Generative Computer Assisted Instruction: an Application of Artificial Intelligence to CAI

DOCUMENT RESUME

ED 078 648 EM 011 178

AUTHOR Koffman, Elliot B. TITLE Generative Computer Assisted Instruction: An Application of Artificial Intelligence to CAI. INSTITUTION Connecticut Univ., Storrs. Dept. of Electrical Engineering. SPONS*AGENCY National Center for Educational Resei,,:cb and Development (DHEW/OE), Washington, D.C.; National ScienCe Foundation, Washington, D.C. Office of Computl!ng Activities. PUB DATE 72 GRANT oEG-0-40895 NOTE 8p.; USA-Japan Computer Conference, 1972.

EDRS PRICE MF7A0.651HC-$3.29 DESCRIPTORS Algorithms; *Arkificial Intelligence; *Computer Assisted Instruction; *Computers; Problem Solving; Programed Texts; State of the Art Reviews IDENTIFIERS Generative Computer Assisted Instruction; Natural, Language Communication

ABSTRACT Frame-oriented computer- assisted instruction (CAI) systems dominate the field, but these mechanized programed texts utilize the computational power of the computer to a minimal degree and are difficult to modify..Newer, generative CAI systems which are supplied with a knowledge'of subject matter can generate theirown problems and solutions, can provide extensive drill and tutoring, and can diagnose learner problems and prescribe remedial feedback. These systems reduce the amount of-instructor effort required to teach new material and encourage students to assume the initiative in studying new topics. Information,-Structure Oriented (ISO) systems have capabilities for natural language communication and are useful in the -humanities-and social sciences...Quantitative systems rely on algorithms and are oriented toward providing students with practice in problem-solving. While the existing generative systems are still experimental it is-expected that their use will become more widespread as-research continues and,as CAI facilities become more available and economical. (PB)

.1 First USAJAPAN Computer Conference, -1972

1 GENERATIVE COMPUTER ASSISTED INSTRUCTION: ANAPPLICATION OF ARTIFICIAL INTELLIGENCE 'TO CAI

Elliot B. Koffman U S. DEPARTMENT OFHEALTH. EDUCATION A WELFARE NATIONAL INSTITUTE OF (University of Connecticut, Storrs, Connecticut) EDUCATION BEEN REPRO THIS DOCUMENT HASRECEIVED FROM DUCED EXACTLY AS ORIGIN THE PERSON OR ORGANIZATIOU ATING IT POINTS OF VIEWOR OPINIONS REPRE STATED DO NOT NECESSARILY INSTITUTE OF SENT OEFICIAL NAIONAL EDUCATION POSITION ORPOLICY

In an ISO system, there must be a capability I. INTRODUCTION for natural language communication between teacher Limited progress has been made in software for and student. The system must be able to determine by computer-assisted instruction:- Frame-oriented CAI the degree of correctness of a student's recionse systems have dominated the field. These systems are comparing it with information stored in its data -e- classically mechanized programmed texts and utilize base. The system:Would be able to understand and student. the computational power of the computer to a minimal answer questions which are posed by the Simmons (1) provides a very comprehensive review of extent. In addition, they are, difficult to modify and research in the areas of natural language communica- tend to provide a fairly fixed instructional sequence. tion and question-answering. Recently, generative CAI systems have appeared. These are systems for CAI which are capable of gener- Carbonell (2,3) has designed an extremely SCHOLAR is an ating their own questions or problems and deriving versatile-system called SCHOLAR. the Hence, they can provide un- example of a 'mixed- initiative" system in that their own solutions. generated limited drill_and tutoring with little effort required student can interrupt the flow-of questions from the course-author to define an instructional by SCHOLAR at any time and interrogate SCHOLAR. - SCHOLAR relies heavily on prior research byQuillian sequence. (4,5) in the area of semantic memory and utilizes a of facts These systems :re supplied with knowledge of semantic network for storage of a data base their-subject matter. Therefore, they can often concerning South America. interpret and answer questions or problems posed by They are also capable of diagnosing A semantic network is an organization of units -the student. mutual the degree of inaccuracy in a student response-and of information in terms of their meaning and providing remedial feedback on an individual basis. interrelationships. Each unit or node in the memory Starting Most of these systems incorporate techniques and contains pointers to other related units. concepts which are outgrowths of research in Artifi- from a given unit, one can trace out a set of rela- tionships by following the diverging trailsof cial Intelligence. pointers. See Figure 1. This paper will describe some of the generative systems which have recently been designed. A gener- For example, the unit Argentina, denoted as an ative tutor under development.by the author, which object, might have attributes such as superconcept, The Values has been used to teach digital computer concepts, superpart, location, capital city, etc. of these attributes would be other units in thenet- will also be examined in detail. work-and, hence, there would be-pointers to thenodes lati- These systems can be divided into two classes. representing the units-country, South America, Also, Those which are oriented towards the humanities and tude and longitude, Buenos Aires, respectively. textual material and those which are more concerned associated with each attribute is a tag indicating unit. with numerical manipulations and quantitativematerial. its irrelevancy with respect to the parent The former have been termed information-Structure both for Oriented systems (ISO) and will be discussed next. The Semantic network is used by SCHOLAR generating and answering questions. The initial context of the questions is determined by theteacher. II. ISO GZNERATIVE SYSTEMS The teacher's input "South America"specifies that node Research in the area of Artificial Intelligence the initial unit to be investigated is the has provided most of the background for the ISO Gen- South America. An attribuce is selected and a ,ques- These systems are intelligent in tion is generated concerning the value;ofthat erative Systems. is 4) the sense that they have knowledge of thesubject- Jtcribute. At the same time, the correct answer matter they are designed to teach. This data base retrieved for comparison with the student response. of facts is stored in memory in an informationnetwork. For example, given the initial context"South This research -is sponsored by the U.S.Office of America" and the attribute "Countries" SCHOLARmight Education under Grant 0EG-0-72-0895. Computer services ask one of the following: were supported by NSF Grant GS-9. What are the countries in South America? 0

