22 
1965 International Conference on Computational Linguistics 
SENTENCE GENERATION BY SEMANTIC CONCORDANCE 
Toshiyuki Sakai and Makoto Nagao 
Department of Electrical Engineering, Kyoto University 
Kyoto, Japan 
~~ '~a~l 
:~..:, ~"<,P.>~ o-~ o 'l 
Sakai & Nagao I 
ABSTRACT 
Generation of English sentence is realized in the following three steps. 
First, the generation of kernel sentence by phrase structure rules; second, 
the application of transformational rules to the kernel sentence; and fina- 
lly the completion of a sentence by the morphophonemic modifications. 
At the first stage of generating kernel sentence, the semantics of words 
are fully utilized. The method is such that a pair of words in the generat- 
ion process (subject noun and predicate verb, verb and object or complement, 
sdJective and modified noun etc.) is selected in accordance with the seman- 
tic categories which are attached to each word in the word dictionary. The 
semantic categories are determined by considering both the meaning of words 
themselves and also the functioning of words in sentences. 
At the stage of transformational rules, sentence is considered not as a 
simple string but as the one having the internal tree structure, and the 
transformational rules are applied to this tree structure. For these two 
stages the generation process is formalized strictly and is realized in a 
computer programming. We have presented in relation to the transformational 
rules a method of sentence generation not from the axiom (from th@ top Gf 
the tree) but from any point, from which the whole tree is constructed. 
We have also proposed that the morphophonemic rules can be presented as 
a kind of operators operating on words in the neighbourhood of a generated 
string. 
I. INTRODUCTION 
At present on behalf of the experience obtained from the research alre- 
ady done, we have fairly exact knowledge about the real difficulties in 
the MT research which are to be solved in the near future. Among these 
problems the most important ones might be how to construct the syntax of 
a language and how to grasp the semantics of sentences of the language. 
There have been many excellent contributions to the problems of syntax, 
but there are still few to the problems of semantics and the interrelation- 
ship between syntax and semantics. We have tried an investigation in this 
area by the method of generation of English sentence. 
The first paper ever published concerning the generation of sentences 
might have been that by Prof. V.H. Yngve of MIT in 1961. We have adopted 
the method once again from the following points of view. 
(i) it acts as a powerful test to the study of sentence structure, 
(ii) " to the study of semantics, 
(iii) " to the study of the relationship between 
syntax and semantics in natural language. 
Generally speaking the sentence generation method, contrary to the 
analysis of a given sentence (which is guaranteed to be a correct one), 
tends to demand the severe construction of syntactic rules and word selec- 
tion rules. It may seem at the present level of mechanical translation 
that the treatment of the sentence property in its entirety is too diffi- 
cult to realize. But if we hope to have the translation as perfect as 
Sakai ~ Nagao 2 
possible, we are necessarily to confront with this problem. 
The quality of a linguistic theory of a language may be best examined 
by the generation of sentences according to the specified linguistic theory. 
Especially the effect of the interrelationship between syntax and semantics 
seems to be clearly exemplified by this so to speak "crude" test. Thus the 
sentence generation method surely responds to this overall treatment of the 
sentence property. This may be considered as a step towards the general 
theory of natural language. 
2. EVALUATION OF THE GENERATED SENTENCE 
2.1 Evaluation of the generated sentence -syntax- 
Several methods are developed for the description of sentence structure. 
We represent the syntax of English by a phrase structure grammar, transfor- 
mational grammar, and mophophonemic rules. The kernel sentence is generated 
by the phrase structure grammar, then some proper transformational rules 
are applied to it, and then the modification of the sentence by the morpho- 
phonemic rules produces the final output. 
These rules should generate "conceivable sentence structures", although 
the actually used sentence structures have several constraints. These are 
for example, 
(i) the depth of the sentence structure 
(ii) the coordination structure 
(iii) the intrinsic unsymmetry of sentence structure --- progressive struc- 
ture, top heavy structure etc. 
In general the rules which are suitable for analysis of a given sentence 
seem to differ from the rules which generate good sentences. The difference 
between these two is the difference between the actual spoken sentences and 
the conceivable sentences. Here we can see man's tendency to the language 
structure. Therefore it will not be worthless to know the frequency ratios 
of the phrases used in the actual sentences. 
2.2 Evaluation of the generated sentence -semantics- 
The next question, and the more difficult one than the former, is the 
determination of what is the proper meaningful sentence. The test for the 
semantic sentence anomaly is far more difficult than the test for the 
grammaticality of the sentence. Here we can think of the following three 
levels of criteria for the right sentence. 
(i) The grammatical sentence which is spoken or written by the average 
person (and the sentence which conveys a concrete concept without 
knowing the circumstances the sentence is spoken). Here "grammatical" 
covers the phonology, phonemics, morphology, syntax etc. These sen- 
tence which are grammatical but which are contradictory in meaning 
and which we do not speak are to be rejected. 
(ii) The sentences which are incomplete in the word usages, inflexions 
and so on but which convey clear understandable concepts. These are 
the so-called corrigible sentences. 
