An English-Like Language For Qualitative Scientific Knowledge

Introduction

It is generally agreed that mathematics is not well suited to describe most biological knowledge, though whether this is an (unfortunate) historical accident or a reflection of how our minds relate to the world is an open question. What appears to be more appropriate, however, are the formalisms which are becoming increasingly well developed through AI research into representing non-quanti­ tative knowledqe. For many reasons, however, some of which are justifiable, these languages (e.g., CONNIVER) tend to remain within AI laboratories. I wish to suggest that this tendency of "growing outward from the AI lab to someday intersect the real world", evidenced by the comparative lack of attempts to apply any developments "within AI" to problems not invented by AI researchers may, in fact, be detrimental to both AI and the rest of the world. With this in mind I have tried to place myself in an "interface" position, to try to bring together the types of knowledge that biolo­ gists need and the tools d~veloped within the AI community for expressing qualitative knowledge.

I am therefore in the process of defining a lan­ guage, called KAL (Knowledge Acquisition Language), which is not a programming language but a lan­ guage intended to be used by biologists or physi­ cians to express as much as possible of the quali­ tative knowledge they use in their everyday work. The development is at the point (December, 1974) of making initial attempts to expose some biolo­ gists to the language, by means of a number of hours of explanation plus a 50 page "user 's manual", written for persons unfamiliar with com­ puters . To parallel KAL's evolution through in­ teraction with biologists, I have begun implemen­ ting some of it, using UCI LISP and CONNIVER, so that it can be developed both "on paper" and as a program, in the question answering system tradi­ tion. It is not anticipated that all of what can be defined on paper will be implemented.

Design Criteria for KAL

The following desiderata were foremost in the design of the language:

1. It must be suited primarily for the expression of that portion of biological (including medical) knowledge which is now expressed in natural lan­ guage text. This excludes all forms of graphical and mathematical expressions, since the problem of representing graphical knowledge (pictures and diagrams) is a flourishing field in AI, and repre­ senting mathematical knowledge is the bread and butter of computers.
2. It must be comprehensible to persons having no experience with computers. Hence, I have chosen to make KAL look as much like a natural language (English) as possible, so that most KAL sentences can be immediately understood by anyone familiar with KAL who knows the technical vocabulary . Most motivated biologists or physicians should be able to learn to write KAL in a reasonable amount of time.
3. The technical vocabulary must be determined by the user, in an open ended (i.e. continually ex­ tensible) manner. Thus, the major parts of speech, used in an appropriate context, should be automa­ tically added to the machine's lexicon. This forces some rigidity on the syntax, but this is not considered detrimental to content, only style. The meaning of any word is given by the totality of sentences in which it has appeared . The skele­ ton of the syntax and semantics are a reserved group of English* words and phrases (keywords) around which the user may invent arbitrary flesh.
4. If it is to be LISP/CONNIVER based, KAL should interface as simply as possible to those aspects of the host languages which are useful, without becoming subjugated by them. One must resist the temptation to the seduced by the so-called "power" of such languages into doing only what they make "natural", and thereby avoiding what would be more appropriate to the task.
5. Deduction, used to answer questions and to pe~­ tially check the consistency of new sentences, should rely primarily on a variety of special me­ chanisms (e.g., set hierarchies, property lists and functional correspondences}, and secondly, on the "problem reduction" paradigm ~ la PLANNER. Question posing is to be in the form of normal declarative sentences having "blanks" to be filled in.
6. Constant attention is maintained while encoding a text for those sentences which cannot be proper­ ly handled in the current (static) version of KAL, particularly those which describe actions or chan­ ge . These are carefully noted in anticipation of a future "dynamic" version.
7. *They could be just as well chosen from many other languages, suggesting the attractive possibility of easy translation for KAL-based knowieriy~.