Session 2.3.1

FILMED FROMrBEST AVAILABLE COPY .

Is it true that Argentina is a country in South this can be cut by a factor of 15 if compiled code America (T/F) is used. A Country in South America is--- (fill in).

There is a certain probability that Another example of an ISO Teacher is the" system he cdntext designed by Wexler (7,8). will change from the initial unit to Wexlei uses an information a related unit. structure that consists of classes, objects within Hence, a subsequent question might be: those classes, and links between objects. Sample Select an Alternative from the list: classes from the-geography of Canada might be Paysardu "Provinces" and "Cities" with "Newfoundland" and Rio De Janerio "Quebec" examples of objects in theformer class Buenos Aires and "St. John's" and "Quebec City" examples of Uruguay Riirer objects in the latter class.There could be a link In the question: between the objects St. John's and Newfoundland and between the objects Quebec City and Quebec. The What is the capital of Argentina? course-author constructs a chain of links in the network by specifying an object together with its class The actual generation of questions is handled name and the next object (with associated claSs name) by a set of procedures which are available for forming in the chain. each of the above question types. Figure2 gives an example of an The question type information structure from this system. to be asked is determined probabilistically.As has been mentioned, the correct answer to a question is derived when the question is generated. The information structure utilized might be Since all though' of as a skeletal representation of relation- expected student responses are numbers or object ships. It is considerably more streamlined than the names, the problem of recognizing a correct student response isminimized. semantic network and, consequently,_ additional data The teacher can specify the must be supplied in order to interpret it. acceptable degree of inaccuracy in numericalanswers and the degree of misspelling allowed for object This data consists of a set of skeleton segments names. and their interpretation strings. A sample segment might be the following: Since SCHOLAR is a mixed-initiative system, the student can interrupt with questions of hisown such STATE *V1 LINK TO CITY *V3 as "Tell me-about Argentina" or "What is the capital city in Argentina". Associated with this segment might be the interpreta- SCHOLAR analyzes the student's tion string question and determines which of the elements in the set (attribute, object, value) are missing.SCHOLAR *V3 is a city in *VI then generates an appropriate reply. This statement pair means that if there isa link in the information network,between VI, a member In the case of "Tell me, about Argentina," of the class "state", and V3,a member of the class SCHOLAR would trace through the portion of the net- "city", then the assertion_V3 isa city in V1 is work originating,at the node for the unit Argentina valid. Skeleton-segments an be combined into skele- and generate a set of statements. SCHOLAR keeps ton patterns of 011 form-"If A then B" where A isa track of the irrelevancy tags to makesure that all statements generated-are pertinent. boolean expression formed from a set of skeleton seg- If the student ments and B is the explanation string for the complete cosies back with "Tell me more about-Argentina," .then skeleton pattern. An example follows: SCHOLAR will generate statements whichare somewhat more remote than the initial statements. _ OSP1 IF (STATE *V1 LINK TO CITY *V2) AND (*V2 LINK TO CAPITAL *V3) It is obvious that the burden on the instructor for specifying a CAI session is minimal. He need AND (*V3 LINK TO *V1) only specify an initial context, the length of the THEN THE CAPITAL OF *V1.IS *V2 session in terms of maximum time and minimum number of questions, the probability of switching to a sub- The skeleton patterns are used by the system to context for question generation, and the degree of generate statements and remedial aid and to verify irrelevancy allowed for a subcontext. student responses. The statement generated would take the form of the explanation string. The major task for the instructor is supplying In skeleton the semantic network. pattern one (SP1) above, either VI or V2 must be Currently, the semantic network specified. is input manually. The other variable would be determined Carbonell is working on an inter- from the network and the appropriate information face to simplify this procedure while Quinlan (5) statement presented to the student. is concerned with the general problem of automatically constructing semantic networks. In order to generate questions, a questioncon- struct must be supplied. The original implementation of SCHOLARwas in BBN LISP- (6) on the XDS -940 time-sharingcomputer. It OQ1 NAME THE CAPITAL OF *P1 utilized approximately 144K 24-bit words. This in- @0K al SP1 (V1 P1, V2 Al)_ cludes 35K words for the LISP system, 90K wordsfor SCHOLAR itself, and 12K words for the semanticnet - The parameter P1 in the resulting question would bea work.The time requiree-to answer a student's member of the class state and could be prespecified question is about a minute. It is anticipated that by the instructor or selected at randorn.The student response, Al, would be considered correct if the three

