PROCESSING OF SYNTAX AND SEMANTICS OF NATURAL LANGUAGE 
BY PREDICATE LOGIC 
Hiroyuki Yamauchi 
Institute of Space and Aeronautical Science, 
University of Tokyo 
4-6-1 Komaba, Meguro-ku, Tokyo 153, Japan 
Summary 
The syntax and semantic analyses of natural lan- 
guage are described from the standpoint of man- 
machine communication. The knowledge based system 
KAUS(Knowledge Acquisition and Utilization System) 
which has capabilities of deductive inference and 
automatic program generation of database access 
is utilized for that purpose. We try to perform 
syntax and semantic analyses of English sentences 
more or less conccurently by defining the corre- 
spondence between the basic patterns of English 
and the extended atomic formula in the framework 
of KAUS. Knowledge representation based on sets 
and logic, the sentence analysis utilizing this 
knowledge are given with some examples. 
i. Introduction 
This paper presents natural language understand- 
ing in man-machine invironments. The syntax and 
semantic analyses program is given almost in 
logical forms of the knowledge based system KAUS 
(knowledge Acquisition and Utilization System). 
KAUS is a logic machine based on the axiomatic 
set theory and it has capabilities of deductive 
inference and automatic program generation of 
database access. 
In natural language understanding, it should be 
required that the syntax analysis is performed 
with association of word semantics. The descrip- 
tions of word semantics are fundamental as well 
as syntax features in language analysis. When 
using natural language to communicate with a ma- 
chine, the understanding of meanings of sentences 
is presupposed to the machine. 
We think of words as representing concept sets 
or property sets, and formalize them into a 
structure called SKELETON STRUCTURE using their 
element-set relationships. In this, we can rep- 
resent semantic categories of words in hierar- 
chical order. The correspondence between natural 
language expressions and system's logical formu- 
las is given in the straightforward manner such 
as 
"X give Y to Z" --- (GIVE X Z Y P) 
where the variables X, Y and I are usually bound- 
ed by their semantic categories. We call this 
set of representations ATOM FORMAT DEFINITION 
SET. Furthermore, causality relations of general 
facts and paraphrases of sentences are given as 
general axioms. Individual facts are also given 
as axioms. We call these sets of representations 
AXIOM SET. Conceptual schemas of databases can 
also be given in the axiom set. The KAUS's 
knowledge base comprises the above three compo- 
nents. 
Utilizing this knowledge base, we try to perform 
the syntax and semantic analyses of natural 
language more or less conccurently in the sense 
that all these processes will be carried out 
through the deductive process in KAUS. At the 
time, the logic program of the analyses written 
in KAUS language is executed by the deductive 
process. 
In the chapter 2, some considerations oh natural 
language(NL) processing are given as prelimina- 
ries. In the chapter 3, the outline of KAUS is 
described, where the knowledge representation, 
the system language and the deductive inference 
rule are presented. In the chapter 4, a rather 
intuitive and straightforward approach to the 
analyses of English sentences is described with 
some examples. A part of the logic program of 
the analyses is also given there. 
2. Some Considerations on NL Processing 
When we are going to construct a natural language 
understanding system, the following four points 
must be clarified: 
i). the main motivation and goal for NL 
processing 
2). knowledge representation suitable for 
NL processing 
3). the aspect of programming language and 
methodology by which the modification, 
updating and extension of the program 
of NL processing can be easily made. 
4). the efficiency of processing 
In the sequel, we will clarify our standpoint 
about the above matters. 
2.1 Motivation and Goal 
At present, the details of the mechanism of hu- 
man language comprehension and knowledge memo- 
rization have not yet been clarified. These are 
still left uncertain. Now, let us consider the 
human process of natural language understanding. 
--389-- 
When we read a sentence, we make an image (mean- 
ing) of each word and match the pattern of the 
sentence predicting what sorts of words will be 
coming to the next. The imagination, predic- 
tion and pattern matching are usually made in- 
voluntarily. Furthermore, if several sentences 
are following, we settle the scene of the dis- 
course and extract the meaning of the current 
sentence with relation to the discourse. That 
is, the word meanings (semantics), the sentence 
pattern (syntax) and the world knowledge are 
fundamental components for the sentence compre- 
hension. 
Our main motivation and goal are to deal with 
the NL processing from the standpoint of man- 
machine communication, and with the above con- 
sideration, to write a syntax and semantics anal- 
ysis program under the knowledge-based system 
~US which has capabilities of deductive infer- 
ence and automatic program generation of data- 
base access. 