(iii) The sentences which are grammatical but carry no concrete meaning 
Sakai & Nagao 3 
if they are not supplemented by tediously long explanations about 
the righteousness of the expression. 
We have here adopted the second criterion for our generation of English 
sentences. That is because we can transform the corrigible sentences into 
the complete ones comparatively easily by checking the concordance of 
gender, number, case etc. Hereafter we are mainly concerned with the sen- 
tence which carries very definite concept, that is to say, the sentence of 
complete semantic consistency. 
3. SENTENCE GENERATION BY SEMANTIC CONCORDANCE 
We concentrate our efforts on the generation of the affirmative active 
declarative sentences. We try to generate this kernel sentence by the 
expansion rules. By the generation of the kernel sentence the attention is 
on the structural balance of the whole sentence, the influence of the choice 
of a word to the other part of the phrase, and their relation to the unified 
concept of the sentence. An expansion rule has a main constituent and the 
other non-main constituents in the expanded part. The latter symbols may 
contain optional elements. 
When an expansion rule is applied to a non-terminal symbol, to which 
there is already given a concrete word, the word is assigned to the main 
constituent of the expanded part. The words to the non-main constituent 
symbols are selected in relation to the main constituent word. A verb or 
a noun is taken as the main constituent of the non-terminal symbol "sentence" 
(initial symbol). 
3. I Terminology 
A set of syntactic word classes (abbr. SWC) 
S = (s,, s2, ..... ) 
A set of words 
W = (w,, w~, ..... ) 
A set of semantic categories 
p = (p,, P~, ..... ) 
A set of non-terminal symbols 
Z = (zo, z,, z~, ..... ), 
M = SuZ - (~,, ~, . .... ) 
A set of P's belonging to a word w 
• PCw) - (p~(.), p~(.), ..... ) 
p~(w) = (p~,,, P~, ..... ) 
A set of words belonging to a syntactic word class 
WCs) = (w~, , ws~ , ..... ) 
z o : axiom 
The type of expansion rules: 
z --~2~ , z eZ 
: string of symbols in M. (lis called a syntactic unit) 
Sakai & Nagao 4 
M(~) = ~z~ : main constituent of string~ . 
N Mj(~) = ~j : non-main constituent of ~ . j is attached 1,2,3,.. from 
left to right to the non-main constituents of string ~ . 
If ~is composed of only one symbol, there is no non-main 
constituent. 
Optional elements in the expansion rules are indicated by a pair of 
brackets attached to the symbols. An optional element in ~ can not be the 
main constituent. 
The type of selection rules: 
s-* w or s ~-~w 
Semi-terminal derivation: 
Expansion rules are applied on non-terminal symbols, to the stage where 
there is no symbol to be expanded. The final string is composed of SWC. 
A set of the derived main constituents for a symbol z: 
A set of all the SWC which can be the main constituent of a non-terminal 
symbol z or the main constituents of a phrase which is generated by succes- 
sive expansions of the main constituents of the original z. 
S(z) = (s,, , s~z, ..... ) 
S(s~ ) ~ (s~) is assumed. 
3.2 The process of generation of a kernel sentence (I) 
We suppose that a sentence has one central thing or concept to be menti- 
oned first of all. This is the main constituent of a sentence. Then a 
second important concept is determined with its grammatical position, 
referring to the central concept already selected. Next a third important 
one is determined likewise, and so on. This process is formally represented 
in the following. 
(i) Zo +~ w(s~ (Zo)) 
s~(zo) is an element of S(zo) which is the set of the derived main consti- 
tuents for zo. W(SL(Zo)) is a word belonging to the set W(s~(zo)). This 
shows the process starts from the selection of a word w for the axiom 
z o , and the sentence is to be constructed with the core word w. 
(ii) z0-~ , if s£(z,)6S(MCZ)) 
The axiom z o can be expanded into the syntactic unit ~ if and only if 
the already selected s{(zo) at the stage (i) is contained in the set of 
the derived main constituents for M(~). 
(iii) M(~)*-*wCsz(zo)) 
The already selected word w(sL(za)) is assigned to the main constituent 
M(2~) of the expanded syntactic unitl. 
(iv) NMk(~)*-~ w~se~CNSk(2))), for an k. 
if a certain condition 
C' PC~A P (~,.t.), P(w:,; :t ""' ) 
is satisfied. 
To each non-main constituents NMk(~) is corresponded each word w~ if 
the semantic categories for the words have a certain relation fz with 
that of the word w assigned to M(~). 
Sakai & Nagao 5 
At the n-th stage of the generation: 
It is supposed that a word is already assigned to the symbol. 
z +~ ,,,(s(z)) 
the n: 
(i) z-~X, if s(z) ~ S(M(X)) 
(ii) M(Z )*-~ w(s(z)) 
(iii) NM~(~)~-~w~4K(se~(NMA(~))) if a certain condition 
./:z (" Pew), P(w:,,), P~w..'~,. ),. .... ) 
is satisfied. 
3.3 Condition~ 
To all the elements of the semantic categories P = (Pl, P2' "'" ), the 
semantic distances are supposed to be defined. 
dll =d(p~, p~) = 0 
The condition ~ may be the following. 