Session 2.3.2

N First USA JAPAN Computer- Conference. 1972

conditions of SP1 were satisfied within the informa- an experimental ISO Generative System. oir system, tion network when Al and P1 were substituted for V2 called IT1 (Interpretive Tutor 1), is similar to and V1 respectively. Carbonell's scholar in that it uses-a-aemantic network to store its data base. Their system; however, If the objects Al and P1 applied to SP1 did not does have the capability to automatically construct lead to a valid "trace" then two types of prompting the semantic network from text supplied by the course- information could be supplied to the student.The -author. The text is the same as that which will be system could search for a correct trace of SP1 using seewby the student. Itis necessary for a linguist V2 e Al and Vi unspecified.Finding a correct trace (or the lesson designer) to monitor this process and would set the value of V1 in the explanation string. occasionally intervene to select from a set of poss- Thus, if the student had_responded "Providence" to ible alternative representations constructed by the the question "Name the capital,of Connecticut", the system. prompting information "The capital of Rhode Island is Providence" would be generated. The system functions as a tutor in the following way. The student reads the text which has been pre- An alternate form of help is available. The sys- pared by the course - author." If he cannot understand tem could substitute the correct value of V2 when a complex sentence, he types in "EXPLAIN" followed V1 e P1 and cycle through the skeleton segments of by the number of -the sentence.The system will then SP1 and their associated interpretation strings. - generate a set-of simpler sentences from that portion of the semantic network derived from the original The result night be sentence. If the student does not know the meaning of a word in a particular sentence, he types in Hartford is a city in Connecticut "DEFINE" followed by that word and the sentence Hartford is a capital number. Th. particular meaning required is ret- One city in a state is the capital rieved from-the lexicon (alto supplied by-the The capital of Connecticut is Hartford linguist) and displayed to the student. Table 1 As-soon as the'student realizes his mistake, he can depicts a result of the "EXPLAIN" command, -

