JP5382698B2 - Automatic exercise system and automatic exercise program - Google Patents

Description

  The present invention relates to an automatic exercise system and an automatic exercise program, and more particularly to a content-independent automatic exercise system and an automatic exercise program that can be applied to a wide range of fields for the purpose of passing on knowledge and know-how related to daily work in factories and offices. .

  As a result of a shift in the style of equipment maintenance from post-maintenance maintenance (BM: Breakdown Maintenance or FBM: Failure Based Maintenance) to time-based preventive maintenance (TBM: Time Based Preventive Maintenance) The frequency of maintenance has decreased, and the stable operation rate of equipment has greatly improved. For this reason, there is a problem that the chances of encountering troubles due to the improvement of the stable operation rate of the equipment are reduced, but the experience of equipment maintenance personnel is insufficient, and knowledge and know-how for troubles are not sufficiently transferred. In particular, the mass retirement of the baby boomer generation with skilled knowledge and expertise (2007 issue) highlights this problem. Therefore, the time and work costs when dealing with sudden troubles and low frequency troubles may be higher than before.

  In order to deal with such problems, factory and office performance information is stored in the database, but this only makes sense when actually used, and in addition, a person with skilled knowledge and know-how Should be able to be used regardless of whether or not they are enrolled. For this reason, there is a demand for the construction of a technology that enables the transfer of knowledge and know-how accumulated in factories and offices even when there is no person in charge with skilled knowledge and know-how.

  Therefore, when looking at this type of conventional technology, for example, the test problem database generation system and test problem creation system of Patent Document 1, the training system apparatus of Patent Document 2, the education apparatus of Patent Document 3, and the like, non-patent Conventional techniques such as automatic generation of problems using a large-scale document collection (on the WWW) of documents 1 and 2 as a knowledge source are known. In addition, in Non-Patent Document 3, automatic generation of exercises that automatically generate and give training exercises in selection format, hole-filling format, and error correction format based on information extracted by analyzing data in the document System AEGIS is reported, and Non-Patent Document 4 reports a system that automatically generates a problem for proofreading by performing morphological analysis and dependency analysis on a technical document.

JP 2003-167506 A JP 2000-338849 A JP2007-233159A

J Sheard and A Carbone. Cadal quiz: Providing support for self-managed learning? In Proc.World Conference on the WWW and Internet, 2000, pp. 482-488, 2000. Etsuo Kobayashi, Shinobu Nagashima, and Mitsuaki Hayase.Programming-free web-based automatic online drill / quiz creator.In Proc.World Conference on Educational Multimedia, Hypermedia and Telecommunications, 2001, pp.990-991, 2001. Kakinuma et al., IPSJ Journal, Vol.46, No.7, pp.1810-1818, 2005 Ohno et al., IEICE technical report, pp.39-44,2007

When considering the transfer of knowledge and know-how of daily operations in factories and offices, the source of knowledge is, for example, daily report information database, specifically, trouble management daily report information that is used in automobile assembly factories. That is, “a set of natural language sentence data independent of specific contents”.
However, all of the techniques disclosed in Patent Documents 1 to 3 and the like depend on specific content. The purpose is to provide interaction and interface to the user, and to reduce the amount of work when the teacher creates teaching materials and problems, so the application scene is limited to specific scenes such as language learning and programming learning. is there. Therefore, the techniques disclosed in Patent Documents 1 to 3 and the like are difficult to apply in order to transfer knowledge and know-how in factories and offices. In addition, knowledge and know-how related to daily operations in factories and offices are not general. Therefore, methods using WWW as in Non-Patent Documents 1 and 2 are not very useful for passing on knowledge and know-how in daily work in factories and offices where technical terms are frequently used.

  In addition, the daily report information database (a kind of trouble management daily report information, that is, “a set of natural language sentence data independent of specific content”) has a huge amount (for example, about 60000 for 8 years of daily report information). ). When handling such an enormous amount of data with the system of Non-Patent Document 3, it is necessary to manually prepare for tagging because the document has a dedicated XML tagged corpus as an analysis target. This advance preparation involves costs that cannot be overlooked because of manual labor. Therefore, actual operation by the system of Non-Patent Document 3 is not realistic from the viewpoint of cost.

  Furthermore, a database operated in a specific environment such as a factory or office such as the above-mentioned daily report information database (a kind of trouble management daily report information, that is, `` a set of natural language sentence data independent of specific contents '') There is a lot of technical terminology, the sentence itself is written with punctuation marks, particles, auxiliary verbs omitted, diaphoresis, etc. There are circumstances. Therefore, the existing general-purpose knowledge dictionary and the method using morphological analysis such as Non-Patent Document 4 cannot accurately analyze the descriptive sentence, and the problem generation ability varies depending on the target document. .

  Therefore, regardless of the type of knowledge source, it can be used universally for the transfer of knowledge and know-how of daily operations in factories and offices, without the need for manual tagging or the specific construction of specific content in advance. Moreover, it is desired to construct an “automatic exercise system that can inherit knowledge and know-how of daily work in factories and offices” that can process a huge amount of natural language sentence data such as a daily report information database.

  The present invention has been made in view of the above circumstances, and important words and causal relationships (e.g., events, causes, treatments, etc.) from natural language sentence data accumulated in a database as knowledge and know-how of daily work in factories and offices. To provide an automatic exercise system and automatic exercise program that can perform automatic problem creation processing, user answer automatic scoring processing, and automatic exercise result aggregation processing based on this Objective. That is, an object of the present invention is to provide an automatic exercise system and an automatic exercise program that can transfer knowledge and know-how of daily work in factories and offices without depending on contents.

