refers to the collection of 
context free rules augmented by LISP programs. 
The role of the grammar is to systematically 
convert input frames into small trees, then com- 
bine and transform them while making sure of the 
grammatical correctness of the generated trees. 
It is not necessary for tree structures to 
accompany sentence generation(McDonald's system 
doesn't use tree structures), but setence gener- 
ation via tree structures has many advantages. 
First of all, lingistic knowledge based on 
transformational theory can be easily imple- 
mented in a computer. Linguistic concepts such 
as subject, object, scope of quantifier, de- 
letfon, raising, e.t.c., are all related to tree 
structure. Also, organizing the system as a 
tree manipulation system is a good way to keep 
its clarity and is helpful in debugging the 
grammar. Suggestive information to improve 
grammar could be obtained by tracing inter- 
mediate trees. Moreover, context free rules to 
construct a derivation tree assures, to some ex- 
tent, the grammatical correctness of the gener- 
ated sentence. The form of a syntactic rule is: 
(<category> <descendents> <Pl> <P2>) 
A rule has four fields. <category> and 
<descendents> form a context free rule: 
<category>=> <descendents>. 
<PI> is a LISP program. It is applied to 
the frames which should be realized as a 
sentence of the <cateogry>. It divides the 
frames into subframes corresponding to each 
<descendents> considering the attached 
conditions. 
<P2> is also a LISP program. It is invoked 
after the completion of <descendents> subtrees. 
Its role is to look at the subtrees and make 
sure of their grammatical correctness. Trans- 
formation is added to the subtrees as necessary. 
Finally, <P2> returns a partial derivation tree 
whose top node is <category>. The rule invoca- 
tion mechanism is explained later. 
Dictionary 
A lexical item in dictionary describes the 
knowledge of each word. As for the predicate, a 
name, a surface expression, semantic features, 
deep cases and their semantic features are in- 
cluded in its description. Similar items are 
included in the noun's frame. The form of an 
item is: 
(<name> <category> <Pi> <unit>) 
<name> and <category> are keys for 
searching the dictionary. In the case of HON (a 
book), the <name> is HON, the <category> is 
noun. 
<Pi> is a LISP program to check conditions 
for lexical insertion. 
<unit> is a frame depicting linguistic 
knowledge of a word. World knowledge can also 
be stored in <unit>. 
The description of a lexical item is at a 
concrete level. Neither lexical decomposition 
nor word description by primitives is adopded. 
Although, with respect to verbs, Japanese has a 
rather systemic way of deriving new words from a 
basic word (for example, from TOB-U (to fly), 
TOB-ASU (to make something fly) or TOB-ERU (can 
fly) are derived.), studies in relations among 
the lexical items seems not to be advanced 
enough for use in a computer at present. 
Generation mechanism 
There are many methods to generate 
sentences. The fill-in-the-blank method is 
easiest. McDonald's system ~'9 derives a sentence 
directly from source data. BABEL ~'s derives a 
sentence indirectly using discrimination nets 
and a syntax net. 
As stated previously, SGS generate a 
sentence via tree structures. Initially, SGS 
receives an orderd triple from a user. Its form 
is: 
<category, input-frames, conditions on the sen- 
tence> 
The system regards the orderd triple as a 
goal. It says "from the input-frames, generate 
a sentence of the category that satisfies the 
conditions". After pushing this triple to the 
bottom of the stack, the system starts the 
generation process described below. 
step i: lexical insertion 
Look at the top of the stack. Let this 
triple be category A, frame Fr-A, condition 
Cond-A . Collect lexical items from the dic- 
tionary that match Fr-A and satisfy Cond-A. 
If no item is found, Go to step 2. Else, 
choose one of the items and return it. 