- halt this sequence. The remaining teaching (unction is similar to The teacher has considerable range in specifying .SCHOLAR'S (though less powerful)-in that the system his instructional program.An instructional_ sequence ginerates TRUE-FALSE and FILL-IN type questions. will consist of a series of information statements followed by one or more questions concerning these The student selects the section of- text on which statements. The entire teaching strategy can be- he wishes to be quizzed.The system then generates of specified by the teacher using appropriate command a set of questions from the corresponding portion statements. More interesting modes of instruction the semantic network.FILL-IN type questions Are are the paramet erized and fully generative mode. produced by omitting a word fiemrthegenerated_sent- ence; TRUE-FALSE questidnwitripteduced by substitu- In the parameterized mode, the teacher plans the ting*(or not substituting in the TRUE case) for-a mein line and remedial sequences of instruction by word in the generated sentence. specifying the order of information statements and questions. However, he can utilize skeleton patterns In summary, three systems have been presented to simplify this task and can either specify para- which make use of information networks. These sys- meters for these patterns or have them selected at tems are experimental in nature and-would not yetbe They are oriented- random. In the fully generative mode, the teacher practical for large-scale use. specifies only which skeleton patterns he wishes towards the "soft-sciences". They combine a question- exercised and the system will generate -a sequence answering or information-retrieval capability with of information statements and questions using theit question-generation; thus, allowing the student to patterns. assume the initiative when he desires to explore interesting points at greater depth. The Wexler There is also the capability for a student to system, especially, has the capability to generate ask questions of the system.The system attempts to prompting information relevant to a student's incorr- recognize the object* and class-names-in the student's ect response. question and then searches for skeleton patterns that would produce traces which contain these-objects. The Once the semantic network has been formed, very actual traces are then presented to the.student in a little information need be supplied by the teacher in sequential fashion. If a student is interested in order for SCHOLAR to generate a rich teaching sequence. the resulting trace produced by a particular skeleton Due to the relative sparseness of its information net- pattern, he can request more traces from this same work, the Wexler system requires additional information pattern. 'n the form of skeleton patterns and question constructs in order to generate a teaching sequence. .How- This system is implemented in ALGOL on a ever, it allows the instructor to assume more ,control Burroughs B5500 with 32K words of core memory._ The over what form this teaching sequence will take. Also, system was operated in a time-sharing environment the information network is somewhat easier to const- which can handle up to 12 users at one time. It was ruct in the Wexler system. found necessary for a single user to utilize the full capacity of the machine in order to obtain reasonable III. GENERATIVE CAI IN TEACHING QUANTITATIVE MATERIAL response times ( 2 minutes or less). The potential.for incorporating generativeCAI Simmons and Melton (9,10) have recently described

Session 2-3-3 in the "hard sciences" is extensive. These courses The problem is then presented to the student and are quantitative in nature and teach techniques of his answers are checked. If he is incorrect, the problem solving. Problem solving competence is often solution process is explained to him.The form of acquired through a process of "learning by doing" in this explanation depends on the type of problem being which a student is required to solve a representative solved. A sample is shown in Table 2. sample of problems. Siklosay (16) describes a methodology for the Algorithms for solution of classes of problems design of a computer tutor which teaches the concepts could be incorporated into CAI systems. In some cases, of set union and intersection.This tutor can ran- _ solution techniques might be sufficiently complex domly generate problems or accept sample sets from that heuristic programs would be necessary. Examples the student. If a 'student's response is incorrect, it of the latter case would be teaching symbolic integra- is analyzed and the appropriate remedial feedback tion (11) or proving theorems (12). In any event, given. The types of errors which can be detected for CAI systems organized_ around a set of algorithms would a set intersection problem are missing elements, have the capability to generate and-solve a wide range elements-are present which'belong to only one of the of problems. sets, elements are present which belong to neither set, etc. Several generative programs have been written by Uhr and his associates-which deal with elementary Hoffman and Seagle (17) describe-a methodology arithmetic or algebra (13,14,15).These-programs for a generative teaching system in which the instr- generate a series of examples for drill-and-practice. uctor-apecifies an individual course for each student. The-difficulty of the problem generated is determined The material to be covered, the problem areas and by the magnitude of the numbers used.The system models to be explored, and the tools which are to be -keeps track of a student's progress and attempts to made available are determined from a data base that provide-him with a suitable-problem. describes the student's background.The student is quizzed on his ability to select the correct tools Lists of correctly answered and incorrectly ans - - -for a problem solution as weP. as his ability to use inked problems are maintained. Ak student will norm- them. ally be presented with a question which he-previously missed. -If-there are no problems which he has yet to The authors claim that a wide varietyrof-pheno- solve-correctly, a new problem will be generated_for mena in physics, economics and various social sciences him.. There_iealso a possibility he will be asked can be deicribed-by analytical models using polynomial to_reviesra problem which is on the correctly ans- expressions. Specifying the details of the process wered list.- to-be modelled is equivalent to setting a aeries of linear constraints on -the polynomial coefficients. If he answers incorrectly, an appropriate reme- By using the techniques of linear programming, suit- dial comment is generated.As an example he might be able sets of coefficients satisfying these constraints told: "Wrong. 3+4 is not equal to 6.Your answer can be obtained. These coeificients would be used in is too small". -- the formulation of the problem to be solved. This procedure has been applied to the generation of Uhr (13) describes a possible extension which economies problems. would define a set of complex operations in terms of a smaller set of primitive operators.In determining An extensive project in the subject areaef an answer to a-complexJproblem, the computer would analytical geometry has been described by Uttal (23). carry out the prescribed set of primitive operators. His system is capable of generating twelve problem If the student's answer was incorrect, the system types which are representative of the problems found could backtrack through its Solution process. It in an analytical geometry course. These-problems would deteriine the point at which the-student's usually involve an expression or graphical representa- _solution diverged from the correct path and explain tion of a particular conic section. The expression the procedure that should have been followed. is also obtained from a general quadratic equation of the form: AX2+BY2+CX +DY +E'O. The Peplinski (15) system generates either linear or quadratic equations.These equations are separa- The required expression is obtained by -letting ted into cheese, such as linear equations with terms certain coefficients to 0 and selecting the others in "X" on the left side only, "X" 2n troth sides at random.The complexity of the equation generated quadratic equations of the form aX4-bm0,aX2-b2:0, depends on-the-constraints imposed on the coefficients. VI AX2+bX+C0, etc. -The student chooses the type of For example, to generate circles centered at the problem he would like. The system sets a series of origin, A0B and C.D.O. indicators depending on the problem type chosen.The indicators tell the system which problem option has Associated with each of the twelve problem types been selected and guide the solution process. For is an answer routine. The routine which determines example, one indicator states that both answers are if a randomly generated point (x,y) falls on the to have the use absolute value; another states that locus represented by a randomly generated equation the problem will require simplification as the coeffi- simply plugs this point into the equation. The cient of the "12" term is not 1.The parameters in expression generator itself is used as the answer the general equation (aX+b) (cX+d)-ale selected at routine when the student is asked to supply the random within the constraints imposed by the indica- equation for a conic section with given standard tors. characteristics.