Noun Phrases and Sets

The user is encouraged to think in terms of set hierarchies. Sets are denoted by noun phrases (NP), in which the noun may be either singular or plural, either denoting the same set. The usual terminal's' rule applies for the plural form unless otherwise stated, as in

MAN IS-SINGULAR-OF MEN

Underlined phrases are keywords. The complement of a set does not exist as a set; the notion of a "constraint" (see below) handles negation, as well as certain union-like constructs. Sets are not intersected directly; for example, if CANADIANS (equivalently, CANADIAN) is a set, one may use CANADIAN as an adjective as well (before a noun) and write either of:

• 2 EVERY CANADIAN MAN~~ CANADIAN
• 2a CANADIAN MEN ARE CANADIANS

These sentences are synonymous, and either creates the subset CANADIAN MEN of CANADIANS. (When dis­ cussing KAL sets we use the plural form by conven­ tion). It is automatically known to be a subset of MEN. IS A (equivalently ARE) is a keyword in­ dicating set-containment whi~causes efficient storage of this fact.

NPs may have keyword quantifiers (e.g., EACH, EVERY, SOME, etc.) at the beginning, followed by arbitrarily many adjectives, a noun, and then pos­ sibly some subot·dinate clauses or prepositional phrases. One either creates a new set by using a new NP, which must be defined to be contained in some existing set, or refers to an existing one. The following sentences illustrate some of these ideas.

3 JOHN~~ MAN

This sentence creates the set MAN if it is new.

• 4 JOHN BOUGHT CAR/1

JOHN is an individual (a singleton), but sets without proper names may not be created (e.g., A CAR), hence we invent the name CAR/1 (the machine may assign the number). Definite articles are used in restricted contexts; e.g. THE and A may be used to refer to a set but not to create one*, as in:

5 THE MAN WHO BOUGHT ~ CAR ~ RICH

The NP here denotes JOHN, since the subordinate clause is used in a pattern matching fashion to locate the X which IS A MAN such that X BOUGHT Y and Y IS A CAR. IS-rn-this context is interpreted as a keyword associating the adjective RICH with the set JOHN. Sets may be second order, i.e., may have sets as instances, and may appear quantified in certain contexts.

*except in certain "generic" contexts.

The user is also encouraged to think in terms of one-to-one and many-to-one links, or functional correspondence between instances of sets (be they individuals or other sets). These links arise in several ways, e.g., through "For All ... There Exists ... "statements, such as

• 6 EACH PERSON HAS ~ HOME
• 7 EACH WOMAN LIKES HER MEN

HAS is a keyword signifying any functional (many­ one) correspondence. Sentence 7 is an example of second order quantification, i.e., a separate sub­ set of MEN may be inferred to exist for each in­ dividual WOMAN. HER is a keyword denoting the existential quantTrfer following the universal EACH.

A second source of correspondence links arises in the notion of structures, i.e. either physical structures (THINGS) or mental ones (IDEAS). The keyverb HAP, for HAS AS PARTS denotes this rela­ tion, as1n:

8 EACH PERSON HAP 1 BODY, 1 SOUL

The comma construction saves writing 2 sentences. If there are BODIES other than human ones, one must write BODY OF A PERSON to distinguish this one. HAP links are-1-1, unlike HAS. One might defineafamily thus: -

• 9 EACH COUPLE HAP 1 MAN, 1 WOMAN
• 10 EACH FAMILY HAP 1 COUPLE, SOME CHILDREN

If it is known that BILL~~ MAN, then

• 11 SMITHS ARE~ FAMILY
• 12 SMITHS HAP (BILL, MARY, {PETER, ANN})

may be understood, with the machine warning of the assumptions it has made: MARY IS A WOMAN and the extensional set {PETER, ANN} ARE CHILDREN. Such references to individuals are-relatively rare in technical literature, however, where by individual I mean of course, not a human in general, but a specifically named (proper noun) member of a set. Thus KAL semantics, unlike those of most natural language systems which are oriented towards dis­ cussing human relations, are biased towards dis­ cussion at the set (generic) level. Its use of articles,~. THE) which we haven't space to dis­ cuss, reflects this.

Adjectives and Attributes

The keyverb IS (or ARE) is used to associate an adjective with a no~ This association may also be (equivalently) effected by placing adjectives before the noun, as in:

13 BIG BLACK BEARS ARE DANGEROUS

which creates the subset of BEARS denoted by DAN­ GEROUS BIG BLACK BEARS, with the adjectives in any order. (The problem of how to deal with the re­ sulting six intervening subsets of BEARS is an annoying one which I have not yet resolved . For­ tunately, three or more IS type adjectives seem to be rare).