2.2 Knowledge Representation 
In natural language analysis, knowledge repre- 
sentation to parse the sentences and extract 
their meanings is most crucial. The various 
kinds of parsers have been revised by experts; 
among them, the ATN parser by W. Woods is wide- 
spread and used elsewhere. It is essentially 
a context-free grammar based one. Further, to 
represent the semantic structure of the sen- 
tences, the frame theory revised by M. Minsky 
is widespread. The system using the ATN or 
ATN-like parser and the frame or frame-like 
structure, present them in semantic networks. 
On the other hand, predicate calculus oriented 
knowledge representation is also available to 
language processing. We will adopt predicate 
calculus oriented representation of knowledge 
for NL processing and database definitions. 
In database applications, predicate logic is of 
greate advantage to define intensional and ex- 
tensional data to which the deductive inference 
rule can be applied. Moreover, it has been 
pointed out that there exist similarities be- 
tween natural language and predicate logic. But 
as the expressive power of first order logic is 
rather restricted, we will extend the formalism 
of first order logic. This formalism will be 
given in the next chapter. 
2.3 Programming Langua$e and Methodology 
To make a computer understand natural language, 
we must program "the understanding mechanism" 
and give "the state of the world" to the machine. 
It should be noticed here that natural language 
contains ambiguous expressions in itself. It is 
usually very difficult how to process ambigui- 
ties of sentences. In general, "ambiguity" may 
be a property of subjects but should not be a 
property of the processing mechan&sm; the solu- 
tion of the subject and the subject itself may 
be given plausibly or uncertainly but not the 
solution mechanism itself. 
With this consideration, we adopt the logic pro- 
gramming method which can be involved in KAUS's 
system language. By this method, we can design 
a sentence analysis program in the top-down 
style. The logic programming with use of the 
KAUS's formal logic is perspicuous, by which the 
modification, updating and extension can be made 
easily. In particular, the correspondences be- 
tween the form of natural language expressions 
and the system's own formulas can be defined 
fairly in the straightforward manner, which are 
referred to in turn by the deductive inference 
and retrieval mechanism to translate the former 
into the latter. 
2.4 Efficiency of Processing 
With respect to the efficiency of processing, 
the program written in KAUS's logic programming 
style may be slightly less comfortable than the 
program written altogether in the procedural 
form becuase of the necessity of the execution 
of deductive retrieval program. However, the 
efficiency would be slightly sacrificed at the 
moment for the sake of clarity of the program. 
3. Outline of KAUS 
In this chapter, we describe briefly about an 
knowledge based system called KAUS(Knowledge 
Acquisition and Utilization System) which was 
realized in accordance with the design philoso- 
phy presented in the paper \[6\] and \[7\]. KAUS is 
constructed on the basis of the axiomatic set 
theory which has capabilities of deductive in- 
ference and automatic program generation for 
database access. It can be also applied to the 
logic programming of the semantic processing of 
natural language. In the following, we focus 
our discussions to the characteristics of knowl- 
edge representation formalism, the deductive 
inference rule and the system language features. 
3.1 Knowledge Representation Formalism 
We think of words as representing concept or 
property sets of things, and organize them into 
a structure on the basis of their set-theoretic 
implications. We will call such a structure 
SKELETON STRUCTURE. Using this formalism, we can 
represent semantic categories of words in hier- 
archical order. For example, the following set 
relations are structurized into Figure i. 
@PERSON- PERSON /\ 
@MaN -- MAN --- WOMAN -- @W~MAN 
@BOY-- BOY GIRL @GIRL 
I I 
#JOHN #MARY 
power set 
-- element 
-- disjoint 
Fiqure I. An Example of Skeleton Structure 
--390 
PERSON D MAN , PERSON ~ WOMAN 
MAN ~BOY , WOMAN ~GIRL , MANn WOMAN = 
MAN ~ #JOHN , WOMAN ~ #MARY (I) 
In the figure, the relation of set inclusion is 
uniformly represented as the element-set rela- 
tions by using the concept of power sets. Then 
the power set of PERSON (we denote it by @PERSON) 
comprises all the subsets of PERSON: 
@PERSON = \[ MAN, WOMAN, BOY, GIRL .... 
.... #JOHN, #MARY .... \] (2) 
Hereupon, let us consider the ordinary first 
order predicate logic at a moment. In the first 
order logic, domains of variables are described 
by using predicates. For example, "every man 
walks" is represented by 
(Vx)\[MAN(x) ...... > WALK(x)\] (3) 
where MAN(x) is used for the domain restriction 
of the variable X. Thus, this restriction can be 
interpreted as "x is an element of MAN". More- 
over, if it were required to answer the question 
"does every boy walks?", the following axiom 
would have to be given. 
(Vx)\[BOY(x) ..... -> MAN(x)\] (4) 
Then, using (3) and (4), the above question can 
be evaluated by Resolution Principle. But such 
descriptions as (3) and (4) are rather cumber- 
some. In place of them, we give the following 
(5) and (6) which have the same interpretation 
as (3) and (4) respectively. 