M(~) <--~w(s(z)) 
w 
i,j,k over all non-main constituent symbols of~ . 
p~ P(w (s(NMk(~)))) 
p p(wCs(z))) 
An example of this process is illustrated in Fig. I. The double line indicates 
the main constituent of a phrase symbol which is written one line above. 
Certain semantic conditions are imposed on the pair of phrase names in a 
phrase, which are underlined. 
3.4 The process of generation of a kernel sentence (II) 
The generation process explained in 3.2 is from the axiom. But there are 
the cases where we want to construct a sentence from arbitrary grammatical 
positions and a given word. For example when we write a complex sentence 
like "The old gentleman whom we saw at the theatre was his father.", the 
main constituent of the subordinate clause is not "gentleman", but the verb 
"saw". So we must generate a sentence from a noun "gentleman" and its 
grammatical position: objective case. 
The process is that first the start point of generation is given by a 
word and its part of speech in a sentence. Next we select a proper rewriti~m 
rule which contains the part of speech of the word selected Just now. Then 
to the remaining elements of the rewritten phrase the proper words are 
assigned, the semantic categories of which coincide with the one of the 
alread~ selected word. This process is continued as far as there remains 
no element which can be rewritten by a phrase. The process is formally 
represented in the following. 
(i) Given z, w, s, where s4-~w, s e S(z) 
(ii) A tree structure whose top symbol is z is constructed by the method 
explained in 3.2. 
(iii) z'--~ ~ , Z = fz,'~ .... z-.. 2z~. 
A phrase z' is selected which contains z as a component of the expan- 
sion rule ~'-~2. 
Sakai & Nagao 6 
(iv) z+~w, Pz:*-~, i = 1,2, ...... ,n 
where certain condition 
is satisfied. 
A proper word is corresponded to each phrase P~. 
(v) Tree structures whose top symbols are Pz~ (i = 1,2, ... ) are const- 
ructed for all ~ by the method mentioned in 3.2. 
(vi) At the stage (iii), z' is newly replaced by z and the same operations 
from stage (iii) to (v) are performed. 
(vii) When z' = zo(axiom) is reached and the steps (iv) to (v) are completed, 
then the whole tree is accomplished under the axiom z o. 
An example of this process is illustrated in Fig. 2. The direction 
indicates the steps the sentence is constructed. 
4. TRANSFORMATIONAL RULES 
4.1 Representation of the rules 
The transformational rules can explain many sentence structures which 
are difficult to treat in an immediate constituent method. For example in 
the sentence "Is he young?", which is a question form of "He is young", 
"is young" becomes discontinuous, separated by "he". This is difficult to 
treat by an immediate constituent method. It is far more natural to explain 
this by the application of a transformational rule concerning question to 
the original affirmative sentence. 
The transformational rules we are now utilizing are classified to three 
types. The type 1 is unary transformations which may be thought of as con- 
verting a sentence from one to another, The type 2 is binary transformations 
which combine two sentences to form a third. And the type 3 is a transfor- 
mation between two phrases. In all these cases we can formally represent 
the transformational rules as the following type. 
z : XI'X~ ....... X~-~y l" X~,-yI" X~y .......... y~.X~y~+~. (I) 
The symbol z which is written on the left side means that this transfor- 
mational rules should be applied to the phrase z. Xi, Xs, ..... , X~are 
either the elements of M or words themselves. Among them we have a special 
symbol ~, which indicates that for this symbol ~, there might or might not 
correspond some term in a~investigated phrase z • That is, ~ expresses an 
arbitrary term. Y,, Y:, ..... , Y~i are vacant, some symbols or words which 
are not equal to X| , X~, ..... , X,. XL, , X~,, ...... , X~ are some symbols 
among Xi, XA, ..... , X~. 
The phrase z may have an internal tree structure, so the transformational 
rule is applied to this tree. An example of this is illustrated in Fig. 3. 
In this figure a noun phrase "the red books" is transformed to another noun 
phrase "the books which are red". This transformation is done by the rule, 
NPI: ~.AD.NQ ---~@-NQ-WHICH BE.AD 
There are problems in the transformational rules such as follows. 
(i) We have no definite criteria as to what kind of sentence structure is 
Sakai & Nagao 7 
to be treated in the scope of phrase structure grammar, and what is in the 
scope of transformational rules. 
(ii) We can name empirically or informally the transformational rules such 
as passive, that deletion, compl~ment/obJect transposition, etc., but to 
represent these rules formally in the form of (1) without contradiction for 
all the sentence structures generated from the specified phrase structure 
grammar, is difficult. 
(iii) Transformations which accompany the changes in the part of speech or 
the morphophonemic forms of words are difficult to treat. 
(iv) A transformational rule can not be applied unconditionally to the 
structure satisfying the rule form, but there are many cases where the 
application of rules depend on the semantics of the sentence. 
4.2 Application of transformational rules 
For the transformational rules of the type 1, we generate a sentence by 
the phrase structure grammar and at the same time memorize the generation 
steps of the sentence by the tree structure representation. Next we apply 
a transformational rule of the type I to this tree. If the rule is found 
to fit to the structure, then another tree is constructed from the original 
tree referring to the transformational rule. 