Usually, one wishes to associate a property with a set under some named relation, usually called an attribute, which may be introduced by the user by hyphenating with IS:

• 14 EACH MAN'S SEX-IS MALE
• 15 EACH APPLE'S COLOR-IS RED

One may also hyphenate IS the other way, as in

• 16 SKY/I IS-LOCATED ABOVE GROUND/1

The use of prepositions like ABOVE (and cases) is described below. Note that THE SKY is not allowed here - one may only the THE ~create a generic noun, not an individual.

Constraints and Incremental Learning

Much of our knowledge comes to us in stages, each being in some way a refinement of what was pre­ viously known. Knowledge acquisition systems ought to be designed to accept information in this way, yet little seems to have been done to disco· ver the underlying processes that adults use, though much has been done by psychologists in es­ tablishing how concepts evolve in the child . Within AI, the thesis of P. Winston (70) is one of the few places in which learning by stages is in­ vestigated, although he does not discuss formal linguistic aspects there to any degree.

In this regard, I have found that it is useful to regard the subject-verb pair, for many verbs, as a function (the verb) - argument (the subject) pair . This is usually useful for verbs which define at­ tributes, and particularly those which h~ve either numerical values or those which may take only a small finite number of values. One can then de­ fine a number of keywords called constraints which specify the allowed "values" for these subject­ verb pairs. For example, instead of 15, we might prefer:

• 17 EACH APPLE'S COLOR-IS ONE-OF (RED YELLOW GREEN) ----

Further examples include:

• 18 EACH ADULT'S AGE-IS GT(20)
• 19 RICHARD li NONE-OF(HONEST CLEVER LOVED)
• 20 CHILDREN LIKE AT-LEAST-ONE-OF(CANDY SPINACH DOCTORS)
• 21 JUST-ONE-OF(BILL JOHN) li~ LIBERAL
• 22 THE PRICE OF A CAR IS GT(THE PRICE OF A BICYCLE) -- - --

Many details surrounding the use of such const­ raints must be carefully spelled out, particularly when they are used in the more general ways illus­ trated here. For example, 20 clearly should mean that each child may have her/his own choice, so that the implicit quantification rules must be stated. It is in these kinds of details that I anticipate the most difficulty both in teaching such a language to non-mathematically experienced persons, and in implementing it. Each contains considerable logical meaning, whose function in the computer must be to assist deductions, and monitor future inputs. By the latter, I intend that each new statement be monitored so that if it is a further constraint on a previous one, or a contradiction, the user is to be notified. Clear­ ly, it is not possible to do this to arbitrary logical depth; some practical compromise will have to be accepted.

While learning through constraints is an essential part of any knowledge acquisition system, it is given a secondary role in KAL, since it was felt that other considerations are more fundamental.

Time, Change and Actions

The problem of representing change has been given surprisingly little attention in the AI literatur~ The series of three papers by Hayes (69,70,71) is essentially the only theoretical discussion of the logic of change in this literature. Philosophical logicians have written at length about such mat­ ters; I have not been able to find anything in an admittedly imperfect search of this literature (e.g., the books edited by Rescher) which would give a hint as to how to write programs.

One turns then to the current "brute force" ap­ proach to change used in AI programming: a situa­ tion or context is a group of "true" statements corresponding to a fixed state of the world, and an event (a change or action) is simulated by ad­ ding to or removing from this group to produce a new "after" situation. Some recent reports using this approach, originally popularized in STRIPS (Fikes and Nilsson, 71) are Sacerdoti (73), Biss et!! (73) and Hendrix et!! (73). It is proposed to follow suit by providing such rudimentary des­ criptions of sequences of events via the context mechanism of CONNIVER. In this respect, the de­ sign of KAL will necessarily be influenced by the implementation language, a potential im· pediment to finding "what you really want" rather than "what's avai 1 ab 1 e".

KAL verbs do not at the moment have any "deep semantic" structure associated with them. There is no attempt to handle time reference, tense, or verb hierarchies; verbs serve merely as a constant in a pattern, along with appropriate keywords. It is felt that the very complex problem of a deep representation of change and the associated natu­ ral laoguage structures could not be given pro­ minence without jeopardizing the many other as­ pects of this work. Thus KAL statements now ex­ press that appreciable part of knowledge which is timeless (without reference to change) . I term this version "static" KAL.