(Vx/MAN)\[WALK(x)\] (5) 
MAN ~ BOY (6) 
where the prefix (Vx/MAN) in (5) which can be 
included in our axiom set described later de- 
notes "for every x~ MAN" The set relation (6) 
can be also given in our skeleton structure. 
Using both (5) and (6), we can derive the answer 
of the above question faster than the ordinary 
first order logic where such representations as 
(3) and (4) are used. This is the merit of the 
skeleton structure and will be much more clari- 
fied in the section of Deductive Inference Rule. 
Atom Format Definition Set In the first order 
logic, constants, variables and functions are 
recognized as terms. Then, an atom is defined 
as P(t~, t2,..., t~) where P is a n-place pred- 
icate symbol and t~ (i = l,...,n) is a term. A 
formula is defined by using such atoms, logical 
connectives and quantifiers. However, when we 
wish to translate a natural laguage expression 
into a predicate form in the above formalism; we 
cannot directly handle phrases and clauses both 
of which are used as verb objects or verb com- 
plements. Therefore, we extend the atom def- 
inition as follows: 
i). a formula can be permitted as a term 
besides constants, variables and func- 
tions 
2). a function can be used as an atom --- 
we call this type of an atom PROCEDURAL 
TYPE ATOM (PTA), while the other type 
atom is called NONPROCEDURAL TYPE ATOM 
(NTA). (note: because of permitting a 
PTA as an atom, we can perceive that 
our logical formulas afford a sort of 
logic programming facilities.) 
The atom format definition set provide us with a 
definition of a correspondence between some syn- 
tactic features bf natural language and our log- 
ical formulas. In addition, PTA definitions are 
also given in the set. In the figure 2, some 
examples are illustrated , where all of the used 
atom formats conform to the following standard 
format definitions: 
NTAI : (V S X, X2 P) (7) 
NTA2 : (A 0 V) (8) 
PTA ,: '(F Y,--- Y~; X~--- X~) (9) 
NTAI is usually used for representing a simple 
sentence, while NTA2 is used for representing an 
Adj.× Noun phrase. In the NTA\] definition, V 
denotes a predicate symbol. S, XI, X~ and P are 
terms, of which S is usually used as the subject 
of V, and P is a formula which denotes the modi- 
fiers concerning time, locus, goal, reason and 
manner of V. Some of the terms may be omitted 
according to the syntactic feature of a sen- 
tence. In the NTA2 definition, (A 0 V) de- 
notes "the Attribute of Object is Value". 
Finally, in the PTA definition, F denotes a 
function name which takes input variables Xt ,.. 
.., X~ and output variables Yi ,..., Y~ • It 
must be remarked here that some of the variables 
in (7), (8) and (9) may be bounded by their do- 
mains or semantic categories, but they were 
omitted for the sake of simplicity. 
"X give Y to Z ..... (GIVE X Z Y P) 
"red X ..... (COLOR X RED) 
"two X' .... (CARDINAL X 2) 
"X on Y" --- (ON X Y P) 
"X be father of Y" --- (FATHER Y X) 
"X + Y = Z" --- '(SUM Z; X Y) 
"X / Y : Z ..... '(DIV Z; X Y) 
"sin(X) = Y ..... '(SIN Y; X) 
Figure 2. Examples of Atom Definition 
Axiom Set The axiom set comprises descriptions 
of the world which specify general facts such 
as causality relations of things and sentence 
paraphrasing, and individual facts such as "John 
gave a red car to Mary". These facts are repre- 
sented by formulas in the standard form 
Q (~Fv G ) (I0) 
--391- 
Where Q represents quantification of variables, 
taking the form (q, V~/d, )(qzvz/d~ )... (qKv,/dK), 
in which q~ denotes a quantifier symbol V or 3, 
vz is a variable name and dz indicates the do- 
main of the variable VZ. F and G represent pre- 
mise and conclusion respectively in the formula; 
namely, F --~ G. More clearly, F is consti- 
tuted from NTAs and PTAs by using logical con- 
nectives A , V ands, and also the same is G 
except that no PTAs are permitted. The inter- 
nal representation of a formula in KAUS is an 
AND-OR tree to whose root node the quantifica- 
tion part Q is attached (see Figure 3). 
Now, we have here several examples* • 
Example I. We can represent a causality rela- 
tion of "give" as follows: 
(Vx/PS)(Vy/PS)(Vz/PO)\[~(GIVE x y z (TIME 
past))v((HAVE y z (TIME present))A 
~(HAVE x z (TIME present)))\] 
This says that "if X ~ PerSon) gave Z ~ Physi- 
cal-Object) to y (~PerSon), then y has Z and X 
has not Z at the moment". 