In order to solve the above problems, an automatic exercise system according to the present invention is:
From a natural language sentence database comprising a plurality of significant fields having a causal relationship, a vector space table comprising a column position information field and a vector space field for each significant field, and an index word field for each significant field And a knowledge collection means for constructing a knowledge database comprising an importance table including an importance field and an index word position information field,
An index word included in a query is extracted as an important word from the importance table as an element of a query vector space, and the query vector space and a vector space of all columns of the question field selected from the significant fields are used. Ranking list creating means for obtaining similarity between the query and all the columns of the question field, and creating a ranking list ranking all the columns of the question field based on the similarity (or in order of the similarity) When,
In the ranking list, a question sentence with a similarity Rα equal to or higher than a threshold value α is applied to a question creation template to create a question sentence, and an answer field selected from the significant field corresponding to the similarity Rα is a correct answer option. The main point is to provide QA pair creation means.

“Causal relationship” means that each field constituting the database has a mutual relationship such as an event, a cause, a treatment, a countermeasure, and an occurrence location, but is not limited thereto.
A “significant field” is a field in which each field of the database is meaningful, such as an event, cause, action, countermeasure, and occurrence location, in other words, a field consisting of natural language sentence data, but is not limited to this. Not.
“Natural language sentence database” refers to a daily report information database and trouble information database in factories and offices, but is not limited thereto as long as it includes natural language sentence data.

“Knowledge database” refers to a database created by collecting knowledge from a natural language sentence database. In the present invention, it is assumed that such knowledge is obtained by collecting knowledge for each significant field. , “Vector space table” and “importance level table”.
The “vector space table” includes a “column position information field” and a “vector space field”. Here, “column position information” refers to position information that can identify the record number of the natural language sentence database of the column in which the natural language sentence data is stored. Therefore, since the “position where the record number can be specified” is sufficient, for example, the position information in the knowledge source table (a table obtained by copying the natural language sentence database for each significant field, the same applies hereinafter) is also included. The “vector space” is an element obtained by extracting an index word included in a column in which natural language sentence data is stored as an important word.

  In addition, the “importance level table” includes an “index word field”, an “importance field”, and an “index word position information field” for each significant field. Here, the “index word” refers to an index word extracted for each significant field by a character string class group generation unit, an index word candidate group generation unit, and an index word extraction unit, which will be described later, included in the knowledge collection unit. . The “importance” refers to a numerical value representing the importance of an index word, and can be obtained by, for example, Equation 2 described later. “Index word position information” refers to position information that can specify the record numbers of the natural language sentence database of all columns in which the index word appears. Therefore, since the “position where the record number can be specified” is sufficient, for example, the position information in the knowledge source table (a table obtained by copying the natural language sentence database for each significant field, the same applies hereinafter) is also included.

  A “query” refers to natural language sentence data randomly selected or extracted from natural language sentence data or a natural language sentence database input by a user when a problem is automatically created. From the viewpoint of the user, the natural language sentence data as the “query” is a question sentence close to a problem that the user wants to create, a word included in the question sentence, or the like.

“Question field” refers to a significant field selected as a field for creating a problem from a plurality of significant fields in the natural language sentence database, but also includes the same significant field in the knowledge source table. In the present invention, any of the event field, cause field, treatment field, countermeasure field, and occurrence location field can be used as the question field, but the event field may be designated as the question field in advance.
The “similarity” is a numerical value obtained by calculating the degree to which the compared natural language sentence data is similar using the vector space, and can be obtained by, for example, Expression 3 described later.
“Answer field” refers to a significant field selected as a field for creating a question (answer) from a plurality of significant fields in the natural language sentence database, but also includes the same significant field in the knowledge source table. In the present invention, any of the event field, cause field, action field, countermeasure field, and occurrence location field can be set as the answer field, but the cause field may be designated as the answer field in advance.
“Problem creation template” refers to a template provided with an embedded field for embedding question fields and / or contents of fields such as events, causes, measures, countermeasures, and occurrence locations.

According to the automatic exercise system according to the present invention having the above-described configuration, the vector space table and the importance level table are created for each significant field from the natural language sentence database by the knowledge collection unit. Then, the index list included in the query is extracted as an important word as an element of the query vector space from the importance table by the ranking list creating means, and the query vector space and the vector space of all the columns of the question field are used. After the similarity between the query and all the columns of the question field is obtained, a ranking list is created that ranks all the columns of the question field based on the similarity (or in order of the similarity). . Thereafter, the QA pair creation means creates a question sentence by applying the question field having the similarity Rα equal to or higher than the threshold α to the question creation template in the ranking list, and the answer field corresponding to the similarity Rα is the correct answer option. It is said.
Therefore, according to the present invention, the problem (question sentence Q and correct answer A) is automatically created without depending on the content. Since the question field having the similarity Rα equal to or higher than the threshold value α is applied to the question creation template, a question according to the user's request is created.

The automatic exercise system according to the present invention further includes:
In the ranking list, the answer field and the correct answer option using the vector space of the answer field corresponding to the similarity Rβ less than the threshold value β, the vector space of the correct answer option, and / or the vector space of other incorrect answer options. And / or when the similarity Rγ with the other incorrect answer option is obtained and the similarity Rγ is less than the threshold value γ, an incorrect answer creating means using the answer field corresponding to the similarity Rβ as the incorrect answer option is provided. You may prepare.

  According to the automatic exercise system having the above-described configuration, the incorrect answer creating means causes the vector space of the answer field corresponding to the similarity Rβ less than the threshold value β, the vector space of the correct answer option, and / or the like in the ranking list. When the similarity Rγ between the answer field and the correct answer option and / or the other incorrect answer options is obtained using the vector space of incorrect answer options of the above, and the similarity Rγ is less than the threshold γ, The answer field corresponding to the similarity Rβ is an incorrect answer option. Therefore, according to the present invention, the problem is automatically created without depending on the content. Since the answer field corresponding to the similarity Rβ less than the threshold β is an incorrect answer option, an incorrect answer option far enough from the correct answer option is obtained, and among them, the correct answer option and / or other incorrect answer options Since an answer field having a similarity γ less than the threshold value γ is regarded as an incorrect answer option, it is avoided that the answer field has the same content as the correct answer option and other incorrect answer options.