Verbs: Static Worlds, Cases and Locations

Now I will describe "static" KAL verb structure. The user is directed to choose verbs in the pre­ sent active plural (e.g., "run", rather than "runs" or "is running", etc.) and to use as ob­ vious ones as possible (e.g., "run" rather than "canter"). Following the verb there are one or more cases, each denoted by certain keywords (usually prepositions), except the "direct object" case, which has no keyword and must come immedia­ tely after the verb.* Several alternate construc-

*For a discussion of the linguistic notion of cas~ see Bruce (73), Kintsch (72), and Fillmore (68).

tions are possible, such as "IS-XXXXED" (e.g., "IS-LOCATED").

The cases are noun phrases, as in:

23 EVERY COMMUTER DRIVE(S) HIS CAR FROM HIS HOME TO HIS JOB VIA HIS STREETS USING HISCAif- --

In 23, the direct object HIS CAR, is associated with COMMUTER through a "For all there exists" link, and occurs as well as instrument. The un­ natural HIS STREETS is necessary to associate a set of streets with each commuter. Any verb suf­ fix in ( ) is to be ignored by the machine; thus there is no check for subject-verb agreement.

24 EACH AMOUNT-OF LIQUID FLOW(S) FROM ITS SOURCE TO ill SINK THROUGH ITS CHANNEL

In 24, the complexities of dealing with mass nouns are involved. This problem is vital to biological knowledge, and has received no attention in the AI literature. I am trying to deal with it by key­ words such as AMOUNT-OF which identifies LIQUID as a mass noun, and creates the equivalent of the individual for discrete nouns. ·

The prepositional keywords which imply a location in fact refer to the implied location associated with each thing; i.e. there is an implicit "physi­ cal" statement:

25 EACH THING HAS ~ LOCATION

underlying all THINGS. Problems associated with underlying physical assumptions are discussed further below.

Implication Statements

Up to now, the only logical connections be~w~en KAL statements (which have all been propos1t1ons) are those which are "built in" to the semantics, such as the implications allowed through set hie­ rarchies and correspondence links. To relate arbitrary statements, implication statements are introduced, either with~ ... THEN ... ,or lft, allowing AND or OR group1ngs of the antecedent. The semantics are-borrowed directly from PLANNER. Thus, if we have

26 IF A PERSON IS RICH OR A PERSON IS HEALTHY THEN THAT PERSON _!i FORTUNATE

then any request to establish _!i FORTUNATE for any subset of PERSONS will redirect the search (esta­ blish the subgoal) for first IS RICH, or failing that, IS HEALTHY. This approach is well estab­ lished:-and is easily handled by CONNIVER. THAT is a keyword binding the following PERSON to the previous one, i.e . , identical substitutions must be made.

Lacking the mechanism to describe actions as se­ quences of static contexts, one may neverthele~s use chains of implication statements to determ1ne if a particular statement, representing a desired situation, can come about. It is felt, however, that this is not sufficiently satisfactory to rep­ resent any but rather trivial actions.

It appears that implication statements should only be used within a situation, just as in the STRIPS approach, predicate calculus deductions do not transcend changes in the world state.

An Example from Basic Neurology

A few sentences will now be presented which have been taken directly from a basic neurology text­ book (Gatz, 71) used by many medical students. After each sentence (in quotes) follows the KAL translation, omitting obvious or trivial sentences for brevity (such as plural definitions), plus a discussion of various features.

"The neuron (nerve cell) is the functional and anatomical unit of the nervous system".

27 EACH NERVOUS SYSTEM HAP SOME NERVE CELLS This sentence introduces four sets: SYSTEMS and NERVOUS SYSTEMS, CELLS and NERVE CELLS, with the obvious set containment. Note that NERVOUS and NERVE are adjectives. The 1 to 1 link from EACH NERVOUS SYSTEM to ITS own subset of NERVE C~LLS is established.