Session 2.34 First USAAl'AN Computer Conference, 1972

To evaluate the correctness of a student response effective teaching of a variety of computer science a"binary branching tree partitioner" nay be utilized. fundamentals through generative CAI. This algorithm is applied to trace algebraic and arithmetic errors by successively decomposing the IV. GENERATIVE CAI IN DIGITAL SYSTEMS expression into smaller left and right branches. If the error is judged to be in the left branch, this The author is developing a generative CAI branch is further decomposed until the source of the systems which will be capable of generating anti solving student error has been pinpointed. a wide range of problems in the area of digital systems. (18,19) This CAI system is currently king The last-system to be discussed here, "An Assoc- used in the introductory computer science course iative Learning Project by Rochart, et al., (24), offered by the Electrical Engineering Department at could also have been included in the section on ISO the University of Connecticut.The course coverage Generative Systems since it uses a simplified form includes an introduction to combinational and seq- of a semantic network to answer student questions. uential design as well as machine and assembly lang- However, it has been designed to teach accounting, uage programming as described in (20). a course which is quantitative in nature. The system is designed to be extremely flexible This system has no capability to generate ques- in that it can completely control the progress of -a tions and, in fact, normally operates as a convention- student through the course, selecting concepts for al frame-oriented system. However, if the student study on an individual basis and generating problems. desires, he can interrupt tie lesson sequence to ask Alternatively, the student can assume the initiative a question. The question is scanned for any keywords and deternine his owtt areas of study and /or supply his (words essential to the subject matter), question own problems.He can also increase or decrease the words, (what, when, where, and how used) and relation amount of explanation and monitoring he receives while words. Stored with each key word is a "property list" Working on a problem. Each of the question words is a property whose value is a statement relating the question word and keyword. In addition, the system also operates in a After the keyword and question word have been identi- "problem-solver" node. In this nodi, the student fied, the appropriate explanatory stateaent is print- specifiei-the concept area and his problem, and the ed out.. An example follows: system will crank out the solution without further interaction. It is anticipated that students in EXAMPLE of Property List for ASSET: later courses and the digital laboratory will utilize WHAT = is tangible for intangibl- property this mode for solving complex minimization problems USE = can be used to generate fu.. ce value and determining the relatives merits of different WHEN = exists if there is value remaining state assignments. at balance sheet tine WHERE = is noted on the left-hand side of When the system is in control of the interaction, a balance sheet it attempti to individualize the depth and pace of REL = (is a form of capital)(is increased instruction presented to each student.A model of in value by credit)(is equivalent to each student is kept which summarizes his past per- property)(is measured by value)... formance in each of the course concepts.In addition, the system is supplied with a concept tree which A student's question such as "What is an asset?" indicates the degree of complexity (plateau) of a would be answered by the statement: "Asset is concept and its relationship with other concepts in tangible or intangible property." the course.