  In the automatic exercise system according to the present invention, for example, the similarity Rα may be the highest similarity in the ranking list, and the similarity Rβ is the most unselected as an incorrect answer option in the ranking list. A low similarity (however, more than 0) may be used.

In order to solve the above problems, an automatic exercise system according to the present invention further includes:
A preparation means for obtaining a question to be asked from the problem database based on the user information database;
A question scoring means for displaying the question on the interface screen, scoring a user answer to the question, and displaying the result on the interface screen together with correct answer information;
Management means for updating the user information database and the problem database based on the result may be provided.

  According to the automatic exercise system having the above-described configuration, the preparation means acquires a problem to be asked from the problem database based on the user information database, the question scoring means displays the problem on the interface screen, and The user's answer to the question is scored, and the result is displayed on the interface screen together with correct answer information. Then, the user information database and the problem database are updated by the management means based on the result. Therefore, according to the present invention, in addition to automatic problem creation, it is possible to perform automatic scoring processing of user answers and automatic tabulation processing of exercise results. If the user uses the automatic exercise system according to the present invention, the knowledge and know-how of daily work in factories and offices can be inherited without a dedicated dictionary or the like.

The automatic exercise means according to the present invention further includes:
An index word included in the correct answer information is extracted from the importance level table as an important word as an element of the correct answer information vector space, and the correct answer is obtained using the correct answer information vector space and the vector space of all the columns of the question field. Find the similarity between the information and all the columns of the question field, create a ranking list that ranks all the columns of the question field based on the similarity (or in order of similarity), and display the contents Reviewing means may be provided.

  According to the automatic exercise means according to the present invention, by the review means, the index word included in the correct answer information is extracted as an important word from the importance degree table as an element of the correct answer information vector space, and the correct answer information vector space, Similarity between the correct answer information and all the columns of the question field is obtained using the vector space of all the columns of the question field, and based on the similarity (or in order of the similarity), A ranking list that ranks all the columns is created and its contents are displayed. Therefore, the user can check a case similar to the mistaken problem, which can be used for establishing knowledge and know-how.

Although the knowledge collection means in the automatic exercise system according to the present invention is not particularly limited, for example, one having the following configuration is preferable. That is, the knowledge collecting means
For each significant field, a suffix group and / or a prefix group is generated by sorting in lexicographic order as a character string class, cutting out a matching portion of the preceding and following character string classes, excluding a character string class whose appearance frequency is less than a threshold T1, String class group generation means for sorting so that the upper string class in an inclusive relation is prioritized over the lower string class, and integration of the same string classes
For each significant field, when there is no inclusion relationship with the previous character string class of the character string class, the character string class is stored as an index word candidate. If the frequency ratio is greater than or equal to the threshold T2, the lower character string class is stored as an index word candidate, while if the frequency ratio is less than the threshold T2, the lower character string class is regarded as the upper character string class and the index word Index word candidate group generation means for storing as candidates,
Index word extraction means for extracting, as index words, index word candidates that overlap among the index word candidates of each group,
For each significant field, the index words included in all the columns are extracted as important words, a vector space having the important words as elements is generated for all the columns, and the natural language sentences in all the columns are generated. A vector space table creating means for creating a vector space table including a position information field capable of specifying a record number of the database and a vector space field;
For each significant field, the importance level of the index word is calculated, and for each significant field, the index word field, the importance level field, and the record numbers of the natural language sentence database of all the columns in which the index word appears are recorded. An importance level table creating means for creating an importance level table having a position information field that can be specified may be provided.

  The “inclusion relationship” refers to a relationship in which one character string class includes the other character string class. In the “upper character string class” and “lower character string class”, the included character string class is the upper character string class, and the included character string class is the lower character string class. For example, “Chain” is the upper character string class and “Cha” is the lower character string class.

  According to the knowledge collection means of the automatic exercise system according to the present invention, the character string class group generation means generates the suffix group and / or the prefix group as a character string class sorted in dictionary order for each significant field. The matching part is extracted from the character string class, the character string class whose appearance frequency is less than the threshold T1 is excluded, the same character string class is integrated, and the upper character string class in the inclusion relationship has priority over the lower character string class To be sorted. Then, the index word candidate group generation means stores the character string class as an index word candidate when there is no inclusion relationship with the character string class immediately preceding the character string class for each significant field. If it is in an inclusive relationship, a frequency ratio is obtained, and if the frequency ratio is greater than or equal to the threshold T2, the lower string class is stored as an index word candidate, while if the frequency ratio is less than the threshold T2, the lower order is stored The character string class is regarded as the upper character string class and is stored as an index word candidate. Then, the index word extraction means extracts index word candidates that overlap in each group among the index word candidates of each group as index words.

Then, the vector space table creating means extracts, for each significant field, the index words included in all the columns as important words, and a vector space including the important words as elements is generated for all the columns. A vector space table including a position information field capable of specifying the record number of the natural language sentence database of all the columns and a vector space field is created. Further, the importance table creation means obtains the importance of the index word for each significant field, and for each significant field, the index word field, the importance field, and all columns in which the index word appears. An importance level table including a position information field capable of specifying a record number of the natural language sentence database is created.
Therefore, as long as the automatic exercise system according to the present invention has a database in which daily report information, trouble information, and the like are recorded in natural language sentence data, knowledge collection can be performed regardless of the type.