Example 2. We can paraphrase "x give z to y" 
to "y receive z from x" • 
(Vx/PS)(Vy/PS)(Vz/PO)(Vp/VMOD) 
\[--~(GIVE x y z p)v (RECEIVE y z x p)\] 
where p/VMOD specifies that p is any formula 
which qualifies the predicates. 
Example 3. We can express the meaning of 
"float" as follows : 
(Vx/PO)(Vy/LQ)(Vu/RNUM)(Vv/RNUM)(Vp/VMOD) 
\[~((SPEC-GR x U)A (SPEC-GR y v)A'(LT u v)) 
v (FLOAT x y p)\] 
This says that "if the specific gravity of × is 
U, the specific gravity of y is V and U is less 
than V , then X float on y". 
Example 4. The fact, "John gave a red car to 
Mary" is represented as follows : 
\[(GIVE #JOHN #MARY CAR (TIME PAST)) 
A (COLOR CAR RED)^ (QF CAR A)\] 
Example 5. The fact, "John drinks coffee after 
John plays tennis" is represented as follows : 
(DRINK #JOHN COFFEE (TIME (AFTER (PLAY #JOHN 
TENNIS (TIME PRESENT))))) 
All of these general/individual facts can be put 
into the axiom set and they will be referred to 
through the deductive retrieval mechanism in 
KAUS. 
*). The real notation implemented by KAUS is 
partly modified (see 3.3 System Language). 
@W, \[~\] --- Q, @w3 
i /\ I 
W ..... F G, , ----~ W3,-- W~2 
/ 
W~,-- W~ ~ G a ~ W4,-- W4~-- W~ 
Fiqure 3. Interconnection betwe.e.ll 
Axioms and Skeleton Structure 
3.2 Deductive Inference Rule 
In the previous section, we have given the for- 
malism of knowledge representation, where the 
concepts and properties of words can be partial- 
ly ordered according to their set-theoretic im- 
plications. Moreover, we have defined the corre- 
spondence between the syntax features of natural 
language and logical expressions somewhat in the 
straightforward manner. We have also described 
how to represent universal/individual facts in 
our axiom set. However, it must be stressed 
here that these three types of knowledge are 
not independently presented in the knowledge 
base but interrelated each other with respect to 
domains of variables of atoms; that is, the bi- 
directed access paths to each constituent of the 
knowledge base are mutually defined (see Figure 
3). By making much use of these access paths, 
the knowledge only necessary for deduction can 
be retrieved efficiently. The deduction is 
performed by the inference rule comprising the 
following four components: 
i). S R: Selection Rule for a literal 
resolved upon 
2). TIC: Test for Implicative Condition 
3). R R: Replacement Rule 
4). T T: Test for Termination of deduction 
SR(Selection Rule) A literal resolved upon is 
selected from the query tree. The selection 
criterions are as follows. First, the left-to- 
right search for a non-evaluated NTA(non-proce- 
dural type atom) is made in the tree. When it 
is found, it is checked whether a constant or 
3-variable is contained in the NTA. If it does 
the case, this NTA is took as a candidate re- 
solved upon. If no such a literal is presented 
in the tree, certain counterplans are executed. 
The details would be omitted here. 
TIC(Test for Implicative Condition) After the 
candidate literal to be resolved upon is decid- 
ed, axioms which contain the literals with the 
same predicate symbol as the candidate literal's 
but with the opposite sign are retrieved in the 
knowledge base. At that occasion, the search 
domain is narrowed to the subset of the axiom 
set by indexing the variable domains of the 
" 392- 
literal resolved upon. The skeleton structure 
is referred to for this purpose. 
Now, let us denote P as such a literal searched 
in the knowledge base and C as a candidate lit- 
eral resolved upon. TIC checks the implicative 
conditions of P ---> C. One of the conditions 
to be checked is the set relation of the P's 
variable domains to the corresponding C's vari- 
able domains. The Table 1 shows this condition. 
Let us consider an example: 
P: (~y/GIRL) (Vx/MAN)(vp/VMOD) 
\[(LOVE x y p)\] (ll) 
C: (Vu/BOY) (~v/WOMAN) (vq/VMOD) 
\[(LOVE u v q)\] (12) 
P says that "a certain particular girl is loved 
by every man". On the other hand, C says that 
"every boy loves some woman" in the ordinary 
sense. In this case, P ---> C can be establish- 
ed because all the corresponding variable pairs 
(X, U), (y, V) and (p, q) satisfy the condition 
in Table i. However, if P and C are 
P: (Vx/BOY)(~y/WOMAN)(vp/VMOD) 
\[(LOVE x y p)\] (13) 
C: (Iv/WOMAN) (Vu/BOY) (Vq/VMOD) 
\[(LOVE u v q)\] (14) 
then, P ---> C can never established in spite of 
satisfaction of Table i, because the most gener- 
al significant unifier does not exist, TIC 
tests this condition, too. 