Examples of this type are: 
~.NP.VT-NPI.~ --~ 1.4.BE-3.BY-2-5 (passive form) 
This book emphasized the recent development clearly. 
--* The recent development be emphasize(d) by this book clearly. 
~-NP.VI2.NPI.~ --~ I.WHAT-DO-2.3-5-? (question) 
Last year John became a doctor of philosophy at thirty. 
--*Last year what do John become at thirty? 
The application of morphophonemic rules to these transformed sentences are 
explained in § 6. 
For the transformational rules of the type 2, we generate first a sentence 
by the phrase structure grammar, with its internal tree structure. Then we 
select a proper phrase name which is a proper branch point of the tree, with 
the word attached to the phrase name. Next we start the generation of ano- 
ther sentence starting with the phrase name and the corresponding word, which 
are selected just now. This generation is by the method explained in 3.4. 
Then the two sentences thus generated have a same word, which is the key 
point in the usual transformational rules of type 2. 
An example of this type is illustrated in Fig. 4. This is a combination 
of two sentences of Fig. I and Fig. 2. The rule applied here is, 
SS: ~-WT1.NPI-~.CM. NPI.~ ~ 1.2.3.WHICH-7.4 
and the generated sentence is 
Several most number computer already precede specialist into trend which 
read the paper. 
This example indicates that a transformational rules can not be applied in 
every case, even if the structure satisfies the rule form. There are many 
other examples of this nature. 
For the transformational rules of the type 3, we have mainly concerned 
with the noun phrases which are the results of the application of transfor- 
mational rules to certain phrases, especially to the sentence form SE. 
For example, 
Sakai & Nagao 8 
NP1.BE.NP1 ~ 1-CM.3-CM 
Kennedy is the president of the U.S. 
--~ Kennedy, the president of the U.S., 
NPI.BE.PP ~ 1-3 
Scientists are in the dome of the south pole 
Scientists in the dome of the south pole. 
This type of transformational rules are incorporated in the generation by 
the phrase structure grammar. 
Another important transformational rules are those which accompany the 
change in the part of speech of words. 
For example, 
VTI.NPI-~ ~ nn -1-PRP-2-3 
apply computer to the MT research 
application of computer to the MT research 
VT1-NPI.@ ~ nn-l. BE GIVEN T0.2-3 
consider the problems of the theory 
--~ consideration is given to the problems of the theory 
We have not investigated yet this type of transformational rules except 
few ones, in which the word dictionary should have information about the 
interchange of the parts of speech. 
5. SEMANTIC CATEGORIES AND THEIR RELATIONSHIP IN SYNTACTIC UNIT 
In the generation process thus defined, each word is determined by the 
selection rule s~--~ w applied to SWC's. How this word selection should 
be done is the semantics here considered. If the selection is done randomly 
without any semantic restriction, completely anomalous sentence will appear. 
To prevent this a new word is to be selected compatible with the already 
selected words which are in the neighborhood. Such semantic selection of 
words will especially Be important in the syntactic relations such as 
subject noun & predicate verb 
subject noun & predicate verb & complement (or object) 
adjective modifier & noun 
noun & noun ! 
adjective & adjective I (coordination structure) 
verb & verb 
etc. 
Selection of a proper word in relation to the other words will eventually 
require the semantic notifications to the words and their mutual relation- 
ship in a sentence. In other words the system of semantic categories is to 
be set up and the meanings of all the words are to be represented in the 
system. 
The construction of a system of the semantic categories may be done 
best by the replaceability relation among words in sentences. For example, 
to the verb "walk", there is a group of words which can be the subject to 
the verb "walk". To the word group thus formed, there will be another word 
group which can be the predicate and has a verb "walk" as its member. This 
word classification has not been tried yet on the whole scale, and indeed 
Sakai & Nagao 9 
this is very difficult to do. So we have done a slightly different way, 
although the fundamental attitude of our word categorization is the repla- 
ceahility of words in sentences. 
We postulate that all the words might be properly characterized by setting 
up a number of key concepts. For example a word "voyage" is categorized as 
journey with the additional images such as amusement, time duration, ocean 
etc. In fact when we speak we actually construct sentences fully aware of 
such additional meanings. 
Thus our aim is to extract such word images and to know how these images 
are mutually connected in such and such sentence structures. So we have 
started the extraction of semantic categories partly taking into consider- 
ation the Roget's thesaurus and some other publications. We have assigned 
the following numbers to the semantic categories of the parts of speech. 
100--299 verb 
300--499 noun 
500--699 adjective 
700--799 adverb 
900-- preposition 
The ten's digit indicates the rough semantic categories in a part of speech 
and the one's digit is to the further classifications. At present the 
number of categories for the verbs ia about 40, for the nouns about 90, 
for the adjectives about 50, etc. For the prepositions we have attached a 
number to each word. These semantic categories are shown in table I. 
Next we have to clarify the connectivity of words in a sentence. To do 
this we have attached several kind of semantic categories to each word in 
the following way. 