In this case,
The frequency ratio is
Gap (C (n-1), Cn) = | log (Ti (n-1) / Tin) | (Formula 1)
(Where C (n-1): upper string class inclusive relation with Cn, Cn: lower string class inclusive relation with C (n-1), Gap (C (n-1), Cn ): C (n-1), Cn frequency ratio, Ti (n-1): Appearance frequency of upper character string class, Tin: Appearance frequency of lower character string class), but not limited to this Not.

In the automatic exercise system according to the present invention, the importance is
Swj = log {(Fj / Dj) +1} (Formula 2)
(Here, Swj: importance, Fj: appearance frequency, Dj: co-occurrence frequency) can be obtained, but the present invention is not limited to this.

…(式３)、により求めることができるが、これに限定されない。 Furthermore, in the automatic exercise system according to the present invention, the similarity is as follows:

(Equation 3) can be obtained, but is not limited to this.

  The description of the configuration and effect of the automatic exercise program according to the present invention will be replaced with the description of the configuration and effect of the automatic exercise system according to the present invention.

Since the present invention has the above-described configuration, important words and causal relationships (for example, events, causes, measures, countermeasures, occurrence points) from natural language sentence data stored in the database as knowledge and know-how of daily work in factories and offices. There is an effect that automatic problem creation processing, user answer automatic scoring processing, and practice result automatic counting processing can be performed based on this.
Since the present invention has the above-described configuration, there is an effect that knowledge and know-how regarding daily work in a factory or an office can be inherited without depending on contents.

FIG. 2 is a system configuration schematic diagram of an automatic exercise system 1 according to an embodiment of the present invention, in which FIG. (A) is an example in the case of a server client system, and (b) in FIG. 1 is a block diagram showing a hardware configuration and a software configuration of an automatic exercise system 1 according to an embodiment of the present invention. It is a flowchart which shows the process sequence of automatic exercise program A00 which concerns on one Embodiment of this invention. It is a flowchart which shows the process sequence of the user management program B00 which concerns on one Embodiment of this invention. It is a flowchart which shows the process sequence of the knowledge collection program C00 which concerns on one Embodiment of this invention. It is a flowchart which shows the process sequence of the character string class group production | generation program C100. It is a flowchart which shows the process sequence of the index word candidate group production | generation program C200. It is a flowchart which shows the process sequence of the importance level table creation program C500. It is a flowchart which shows the process sequence of the problem preparation program D00 which concerns on one Embodiment of this invention. It is a flowchart which shows the process sequence of the exercise program E00 which concerns on one Embodiment of this invention. It is a figure which shows an example of the daily report DB5 which can apply this invention. It is a figure which shows an example of vector space table Tb9A among knowledge DB9. It is a figure which shows an example of importance level table Tb9B among knowledge DB9. It is a figure which shows an example of problem DB10 created by this invention. It is a figure which shows an example of user information DB11 used by this invention. FIGS. 4A to 4F are diagrams showing examples of screens displayed during execution (A01 to A03) of the automatic exercise program A00. An example of the knowledge source table Tb13A created for each significant field of the daily report DB5 is shown. It is a figure explaining suffix generation (before sorting). It is a figure explaining suffix generation (after sorting). It is a figure which shows an example of index word candidate table Tb14A. It is a figure for demonstrating the index word candidate group production | generation program C200. It is a figure which shows an example of the query designation | designated screen G15 currently displayed by D09. It is a figure which shows an example of the ranking list | wrist RL created by D11c. It is explanatory drawing of the process compared with threshold value (alpha), when selecting Q-A pair (pair of question sentence Q and correct answer A) based on the ranking list | wrist RL in D12-D14. It is explanatory drawing of the process compared with threshold value (alpha) like FIG. 17A. It is a figure which shows an example of the question creation template TP. It is a figure which shows the example which selected the correct answer A from the answer field and registered Q-A pair. It is a figure which shows the example which selected the incorrect answer B from the answer field, and registered the incorrect answer option. It is explanatory drawing for demonstrating the judgment process whether the similarity degree of the selected incorrect answer choice and the already selected incorrect answer option exceeds threshold value (gamma). It is explanatory drawing which shows an example in which the problem was completed. An example in which an incorrect answer B is selected from the answer field and an incorrect answer option is registered is shown (modified example with specified similarity). It is a figure which shows an example of the problem display screen G23. It is a figure which shows an example of the result display screen G24. It is a figure which shows an example of the result display screen G25. It is a figure which shows an example of the review screen G26. It is a figure showing an example of end screen G27. It is a figure which shows an example of the reply error warning screen G28.

  Hereinafter, an automatic exercise system and an automatic exercise program according to an embodiment of the present invention will be described in detail with reference to the drawings.

[Automatic exercise system 1]
FIG. 1 is a system configuration diagram schematically showing a hardware configuration of the automatic exercise system 1, and FIG. 2 is a block diagram showing a hardware / software configuration of the automatic exercise system 1. In these drawings, the automatic exercise system 1 analyzes a plurality of significant fields (for example, events, causes, and treatments) having a causal relationship by analyzing natural language sentence data of an arbitrary database (for example, daily report information database) in a factory or office.・ A knowledge collection process (see FIG. 5A) that collects knowledge information such as location information L, index words, importance, vector space, etc. This is a system for performing a problem creation process (see FIG. 6) for creating a problem by using, an exercise interface (see FIG. 7) for providing a practice interface to the user, and scoring the user answer input.

  The automatic exercise system 1 includes an automatic exercise program A00 and the like for executing the flowcharts shown in FIGS. 3 to 7 on a server 2 (automatic exercise device 2) (see FIG. 1A) and a personal computer 3 (automatic exercise device 3). ) (See FIG. 1B), it can be constructed by either a server client method or a stand-alone method. In addition, the server 2 or the personal computer 3 is connected to the daily report DB 5 accumulated by the DB system 4 of the factory or office ("DB" means a database, the same shall apply hereinafter), the telephone line 7 and / or the LAN 8, etc. It is connected so that it can communicate. Thereby, the server 2 and the personal computer 3 can receive and read data from the daily report DB 5, and can transmit and write data to the daily report DB 5. Details of the daily report DB5 will be described later.

