AUTOMATED PROCESSING OF MEDICAL ENGLISH 
Int ro duct ion 
The present interest of the scientific community in automated 
language processing has been awakened by the enormous capabilities of 
the high speed digltal computer. It was recognized that the com- 
puter which has the capacity to handle symbols effectively can also 
treat words as symbols and language as a string of symbols. 
Automated language processing as exemplified by current 
research, had its origin in machine translation. The first attempt 
to use the computer for automatic language processing took place in 
1954. It is known as the "IBM-Georgetown Experiment" in machine 
translation from Russian into English. (I ,2) 
The experiment revealed the following facts: 
a. the digital computer can be used for automated language pro- 
cessing 2 but 
b. much deeper knowledge about the structure and semantics of 
language will be required for the determination and semantic 
interpretation of sentence structure. 
The field of automated language processing is quite broad; it 
includes machine translation, automatic information retrieval (if 
based on language data), production of computer generated abstracts, 
indexes and catalogs, development of artificial languages, question 
answering systems, automatic speech analysis and synthesis, and 
others. 
Approaches to automatic information retrieval, quantitative 
studies of generic relations between languages and style analysis~ 
have been based to a great extent on statistical considerations~ such 
as frequency counts of linguistic units (phomemes, morphemes~ words, 
fixed phrases). In each of these approaches linguistic analysis was 
considered to be a useful but insufficient method for automated 
information processing because of the many unresolved problems in 
language analysis. 
Implementation of statistical techniques for automated indexing, 
classification and abstracting has proved useful despite certain 
limitations caused by our lack of knowledge of language. 
Some of the major problems in language processing are: 
a. There is no method for Storing in a computer the speaker's 
knowledge of the universe. 
b. Syntactic and semantic ambiguities pose severe difficulties 
for implementation. 
c. Difficulties associated with the design of a general purpose 
formal semantic language (intermediate language) into which 
an input natural language could be mapped by an algorithm. 
d. Recognition and interpretation of logical inferences which 
are contained implicitly in natural language. 
e. Lack of computer-oriented dictionaries and microglossaries. 
In recent years mathematically oriented studies'of the nature 
of natural languages have been directed to the development of formal 
models of grammars, such as context free r context sensitive and 
transformational grammars. Formal characteristics of these models of 
grammars, their generative power, and their adequacies and inadequa- 
cies may be found in the literature (3, 4, 5, 6, 7). 
Several noted scientists, such as Bar-Hillel (8), have expressed 
a pessimistic view in regard to practical implementation of machine 
translation. 
Nevertheless, there is merit in continuing efforts for more 
fundamental research in the area of formal and applied linguistics 
and computer applications. Even if we are not able to resolve all 
the problems in language processing at once, limited goals can be 
attained and tested for validity by design of a model for language 
processing within a restricted language domain, such as medicine. 
Some Characteristics of Medical English 
Aware of the many problems associated with automated processing 
of natural language, we have limited our efforts, for the present, to 
the language domain used in pathology diagnoses, a subset of Medical 
English. 
Medical diagnosis may be described as the process used by the 
physician to determine the nature of disease, or as the art of dis- 
tinguishing one disease from another. The name which is assigned to 
a disease ~plies the unique configuration of signs and symptoms 
believed to be characteristic of the condition which has been diag- 
nosed. The diagnosis can be regarded as a summary of the more com- 
plete medleal document in a conventionalized medical style. 
Medical diagnoses are characteristically free of verb phrases. 
The copulative verb "to be" is frequently implied by the use of 
comma. Often~ the pseudosentence structures appear to be grammati- 
cally illogical. Nevertheless~ these structures carry semantic mean- 
ing and are generally understood by others in medicine. Modifiers 
frequently occur in discontinuous sequence with the nouns they modify. 
Anaphoric expressions are eo~auonplace. 
The terminology consists of a mixture of Latin~ Greek and Eng- 
lish derivatives. Not uncommonly~ diagnostic statements exhibit 
features of all three languages. Evidence of heterogeneous linguis- 
tic origin is also found in single word forms. The language is rich 
in the use of compound word forms which are segmentable into single 
constituents. 