Verb: 
PI: 
P2: 
P3: 
P4: 
PS: 
P6: 
Noun: 
PI: 
categories intrinsic to the verb. 
categories which can modify the verb (additional images of the verb) 
categories which can stand as subject to the verb. 
categories which can stand as object or complement to the verb. 
special prepositions following the verb if any. 
grammatical indication as to the form of the verb. 
categories intrinsic to the noun. 
P2: categories which can modify the noun (additional images of the noun) 
P3: grammatical indication as to the form of the noun. 
Adjective: 
P1 : 
P2: 
P3: 
Adverb: 
Pl: 
categories intrinsic to the adjective. 
categories which can modify the adjective (additional images of 
the adjective) 
prepositions following the predicative adjective. 
categories intrinsic to the adverb. 
This expresses, for example, that a verb can take a noun for the subject 
whose semantic c~tegories P1 belong to the P3 of the verb, and can take a 
noun for the object or complement whose semantic categories PI belong to 
the P4 of the verb, etc. These are the conditions/~ introduced in 3.3. 
Examples of words having these connectivity informations are shownin table 2. 
Sakai & Nagao 10 
The generation process is thus first the selection of a verb, and then 
the determination of subject, object or complement referring to the semantic 
concordance mentioned here. Therefore each expansion rule of the phrase 
structure grammar has the indication such as 
SE--~NP< P1,P3~.VT__~I.NP1 <P1,P4>-PRP< Pl,P5 >-NPI 
NP--~ ADJ< P1,P2 >-NQ 
Inthis example VT1, NQ which are underlined are the main constituents to 
which some concrete words are selected. So for NP, NP1, PRP and ADJ, proper 
words are to be assigned having the semantic concordance between the pair 
of categories bracketed by < > . The first element in < > is the category 
of the phrase to the left of < , and the second element is that of the main 
constituent. Therefore a word is assigned to NP, whose semantic categories 
P1 have the same term in the P3 of VTI, and so on. 
But there are many grammatical phrases where we can not tell what kind 
of semantic relationship are to be established. We have not attached the 
semantic relationship to the phrases like, 
main verb : adverbial phrases preceded by preposition 
main sentence : subordinate clause 
noun : its adjectival clause 
etc. 
It is also difficult to find out the semantic relationship between the nouns 
of the form, 
NO + NO, NO + NO + NO, NO of NO 
such as, 
machine translation, information processing machine 
generation of sentences. 
However when these phrases are given from suitable transformations of 
another phrases such as 
solution of a problem .-.solve a problem, 
generation of sentences.--, generate sentences, 
we can establish the semantic relationship of these two noun in the phrase 
before the transformation is applied. In this case we have to know the noun 
form of a verb or its vice versa. This information is contained in the word 
dictionary as P6 for the verb and as P3 for the noun. 
The semantic relationship here introduced is essentially the connectivity 
between two words in a phrase, so there is a possibility of generating absurd 
sentences. To prevent this we have to know more minute mutual influence of 
meanings among words in a sentence. 
6. MORPHOPHONEMIC RULES 
We want to propose that the morphophonemic rules can be represented by 
a kind of operators operating on the words in the neighborhood. We include 
negation action, tense, case etc. to this level. We take the operators 
such as follows. 
(not), ~r (present tense), ~s (past tense) 
inf (infinitive), ~ (third person singular), pl (plural) 
su'~ (subjective casej, ob~ (objective case) 
Sakai & Nagao 11 
ing (ing-form of a verb), ed (past participle) 
nn (nominalization of a verb)~- etc. 
The functions of these operators are, 
n + verb ~ do + not + verb 
m n + aux. verb---~ aux. verb + not 
w pr + verb--, verb (present form) 
ps + verb --~ verb (past form) 
ing + verb--~verb (gerund or present participle) 
ed + verb --. verb (past participle) etc. 
Besides these, for the verb BE(R) and HAVE(~), the following three steps 
are to be applied in this order. 
(I) ~ + ~ + verb---, be + being + e d + verb 
(2) ~ + verb---, be + in~ + verb 
(3) h + be --. have been 
For example in the generation of a sentence shown in Fig. 5, the sentence 
obtained initially is 
# the father sg sub~ ~ enjoy fresh breeze sg obj # 
Here the following operationsare applied. 
s_~ oh__/---, ob~ s_~, s_~# ---~ # 
noun ~ ---. noun, ~ enjoy ~ be + in~ + enjoy 
in~ + enjoy---~ enjoying, ~ + be ~ have been 
Sg sub--~ subs~, sg have---~ has 
noun sub ~ noun 
And we can get the final form, 
# the father has been enjoying fresh breeze # 
These operators can appear both in the phrase structure grammar and in 
the transformational rules, but the operations of these are supposed to be 
done after the application of transformational rules. But there occur many 
complicated situations for the sentences after the application of transfor- 
mational rules and to what extent we can go on this line remains to be seen 
in the future. 
7. EXPERIMENTS 
We are doing experiments on sentence generation by a medium size compu- 
ter KDC-I installed in our university. The word size is about 450 (verb 
about 140, noun about 170, adjective about 80, adverb about 40, etc.) The 
phrase structure rules are about 80 of the form z--~F,.~ .... F~(max. of 
n = 5 at present). The rules include the optional terms so that effectively 
the number of rules increases. 