Because back-track may occur in later process, 
preserve the unchosen items. Remove the top 
element from the stack. Go to step 1. 
22 
step 2: subgoal expansion downward 
If subtrees under category A are com- 
pleted, go to step 3. Else collect rules of 
the form A descendents P1 P2 from the 
grammar. Select one of them. Suppose the 
selected one is <A (B C) Pl P2>. Execute 
program Pl to create the subgoals, P1 tries 
to divide Fr-A into Fr-B and Fr-C. Pl also 
converts Cond-A to Cond-B and Cond-C 
respectively. If this division is successful, 
push the resulting subgoals <B Fr-B Cond-B> 
and <C Fr-C Cond-C> onto the stack. Go to 
step i. If division is unsuccessful, try 
another rule. If all the tried rules fail, 
start back-tracking. 
step 3: tree building upward 
This stop treats the case where subtrees 
under category A are completed. Execute 
program P2 in the rule <A descendents Pl P2> 
which was used to divide Fr-A at step 2. P2 
tries to confirm the grammatical correctness 
of the completed subtrees using interpretation 
of them. If one of them is found to be 
ungrammartical, start back-tracklng. Else 
transform them as necessary and provide data 
for later interpretation of the completed 
tree. Combine the category A and subtrees to 
complete the partial derivation tree cor- 
respoinding to the goal <A, Fr-A, Cond-A> on 
the top of the stack. Remove this triple from 
the stack. If the stack is empty, collect the 
terminals of the tree in left-to-right order, 
give morphological inflection to the sequence 
of terminals and print them. Otherwise, go to 
step i. 
Categor U A Cond-A 
I FP-A I 
step I. I 
I I A I 
/ \ I 
texicat item I 
/ \ I 
....... I step 2, 
I I Category B Cateqory C 
Con~-B Con~-C 
1 I 
/ \ / \ 
/ \ / \ 
I step 3. I 
Figure 2 
I 
A / \ 
/ \ 
B C 
/ N / X 
/ \ / \ 
Generation Protons 
Simplified Syntax of Japanese 
This section gives a brief account of the 
simplified Japanese which SGS tries to deal with. 
A Japanese simple sentence consists of 
three parts, as is shown below. It is important 
to notice that these parts assume different 
functionalities. 
Part A to Part B 
<PP>~--)<UERB><CAUSATIUE>(PASSIUE)->{ASPECT)---> 
I m>(TE-MIRU..}- 
I ....... <ADJECTIUE>,nlt .................. I 
Part B to Part C 
....... )---(HAI, DA, RASll .... }---) 
Part C 
....... )---(KA, HA, HE, RO .... ) 
Figure 3, Slmptitied Syntax o£ Ja?anese 
The first part, A, expresses epistemic 
moaning of the sentence. It begins with several 
propositional phrases (typically two or three) 
in unspecified order. A prepositional phrase is 
derived by the rule <PP ~> <NP><P>, where <NP> 
is a noun phrase, and <P> is a post positional 
particle Particles belonging to <P> are GA, NO, 
NI, WO, DE, etc.. They work as surface case 
markers. 
After a sequence of <PP>s, there comes a 
verb, an adjective, or nothing. A verb can be 
followed by SERU'SASERU (causative particles) or 
RERU RARERU (particles of passive or spontaneity 
etc.). These particles are connected to a verb 
so tightly that thay work as a single word. 
Words for <ASPECT>, TE-MIRU(to try), TE-AGERU 
(to indicate a speaker's attitude to the hearer 
in which a speaker kindly does something for the 
hearer), TE-KURERU(opposite to TE-AGERU), etc. 
are the last constituents of part A. These are 
all verbs. 
The second part, B, indicates a speaker's 
attitude to. the proposition expressed by part A. 
This part contains DA(affirmative), 
NAI(negative), DAROU(guess), RASII(conjecture), 
etc.. These are all particles. 
Expressions in the last part, C, are meant 
to cause some effect on the hearer. Among them 
are KA(interogative), NA(prohibition), 
NE(suggestion), RO(imperative), etc.. 
A predicate(verb, adjective) has a case 
structure. For example, OK-U (put) has three 
cases: CACT(causal actant), THEME, and LOCUS. 
Each case is accompanied by specific particles. 
CACT and GA, THEME and WO, LOCUS and NI or DE 