The distinctive semantic features of diagnostic statements may 
be categorized as follows: 
• anatomic site affeeted~ or body system involved in 
the disease process; 
• disease condition, including structural changes ranging 
from gross observations to intracellular ultrastructural 
changes; 
• causative agent of the abnormality; 
• disease manifestations~ including physiological and 
chemical changes~ observable manifestations, and 
symptoms reported by the patient; 
• therapeutic agents or processes used; 
• causal relationships among disease entities; 
• method or souree of diagnosis. 
Two or more of these distinctive semantic features may be com- 
bined.in a single conceptual unit, e.g., "measles" implies both the 
specific infectious disease manifested, and the etiology, the rubeola 
virus; while "pneumonia" describes the inflammatory disease process 
or condition, as well as the anatomic site affected,, lung. 
On the other hand, the precise designation of the loeation at 
which a disease entity has manifested itself may require a complex 
statement for adequate description of the semantics relative to ana- 
tomic site affected, e.g., a lesion may be found in the "apicoposte- 
riot segment in the upper division of the upper lobe of the lung." 
After mentioning some of the pecularities of Medical English we 
will turn to the description of the system for automated processing 
of Medical English which is now under development at the Division of 
Computer Research and Teehnology, National Institutes of Health. (9, 
I0, II, 122 13) 
Systematized Nomenclature of Pathology 
In any type of automated language processing at least two basic 
components are required: 
a. Lexicon 
b. Grammar 
Experience in machine translation revealed that commercial dic- 
tionaries consisting mostly of word lists are not suitable for auto- 
mated language processing. Their main disadvantages are that they 
are out-of-date, incomplete and inaccurate for the purpose of mor- 
phological~ syntactic and semantic analysis. 
In our work we have been using as a lexicon base the System- 
atized Nomenclature of Pathology (SNOP) (14), the structure of which 
is described below: 
SNOP is a special purpose lexicon created by pathologists to 
assist them in the organization and retrieval of information. The 
SNOP language consists of a relatively rich word vocabulary and a 
primitive grammar. 
A term or conceptual unit is listed in only one of the four 
semantic categories of the vocabulary and is assigned a unique numer- 
ical code within the given information class. 
The four semantic categories of the SNOP are: 
Topography (T) - the body site affected 
Morphology (M) - the structural changes resulting from 
disease 
Etiology (E) - the causative agents (micro-organisms, 
drugs and chemicals) 
Function (F) - the physiological manifestations associ- 
ated with disease, including symptoms and 
a limited number of specific infectious 
diseases. 
The conceptual units in each category are information content 
words which are used by the pathologist to convey in a condensed 
form the concept being described. The basic syntactic structures of 
the dictionary entries are: single nouns, adjective phrases, attri- 
butive noun phrases and participial phrases. 
The code of dictionary entries is divided into four separate 
and independent fields T, M, E, F. Within a given field, terms are 
assigned a four digit number. The first digit refers to the section 
of the field, while the other three digits indicate progressively 
finer subdivisions. These groupings reflect natural relations among 
the terms insofar as possible. The code structure permits selection 
of more specific terms by moving down the list and of more generic 
terms by moving up the list (see Appendix I). 
Semantic linkage pointers are also incorporated in the diction- 
ary which enable the cross-referencing of information within the same 
or another category. 
We intend to extend the semantic codes and also to allow the 
insertion of additional semantic information. The individual codes 
will be linked together to represent the content of the messages and 
messages will be linked to represent the content of a diagnosis or 
other relevant medical document. 
Grammar 
Automated processing of Medical English (APME) consists of a 
series of computer programs which given as input a body of medical 
text, will produce as output~ a linguistic description and semantic 
interpretation of the given utterance. 
In the initial stage of our research, our attention was focused 
on patholoEy diagnoses since this domain of discourse intersects with 
all other medical specialty areas. Consequently, research work in 
this area should be applicable to other medical specialties. 
The APME parsin E algorithms consist of series of interrelated 
linguistic and programming operations which are described briefly as 
follows : 
Morphological Analysis embodies the identification and transfor- 
mation of terminal morphemes and recognition of a limited set of pre- 