28 ~ NEURON IS-SAME-AS ~ NERVE CELL IS-SAME-AS is a keyverb denoting synonomy. The phrase NERVE CELL thus becomes interchangeable with NEURON.

29 THE FUNCTIONAL UNIT OF THE NERVOUS mlEM IS-SAME-AS THrN'tlm"ON

The sets UNIT and FUNCTIONAL UNIT are introduced, the latter being related to each NERVOUS SYSTEM by HAP (implied by OF THE). FUNCTIONAL UNIT OF_ THE NERVOUS SYSTEM becomes another name for NEURON.

"Each consists of a cell body (cyton) and one to several dozen processes of varying length called nerve fibers".

30 EACH NEURON HAP 1 CELL BODY, l TO 40 l'ROCESSES - -

In contexts which take a numerical value, a looser constraint may be used, such as upper and lower bounds. Of course these constraints ideally should not be sharp; the use of such constraints is a rudimentary attempt to introduce fuzzy know­ ledge. Each neuron no.w has associated, through l-1 links, an individual CELL BODY (a subset of BODY) and a subset of PROCESSES not exceeding 40 in size. The comma construction saves writing two sentences. To continue with this sentence:

• 31 THE CELL BODY OF A NEURON IS-SAME-AS THE ffiON -- -
• 32 THE PROCESSES OF A NEURON ARE-SAME-AS NERVE FIBERS --
• 33 THE LENGTH OF A PROCESS IS FROM 0 TO TSOCM. -- -- -

Only 33 requires comment. Every PROCESS is de­ fined to have associated with it (HAS) a LENGTH, which is known from the remainder or-the sentence to be a numerical attribute. It may therefore appear hyphenated with IS, as, for example, in:

34 PROCESS/1 LENGTH-IS 200 CM

If sentence 34 were entered, the violation of tht constraint established in 33 should be noted.

Numerical attributes may carry units.

"Dendrites are short branching fibers which nor­ mally receive impulses at their peripheral termi­ nals and conduct them toward the nerve cell body".

• 35 EACH CYTON HAP SOME PERIPHERAL TERMINALS
• 36 IF A DENDRITE RECEIVE(S) SOME IMPULSES AT SOME-OF ITS PERIPHERAL TERMINALS THEN THAT DENDRirr-CONDUCT(S) SOME IMPULSrs-- TOWARD ITS CYTON. --

Here we run into the problem of expressing a dyna­ mic fact (a sequence of events) in static KAL. One must use an implication statement in order to obtain the desired cause-effect relation between the two statements, as discussed earlier. The machine can only use this statement "backwards" (i.e. to search for an appropriate instance of the antecedent to establish an instance of the conse­ quent) since it functions as a CONSEQUENT theorem. What is really intended by this sentence is more like a "demon", "triggering" the assertion of the consequent should the antecedent be asserted. Rather than hastily introduce such machinery (a temptation when using a language like CONNIVER) to appear to have some capability of handling sequen­ ces of events, I prefer to first gain more exp&­ rience with the static worlds from which dynamic worlds are probably to be built, and then to int· roduce problem-determined language (rather than CONNIVER-determined) for events.

Although the two occurrences of DENDRITE are bound by THAT, the two subsets of IMPULSES are not nece­ ssarily the same. We are trying to treat the notion UIPULSES as if they are the THINGS when, in fact, they are a sequence of events. One can get away with this to some extent; a proper solution will have to be found in the dynamic version of KAL .

"The term axon, in a strict sense, applies to a single long fiber conducting impulses away from a nerve cell body".

• 37 X IS A NERVE FIBER AND
• X LENGTH-IS GT(l CM}AND
• X CONDUCT(S):SOME IMPULSES AWAY-FROM ITS CYTON -- --

IFF X IS-SAME-AS AXON

Here we have an appropriate use of an implication statement, in fact an equivalence. (There remain~ however, the necessity of treating the conducting of impulses as a static phenomenon). The reader may wonder: "Why not just have three statements about AXONS?" The distinction lies in the manner in which the fact retrieval algorithm ought to treat the two possibilities. Given 37, the algo­ rithm knows these three facts to be tightly asso­ ciated, i.e., should facts involving AXONS appear during some discourse, these three facts are to be immediately given prominence (over other facts about AXONS) in subsequent processing. The same should occur if, say, two of the three appear, i.e., the third might be sought, to see if we are talking about AXONS.* Given the three facts inde­ pendently, no such connection would be inferred . The use of X as a bound variable, while looking less like English than previous statements, makes possible the neat conjunction of conditions with­ out messy "English-like" syntax. One should only carry along such syntax as long as its clarity to complexity ratio is rewarding. Statements such as these, involving considerable complexity of at least one side, pose difficult problems for the question-answering algorithm, and should be good counter-examples against a pure top-down (problem reduction, or undirectional heuristic search) pro­ cedure.