are usually used in pairs. The case system is a 
basic linguistic structure in itself, but the 
primary objective of SGS is not the study of 
case system in Japanese, so SGS utilizes the 
case system of EXPLUS. 
Syntactic rules governing the connection of 
particles following a predicate are said to be 
described by a regular grammar. 
As for tense representation, TA is used to 
indicate the past or perfect tense. TA can be 
inserted between either part A and part B, or 
23- 
part B and tense systems are discussed in the 
following sections. 
A compound sentence is composed of simple 
sentences. A relative clause in Japanese is 
derived by the rule <NP> ~<S><NP>. This rule 
yields a left branching structure peculiar to 
Japanese in centrast to English. In this paper 
an example of a sentence using a relative clause 
is shown with discussions. However, sentences 
with coordinate structures are not treated. 
From a transformational stand point, em- 
bedding structures are important. A causative 
or passive sentence is typical of embedded 
structures. A generation example of a causative 
passive sentence is shown later. But how a pas- 
sive and causative sentence is derived from the 
initial structure is not definitely solved. 
Aspect z,2 
In order to achieve temporal represen- 
tation, treatment of tense and aspect is 
inevitable. First we discuss the Japanese 
aspect system which brings a lot of insights 
useful to computational linguistics. 
The basic role of aspectual representation 
is the distinction between perfect and 
non-perfect. It seems to be common to many 
languages. However, actual languages provide 
mechanisms for aspectual representation de- 
veloped beyond this distinction. 
In Japanese, many types of aspects are 
realized by using aspectual particles following 
a verb. TE-IRU and TE-SIMAU are most typical. 
For instance, YON-DE-IRU (YON is a contracted 
form of the verb YOM-U (to read)) means the 
repeat of reading or the experience of reading. 
YON-DE-SIMAT-TE-IRU means being in the state 
after the achievement of reading. Several 
primitive aspects are shown in fig. 4. 
Istetive --simpte, resut~atlve, progreesive 
I ..nil, TE-IRU. aspect-linchoatlue ..$1-KAKERU, SI-HAZIMERU, etc.. 
Icomptetive ..TE-SIMAU, ~I-OWARU, etc. Iothers ..TE-ARU, TE-YUKU, TE-KURU, etc., 
l+stativeI+durative ..ADJECTIUE, DA, etc. I 
l-durmtive ..NIRU(resemble),e~c. 
pred~ca~el l-s~atlvelcross c\[as=Ificatlon by 
l+-dura~ive,+-resutta~ive,etc. 
Figure 4. Cte~sificatton o~ Rspects and Predica~es 
There are stative, inchoative, completive, 
and other aspects. The stative aspect is sub- 
classified into three subclasses. TE-IRU per- 
forms an important role in establishing these 
subclasses. 
Verbs in Japanese are classified according 
to the aspectual meaning of the combination of 
the verb and aspectual particles. As a result, 
aspect features are assigned to a predicate and 
an aspectual particle. For example, \[-durative, 
+resultive\] is assigned to OK-U. \[+stative, 
+durative\] is assigned to an adjective or a 
cupulative expression DA, and so on. With 
regard to particles, \[+stative\], \[+completive\] 
are assigned to TE-IRU and TE-SIMAU 
respectively. 
Once aspect features are assigned to the 
predicate and the particles, an interpretation 
of the aspect of a composite predicate is 
mechanically deduced by looking only at the 
aspect features of each consistuent. The aspect 
of YON-DE-IRU, for example, is obtained in such 
a way that the aspect features of YON-U(read) 
and TE-IRU are examined first. YOM-U has 
\[-stative, +durative, +resultative\] and TE-IRU 
has \[+stative\]. Then the features are syn- 
thesized in obedience to 'synthesizing rules of 
aspects' In this case the result is 
\[+durative, +resultative, +stative\]. It allows 
twe interpretations, which is compatible with 
the aspectual ambiguity of YON-DE-IRU. One 
interpretation, based on the combination 
\[+durative, +stative\], is the progressive 