fixes. The input to the morphological analysis and subsequent mor- 
phosyntactic transformations is the unedited Medical English of a 
pathologist. The transformation procedure consists of a set of rules 
by which the adjective forming suffixes are substituted by a set of 
nominalizing suffixes (adjective to noun transforms) or plural nominal 
suffixes are substituted by their singular allomorphs (plural to 
singular noun transforms) and, finally~ transforms of nouns to their 
synonymous or near-synonymous forms (i.e., DIS~TERY--~DISENTERIA). (12) 
The main reason for the transformation of terminal morphemes is 
to provide a means for the successful retrieval of word forms which 
occur in the text in a derivative form but are listed in the SNOP die- 
tionary in an alternative form. 
We have been preparing rules for morphosemantie segmentation of 
composite word forms which aremainly derived from Greek or Latin. 
The decomposition implies the recognition of constitutents, the order 
of constituents, how they are intertwined and the assignment of 
semantic value to productive components. 
For example, the composite 'BRONCHITIS' is segmented into two 
components CI-T and C2-M, namely: 
a. Terminal morpheme-ITIS (C2-M) which is the semantic marker 
of inflammation process (semantic category 'M'), and 
b. The stem morpheme BRONCH(US) (CI-T)~ the site of the body 
where the inflammation process occurs (semantic category 
'T'). The semantic structure of the class of word forms 
such as 'BRONCHITIS' is expressed as 
W (T, M) -+ (CI -T) + (C2-M) 
where the order of semantic components CI-T and C2-M is 
fixed and cannot be reversed. 
It is expected that the analysis of the semantic structure of 
components into their semantic constituents and the formalization of 
the structural and semantic relationships among them will be useful 
for the preparation of computer-oriented medical micro-glossaries. 
S~ntactic and Semantic Analysis 
Syntactic and semantic analysis of medical text involves the 
determination of sentence structure and its semantic interpretation. 
This problem can be resolved by the implementation of trans- 
formational rules by which different types of noun Phrases 
will be transformed into a set of semantically equivalent 
phrases. For example: 
NP-~ N 1 + N 2 (muscle atrophy)--~ N 2 + OF + N 1 (atrophy of 
muscle)~ N 2 + COMMA + N 1 (atrophy, muscle) ~ (N2--> A) ÷ 
N 1 (atrophic muscle)~ (NI--> A) + N 2 (muscular atrophy). 
Two tree diagrams are required for the representation of cumu- 
lative derivations in an abbreviated form. 
NP 
N / ~N (muscle atrophy) 
(A ---+N) N (muscular atrophy) 
(A --~ N) of N 
colm~a 
2. 
(atrophic muscle) 
(atrophy of muscle) 
(atrophy, muscle) 
The problem of paraphrasing is closely related with auto- 
matic recognition of synonymous or nearly-synonymous expres- 
sions which are not found in the lexicon. Even if we would 
be able to assign an approximative value to so-called "unknow~ 
terms" by the implementation of deductive rules, the physi- 
cian o~ the expert in the related field will have to make the 
decision about the synonymity of the term in question. 
12 
3. 
4. 
In both cases it is assumed that the semantic content of 
the message is understood. Otherwise, if the transforma- 
tional rules were based only on the formal structure of 
noun phrases they could generate phrases which are not com- 
patible with the semantic content of the message. Conse- 
quently, the transformational or paraphrasing rules should 
not be based only on syntactic features but also on semantic 
features of components of the given semantic unit~ i.e., on 
the deep structure of the message. 
The boundaries of the semantic unit do not correspond, 
necessarily, to the boundaries of the noun phrase. For 
this reason, we prefer the term "kernel phrase" instead of 
noun phrase. However, it seems reasonable to identify first 
the semantic correlations within the frame of the syntactic 
construction (noun phrase, adjectival phrase, verb phrase, 
etc. ) . 
The noun phrase structure grammar comprises - among others - 
the analysis and synthesis of adjectival phrases and attri- 
butive noun phrases. 
The adjectives were classified in terms of their semantic 
compatibility with semantic classes of head nouns. The fol- 
lowing properties of adjectives and nouns were considered: 
Ad j act ives : t 
a. Feature of transformabillty (A-# N; A + ed--) N; A + ing.-) 
N) 
b. Semantic class membership (Topography, Morphology, 
Etiology, Function, General) 
c. Admissible semantic correlations with semantic classes 
of head nouns NPx --> Ax + Nx where (x) denotes the type 
correl at ion. 
Noun._.._.~s: 
a. Feature of transformabillty (N-->N; N-+A); 
b. Semantic class membership (T,M,E,F,G as above) 
For example: 
ACUTE STAPHYLOCOCCAL PNEUMDNIA 
A-TR/G/M + A-TR/E/M + N-T'R/M-~(NP) M,E where 
A-TR/G/M (acute) = Adjective (A), transformable 
(TR), 'General' (G), semantic correlation with N(M) ; 
A-TR/E/M = Adjective, transformable, semantic 
class membership-'Etiology' (E), modifying N(M) ; 