In addition, if there are two keywords recognized The system uses this information to determine in the question, a statement (or set of- statements)- how quickly a student should progress through the expressing their relationship will be retrieved tree of concepts, the particular path which should be and printed out.This statement will be stored with followed, the degree of difficulty of the problem to each of the keywords under the property "reiationshlp". be generated, and the depth of monitoring and explana- If-the keysiords are not directly related, a path tion of the problem solution. through intermediate keywords-is searched for. -Tor example, if the keywords are "assets" and "wealth" The course is organized as a set of solution the path would be "WEALTH IS ECONOMIC VALUE.VALUg algorithms. Normally there is a single algorithms for MEASURES AN ASSET." each major concept of the course. The algorithms solve the problems much as a eV:dent would, breaking The authors feel that such a system, while not each problem down into a series of sub-tasks. After generative, still presents a useful alternative to each sub-task is accomplished, a decision is made _frame - oriented CAI. It providet the student with whether or not to question the student on this part the-capability to ask questions when they arise and of the problem solution.This decision is based on to temporarily divert from the-normal sequence of the student's current level of achievement (a real frames. number between 0 and 3) in the concept and the difficulty of the sub-task. The following describes a generative tutor which is utilized in conjunction with an intelligent monitor If the student is questioned, then his answer of student progress. The objective is to provide is Compared with the system's solution. If the stu- a greater degree of individualization of instruction dent is correct, his level is increased; if he is than is currently available as well as enable incorrect, be will receive a remedial comment

Session 2.3.5 explaining the correct solution procedure and his In order to generate problems, a probabilistic level for that concept will be decreased. The high- grammar must be specified for each problem class. er a student's level, the fever questions he will be A probabilistic grammar is a formal language in asked. When the student reaches a level of 3 in a which each production rule is assigned a probability concept, the system will solve subsequent problems of being applied. These probabilities are defined dealing with this concept for Mn. Table 3 presents in such a way that the production rules leading.to examples of different degrees of interaction witha student in an octal addition problem. sore difficult problems have higher probability as The first a student's level increases. character of all student inputs is understored. In thititay,_the difficulty of the problem generated is tailored .to 0 fit each individual student. The magnitudes of the increment and'decrement for correct and incorrect answers respectivelyare The system has been implemented on the IBM 360/65 also individualized to fit the student's pastper- at the University of Connecticut. formance in a concept. The mechanism for accomplish- It is programmed ing this is described elsewhere (21). in the Conversational Programming System (CPS) (22) CPS is a dialect of PL/I and includes some string Since there are a large number of concepts processing features which have been extreeely us,- ful in programing this system. available for study, the system attempts to select the_next Concept in such a way as to make optimal V.CONCLUSIONS use of the student's time.The goal is to pace the student through the concepts quickly enough so that he does not become bored or unmotivated and yet This paper has attempted to survey wwide range not of generative CAI systems. so fast that he becomes unduly confused. The purpose of these systems is to reduce the amount of effort required There is no set_order in which the concepts are by the instructor to produce new instructionalsat - trial. selected nor is there a set level of achievement Also, these systems are intended to free up which every student must exceed in order to advance. the instructional process and, provide the student with the capability to assume the initiative and The algorithm attempts to individualize concept selec- investigate topics which interest his. tion through examination of the student's performance record. Some of these systems have extensive capabilities for natural language communication. Each student is assigned &master average when They are, thus, able to interpret a student's questions and generate he first logs onto the CAI system.This could be a &meaningful, well-phrased response. These systems function of his I.Q. or class standing.Currently, are provided with a data base of factual information each student is arbitrarily assigned an initial and would be applicable to instruction in the "soft easter average of 2. His master average is updated - sciences". after the completion of a concept.

A student's master average controls the speed The quantitative systems have less of a need for with which he jumps from one plateau of the concept understanding natural language, though, they could be provided with this capability. tree to the next. In order to jump to the next Their "data base" consists of a set of algorithms for problem genera- higher plateau, the average of his levels of achieve- tion and solution. They are oriented towards-provid- ment in all concepts at and below the current plateau ing a student with practice in problem solving. must exceed his master average. Consequently, the lover a student's master average, the faster be will progress. A generative system for teaching digital computer concepts has been described in detail. This system Once the student's plateau has been determined, includes an intelligent executive routine which attempts to individualize the path taken by each the system chooses one concept from among the candi- student through the tree of course concepts. dates at that plateau. Each concept is evaluated In based on a number of factors such as the time elapsed addition, the instruction and sonitoring provided by since its last use, the "stability" of its current each solution algorithm are dynamically adapted to suit a student's current level of knowledge. level, the sign and magnitude of its most recent The system can also operate as a problem-solver and will level change (negative changes are weighedsore provide portions of the solution process which have heavily), and its relevance to the other conceptsas determined by the number of branches of the tree already been mastered by -a student. connected to it. The highest scoring concept is All of the systems discussed are somewhat- experi- selected for presentation to the student.The stu, !mental in nature. dent always has the option of-vetoing this selection It-in expected that the use of generative CAI wall increase as research in this area and choosing his own concept or accepting thesys- tem's second best choice. continues and facilities for CAI become more available and sore economical. After a concept has been decided upon, the REFERENCES system must generate a problem within this concept area unless the student prefers to supply one. The 1. Simmons, R. F., rtiatural Language Question- Answer- system attempts to tailor the problem difficulty to suit the student's level in that concept. ing Systees:1969," Colmunications of the ACM, Vol. Table 4 13, Ho. 1, January 1970,_pp. 15-30. gives some examples of problems which may be generated .