interpretation---being in the state of reading. 
The other interpretation, based on \[+re- 
sultative, +stative\], is the experiencing 
interpretation--- being in the state of after 
reading. These aspectual ambiguities are 
resolved by context or adverbials. Similarly, 
the aspect of YON-DE-SIMAT-TE-IRU is obtained in 
the same way. 
It is easy to see the advantage of 'aspect 
description by aspect features'. It enalbles us 
to treat the (Japanese) aspect mechanically in 
both directions -- sentence understanding and 
sentence generation. However, though a great 
deal of progress has been made in the study of 
Japanese aspects, we have not yet devised a 
satisfactory system for aspect description by 
aspect features. 
Tense i,a 
It is well known that TA stands for not 
only past tense but also the speaker's confir- 
mation, recollection, or immediate requirement. 
Consequently, we can not simply say that TA 
indicates past tense. Instead there are a 
number of evidences suggesting that TA indicates 
the perfect as well. As will be explained in 
the following, treating TA as a perfect- 
indicator leads to a succsinct description of 
tense interpretation in Japanese. This fact it- 
self, in the author's opinion, is the strongest 
evidence for TA as a perfect-indicator. 
\[+perfect\], therefore, is assigned to TA. It is 
also assigned to a predicate accompanying TA. 
If a predicate does not accompany TA, \[-perfect\] 
is assigned. Some definitions are needed before 
stating tense interpretation in Japanese. 
Definition: speech time is the time when a 
speaker speaks, and event time is the time 
occupied by the events(facts) refered to by a 
sentence or a clause. 
With this definition, the principle of tense in- 
terpretation in Japanese is stated as follows. 
--24 - 
• A sentence of a clause containing a 
predicate of +perfect(-perfect) refers to the 
events or facts previous(not previous) to the 
standard time. 
The standard time of a simple sentence or 
a main clause is the speech time. The 
standard time of a subordicate clause is the 
event time refered to by the main clause. 
In short, TA asserts something has occured 
previously. Detailed tense interpretation using 
the aspect feature 'stative' is summarized in 
figure Fig. 5 which is hereafter called 'the 
principle '. 
mlmal • II m m m im m m u ml i m R m m m am m ml o m i ~m i im m Imml w ml I m i 
aspect of I +-perfec~ t interpretation at ~he predicate 
t I standard ~ime(.presen~} 
iIiIIi IIIiII1~1111 i~1111 iii iIiIIiIIiiii IiiiiIi i ii 
t -perfect I present s~a~e +~ative I ..................................... 
l +perfect I pes~ stabs 
t -perfect ! presen~ or future action ~ruth, habit I 
-statlve I ..................................... I +perfect I Dast action, event 
habl% I 
Illllllll II Ill Ill lllllllllllllllll I!1 Ull lllllllllll I 
Figure 5. Principle of Tense Interpretation 
The principle is applicable to any simple 
sentence and the majority of complex sentences. 
However some complex sentence has exceptional 
tense interpretation. Consider the next sen- 
tence in which the conjunctive TOKI is used. 
KAKI-WO TABE-TA TOKI KANE-GA NAT-TA. 
(a persimmon) (ate) (a bell) (rang) 
When I ate a persimmon, a bell rang. 
According to the principle, TA of TABE-TA 
assures that eating-a-persimmon preceds bell- 
ringing. But, unfortunately, such is not the 
case. The fact implied by the sentence is the 
simultaneity of eating-a-persimmon and 
bell-ringing. 
Such an exception may be ascribed to the 
peculiarity of the conjunctive TOKI. Since TOKI 
is also a noun and means time. TOKI used as a 
conjunctive is apt to connote 'at the itme 
when'. Exceptions to tense interpreation seem 
to depend on the conjunctive in the case of an 
adverbial clause, or the head noun in the case 
of a relative clause. Therefore case studies of 
tense interpretation are needed. 
Tense interpretation of the sentence type 
SI-- conj --$2 concerning Japanese tense con- 