N-TR/M = noun, non-transformable (TR), belonging to 
semantic category 'Morphology' (M). 
(NP)M,E = adjectival kernel phrase belonging to the 
semantic category 'M' (morphology) and category 'E' (etiol- 
ogy). 
The semantics assigned to adjectivals reflect, to a greater 
or lesser extent~ the semantic and syntactic relations between 
the adjectivals and the head nouns that they modify. Rules 
were written for strings of adjectives which are preposed or 
14 
postposad to the head noun as immediate constituents or in 
discontinuous sequence. Analogous criteria were used to pre- 
pare rules for recognition of attributive noun phrases. (13) 
Z. Harris suggested that it might be possible to establish, what 
he calls, "classified relationships" between adjective and noun in 
advance r i.e., to have dictionary pairing of words and their classi- 
fiers for a scientific language because the metaphorical usage is 
highly restricted in scientific discourse. It seems reasonable that 
adjectival phrases or appositional noun phrases which are perceived 
as single conceptual units should be listed and treated in the same 
way as compounds. 
Any treatment of ~hglish adjectives shows that the attribution 
and predication is far from simple. For instance, an adjective may 
modify a noun whleh is not present explicitly. An example is the 
phrase "parathyroid (gland) adenoma" meaning "parathyroid gland ade- 
home" where the adjective "parathyroid" is semantically related to 
the absent noun "gland." i 
The complexity of the analysis of adjectival phrases becomes 
apparent in attempting to semantically classify adjectives. Many 
adjectives may be members of two or more classes~ i.e., the given Axy 
can modify either N x or Ny, e.g.~ general adjective 'ab~ormal' can be 
correlated semantically with N T or N~ or N E or ~F' thus, yielding mul 
tiple readings. Even further refinement of the classification of 
adjectives will not completely solve the difficulty. 
16 
Semant ics 
It was already mentioned before that our goal is the successful 
mechanical recognition, interpretation and subsequent storage of 
Medical English for meaningful and timely retrieval of medical data 
in accordance with the needs of the user. 
When the boundaries of kernel phrases are identified by the 
implementation of syntactic and semantic algorithms the next step is 
the establishment of semantic correlations among the major semantic 
units of the utterance. The semantics of kernel phrases are conveyed 
by a set of relational predicates. 
Relational predicate having a propositional function f(x) des- 
cribes the type of semantic relationship among the major kernel 
phrases as mentioned before. 
For example, the statement: "Pneumonia~ due to staphylococcus" 
can be formalized as 
\[(RI) (M,E) 1 
where R 1 is the causative relational predicate "due to," "pneumonia" 
is the kernel phrase belonging to the semantic category 'M' (morpho- 
logy) and "staphylococcus" being a member of the category 'E ~ (caus- 
ative agent). The relational expression "due to" can be substituted 
by other equivalent expressions such as "caused by" or "resulting 
from," since they designate a similar relationship between the same 
semantic categories. The notion of argument pair(M,~ is not limited 
to single constituents as in the example above but it refers to kernel 
phrases in which the number of constituents is variable. For example~ 
the statement mentioned above can be expanded as follows: "Acute and 
severe pneumonia caused by staphylococcus albus" where the left-bound 
phrase "acute and severe pneumonia" (M) and the right-bound phrase 
"staphylococcus albus" (E) are related by the relational predicate 
"caused by" (RI). Consequently, the statement \[(R 1) (M,E)\] holds in 
both cases. 
The classification of relational predicates RI, R2 - - R n will 
reveal different types of relations among the semantically marked 
kernel phrases belonging to the semantic categories T~M,E,F or per- 
haps to other categories which may result from further semantic anal- 
ysis. The precision of relational predicate rules may be increased 
by refinement of admissible co-occurrences of subclasses of the basic 
semantic constituents of kernel noun phrases. 
For example, statement: \[(Rx) (M 12"*, T II**)\] is to be 
interpreted that there is a certain relationship Rx (not defined here) 
between the subclass MI2** (fractures) and subclass T II** (bones) 
where ** implies any member of the respective subclass. 
The syntactic structure of relational predicates consists either 
of a single functional element~ such as the preposition 'in,' or com- 
pound expressions~ such as 'due to~' or it can be a punctuation mark 
such as colm~a° 
The semantic rules can be regarded as the axioms of the system 
and therefore theorems derived from these axioms will describe various 
properties of Medical English which can be tested for truth or falsity 
for completeness, and for ambiguity. 
18 
The body Of theorems will constitute the grammar of a formal 