The transformational rules are now only fundamental ones such as, 
negation, passive transform, several kind of question form, nominalization 
of phrases, several kind of binary transformations, etc. 
We treat the sentences such as 
S+V+O,+O~ 
S+V+O+C 
by the binary transformation from two sentences of the form, 
Sakai & Nagao 12 
S + V + X : X z 01 have 02 
S + V + X : X = 0 be C 
S+V+Oi+O~ 
S+V+O+C 
Examples of generated sentences: 
Working to set for new present-day of area be abroad. 
To correspond to only research, different monograph work, be above such 
its branch of difficulty. 
To be being between selection invitation to assemble, a completely immense 
technique be to know desirable boundary. 
Of widely digital elimination of number, view of summary question which 
arise to be automatic here satisfy his directly attractive 
subject at only impossible reference. 
We think that this study of sentence generation of English suggests 
something to the translation technique to English from Japanese which has 
no sentence structure like English. 
In parallel to this study we are now experimenting just the reverse pro- 
cedure of sentence generation. That is, we are trying to decompose a given 
sentence into several semantic units or set of kernels. This process will 
clarify in a subjective sense the amount of information contained in a 
given original sentence. Moreover this will contribute very much to the 
field of information retrieval and also to the clarification of the logic 
of a discource, such as question-answer problem. 
(i) 
(ii) 
(iii) 
(iv) 
(v) 
Sakai & Nagao 13 
TABLE I-1. SEMANTIC CATEGORIES (VERB) 
Hand(110) 
Eyes etc.(120) 
!ntel~igent(130) 
behaviour 
Mental(140) 
Spiritua!(160) 
Mea___!(170) Soci____~_~(180) 
(111) 
(112) 
(113) 
(121) 
(122) 
close, open, cover, fill, hold, drop, mark, 
plant, put, fire, take, draw, make 
keep, carry, use have, get, give, help, bear, 
raise, hold, take 
connect 
(131) (152) 
hear, find, see, look, watch 
say, speak, talk, tell, call, laugh, order, 
read, sing, state, cite, pronounce 
(153) 
(134) 
(135) 
ask, answer, hear, find, order, cite, address 
add, get, receive, send, need, select, treat, 
eliminate, accept, arrange 
learn, read, find, write, see, process, apply, 
program, compute, speciaroze, compare, Judge, 
indicate, understand, translate, know 
deal, treat 
interest, experience 
(141) feel, remember, think, reflect 
(142) enjoy, thak, love, fear, like, 
(143) wish, hope, care, want 
(144) stimulate, attract 
(161) 
(162) 
believe, know, mind, think, mean 
attempt, aim, intend 
drink, eat 
(181) build, found, publish, generate, develops, assemble, 
bridge 
(182) kill, care 
(183) pay, receive, buy, get, give, save, need, present, 
provide, precede, concede 
(184) drive, fly, sail, ride, guide 
(185) follow, lead, elect, participate 
(186) live, work, make, cooperate, exist 
(187) contribute, serve, help, save, pronounce 
(188) consist, exist 
(189) correspond, concern, 
Body action(190) 
(191) (19a) 
(193) (194) 
(195) 
(196) 
Chan~e of state(230) (231) 
(232) 
(a33) 
play, show, try 
sit, stand, start, stop, put, set 
visit, call meet, show, see, appear, address 
sleep, rest, remain, wait 
walk, move, pass, leave 
come, go, reach, run, stay, arrive 
change, turn, arise, remain 
start, go, come, drop, leave, begin 
extend, follow, increase, form 
Sakai & Nagao 14 
Natural phenomena(250) 
(251) 
(252) 
rain, blow, cover, drop 
burn, fire 
TABLE I-2. SEMANTIC CATEGORIES (NOUN) 
Human beings(300) 
man (301) 
family (302) 
social (303) (304) 
(305) Nature(340) 
celestial (341) 
atmospheric(342) 
geographic (343) 
minerals (344) 
water (345) 
Large things(360) 
movable (361) 
building (362) 
parts of 362 
(363) 
place (364) 
Articles(370) 
books (371) 
foods (372) 
furniture (373) 
playthings (374) 
(375) Mental action(380) 
thinking (381) 
feeling (382) 
(383) 
(38~) 
(385) 
(386) 
Action(390) 
(387) 
(391) (392) 
(393) 
(394) 
(395) 
boy, child, girl, person, man, woman, 
Jack, Betty, Nelson 
brother, company, family, father, friend, 
people, sister, son, mother 
king, soldier, god, president 
specialist, reader, scientist, professional, 
(group), blind, editor, debutant 
generation, group 
sun, moon, earth 
rain, wind, air 
river, hill, mountain, road, land, field, sea 
Mt. Fuji, Kaatskill, Appalachia, Lake Biwa 
rock, stone, gold, silver 
water, sea, rain 
bus, train, car, ship 
house, church, school, Kyoto Univ. 