  In the case of the server-client method (see FIG. 1A), the automatic exercise system 1 includes a server 2 and a client 6 as shown in FIG. The server 2 includes a storage unit 2a (also referred to as ROM 2a and RAM 2a), a control unit 2b, an input unit 2c (also referred to as keyboard 2c and mouse 2c), an output unit 2d (also referred to as screen 2d and speaker 2d), and various devices. The storage unit 2a is composed of a ROM and a RAM, and the RAM is an automatic exercise program A00 (see FIG. 3), a user management program B00 (see FIG. 4), knowledge, which describes the procedures for executing the flowcharts of FIGS. A collection program C00 (see FIGS. 5A to 5D), a problem creation program D00 (see FIG. 6), and an exercise program E00 (see FIG. 7) are installed, as well as a knowledge DB 9 (see FIG. 9) and problems DB10 (see FIG. 10), user information DB11 (see FIG. 11), and other various programs and data are stored. The ROM stores programs and various data for controlling the server 2 and various devices. Details of the knowledge DB 9, the problem DB 10, and the user information DB 11 will be described later.

  The control unit 2b includes a CPU (central processing unit) and an OS (operating system), calls the programs A00 to E00, sends a control command to each unit of the device, and executes a processing procedure described in the program. It functions as automatic exercise means (that is, user management means, problem creation means, exercise means). Details of these processing procedures will be described later. Further, the control unit 2b executes various programs based on user input, sends various commands to each unit of the device for execution, executes other various programs, exchanges various data, and controls / communications as an operating system. Controls functions and other parts of the device.

  The input unit 2c is not particularly limited as long as it is used for data input, and is configured by a keyboard, a mouse, a pointing device, etc. as input means. The output unit 2d is not particularly limited as long as it presents data by image, print, or sound, and is configured by a screen, a speaker, a printer, or the like as output means.

  The client 6 includes a storage unit 6a (also referred to as ROM 6a or RAM 6a), a control unit 6b, an input unit 6c (also referred to as keyboard 6c or mouse 6c), an output unit 6d (also referred to as interface screen 6d), and various devices. The storage unit 6a includes a ROM and a RAM. The RAM stores various programs and data, and the ROM stores programs and various data for controlling the client 6 and various devices. The control unit 6b includes a CPU (Central Processing Unit) and an OS (Operating System), and executes various procedures based on user input in addition to executing procedures for using the automatic exercise program A00 by connecting to a server. Various commands are sent to and executed by various parts of the apparatus, various other programs are executed, various data are transferred, control / communication functions as an operating system are controlled, and other parts of the apparatus are controlled.

  The input unit 6c is not particularly limited as long as it is used for data input when the automatic exercise system 1 is used in the server client system, and is configured by a keyboard, a mouse, a pointing device, and the like as input means. The output unit 6d serves as data presentation means (e.g., display means, printing means, notification means) when the automatic exercise system 1 is used in a server client system, and presents data by image, printing, or sound. If it is not limited, it is comprised by an interface screen, a speaker, a printer, etc.

  In the case of the stand-alone system (see FIG. 1 (b)), the personal computer 3 shown in FIG. 1 only needs to correspond to one obtained by replacing the reference numerals 2a to 2d in FIG. 2 with reference numerals 3a to 3d. Although not shown, the storage unit 3a (also referred to as ROM 3a, RAM 3a), control unit 3b, input unit 3c (also referred to as keyboard 3c, mouse 3c), output unit 3d (also referred to as screen 3d, speaker 3d), various devices Prepare. Among these, the storage unit 3a and the control unit 3b have configurations corresponding to the storage unit 2a and the control unit 2b of the server 2, respectively, and the input unit 3c and the output unit 3d are respectively the input unit 6c of the client 6. A configuration corresponding to the output unit 6d is provided. The storage unit 3a may be shared, and the automatic exercise program A00 may be installed in each of several personal computers 3.

[Daily DB5]
FIG. 8 shows an example of the daily report DB5. The daily report DB 5 is a database that serves as a knowledge source for problem creation by the automatic exercise system 1, and stores knowledge and know-how related to daily work in factories and offices in natural language sentence data. In other words, as shown in the figure, the daily report DB5 is a daily report of work reports (here, daily trouble management work reports for factories) that are reported by field personnel in natural language sentence data. The “language sentence database” is not limited to this as long as it includes natural language sentence data.

As shown in FIG. 8, the daily report DB5 includes a plurality of significant fields having a causal relationship, an event field (event significant field), a cause field (cause significant field), a treatment field (treatment significant field), a countermeasure field (a countermeasure significant field). )-If the record that consists of the occurrence location field (occurrence location significant field) is used as a unit and the data constituting each field is composed of natural language sentence data or data that includes this part, the type of significant field The contents are not particularly limited.
Here, the “causal relationship” means that there is a mutual relationship such as an event, a cause, a treatment, a countermeasure, and an occurrence location, but is not limited thereto. In addition, from the above, “significant field” means the event field, cause field, action field, countermeasure field, and occurrence location field that make up the daily report DB5, but it is a meaningful field, in other words, natural language. The field is not limited to these as long as the field is composed of sentence data.
As described above, the automatic exercise system 1 can use such a daily report DB as a knowledge source for problem creation, so it does not depend on content, can be applied to various factories and offices regardless of the type of industry, and a specific database is stored in advance. do not need.