Session 2-34 First/SA-JAPAN Computer Conference. 1972

2. Carbonell, J. R. "AI in CAI: -An Aritificial 17, Hoffman, R. B., and Seagle, J. P.,"A Problem Intelligence Approach to Computer- Assisted Oriented Computer-Based Instructional Procedure," Instruction,"IEEE Transactions on-Man-Machine Proceedings of 24th ACM National Conference: Systems, Vol. MKS-11, No.4, Dec. 1970pp.190-202. pp. 97-110, 1969.

3. Carbonell, J. R., "Mixed- Initiative Man-Computer 18. Koffman, E. B., "Individualizing Instruction in Instructional Dialogo04," Bolt Beranek and A Generative CAI Tutor,"Communications of the Newman Report No. 1971, Cambridge, Mass., May ACM, June, 1972, pp. 472-473. 1970. 19. Koffman, E. B.i "A Generative CAI Tutor for 4. Quinlan, M. R., "Semantic Memory," in Minsky, Computer Scien e Concepts," Proceedings of the M. (ed) Semantic Information Processing,M.1.T. AFIPS 1972Spring Joint Computer Conference. _ Press, Cambridge, Mass. 1968, PP. 227-270. 20. Booth, T. L., Digital Networks and Computer 5. Quillian, M. R., "The Teachable Language Cour Systems, Wiley and Sons, New York, 1971. prehender: A Simulation Program and Theory of Language," Communications of the ACM, Vol. 12, 21. Hall, J. N., "Generation:A Computational No. 8, Aug. 1969, pp. 459-476. Tutor," M.S. Thesis, Department of Electrical Engineering, University of Connecticut, May, 6. Bobrow, EL G., Murphy, D. L. and Teitleman, W., 1971. "The BBN-LISP System,"Bolt, Beranek and New- man, Cambridge, Mass. 1968. 22. IBM Corporation, "Conversational Programming System (CPS) Terminal User's Manual," IBM 7.Wexler, J. D., "A Generative, Remedial and Query Report GH-20-0758-1, 1970., -System for Teaching by Computer,"Ph. D. Dissertation, Univ. of Wisconsin, Madison, 23. Utall, W. R., Pasich, T., Rogers, M. and March 1970. Hieronymus, 2., "Generative Computer Assisted Instruction," Communication No. 243, U.Aaf. 8. Wexler, J.-D., "Information Networks in Genera- Michigan, Mental Health Research Institute, tive Computer-Assisted Instruction," IEEE 1969: Transactions on Man-Machine Systems, Vol. MMS- 11, No. 4, nee. 1970, pp. 181-190. 24. Rockart, J. E., Scott Morton; M.S. and Zannetos, Z. S., "Associative Learning Project-Phase-1 9. Simmons,R. F., "Natural Language for Instructional System," Working Paper, A.-P- Sloan Schoiol of Communication,"In Artificial Intelligence and Management, M.I.T., Jan 1972. Heuristic Programming, Edinburgh Univ. Press 1971, pp. 191-198.

A 10. Melton, T. R., "IT1: An-Interpretive Tutor," Table 1 Technical Report No. NL-4, Dept. of Computer Sciences and CAI Laboratory, Univ. of Texas, 1971 Use of "EXPLAIN"

11. Moses, J., "A Program for Drilling Students in SENTENCE:" The first plant to appear on a newly formed Freshman Calculus Integration Problems," Project tropical island is the stately and graceful MAC Report MAC-M -369, M.I.T., March 1968. coconut palm.