Related Work and Discussion

Several other projects bear similarities to KAL. W. Martin and coworkers at MIT are developing a language OWL (formerly MAPL) in which to discuss the "world of business". Like KAL, it is set oriented, CONNIVER based, and uses a restricted syntax laced with keywords, though it is difficult for non-LISP programmers to read. It also inclu­ des second order sets, functional correspondence links, and case structure. They do not say they work with actual business texts (not surprisingly), nor do they claim running programs. MAPL-OWL ap­ pears to be the project most similar to KAL.

Another similar undertaking has been described by Isner (72) and Pople and Werner (72), in which neurological knowledge is encoded in a simplified English, for input to a question answering program written in the LISP extension GOL (Pople, 72), which paralleled some of the innovations of PLAN­ NER. The emphasis seems to have been on using actual scientific knowledge, and on getting a pro­ gram running. There is no discussion of the syn­ tax of the language, nor of the semantic structures needed for such text. The examples given however are among the most extensive that have appeared of this type of NL programming.

A great deal of careful work has been done by the Swedish group (Sandewall, 70,72; Makila, 72,74; Palme, 70,72,73) in defining various formalisms for representing NL semantics, though they consider general NL. In particular, PCF-2 is a predicate calculus-like notation for a broad range of NL constructs, though it does not appear to have been implemented.

There have been numerous other well known attempts which space does not permit discussing. Most of these try to consider NL in its full generality, particularly in its verb structure. Any of these programs, however, will accept only relatively few NL sentences, and the problem of defining just which these are is often ignored, or if attempted, is too complex to be readily understood. The re­ sult of attempting such generality would then ap­ pear to be that for anyone not deeply involved in the programming, writing a body of text acceptable to the program is virtually impossible. Since KAL has a relatively rigid syntax, and semantics res­ tricted to scientific needs, their description is sufficiently succinct that nonspecialists in com­ puters may learn it, increasing the possibility of such persons being able to communicate with a machine.

*(i.e., we're in the AXON frame, to use the current jargonology).

The main advantage of the "artificial" language approach, however, is that it focuses attention on the semantic essence of each sentence being trans­ lated. While encoding text, one must constantly ask which KAL construct is appropriate, noting those sentences which cannot be adequately repre­ sented. This exercise is an essential part of the develo~nent of artificial languages, and tends to be neglected when the designers invent their own text.

I have found at least four common types of senten­ ces which do not translate well (or at all) into KAL. Obviously, the most serious limitation is with dynamic descriptions, as we have noted. The assertion of a proposition should not be confused with the simulation of the event it describes, nor should implication statements serve as causal links between statements (which, not being situa­ tions, shou 1 d not be "caused"). A second conmon problem, also illustrated in the example, is in. handling fuzzy information- recall how I substl­ tuted a sharp value for the adjective "long". A first step toward alleviating this problem would be to introduce fuzzy sets and truth values, which has been done by Le Faivre (74) in his language FUZZY, a CONNIVER-like LISP extension. A third problem concerns complex spatial relationships, a frequent requirement in physiology. Probably, the blocks world approach of listing coordinates will not be feasible for tortuous structures. Until more subtle methods of describing complex objects are developed, such knowledge will have to be ex­ pressed in propositions with simple prepositions where possible, and ignored otherwise. The fourth problem involves imbedded sentences - sentences which refer to other sentences. I have made no attempt to consider this problem in KAL; Moore (73) has discussed some possible approaches.