[Knowledge DB9]
9A and 9B show an example of the knowledge DB 9 created by collecting knowledge from the daily report DB 5 (natural language sentence database) by executing the knowledge collection program (see C00 in FIG. 5A). As a specific example obtained by collecting knowledge as described above, FIG. 9A shows an example of the vector space table Tb9A, and FIG. 9B shows an example of the importance degree table Tb9B. These tables are created for each significant field, and vector space tables Tb9A1 to Tb9A5 and importance level tables Tb9B1 to Tb9A1 for all significant fields (event field, cause field, action field, countermeasure field, and occurrence location field) of the daily report DB5. Tb9B5 is provided. The processing procedure of the knowledge collection program C00 will be described later.

The vector space table Tb9A is created for each significant field (event, cause, treatment, countermeasure, occurrence location) of the daily report DB5.
(1) Field of position information L (position information that can specify record numbers of all columns in which index words appear in the natural language sentence database (daily report DB5), record number of knowledge source table Tb13A, and so on) Location information field)
(2) A vector space field (vector space field) including all important words included in each column of the significant field as elements.

In addition, the importance table Tb9B is created for each significant field (event, cause, action, countermeasure, occurrence location) of the daily report DB5,
(1) Index word field (index word field) included in each significant field,
(2) The importance level field (importance level field) of the index word obtained by Equation 2 described later,
(3) a position information L field (position information field) in which the index word appears.
Here, the “index word” means a character string class group generation program (see C01), an index word candidate group generation program (see C02), an index word extraction program (C03), which are called from the knowledge collection program (see C00), which will be described later. Refers to the index word extracted for each significant field by executing (see). Further, the “importance” refers to a numerical value representing the importance of an index word, and can be obtained by, for example, Equation 2 described later.

[Problem DB10]
FIG. 10 shows an example of the problem DB 10. The problem DB 10 is created by executing the problem generation program D00, and a part thereof is updated by executing the exercise program E00. The question DB 10 has (1) question number, (2) question sentence Q, (3) difficulty, (4) total number of questions, (5) total correct answer number, (6) correct answer A, (7) incorrect answer B ( B1 to B4) and a database composed of fields such as answer scores.

[User information DB11]
FIG. 11 shows an example of the user information DB 11. The user information DB 11 is created by executing a user management program B00 described later, and a part thereof is updated by executing the exercise program E00. The user information DB 11 is composed of user recognition information and user result information. (1) Number, (2) User name, (3) User ID, (4) Password (not shown), (5) Number of trials (total exercises) The number of trials), (6) the number of correct answers (total number of correct answers), and (7) a database consisting of fields such as cumulative scores 1 to n for each exercise trial.

[Operation of Automatic Exercise System 1 and Processing Procedure of Automatic Exercise Program A00]
The operation of the automatic exercise system 1 according to an embodiment of the present invention and the processing procedure of the automatic exercise program A00 will be described with reference to flowcharts and the like.

[Automatic exercise program A00]
In the automatic exercise program A00 of FIG. 3, when the control unit 2b detects a start input from the user (A01), the control unit 2b proceeds to A02, calls and executes the user management program B00, and functions as a user management unit. That is, the control unit 2b displays the user authentication screen G12A of FIG. 12A and performs user management processing (user authentication and user registration) (A02). When the user authentication is successful, the control unit 2b proceeds to A03, displays the user process selection screen G12B of FIG. 12B, and waits for user input (A03).

When detecting that the knowledge collection button B12B1 is pressed in A03, the control unit 2b functions as a knowledge collection means by calling and executing the knowledge collection program C00, and for each significant field, the vector space table Tb9A and the importance level A table Tb9B is created and stored in the knowledge DB 9 (A04). Details of the knowledge collection process will be described later.
Further, when the control unit 2b detects pressing of the question creation button B12B2 in A03, the control unit 2b functions as a question creation unit by calling and executing the question creation program D00 to create the question illustrated in FIG. The problem is stored in DB10 (A05). Details of the problem creation process will be described later.
Furthermore, when the control unit 2b detects that the exercise button B12B3 is pressed in A03, the control unit 2b functions as an exercise means by calling and executing the exercise program E00, and reads out the problem from the problem DB 10 and performs a predetermined number of problem assignment processes. The question is given until completion, and correct / incorrect determination of the user answer and analysis of the determination result are executed (A06). When the knowledge collection process (A03), the problem creation process (A04), or the exercise process (A06) ends, the control unit 2b displays the user process selection screen G12B and waits for user input (A03). The control unit 2b executes the automatic exercise program A00 by the above procedure.

  By using the automatic exercise system 1 or the automatic exercise program A00, the user can acquire knowledge and know-how related to daily work in factories and offices. The automatic exercise system 1 and the automatic exercise program A00 are useful for obtaining on-site know-how and special skills that are frequently used on-site technical terms, regardless of the field or industry.

[User management program B00]
When the control unit 2b executes A01 in FIG. 3 and proceeds to A02, calls the user management program B00 in FIG. 4 and proceeds to B01, the control unit 2b functions as a user management unit. When the control unit 2b detects that the authentication button B12A1 is pressed in the state where the user ID “ise” and the password “******” are input as shown in the user authentication screen G12A in B01 (B01 → B01a: no ), Refer to the user information DB 11 (B02). Then, the control unit 2b determines whether or not the corresponding user has been registered, and when determining that the user has been registered (B03: yes), refers to the number of trials and the number of correct answers in the user information DB 11 as user comprehension information (B04). . Then, the control unit 2b displays the user process selection screen G12B, and waits for user input of any one of the knowledge collection button B12B1, the problem creation button B12B2, the exercise button B12B3, and the end button B12B4 (B05 (A03)) ).

  On the other hand, if B01 detects that the authentication button B12A1 is pressed without being correctly entered in the user ID input box and password input box shown on the user authentication screen G12A in B01 (B01 → B01a: no), refer to the user information DB11 in B02 After that, it is determined that the corresponding user is not registered (B03: no), the authentication error warning screen G12C of FIG. 12C is displayed (B06), and user input is accepted (B07). The control unit 2b proceeds to B08 when detecting the pressing of the new registration button B12C1 (note that the control unit 2b also proceeds to B08 when detecting the pressing of the new registration button B12A2 in B01a).