12. Robinson, J. A., "A Reviescof Automatic Theorem- EFFECT OF EXPLAIN COMMAND: -Proving". In Schwartz, J. T. Proceedings of The plant is the coconut palm 19th Symposium in Applied Mathematics, American The plant appears first on a tropical island Mathematical Society,,Providence1967, pp 1-18. The tropical island is newly formed The coconut palm is stately and graceful 13.Vhr, L., "Tiaching Machine Programs That Generate Problems as a Function of Interaction With Students,"Proceedings of 24th ACM National Table 2 Conference: 125-134, 1969. Remedial Feedback for Algebra Problem 14. Wexler, J. D., "A Teaching Program That Generates Simple Arithmetic Problems,"International 4X**2-25*0 Journal Man-Machine Studies, Vol. 2, No. 1, YOUR ANSWERS ARE X = 2 1/3, X = -2 1/3 January 1970, pp. 1-27. GENERATED ANSWERS ARE X . -2 1/2, X g. 2 1/2 15. Peplinski, C., "A Generating System for CAI Teach- ing of Simple Algebra Problems," Technical Re- YOUR HAVE MADE AN ERROR.. THE PROBLEM IS port 24, Computer Science Dept., University of Wisconsin, Madison, May 1968. 4X**2 -25 g. 0 THE EXAMPLE IS A DIFFERENCE OF TWO SQUARES. AND 16. SiklOssy, L., "Computer Tutors That Know What They Teach,"Proceedings Fall Joint Computer THUS FACTORS INTO THE SUM AND DIFFERENCE OF THE Conference 1970, pp. 251-255. SQUARE ROOTS,_AS FOLLOWS

Session 2.3-7 a

(ZX + 5) (2Y. -5) =0 Table 4 THEN 2X = -5 and 2X = 5. Sample Problems Generated DIVIDING TO SOLVE, S * -2 1/2 or X= 2 1/2 1. CalculateA3.E - 2.B4 in the base16 using- complementsubtraction. Table -3 2. Derive thetruth table for (((P(3)Q)f R)V(Q 4 R)) Different Degrees of Interaction

3. Design acombinational logic circuitsuch that: Problem: Form the sum of 75.3, 2.7 in the base 8. Input Ais 2 bits long. LEVEL RANGE 0-1 Input Bis 2 bits long. There isa control bit I such thatif I = 0, Modify 2.7-for addition to 75.3 the outputX = A + B; if I = 1, theoutput X = A *B. 02.7

Correct.The rightmost digit is digit 11. 4. Minimizea function which has 0,1, 3, 4, 10 as minitermsand 2, 8, 14 as don'tcares using What is the decimal sum of digit #1 KarnaughMaps. of 75.3 and digit 11 of 02.7?

5. Design amodulo 6 counter with-asingle input line X. No.Their decimal sum is 3+7 or 10. If X = 0the count increases by 1, What is digit #1 of the if Xs1 the the base 8? count decreasesby 2. .1 The outputshould be 1 when the countis O. No. The decimal number 10 is equivalent to 12 6. Design asequential circuit-to recognizethe input in base 8. Digit 11 is 2. sequence0011. What is the carry to digit 12? 7. Derive the 1 transition table for asequential network with2 Flip-Flops such that31 = (Y2 V X), Correct. What is thR decimal sum of digit 12 K1 = (Y1V Y2): S2 = (-.4X A (Y1 VY2)), R2 = X. of 75.3 and digit 12 of 02.7 and the carry from digit 11? COUNTRY (SUPERCONCEPT (STATEINDEPENDENT)) (EXAMPLES ._ARGENTINA LEVEL RANGE 1-2 The rightmost digit is digit il. SOUTH AMERICA) What-is digit 11 of the sum in the base 8? (SUPERCONCEP .CONTINENT) URUGUAY (COUNTRIES . URUGUAY 2 . ARGENTINA .) Correct. What is the carry to digit 12? 0 No. 347 = 12 in the base 8. .ARGENTINA (SUPERCONCEPT . COUNTRY) The carry to digit #2 is 1. (LOCAT . SOUTH AMERICA) BORDERING COUNTRIES) What is digit 12 of the sum in the base 8? (BRAZIL . URUDPAY) /

LEVEL RANGE 2-3 Figure 1 Part of a.Semantic Network on South America What is the complete sum of 75.3, 02.7 inthe base 8? 27.2 CLASS OBJECTS No. The sum of 75.3, 02.7 is 100 .2 in thebase 8. NEWFOUNDLAND QUEBEC Providence LEVEL RANGE 'L 3 NS *JEBEC The sum of 75.3, 02.7 is 100.2 in the base 8. City

Capital 525 000 57853 505000 Population 1961 Year. Figure 2 Part of an Information Network

Session2.3.8