RR(Replacement Rule) Let F be an axiom contain- 
ing P and let G be the query clause containing 
C. RR generates a new query clause R after 
substitution o has been done. That is, Co is 
replaced with (Fo - Po), resulting in 
R: QR (G~ - C~) .(Fo - P~) (15) 
where • denotes that (Fo - Po) should be attach- 
ed to the place where C was resided. Substitu- 
tion o is defined in Table I. 
TT(Test for Termination) When all the nodes in 
the query tree have been evaluated, deduction 
is terminated. Though we have said that all 
NTA nodes in the query tree are evaluated by 
means of the above SR, TIC and RR, we have not 
said almost anything about PTAs. We must now 
mention about them. PTAs in the query tree are 
in general, ready to be evaluated just at the 
time when all its input variables have been 
filled up with their values. Then, the evalua- 
tion is performed by calling the subroutines 
attached to PTAs. 
3,3 System Language 
The system language provide us with fucalties 
of constructing knowledge base and relational 
database. Besides that, it can be used as sub- 
stitute for logic programming and query repre- 
Table i. Implicative Condition 
QF~ Q~ QA~ XR~ CONDITION 
V V V Xp~ n X~ X~ n X~ :~ 
V 3 3 X~,. X~: ~ Xz; 
3 V 3 X~ XP,. ~ XZ~ 
--- V V X~ ......... 
--- 3 3 X~ ......... 
V --- V X~ ......... 
3 --- 3 Xr~ ......... 
3 3 (non-implicative) 
P: Qp(Xp, Xr, ~ " -CT: negated form of 
C: Q~ (X~, X~, X~ query C 
R: QR(X~, ~= ~ R: replaced form of 
sentations. The syntax of the system language 
has already been supposed in the section 3.1. 
At this section, we present the really imple- 
mented features of the language briefly. De- 
tails are excluded because this is not the pur- 
pose of this paper. 
The syntax of the system language is based on 
the tuple (V0, Vi,... , V n) where V 0 is either a 
system command name, a predicate symbol or a 
variable whose domain is a PREDICATE; Vi(i ¥0 ) 
is a term as which a string and a numerical con- 
stant and a formula are permitted. The system 
language has the following characteristics that 
are not included in the first order logic: 
i). Variables can be explicitly specified. 
For example, \[AX/MAN\]\[EY*/WOMAN\] 
where A and E denote the universal and 
existential quantifier respectively, 
and the symbol * attached to Y denotes 
that Y is a query variable. 
2). A predicate symbol itself can be a 
variable. For example, 
\[EX*/PRED\]($X, #JOHN, #MARY)? 
where the symbol $ attached to X denotes 
that X is a variable and the symbol ? 
denotes that the expression is a query. 
3). A recursive expression can be permitted. 
4). A procedural type atom PTA --- a func- 
tion --- can be permitted as an atom. 
For this reason, an aspect of logic 
programming is obtained. 
4. Syntax and Semantics Analyses 
In syntax and semantics analyses, word meanings, 
sentence patterns and world knowledge are fun- 
damental. Characteristics of the sentence anal- 
ysis partly depend on knowledge representation 
used. As described in the chapter 3, we repre- 
sent knowledge in the framework of sets and 
logic in KAUS. The characteristics of our sen- 
tence analysis program are that, during the 
analysis, we use the rather direct correspond- 
ence between the basic sentence patterns (syn- 
tax) of natural language and the extended atomic 
formulas in KAUS, and that the pattern matching 
method can be used together with the deductive 
- 393 
inference rule. A more characteristic is that 
words in a clause are put into four groups pre- 
serving the word order, each of which contains 
subjects, the main verb, direct objects/comple- 
ments and indirect objects/complements respec- 
tively. In this chapter we present the English 
sentence analysis program using the above method. 
4.1 Descriptive Presentation of the analysis 
The analysis takes the following steps: 
i. The parts of speech of each word in the in- 
put sentence and atom format definitions at- 
tached to each predicative/attributive word 
such as verbs, adjectives, prepositions and 
the others are fetched from the knowledge 
base. 
2. The input sentence is partitioned into a set 
of clauses by considering the positions of 
conjunctions, relative pronouns and punctua- 
tion marks. 
3. Words in a clause are put into four groups 
preserving the word order in the clause, each 
of which may contain subjects, the main verb, 
direct objects/complements respectively. For 
this purpose, the correspondence relation be- 
tween the NL syntax and the atomic formula 
syntax is utilized. 
4. Each phrase in the four groups is decided 
whether it indicates qualification of nouns 
or of predicates. 
5. If there are AND/OR conjunctions in a clause 
it is decided whether they are in the scope 
of a preposition or not. 
6. After all of the words in a clause have been 
interrelated each other, the remaining 
clauses are processed by repeating 3 to 5. 