  When the control unit 2b displays the new user registration screen G12D of FIG. 12D in B08, accepts user input, and detects that the registration button B12D1 is pressed while the user information is input in the input box BX of FIG. (B09), it is determined whether or not the input user information satisfies the specified condition (B10) .If it is determined that the condition is satisfied (normal input) (B10: yes), the process proceeds to B11 and input The user information is stored in the user information DB 11 to be newly registered (B11). As a result, the new user registration screen G12E shown in FIG. 12E is displayed, and pressing of the return button B12E3 shown in FIG. 12 is detected. Then it returns to B03.

When the controller 2b detects that the return button B12C2 on the authentication error warning screen G12C is pressed in B07, the controller 2b returns to B01. If the control unit 2b determines in B10 that the user input on the new user registration screen G12D is not a normal input, the control unit 2b displays a registration error warning screen G12F in FIG. 12F (B13). When detecting that the return button G12F1 is pressed in B13, the control unit 2b performs error processing (B14), deletes the input user information, and then returns to B07.
The control unit 2b executes the user management program B00 by the above procedure, and when user authentication is successful (B03: yes), the process proceeds to B05 (i.e., A03) via B04, and waits for user input. In response, the following knowledge collection program C00, question creation program D00, exercise program E00, etc. are executed.

[Knowledge collection program C00]
The knowledge collection program C00 includes character string class group generation processing (C01), index word candidate group generation processing (C02), index word extraction processing (C03), vector space table creation processing (C04), importance level table creation processing (C05). ) Is provided for causing the control unit 2b to execute. These processes are collectively referred to as knowledge collection processes (C01 to C05). When performing each process, the control unit 2b functions as a character string class group generation unit, an index word candidate group generation unit, an index word extraction unit, a vector space table generation unit, and an importance level table generation unit. Cases that function to perform each of these processes are collectively referred to as knowledge collection means.

  This will be described with reference to the flowchart below. The control unit 2b executes A01 and A02 in FIG. 3 and proceeds to A03. When the control unit 2b detects pressing of the knowledge collection button B12B1 in A03, executes A04, that is, calls the knowledge collection program C00 and proceeds to C01 in FIG. 5A. , Function as knowledge collection means. Thereby, as the knowledge collection process, the control unit 2b obtains the importance level table Tb9A and the vector space table Tb9B illustrated in FIG. 9 for each significant field from the natural language sentence database including a plurality of significant fields having causal relationships. Process to build knowledge DB9 provided.

  In C01, the control unit 2b functions as a character string class group generation unit by calling and executing the character string class group generation program C100 of FIG. 5B. As a character string class group generation process, a suffix is added for each significant field. Group and / or prefix group is created by sorting in lexicographic order as a character string class, extracting the matching part of the preceding and following character string classes, excluding character string classes whose appearance frequency is less than threshold T1, and integrating the same character string classes And the sorting process is performed so that the upper character string class in the inclusive relation has priority over the lower character string class. In the following description, the suffix group will be described. However, the control unit 2b performs the same processing as the suffix group for the prefix group.

  When the control unit 2b proceeds to C101 of the character string class group generation program C100 via C01, the position information L and the natural language sentence data of all the columns of each significant field are obtained from the daily report DB5 for each significant field. A knowledge source table Tb13A (including an event knowledge source table, a cause knowledge source table, a treatment knowledge source table, a countermeasure knowledge source table, and an occurrence location knowledge source table) illustrated in FIG. Processing to be stored in 2a is performed (C101).

  In C102, for each knowledge source table Tb13A, the control unit 2b converts the text constituting the read natural language sentence data from all the characters existing in the text, as illustrated in the table Tb13B of FIG. 13B. Recognize that it is a set (n-gram set) of partial character strings (suffix: prefix) up to the end, and create a table in which the set (suffix) is mechanically arranged for each position information L. In C103, for each knowledge source table Tb13A, the control unit 2b sorts the suffixes in lexicographic order for each natural language sentence data having the same position information L. In C104, as illustrated in the table table Tb13C of FIG. 13C, the control unit 2b creates a table in which all the suffixes are integrated for each significant field and the whole is sorted in dictionary order for each significant field.

  In C105, the control unit 2b sets predetermined values for the counter i and the suffix total number j. The counter i is set to 0 as an initial value, and the total number of suffixes “sorted and integrated” in C104 is set to the total number of suffixes j. In C106, the control unit 2b compares the “i-th suffix” with the “(i−1) -th suffix” and determines whether or not there is a matching portion. When determining that there is a matching part (C106: yes), the control unit 2b proceeds to C107, cuts out the matching part as the character string class Ci, and stores the cut out character string class Ci in association with the position information L. Incidentally, as apparent from the above, “position information L” specifies the record number of the knowledge source table Tb13A that is the origin of the extracted character string class Ci, that is, the record number of the daily report DB5. Note that here, even if there is a single character, the character string class Ci is extracted.

  For example, if the (i-1) -th suffix is “empty DP chain broken” and the i-th suffix is “empty DP chain”, the duplicate “empty DP chain” is the i-th character string class Ci. Cut out. However, even if there is an overlap, if the (i-1) -th suffix is "empty DP chain break" and the i-th suffix is "DP chain", it will not match from the initial letter. In such a case, it is determined that there is no matching part.

  Then, after performing C107 or determining that there is no matching part in C106, the control unit 2b increments i to make the processing target the next suffix in C108 and whether or not all suffixes have been processed in C109. If it is in the middle of the determination, the process returns to C106 to perform the process for the next “(i + 1) th suffix”. The control unit 2b creates a table composed of the position information L and the extracted character string class by executing C105 to C109, and stores the table in the storage unit 2a. Since the suffix sorting method to be processed is not particularly limited, the i-th and (i + 1) th may be compared or the i-th and (i-1) -th may be compared depending on the sorting method. And (i + 1) th may be compared at a time.