junctives rOKI(when), MAE(before), ATe(after) is 
summarized in Fig. 6. 
S1 is a subordinate clause. $2 is a main 
clause. The aspect feature +-stative is a fea- 
ture belonging to the predicate of $2. 'ap- 
plicable' means that the principle is 
applicable. 'simultaneous' means that the tense 
interpretiation is exceptional and the simul- 
taneity of the events refered to by S1 and 82. 
In the case of relative clauses, a tense 
interpretation table like the above can be 
similarly constructed, but the situation is 
worse in the case of adverbial clauses. There 
....... at" ..................................... I S2 I interpretation 
mll sllam • • • II • mlnm • Rml • wma IO u I mill I W ~lll m li flu mm m ml elm 
t -perfect: +perfect I t , -perfect : slmul~aneous 
-stettve+ ..................... or 
I +perfect i epp(icable I +perfec~l ........................... 
I I -perfect i applicable TOKI .............................................. 
+perfect i -perfec~J .......... I 
t -perfect t simultaneous ++ststluel ..................... 
I I I +perfect I I I +per feet I ........................... 
I i I -perfect I ungrsmm~et 
.................................................... 
I i t +perfect 1 MQE J-stattvel -perfectl .......... I applicable 
I ~ I -perfect I 
................................ .._. ..... . .......... 
I I I +perfect I QTO {-stative: I +perfectl .......... I applicable 
I -perfect l 
I Ill i III Im i ill • • I I iii l mm ill III Ill I II I ii ii i ii II I I I l n II n I l iii l I iii 
Figure G, ConJunctives and Tense InterpretetLon 
also exsist complex sentences requiring tense 
interpretation opposite to the principle. 
TAROU-GA TABE-TA KEIKI-WA HANAKO-GA TUKUT-TA. 
(TAROU--name)(ate)(a cake)(HANAKO--name)(made) 
The cake that TAROU ate was made by HANAKO. 
The main clauses is HANAKO-GA TUKUT-TA 
(HANAKO made a cake). The relative clause is 
TAROU-GA TABE-TA(TAROU ate the cake). Both 
clauses include TA, so the prediction by the 
principle is that the event TAROU-GA TABE-TA 
preceds the event HANAKO-GA TUKUT-TA, which is 
exactly opposite to usual tense interpretation. 
Because of these difficulties, SGS did not 
go far with respect to aspect and tense 
interpretation. Obviously further investigation 
from a linguistic point of view is needed for 
mechanical aspect-tense interpretation. 
Generation Examples 
Relative Clause 
There are two types of relative clause. 
One is the TAROU-GA TABE-TA KEIKI(the cake which 
TAROU ate) type. The other is the TAROU-GA 
KEIKI-WO TABE-TA ZIZITU(the fact that TAROU ate 
a cake) type. 
The example shown is the former type. 
Meaning structures consist of two propositional 
frames P0000Ol and P000002. Note that they have 
a common filler (NO00002.1TA). 
Initially specified are the top category 
SS, the arrangement of propositional phrases-- 
first THEME then LOCUS, and the surface subject 
--THEME. These inputs are goals saying 
"generate a sentence from the frames shown in 
figure 7. As to the sentence, its category must 
be SS(simple sentence), its surface subject must 
be THEME--ITA(a board), and THEME must be to the 
left of LOCUS". 
On receiving these inputs, the system 
starts rule invocations. The invoked rule 
-25 
selects a frame suitable for a main clause. 
Priority of the selection is given to the frame 
which includes a REL-TM(relative time) slot 
filled with "HATUWA"(speech time). In this 
example, P000001 is selected. It states that 
TAROU-GA ITA-WO TATEKAKE-RU(TAROU leaned a board 
somewhere). P000001 being selected, the system 
continues invoking rules in order to translate 
P000001 into a main clause. 
During the course of rule invocations, the 
generation process reaches the stage where the 
THEME slot is treated. Because the THEME slot 
and its filler--(NO00002.ITA), are always sup- 
posed to correspond to a noun phrase, rules of 
the form <NP>~>... are invoked one by one. 
As (N000002/ITA) is shared with another 
frame, P000002, which states that HANAKO-GA ITA- 