intermediate language having the semantic classes T, M, E, F and 
others to be defined, as its vocabulary. 
The axioms of the system are based upon the distribution of 
semantically significant elements. The key components consist of 
operators and their operands. The operands~ linguistically related 
to noun phrases, are members of particular semantic categories, and 
the operators are theorems describing the relationship among the 
operands • 
Relationships among syntactically connected parts of sentences 
will be tabulated and associated with the meaning of semantic cate- 
gories in the SNOP dictionary. The relationship between syntactic 
~nd semantic units is expressed by a set of linguistic operators which 
lave the function of relational predicates in the intermediate language. 
Conceptual analysis and preparation of algorithms for trans- 
lating restricted Medical English into well-formed expressions will 
equire definition of the functional form of relational predicates in 
:he intermediate language and their syntactic features as linguistic 
)perators in natural language. 
We assume that heuristic-type rules for automatic problem-solving 
~ill be expanded beyond the present logical framework and combined 
~ith linguistic techniques for automatic processing of natural lan- 
l .Tuage. This will lead to the development of a more powerful theory of 
)roblem-solving, in general (15, 16). 
Conclusions 
The development of a methodology for machine encoding of diag- 
nostic statements into a file, and the capability to retrieve informa- 
tion meaningfully from data file with a high degree of accuracy and 
completeness is the first phase towards the objective of processing 
general medical text. 
The results of morphological, syntactic and semantic analysis 
are translated via computer programs into the existing intermediate 
language which comprises four information fields: Topography, Mor- 
phology, Etiology~ and Function, as they are listed in the SNOP die- 
tionary. Once the units are identified and mapped into the inter- 
mediate language, they could be replaced interlingually by equiva- 
lent units of another syntactic and semantic system. The replacement 
could be implemented mechanically through the selection of the equiv- 
alent syntactic structure, in our case, SNOP noun phrases in their 
intermediate language form. 
The most amazing aspect of language is the fact that despite its 
enormous complexity human beings are able to use it with success as a 
communication tool. If we are ever able to discover and describe the 
process of human thought, we will be closer to the resolution of many 
problems associated with the formalization and subsequent automatiza- 
tion of natural language. It is not our intention to tackle all the 
problems inherent in natural language. We believe that we will be 
able to refine our algorithms and further develop a system which will 
process medical text by applying the formalized linguistic analytic 
procedures for the storage of data in such a way that the users' 
cequirements can be met. 
q-Ol6lT TOPOGRAPHY-NUMERICAL (contlnuad) 
SECTION 2: RESPIRATORY TRACT 
20-RESPIRATORY TRACT i 
2000 Respiratory tract, NOS 
201-Upper Respiratory Tract 
2010 Upper respiratory tract, NOS 
Nose, accessory sinus and 
2020 
nasopharynx combined sites .2138 
2139 
.202 -Lower Respiratory Tract 
Lower respiratory tract, NOS 
Larynx, trachea, bronchi and lungs combined 
sites 2200 
21 - NOSE 
ZlO0 Nose, NOS 
2101 Mucous membrane of nose 
2102 Respiratory region of nose 
2102 Olfactory region of nose 
2104 Nasal gland 
2100 Olfactory gland 
2100 Cavernous plexus of nose 
2107 Lamina proprla of nose 
2100 Nasal meatl 
211-External Nose 
2110 External nose 
z111 Roof'of nose I 
2112 Dorsum of nose 
2113 Apex of nose 
2114 Ala nasl 
2115 Nasal septum, mobile portion 
212 -Nasal Cartilage 
2120 Nasal cartilage 
2121 Greater alar cartilage 
2122 Lateral nasal cartilage 
2123 Nasal septal cartilage 
2124 Lesser alar cartilage 
2125 Vomeronasal cartilage 
213,214 -Internal None 
2130 Internal nose 
2131 Nares 
2132 Nasal vestibule 
2133 Nasal fossae 
2134 Nasal septum, NOS 
2135 Cboanae 
2136 Nasal turbinate, NOS 
2137 Inferior nasal turblnate 
Middle nasal turblnate 
Superior nasal tnrblnate 
2201 
2202 
2203 
2210 
2211 
2212 
2813 
2214 
2215 
2220 
2221 
2222 
2223 
2224 
2225 
2230 
2231 
||2| 
22 - ACCESSORY SINUS 
Accessory sinus, NOS 
Paranasal sinus 
Accessory nasal sinus 
Mucous membrane of accessory 
sinus 
Accessory sinus gland 
Lamina propria of accessory 
sinus 
221 -Maxillary Sinus 
Maxillary sinus, NOS 
MaXillary antrum , 
Bight maxillary sinus 
Left maxillary sinus 
Mucous membrane of maxillary 
sinus 
Maxillary sinus gland 
Lamina propria of maxillary 
Sinus 
222-Frontal Sinus 
Frontal sinus, NOS 
Bight frontal sinus 
Left ,frontal sinus 
Mucous membrane of frontal sinus 
Frontal sinus gland 
Lamina proprla of frontal sinus 
223-Ethmoid Sinus 
Ethmold sinus, NOS 
Ethmold antrum 
Right ethmoid sinus 
Left ethmoid sinus 