7. After all of the clauses have been processed, 
substantiation of each personal and demon- 
strative pronoun is established. 
8. In consequence, the extended formula deduced 
from the input sentence is obtained. 
To get comprehension of the above description, 
let us consider the next example: 
IN HIBIYA-PARK, JOHN MET JACK AND MARY. (16) 
This is the case that the sentence comprises 
only one clause. On this sentence, the basic 
sentence pattern of "meet", that is, "PERSON1 
meet PERSON2", is fetched from the atom format 
definition set. The atom format definition of 
the preposition, "in" is also fetched; that is, 
"THING in PLACE". But this will not be used in 
this case. Then, grouping of words in the sen- 
tence is made according to the pattern "PERSON1 
meet PERSON2", resulting in 
group 1 : 
group 2 : 
group 3 : 
group 4 : 
\[ IN HIBIYA-PARK, JOHN* \] 
\[ MET \] 
\[ JACK* AND MARY* \] 
\[ ¢ \] ; an empty group 
where the words marked with the star symbol * 
denote the instances of PERSON1 and PERSON2. 
This can be established by using the skeleton 
structure in which JOHN, JACK and MARY are de- 
fined as elements of PERSON. In the next place, 
the phrase "IN HIBIYA-PARK" is decided as indi- 
cating qualification of the verb "MET". The 
conjunction "AND" is then determined as to be in 
the scope of the conjunction of direct objects 
of "MET". As the final result, we obtain the 
following formula: 
\[(MEET JOHN JACK (TIME PAST)~(PLACE (IN 
HIBIYA-PARK)))A(MEET JOHN MARY (TIME PAST) 
A(PLACE (IN HIBIYA-PARK)))\] (17) 
Let us consider a more complex sentence which 
contains a relative clause and a personal pro- 
noun in it: 
JOHN GAVE A RED CAR TO MARY WHO HE LOVES. (18a) 
JOHN GAVE A RED CAR TO MARY WHO LOVES HIM.(18b) 
In this case, both of the sentence (18a) and 
(18b) are split into the two clauses respective- 
ly; among which, each nucleus clause of (18a) 
and (18b) is the same, that is, "John gave a red 
car to Mary"; but the relative clause in (18a) 
means that "John loves Mary" while that in (18b) 
means that "Mary loves John". On the nucleus 
clause, the basic sentence pattern "PERSON1 gave 
THING to PERSON2" for "give" and "red THING" for 
the attributive adjective "red" are fetched from 
the knowledge base. Then, grouping of words in 
the clause is made according to the basic sen- 
tence pattern, resulting in 
group 1 : \[ JOHN* \] 
group 2 : \[ GAVE \] 
group 3 : \[ A RED CAR* \] 
group 4 : \[ TO MARY* \] 
In the next place, the pattern "red THING" is 
applied to "RED CAR" in the group 3, and in 
consequence, the formula (COLOR CAR RED) is de- 
rived, We translate the indefinite article "a" 
into the formula (QF CAR A) to denote that "car" 
is qualified by "a" in the clause. Thus, all 
the semantic relations among words in the nucle- 
us clause "John gave a red car to Mary" have 
been derived: 
(GIVE #JOHN #MARY CAR (TIME PAST)) 
A(COLOR CAR RED)A(QF CAR A) (19) 
The relative clause in (18a), "who he loves", 
is transformed to (LOVE HE MARY (TIME PRESENT)), 
and "who loves him" in (18b) is transformed to 
(LOVE MARY HE (TIME PRESENT)). This can be 
attained by making use of the basic sentence 
pattern, "PERSON1 love PERSON2". In case of 
"who he loves", the following four word groups 
are initially generated by using this pattern. 
--- (WHO) HE LOVES --- 
group 1 : \[ (WHO) HE* \] 
group 2 : \[ LOVES \] 
- 394 
group 3 : \[ # \] ; an empty group 
group 4 : \[ \] ; not to be used 
Then, taking account of the state of the group 1 
and the group 2, the antecedent of "who" is de- 
cided to be the direct object of "love". In the 
last place, the personal pronoun "he" is sub- 
stantiated by"John". The similar discussion 
can be given to the case, "who loves him", and 
therefore, it is omitted here. The final results 
of the analysis of (18a) and (18b) are 
(GIVE #JOHN #MARY CAR (TIME PAST)) 
A(COLOR CAR RED)^ (QF CAR A) 
^(LOVE #JOHN #MARY (TIME PRESENT)) (20a) 
(GIVE #JOHN #MARY CAR (TIME PAST)) 
A (COLOR CAR RED) A (QF CAR A) 
A(LOVE #MARY #JOHN (TIME PRESENT)) (20b) 
So far, we have been concerned with declarative 
sentences. With regard to interrogative and 
imperative sentences, we transform them to de- 
clarative sentences prior to the successive 
analysis of them. Further, passive sentences 
have not been taken into account of hitherto. 