WO OI-TA(HANAKO put a board), <NP>-~<SS><NP>, a 
rule for a relative clause, eventually is 
invoked. It produces a relative clause--HANAKO- 
GA OI-TA ITA(a board which HANAKO put 
(eMT~-- 
((IDENT • (P@ee@e2 PROPOSITION)) 
(LINK 
(THEME (N@QO@e2 . ITA)) 
(CACT (N@OeO@3 . HRNAKO)) 
(REL-PTM (Peeeeel . PROPOSITION))) 
(SELF . (a ITU-DOKO with REL-PTM) (a MODALITY)) 
(REL-PTM - (KAHRYOU (50RE-WA (P0@@0@I . PROPOSITION) TOKI))) (SF 
- ~UASPC ~UTI) (PREDICATE • OK-U UERB) 
(CACT • (N0@@0@3 . HANAKO)) 
(THEME • (Heoeee2 . ITA))) 
((IDENT - (P000001 . PROPOSIT.ION)) 
(LINK • (THEME (N000002 . ITA)) (OACT (N000001 , TAROU))) 
(SELF - (a ITU-DOKO with REL-TM) (a MODALITY)) 
(REL-TM - (KANRYOU (SORE-UA 'HATUWA' TOKI))) (SF 
• XVASPO ~VTl) 
(PREDICATE - TATEKAKE-RU VERB) (CACT 
• (N000001 . TAROU)) (THEME 
- (N000002 . ITA)))) 
(QSTM-- 
((IDEHT - (N000003 . HANAKO)) (SELF • (a HITO)) (SF • XANIMAL)) 
((IDENT - (N000002 . ITA)) 
(SELF • (a SYAHEIBUTU) (a HEIMEN)) 
(SF • ~ARTOBJ)) 
((IDEMT - (N@@@@@I . TAROU)) (SELF - (a HITO)) (SF • ~ANIMAL))) 
somewhere). It first builds a tree for the 
sentence HANAKO-GA ITA-WO OI-TA from P000002 and 
completes the realtive clause by moving the 
position of ITA to the end of the sentence. 
Generally speaking, complex noun phrase re- 
strictions should be considered, but they do not 
work here. After the completion of the relative 
clause concerning (N000002.ITA) with a corre- 
sponding derivation tree, SGS tries to complete 
the main clause, but, since the rule invoked for 
the main clause allows only CACT--TAROU as a 
surface subject, it can not satisfy one of the 
initial goals (S-SUBJ = THEME). So backtrack 
occurs. 
Finally, the alternative rule 
<SK>=><SK><RAREi> is invoked. It generates a 
passive sentence whose subject is THEME--ITA, 
and the rest of the specifications are also 
satisfied. '-*-' in the derivation tree in- 
dicates a non-exsistent filler of the obligatory 
case in the given frame. 
A passive sentence treated 
by SGS is 'a pure passive sen- 
tence' which does have a counter 
part in English. There is also 
another type called 'an 
adversitive passive sentence'. 
This type is too subtle to treat 
mechanically. Therefore we con- 
sider only pure passive sentence 
and the rules for them. 
Causative Passive Sentence 
Japanese causative sen- 
tences, which are identified by 
the occurence of VERB + SERU. 
SASERU, often admit two types of 
interpretation. Consider the 
next sentence. 
-- Initlat-CATEG -- 
$ SS 
-- Inltlat-COND -- 
$ (5-SUBJ • THEME)(SPAN-SEQ • THEME LOCUS)() 
SS 
I SK ....................................... TENSE 
I 
SK .................................... RAREi 
I THEME .................... LOCUS ...... CAOT ......... VERB 
I NP ................. PPK 
I I 
SS ................ NP I 
I 5K 
............ TENSE 
I CACT 
...... LOCUS--VERB 
I I 
NP---PPK NP--PPK 
I I I NOUN NOUN I 
I I I 
HANAKO GA -*- HI OK-U TA 
I I 
I I 
I I 
I I 
I I 
I I 
I I NOUN I 
I I ITA GA 
NP--PPK NP .... PBK 
1 1 I 1 NOUN I NOUN I 
I I I I 
-*- NI TAROU NIYORI TATEKAKE-RU RAREI 
OUT-PUT- HANAKOGA OITAITAGA TAROUNIYORI TATEKAKERARETA 
Figure 7. Sentence with a Relative Clause. 
TA 
--26- 
(eMTR-- ((IDEMT • (P000068 . PROPOSITION)) 
(LINK - (THEME (H000140 , HASIGO)) (CACT (H000141 . HANAKO))) 
(SELF - (a ITU-DOKO) (a MODALITY wlth TENSE)) 
(TENSE • KANRYOU) 
(SF • XVASPC XUTI) 
(PREDICATE " TATEKAKE-RU VERB) 
(CACT - (N000141 . HAHAKO)) 
(THEME • (N000140 . HASIGO)))" 
((IDENT • (POOOOG7 . PROPOSITION)) (LINK • (THEME (POBOeGB , PROPOSITION)) (REL-TM (POOOOBB , PROPOSITIOH))) 
(SELF - (a ITU-DOKO with REL-TM) (a MODALITY)) 
(REL-TM • (SORE-~A (Pe000SB . PROPOSITION) MAE)) 
(TENSE • (KANRYOU (SORE-WA 'HATUQA' TOKI))) 
(SF - ~UASPC ~VT2) 
(PREDICATE • S~SE-RU VERB) 
(THEME " (PeeeesB . PROPOSITION)))) 
(@STM-" ((IDEMT • (H000141 . HANAKO)) (SELF • (a HITO)) (SF • XAMIMAL)) 
((IDENT • (N000140 . HASIGO)) (SELF • (m BUTTAI)) (SF " ~ARTOBJ))) 
-- InitisI-CATEG -- $ 