References 

DOSTERT, L.E.: Machine Translation and Automatic Language Pro- 
cessing: Vistas in Information Handling; Spartan Books; Washing- 
ton 1963 

GARVlN, P.L.: The Georgetown - IBM Experiment of 1954: An 
Evaluation in Retrospect: Papers in Linguistics in Honor of 
Leon Dostert; Ed. by W. M. Austin; Mouton Co.; The Hague, 1967 

SEDELOW, S.Y., SEDELOW, W.A. JR.: Stylistic Analysis; Auto- 
mated Language Processing; John Wiley and Sons~ Inc.; New York 
1967 

4. CHO~KY, N.: Aspects of The Theory of S~ntax; MIT Press; 
Cambridge, Mass.~ 1965 

5. Automatic Language Processing Advisory Committee, LANGUAGE and 
MACHINES - COMPUTERS IN TRANSLATION AND LINGUISTICS : National 
Academy of Sciences, Washington 1966 

BOBROW~ D.G.: S~n~tactic Theory in Computer Implementations; 
Automated Language Processing; John Wiley and Sons, Inc.; 
New York 1967 

Seminar on Computational Linguistics; Edited by A. W. Pratt, 
A. H. Roberts and K. Lewis; U.S. Department of Healthy Education 
and Welfare; National Institutes of Health; Public Health Service; 
Publication No. 1716; U.S. Printing Office 1967 

BAR-HILLEL, Y.: Language and Information; Addison-Wesl~y, Read- 
ing~ Massachusetts 1964 

PRATT, A.W., THOMAS, L.B.: An Information Processing System for 
Pathology Data; Pathology Annual-66; Appleton-Cen%ury-Crofts 
Publishers; 1967 

DUNHAM, G.: Pathology Diagnoses Language Encoder; Division of 
Computer Research and Technology; National Institutes of Health; 
Internal Report. 

EPSTEIN, M.: A STstem for Pathology Data Processing; Division 
of Computer Research and Technology; National Institutes of Health; 
Bethesda, Maryland 1969 

PRATT, A.W., PACAK, M.: Identification and Transformation of 
TerminalFbrphemesin Medical English; Methods of Information in 
Medicine; Vol. 8; No. 2, April 1969; F. K. Schatauer Verlag, 
Stuttgart, New York 

PACAK, M., DEFRANCESCO, H: Adjectival Phrases in Medical 
~; Division of Computer Research and Technology; 
National Institutes of Health; Bethesda, Maryland; in pre- 
paration. 

Systematized Nomenclature of Pathology; College of American 
Pathologists, Chicago~ Illinois 1965 

GARVIN, P.L.: The., Place of Heuristics in the Fulcrum Ap. p,,.roach, 
to Machine Translation; Lingua, Vol. 21; North-Holland Publish- 
ing Company; Holland, 1968 

SLAGLE, J.R. : E~eriments with a Deductive Question-Answerin ~ 
~; Comm. ACM 8, 12, 1965 