door, window, room 
road, street, garden 
book, picture, paper, story, Newsweek, Bible, 
handbook, library, monograph, article, summary 
literature, supplement, journal, proceeding, 
report, volume, tewt-book 
food, egg, bread, milk, corn, salt, pepper, water, bear 
table, box, bed, dress 
ball, teniss 
processor, machine, computer 
reason, idea, hope, mind, thought 
love, life, fear 
aim, end 
readiness 
gratitude, thank, patience, acknowledgement 
knowledge, thought, view, opinion, reference, aspect, 
conception, comment, consideration, understanding 
sense, art, view 
life, love 
war 
question, answer, speech, call, order, judgement, 
citation, problem example 
death, existence 
visit, watch, indication, advance, access, response, 
change 
Sakai & Nagao 15 
Abstract(400) 
(~Io) 
Sgci@l terms(420) 
(430) 
(~7o) 
(396) 
(397) 
(398) 
(399) 
(4oi) 
(4o2) 
(403) (4o~) 
(4o5) 
(~o6) 
(~o7) 
(411) 
(4~2) 
(413) (414) 
(415) 
(421) 
(422) 
(423) (424) 
(425) 
(431) 
(43a) 
(433) (43~) 
(435) (436) 
(471) (4?2) 
(473) 
work, help, treatment 
voyage, play, trip 
research, work, study, aid,, contribution 
selection, processing, use, application, programming, 
elimination, addition, presentation, publication, 
specialization 
name, word 
thing, matter, something, state 
way, form, mean, point 
case, matter, measure, course, use 
cause, change, end 
color, sight 
sound, voice 
group, 
gap, point, boundary, link 
interest, value, validity 
limitation, difficulty, success, problem 
kind, respect, branch, feature 
money 
bank, company, shop 
law, right 
city, country, street, town 
world, country, Japan, America, nation 
data, language, word, German, English, French, reference 
technique, science, manner, way, methodology, 
originality 
particular, detail, summary, assembly 
translation, generation, development, appearance 
comparison, arangement, cooperation 
information, fact, source, cansequence 
subject, material, list, listing title, 
field, area, circle, center 
course, state, live, manner, way, stage 
context, implication, content, indication, source 
TABLE 1~. SEMANTIC CATEGORIES (ADJECTIVE) 
Qualifyin~ 
state of the object 
outward(530) 
color (531) 
shape (532) 
length (533) 
height (534) 
extent (535) 
size (53~) 
(537) 
black, blue, gree~, red, white 
round, plain 
long, short 
deep, high, low 
wide, narrow 
large, little, small, big, least 
single, individual, numerous, multiple 
Sakai 2, Nagao 16 
subjective(550) 
beauty (551) 
fair (552) 
free (553) 
full 
(555) internal(570) 
material (571) 
weight (572) 
hardness (573) 
temperature (574) 
new, old (575) 
soft, (576) 
social state(590) 
wealth (591) 
position (592) 
fact (593) 
restruction (594) 
(595) 
(596) 
(597) mental state(600) 
sentiment 
intimacy 
time(620) 
valuation(640) 
good, bad 
abstruct(660) 
(601) 
(602) 
(603) 
(604) 
(605) ( 
(621) 
(622) 
(623) 
(641) 
(642) 
(643) 
(661) 
relation(670) 
(680) 
(662) 
(663) (664) 
(665) 
(671) 
(672) 
(673) 
(674) 
(675) 
(681) 
(682) 
(683) 
beautiful, pretty 
fair, clear, fine 
free, fresh 
full, complete 
Undisputed 
gold, 
light, heavy 
hard, soft 
cold, warm, hot 
new, old, fresh 
soft, swee, fresh 
rich, poor, cheep, modest 
great, deep 
%rue, correct, natural 
free, strict, major 
skilled, detailed 
commercial, available 
neighboring 
glad, happy, sad 
dear, ready, desirable, familiar 
aware, wary, afraid 
active, attractive 
proud 
fast, quick 
ready 
late, recent, current, present, final 
good, bad, right, modest, valuable 
very, immense 
natural 
general, collective, common, particular, 
special, comprehensive, standard, specific 
next, certain 
present, absent, conventional 
related, informed, concerned 
complex, detailed, bound, complate 
right, left 
close, near, intermediate 
direct, conclusive, extensive 
individual 
possible, impossible, necessary, enough 
digital , linguistic, automatic spectral, t 
technical, scientific 
leterary, professional 
world wide 
Sakai & Nagao 17 
TABLE 2. Semantic categories attached to words 