$5 
-- Initial-COMD -- 
$ (S-SUBJ • THEME)() 
5S ( 
SK .................................. TENSE 
i SK ............................... RARE% 
i ) 
CACT ....... THEME .................. SK ............... SASE-WO .I 
) J ) I I I i 
I I THEME ..... LOCUS ...... VERB I 
I 1 1 I I t ' 
MP .... PPK MP---PPK HP---PPK NP--PPK I I 
I I t I I I I I I ~ ' 
NOUN L NOUN t NOUN ~ NOUN I I I I t I I I ~ I I I 
-%- HIYORI HAHAKO GA HASIGO WO -%- HI TATEKAKE-RU 5ASE-WO RARE% TA 
OUT-PUT- HANAKOGA HASIGOWO TATEKAKESASERARETA 
Figure 8. Causative Passive Sentence. 
TAROU-GA HANAKO-WO YUKA-SERU. 
(to go) 
= THEME)--initial goal-- 
After the com- 
pletion of the deri- 
vation tree and the 
(embedded) sentence 
corresponding to 
P000068, the rule 
mentioned above notices 
the tense of P000068 as 
being +perfect. This 
featre would entail an 
occurence of TA in 
front of SASERU on the 
surface level. But 
word order such as 
...TA SASERU... is 
ungrammatical so TA is 
supressed. 
A causative sen- 
tence corresponding to 
P000067 is built by 
raising CACT of the 
embedded sentence. The 
raised CACT--HANAKO 
changes to THEME. The 
resulting tree 
structure is, roughly 
speaking, \[-*- HANAKO 
\[HASIGO TATEKAKE-RU\] 
TA\]. The symbol -*- 
means non-exsistent 
filler. 
Owing to this 
structure, a passive 
sentence whose subject 
is THEME--HANAKO can 
be derived and (S-SUBJ 
is satisfied. 
One interpretation is that TAROU forces 
HANAKO to go. The other is that TAROU permits 
HANAKO to go. Ambiguities can be resolved by 
adverbials or context. These ambiguities bring 
difficulties to the treatment of causative sen- 
tences, but, for simplicity, SGS deals with only 
the former type. 
The example above is a causative-passive 
sentence. User's specifications are of the 
category SS and (S-SUBJ = THEME). The initial 
meaning structures consist of two propositional 
frames. The generation process begins by 
choosing a HATUWA frame to serve as an orign of 
time relations in the given frames. The chosen 
frame, P000067, includes a predicate slot con- 
taining SASERU. It will produce a causative 
sentence. 
While SASERU is a causative particle, it 
behaves as a verb in the deep level. It is a 
verb which takes a sentencial object whose case 
is THEME. Therefore the invoked rule responsi- 
ble for completing a causative sentence 
searches for a sentencial object. P000068 is 
the frame for a sentencial object. It states: 
HANAKO-GA HASIGO-WO TATEKAKE-TA. 
(a ladder) (leaned) 
HANAKO leaned a ladder. 
Conclusion 
Sentence generation is a basic task for an 
intelligent system, such as a consultant system 
or a Q.A. system, e.t.c.. SGS, though it is far 
from being satisfactory, is one step closer to 
an intelligent sentence generation system. The 
next step should be manifold. SGS admits 
various improvements. 
During the generation process, diverse mes- 
sages are exchanged between invoked rules so 
that messages tend to get out of control. 
Greater regulation is needed. 
As for the dictionary, it would be in- 
teresting to incorporate 'lexical 
decomposition'. Introducing 'lexical decom- 
position' can be helpful in organizing lexical 
items in a dictionary. However it requires a 
more refined method of lexical insertion. 
Linguistic knowledge should be thoroughly 
investigated and digested. Though the aspect- 
tense system in Japanese has been investigated 
to some extent, it is not obvious whether the 
description of aspect-tense system by features 
is sufficient to represent temporal knowledge. 
Presently, SGS lacks the ability to con- 
tinuously produce sentences. In order to form a 
paragraph the problem of coreference mechanism 
--2?-- 
must be solved. Japanese is so rich in ellipsis 
it is necessary to reveral and implement the el- 
lipsis system. 
ACKNOWLEDGEMENT: The auther is grateful to 
Mr. Tanaka, Chief of Machine Inference Section 
of Electrotechnical Laboratory and, other 
members of the section, for helpful discussions. 