The passive voice is especially used in sentences 
in which it is unnecessary or undesirable to 
mention the agent, though the agent may be ex- 
pressed by means of an adjunct with "by,'. The 
verbal meaning of the passive voice may also be 
brought out by adjuncts expressing other adverb- 
ial relations, such as time, manner, cause or 
instrument. A passive sentence may be analyzed 
with adaptation of the special treatment of the 
verb "be" followed by a past participle. For 
example, 
JOHN IS LOVED BY MARY. 
---> \[JOHN\] IS \[LOVED\] \[BY MARY\]. 
---> MARY LOVES JOHN. 
---> (LOVE #MARY #JOHN (TIME PRESENT)) (21) 
THE BOOK IS WRITTEN IN ENGLISH. 
---> \[THE BOOK\] IS \[WRITTEN\] \[IN ENGLISH\]. 
---> \[$X\] WRITE THE BOOK IN ENGLISH. 
---> (3x/PS)\[(WRITE x BOOK (MANNER 
(IN ENGLISH))A(TIME PRESENT)) 
A(QF BOOK THE)\] (22) 
where the \[ \] is used to denote special atten- 
tion to readings. The special treatment of the 
verb "be" is not only introduced to passive sen- 
tences but also to the other fragments shown in 
Table 2. For example, a sentence pattern with 
the formal subject Of "be" is treated as 
IT IS EASY TO PLEASE JOHN, 
---> \[TO PLEASE JOHN\] IS \[EASY\]. 
---> (GRADE-DIFFI (TOINF (TO PLEASE #JOHN)) 
EASY) (23) 
IT IS SNOOPY THAT HAS STOLEN THE FISH. 
---> \[THAT HAS STOLEN THE FISH\] IS \[SNOOPY\]. 
---> (STRESS-VAL (STEAL #SNOOPY FISH 
(TIME PRES.PERFECT)) #SNOOPY) 
A(QF FISH THE) (24) 
It should be denoted here that both of (23) and 
(24) have been transformed to a similar form in 
terms of (A 0 V) atoms. 
Table 2. Treatment of the Fragment of 
"X be Y " 
X be Y Atomic Formula 
N be Adj. (ATTR N Adj.) 
N~ be Prep. N~ (PREP N, N2) 
N, be ReI.N of N~ (REL.N N~ N,) 
ReI.N of N, be N~ (REL.N N, N~) 
N, be N~ (ATTR N, N~) 
THERE be N, Prep. N~ (PREP Nf N~) 
IT be Adj. TO-INF (ATTR TO-INF Adj.) 
IT be Z THAT-CLAUSE (ATTR THAT-CLAUSE Z) 
N be P.P. Adjunct (PRED S U V P) 
note. N : Noun 
Adj. : Adjective 
Prep. : Preposition 
ReI.N : Attributive noun 
TO-INF: To-infinitive 
P.P. : Past Participle 
4.2 Logic Program 
By using a logic programming method, we can 
clearly show what should be done in the prog- 
ram depending on an approach taken for the sen- 
tence analysis, in which modification, updating 
and extension can easily be made. In the pre- 
vious section, we have shown our approach by 
which a sentence may be analyzed rather intui- 
tively and straightforwardly in the framework 
of KAUS. According to this method, we present, 
in this section, a part of the program by which 
an input clause is analyzed yielding an extended 
formula in KAUS. The execution of the program 
is undertaken by the deductive retrieval process 
with use of the merits of our knowledge repre- 
sentation. In the following representation of 
the program, we made convention that the top 
formula is concluded if all of the successive 
formulas indented are evaluated (proofed). 