Expressing the underlying laws of physics (which include the "frame problem") that a machine must know to make decisions which seem trivial for even animals, is well recognized as one of the most complex tasks in representing knowledge in compu­ ters. One such law (statement 25) was briefly touched upon earlier. This is a static law, and is expressable in a KAL statement. There are many such laws, usually implications involving preposi­ tions. One may imagine, for example, a transiti­ vity law in KAL, the kind one finds in blocks worlds. The problem, of course, is to define for the machine all the laws it will need.

It is doubtful that this can be done in any other way than the way it is done now, at least for the foreseeable future. That is, one must postulate a set of laws and monitor the machine's actions to detect errors, omissions and contradictions. (Note that questions of logical consistency be~ome extremely difficult for languages suth as KAL 1n which we have no metatheorems). I suggest, though I do not yet have the experience to confirm it, that debugging a system which operates on KAL statements should be far easier than one using a less transparent notation, in which the human must perform a task which is very unnatural: manipula­ tion of abstract symbols. Since the meaning of a KAL statement can be comprehended at a glance, one can quickly sift through a page of definitions . looking for suspicious ones. Thus the syntact1c and semantic "distance", if I may coin such a term, between the CONNIVER expressions and the correspon­ ding KAL sentences should be many times "less" than the distance between more general Nl senten­ ces and LISP encodings of the equivalent predicate calculus, to take an extreme example. This trans­ parency of notation should be of particular value when it comes to dealing with dynamics, which the state of the AI programming art suggests are to be modelled by situations as defined earlier, since . the number of statements that the human must monl­ tor is greatly increased. Rather than conjecture further in this direction, I would like to discuss one additional point.

There is a whole range of knowledge that cannot be expressed even in a dynamic KAL without enlarging it so much as to make it a complete "progranming" language. I have in mind the so-called "procedu­ ral knowledge", which amounts to introducing the notions of "execution", "flow of control", "variables", etc. If a language is to be prac­ tically usable by nonspecialists in computers, it must not become as complex as the most complex "progranming" language (e.g. CONNIVER). Hence . there exists a very murky area of language defl­ nition which would be adequate for procedural knowledge without becoming incomprehensible to non-computer specialists.

There is one possible direction in which a lan­ guage like KAL might be extended without making it too complex, and which I intend to try to integ­ rate with the dynamic extension. I have in mind the type of facility for simulating dynamics pro­ vided by the discrete simulation languages such as GPSS and SIMULA. Persons relatively unsophisti­ cated in computer science seem to derive consi­ derable benefit in describing dynamic phenomena in these languages, and in interacting with "simula­ tion models". It would seem that a synthesis of the AI approach to dynamics and the "simulation" approach is due.

Another possible advantage of an artif~cial l~n­ guage for science should be.mentioned 1n clos1ng : Scientific discourse, both 1nformal and formal, 1s often alarmingly vague and verbose. Should a l~n­ guage like KAL become well enough developed, sc1en· tists might be persuaded to express their findings and conjectures directly in it, eliminating much volume and ambiguity. The development of global computer networks would make possible instant sharing of such knowledge, and the pr~blem of having the surface structure of KAL-l1ke language look like most Nls is, of course, trivial. Thus scientists of many tongues might conmunicate di­ rectly and clearly. The alternative for such net­ works is to allow arbitrary NL to flow through them, a measure sure to cause c011111unication block­ ages as severe as our present paper system.

But that is too futuristic.

Epilogue (June 1975)

In the paper, at the time of writing (Dec. 74) it was indicated that the intention was to attempt to implement some of KAL in CONNIVER. Subsequent ex­ perience actually trying to use CONNIVER resulted in the following conclusions:

CONNIVER is not sufficiently well documented and debugged to make using it at "remote" sites (where "remote" probably means "off the ARPA net") a re­ warding experience. CONNIVER is possibly adequate for certain toy world problems, but is too ineffi­ cient to be a suitable language in which to imple­ ment any practical natural language system. This was no surprise, since others had made similar ob­ servations. It was felt, though, that the ex­ perience was worthwhile so as to better understand the concepts and the pitfalls. A better alter­ native may exist in the language 2.pak (see the paper by L.F. Melli in this Proceedings).

Acknowledgment

The author gratefully acknowledges the support of the I.W. Killam Memorial Fund of the Montreal Neurological Institute.