REFERENCES: 

\[I\] Ota Akira: "Comparison of English and 
Japanese, with special Reference to Tense 
and Aspect", Studies in English 
Linguistics, Asahi Press, 1972. 

\[2\] eta Akira: "Tense Correlations in English 
and Japanese", Studies in English 
Linguistics, Asahi Press, 1973. 

\[3\] Huber,F.: "On the Generation of English 
Sentence", IEEE Trans. of Computers, 
25:90-91, 1976. 

\[4\] Goldman,N.M.: "Computer Generation of 
Natural Language From a Deep Conceptual 
Base", Stanford AIM-247, Jan. 1974. 

\[5\] Goldman,N.M.: "Sentence Paraphrasing 
from a Conceptual Base", Comm. Assoc. for 
Computer Machinery, 2, 18, 1975, 96-106. 
Academic Press, 1975, 41-58. 

\[6\] Hutchins,W.J.: "The generation of syn- 
tactic structures from a semantic base", 
North-Holland, 1971. 

\[7\] Kuno Susumu: The Structure of Japanese 
Language, MIT press, 1973. 

\[8\] McDonald,D.: "Preliminary Report on a 
Program for Generating Natural Language", 
IJCAI4, 1975, 401-405. 

\[9\] McDonald,D.: "A Framework for Generation 
Grammars for Interactive Computer 
Programs", AJCL, Microfiche 33:4, 1975. 

\[i0\] Schank,R.C.: "Conceptual Information 
Processing", North-Holland, 1975. 

\[ii\] Self,J.: "Computer generation of 
Sentences by Systemic Grammar", AJCL, 
Voi.12-5, Microfiche 29, 1975. 

\[12\] Shapiro,S.C.: "Generation as Parsing from 
a Network into a Linear String", AJCL, 
Microfiche 33 : 45, 1976. 

\[13\] Simmons,R. and Slocum,J.: "Generating 
Egnlish Discourse from Semantic Networks", 
CACM, Voi,15, No.10, 1972. 

\[14\] Tanaka et al.: "EXPLUS-A Sementic Parsing 
System for Japanese Sentences", Third USA- 
JAPAN Computer conference, 1978. 