VERB 
add 
answer 
bear 
begin 
Pl : 132 
P2 : 730 760 
P3 : 300 
P4 : 369 420 
P5 : 990 
Pl : 131 
P2 : 730 740 
P3 : 300 
P4 : 300 393 395 396 
Pl : 112 
P2 : 722 741 
P3 : 300 320 340 360 
P4 : 381 400 
P5 : 960 
Pl : 232 260 
P3 : 300 
P4 : 390 
believe PI : 160 
P2 : 743 744 
P3 : 300 
p4 : 300 380 400 
bring Pl : 196 
P3 : 300 320 341 361 380 
P4 : 300 
P5 : 980 
build Pl : 181 
P3 : 300 
P4 : 346 361 364 368 
hear P1 : 121 131 
P3 : 300 320 
P4 : 390 380 400 412 413 414 418 
help P1 : 112 186 
P3 : 300 341 342 360 380 
P4 : 300 320 380 390 400 
sea P1 : 193 121 150 
P3 : 300 320 
P4 : 300 310 320 330 340 360 380 
think PI : 160 130 
P3 : 300 425 
P4 : 380 390 402 403 405 425 
enjoy Pl : 142 
P3 : 3oo 320 425 
P4 : 340 372 391 395 397 463 
NOUN 
answer P1 : 393 
P2 : 510 520 550 590 640 660 
apple PI : 330 
P2 : 510 520 531 532 536 551 575 
bank P1 : 368 
P2 : 536 575 591 592 640 660 
bird P1:322 
P2 : 510 520 530 551 578 
book P1 : 362 
P2 : 510 520 530 551 572 641 
change Pl : 413 
P2 : 520 536 550 590 620 640 
country P1 : 384 385 
P2 : 510 520 536 551 590 
flower P1 : 330 
P2 : 531 536 551 571 576 591 
hour P1 : 410 
P2 : 597 594 620 640 660 
order PI : 393 
P2 : 536 553 573 621 640 660 
fear P1 : 391 395 
P2 : 534 536 578 592 593 
love P1 : 382 383 
P2 : 550 574 576578 593 595 
right P1 : 388 411 
P2 : 554 592 593 594 640 
sight PI : 405 
P2 : 534 535 551 574 597 660 
ADJECTIVE 
old PS : 575 581 623 
fresh PI : 575 576 553 
strong PS : 573 582 
great P1 : 592 
plain P1 : 532 596 
free P1 : 553 594 
right Pl : 538 593 641 
Sakai & Nagao 18 
SE(precede) fl 
( compute r) NP WTI( p recede ) NPI ( specialist ) 
AVV VT1 
J i (compu~~ already precede 
(several)AD1 ~ NQ(computer) 
/if (most)AD 
J ~- NO(number) NO(computer) 
/ / several most I computer 
i number 
(speciali~Q j~t rend) 
NO PRP NPI ( t rend) Jl 
specialist 
into I trend 
Several most number computer already precede specialist into trend. 
Fig.1 Sentence generation from the axiom and a verb "precede" 
SE(read) ? 
SK(read) 
(specialist)NP ? 
(specialist) ~-~ 
WTl(read) fl 
Y~l ( r e ad ) 
I 
(the)AD/l ~ read N~(specialist) 
ART NO L I 
the specialist 
NPl(paper) 
(the)ADl ~(paper) 
A~T NO I f 
the paper 
The specialist read the paper. 
Fig.2 Sentence generation from NP1 and a noun "specialist" 
Sakai & Nagao 19 
SE II 
SK 
NP BE NPI 
iI i /~\ NPI be 
I) ADI AD NQ PNO R 
\[\[ l ART A~J I 
./ ~o 
these t I 
the red book 
These be the red book. 
SE fl 
SK 
NP BE NPI 
NPI e 
PNO I 
ART NO which be ADJ 
the red book 
These be the book which be red, 
NPI : O.AD.NQ ~ 0.NQ.WHICH BE-AD 
Fig.3 Transformation in a phrase. 
I ~I ~ ART 
0 j °o. 71 ) 
NO computer several I 
number 
I 
NP 
NpIii \] ! alreadypre ede NO 
SS 
WTI NPI ~\[ICH 
PP 
) Jl specialist 
\[ into NO P 
trend 
J\ \[ WTI NP1 
VTII ADIII il 
read ARTI N~ 
the paper 
Several most number computer already precede specialist into trend 
which read the paper. 
SS : ~.WT1,NPI,~.CM,NPI.@ ---~I-2o3-WHICH.7-4 
Fig.4 Transformation between two sentences. 
Sakai & Nagao 20 
#S/:# II 
SK 
NP ÷ sub ~/TI 
NPI AU VTi 
J\ " , AZ enjoy 
I JI 'v" 
the father 
~ obj 
AD Nq 
I\[ II 
ADJ NO + sg I 
! 
fresh breeze 
# the father sg sub h b enjoy fresh breeze s_~ob~j # 
Fig.5 Application of morphophonemic zules as operators. 

REFERENCES 

V.H. Yngve: Random generation of English sentences, 
1961 International Conference on Machine Translation of 
Languages and Applied Language Analysis, London, 1961 

C.J. Fillmore: The position of embedding transformations 
in a grammar, The Ohio State University Rf Project, 
POLA No. 3, 1963 

J.J. Katz & J.A. Fodor: The structure of a semantic theory, 
Language, Vol.39, No.2, 1963 

S. Ceccato: Linguistic Analysis and Programming for Mechanical 
Translation, Gordon and Breach Science Pub. 1961 

M. Nagao: An Approach to The General Theory of Natural Language, 
Preprints for Seminar on Mechanical Translation, 
U.S.-Japan Committee on Scientific Cooperation, Pannel II, 
April 1964 