Program 
(LOGICAL-FORM EF S) 
'(GET-VINF VP MVERB ; S) 
'(GET-DON SDOM XDOM YDOM ; VP S) 
'(CREATE-GROUP SGP VGP XGP YGP ; 
SDOM VP XDOM YDOM S) 
(ARGUMENT SUBJ SDOM SGP) 
(ARGUMENT XOBJ XDOM XGP) 
(ARGUMENT YOBJ YDOM YGP) 
(NOD-ATOM SHOD SUBJ SGP) 
(NOD-ATOM XMOD XOBJ XGP) 
(NOD-ATOM YMOD YOBJ YGP) 
(VMOD-ATOM VMOD VGP SGP XGP YGP) 
(KERNEL-S KS VP MVERB SUBJ XOBJ YOBJ VMOD) 
'(CNCT-LOGICAL EF ; KS SHOD XMOD YMOD) 
'(ALL-MARKED ; S) 
395- 
(ARGUMENT EARG EDOM EGP) 
'(MARK EGP ; EDOM EGP) 
'(UNMARK-PREP-OBJ EGP ; EGP) 
'(MAKE-SET EARG ; EGP) 
(ARGUMENT EARG EDOM empty) 
'(SEARCH REL ; INPUT) 
'(NON-EQ REL WHERE) 
'(ANTEC X ; REL INPUT) 
'(PRE-W Z ; REL) 
'(SYN-CAT K ; Z) 
'(NON-ELM K PREP) 
'(PUT-ELM EARG ; X) 
(NOD-ATOM EMOD EARG EGP) 
'(DET-MOD DET ; EARG EGP) 
'(ADJ-MOD ADJ ; EARG EGP) 
'(NPREP-MOD NPREP ; EARG EGP) 
'(MAKE-EF EMOD ; DET ADJ NPREP) 
(VMOD-ATOM VMOD VGP SGP XGP YGP) 
(TENSE-NOD TENSE ; VGP) 
(AUX-MOD AUX ; VGP) 
(VPREP-MOD VPREP ; VGP SGP XGP YGP) 
(ADV-MOD ADV ; VGP SGP XGP YGP) 
(TO-INF-MOD INF VGP SGP XGP YGP) 
(MAKE-EF VMOD ; TENSE AUX VPREP ADV INF) 
__= 
********************************************* 
GP DON 
/ \ J  EDOM VGP EGP VP /i\ 
SDM~XDOM YDOM SGP XGP YGP 
ARG " ARG 
MVERB I 
su J XO " OBJ 
VMD~ EMOD Jl 
SMOD XM~D~YMOD 
********************************************* 
IN HIBIYA-PARK, JOHN MET JACK AND MARY. 
VGP = \[MET\] 
SGP = \[IN HIBIYA-PARK, JOHN\] 
XGP = \[JACK AND MARY\] 
YGP = \[@\] ; not used 
VP = \[VPI\] cf. VPI: Verb × Direct Object 
SDOM = \[PERSON\] 
XDOM = \[PERSON\] 
YDOM = not used 
MVERB = \[MEET\] 
SUBJ = \[JOHN\] 
XOBJ = \[JACK, MARY\] 
YOBJ = not used 
EF = \[(MEET #JOHN #JACK (TIME PAST) 
A(PLACE (IN HIBIYA-PARK))) 
A (MEET #JOHN #MARY ....... )))\] 
Figure 4. A Sample Process by the Program 
The first program denotes that an input clause S 
is translated into an extended formula EP by 
evaluating the successive thirteen indented for- 
mulas, of which the ARGUMENT formulas are in 
turn to be replaced with the premises of the 
second or the third ARGUMENT program, or the 
other ARGUMENT programs that are not exhibited 
here, by the deductive inference rule. Then, 
these premises are in turn to be evaluated. The 
third ARGUMENT program exhibited above is con- 
cerned with a relative clause except where- and 
prep.-relative-noun clauses. The other NTAs in 
the first program may be evaluated in the simi- 
lar way. The figure 4 shows a sample process of 
the program, where the main terms used in the 
program are also shown by categorizing them 
hierarchically in the skeleton structure. 
Conclusion 
We have described the syntax and semantics anal- 
yses of natural language (English) within the 
framework of the intelligent man-machine system 
KAUS. The more the volume of data related to 
the sentence analysis is enlarged in the knowl- 
edge base and the sentence analysis program it- 
self is extended, the more the class of accept- 
able sentences will be broadened. This may be 
ensured by the method described hitherto. 
Acknowledgement 
The author is indebted to associate professor 
Setsuo Ohsuga of University of Tokyo, who gave 
the author helpful suggestions in the presenta- 
tion. 

REFERENCES

\[i\]. Bertram Bruce, Case Systems for Natural 
Language, Artificial Intelligence 6, 1975. 

\[2\]. C.L. Chang & R.C.To Lee, Symbolic Logic 
and Mechanical Theorem Proving, Academic 
Press, 1973. 

\[3\]. Hendrix G.G. et al., Developing a Natural 
Language Interface to Complex Data, Proc. 
3rd VLDB, Part II, pp.37-58, 1977. 

\[4\]. Herve Gallaire & Jack Minker (edited), 
Logic and Databases, Plenum Press, New 
York, 1978. 

\[5\]. Robert F. Simmons, Some Relations between 
Predicate Calculus and Semantic Net Repre- 
sentations of Discourse, 2nd IJCAI, 1971. 

\[6\]. S. Ohsuga, Semantic Information Processing 
in Man-Machine Systems, Proc. IEEE, 1977. 

\[7\]. S. Ohsuga, Toward Intelligent Interactive 
Systems, Proc. The IFIP W.G. 5.2 Workshop 
Seilac II on Methodology of Interaction, 
North Holand, 1979. 