  In C110, the control unit 2b calculates the appearance frequency Ti of the character string class cut out from the suffix, and includes a table of the cut-out character string class field, the position information L field, and the appearance frequency Ti field. Is created and stored in the storage unit 2a. Here, the “appearance frequency Ti” is counted for each significant field, including the character string class C that has not been cut out. Then, the control unit 2b proceeds to C111.

In C111, the control unit 2b sorts the extracted character string classes in dictionary order. At this time, the control unit 2b sorts those with the same first character so that the upper character string class is in the previous array and the lower character string class is in the subsequent array. For example, the control unit 2b sorts so that the first array is “chain” and the second array is “Cha”.
In C111, the control unit 2b further integrates the same character string class. For example, when there are a plurality of “chains”, the appearance frequency Ti and the position information L are kept, and these are combined into one. Then, the control unit 2 identifies the field of the integrated character string class and the field of the position information L (the record number of all knowledge source tables Tb13A in which the extracted character string class appears, that is, the record number of the daily report DB5) A table including an appearance frequency Ti field and an exclusion flag field is created and stored in the storage unit 2a.

  Next, in C112, the control unit 2b determines whether the appearance frequency Ti of the character string class is less than a predetermined threshold T1, and excludes the character string class when the appearance frequency Ti is less than the threshold T1. To do. Here, the threshold value T1 can be arbitrarily set, but may be a constant, or may be a ratio to the total number of appearance frequencies of all character string classes. The reason for excluding character string classes with a small appearance frequency Ti that does not satisfy a certain threshold is based on the fact that such character string classes are often meaningless in many cases. Specifically, the character string class is excluded by setting a flag in the corresponding column of the above-described exclusion flag field. For example, when “T1 = 5”, a flag is set in the exclusion field with an appearance frequency of less than 5. Note that the above table may be re-created without setting the flag and excluding the same character string class to be excluded.

  With the above procedure, the control unit 2b performs the character string class group generation process. The control unit 2b performs the character string class group generation process in the same manner by generating a prefix group. For processing of prefix groups, the control unit 2b sets a set of partial character strings (prefix: suffix) from the top of the text to all characters existing in the text, which constitutes the read natural language sentence data ( The process is the same as above except that it is recognized as an (n-gram set). Therefore, the description of the suffix group is replaced with the description of the prefix group.

The control unit 2b, when proceeding to C02 in FIG. 5A, calls and executes the index word candidate group generation program C200, functions as index word candidate group generation means, and performs index word candidate group generation processing for each significant field. For each group (suffix group, prefix group), the character string class C (i-1) immediately preceding the character string class Ci generated in C01 and the inclusion relationship (one character string class is the other character In the case where “Chain” is not in the upper character string class and “Cha” is in the lower character string class, for example, the character string class Ci is stored as an index word candidate. If there is an inclusive relationship, the frequency ratio is
Gap (C (n-1), Cn) = | log (Ti (n-1) / Tin) | (Formula 4)
(Where C (n-1): upper string class inclusive relation with Cn, Cn: lower string class inclusive relation with C (n-1), Gap (C (n-1), Cn ): C (n-1), Cn frequency ratio, Ti (n-1): Appearance frequency of the upper character string class, Tin: Appearance frequency of the lower character string class), and the frequency ratio is equal to or greater than the threshold T2. In this case, the lower character string class is stored as an index word candidate, while if the frequency ratio is less than the threshold T2, the lower character string class is regarded as the upper character string class and stored as an index word candidate by the longest match method. I do. In the former case, the lower character string class “Cha” is stored as an index word candidate as it is. In the latter case, the lower character string class “Cha” is stored as an upper character string class “Chain”. In the following description, the suffix group is taken up and explained. However, the prefix group is processed in the same manner as the suffix group.

  When the control unit 2b calls the index word candidate group generation program C200 of FIG. 5C in C02 and proceeds to C201, it sets the counter i to 0 and sets the total number j of character string classes. At this time, the total number j in which the character string class for which the exclusion flag is set is excluded is set.

In C202, the control unit 2b determines that the "character string class Ci to be processed (i-th character string class)" is the previously read character string class C (i-1) It is determined whether or not it is in an inclusive relationship with ((i-1) th character string class).
If it is determined that the control unit 2b is not in an inclusion relationship (C202: no), the control unit 2b proceeds to C205, and stores the i-th character string class Ci in the storage unit 2a together with the position information L as an index word candidate. Specifically, for example, the control unit 2b performs the index word candidate table Tb14A including the index word candidate field illustrated in FIG. 14A and the position information L field for each significant field.

  On the other hand, if the control unit 2b determines that there is an inclusion relationship (C202: yes), the appearance frequency Ti (i) of the character string class Ci (i-th) in the inclusion relationship in C203 and the character string class C (i-1 ) ((i-1) th) appearance frequency Ti (i-1) is used to determine the frequency ratio Gap according to the above equation 1, and it is determined whether the frequency ratio Gap is less than the threshold T2. As described above, the string class is sorted so that the upper string class is the first and the lower string class is the later array for those with the same first character, so the (i-1) th The character string class C (i-1) is the upper character string class, and the i-th character string class Ci is the lower character string class.

When the control unit 2b determines that the frequency ratio Gap is equal to or greater than the threshold value T2 (C203: no), the index word candidate field and the position information L field are used together with the position information L using the lower character string class Ci as the index word candidate. Is stored in an index word candidate table Tb14A consisting of
On the other hand, when determining that the frequency ratio Gap is less than the threshold T2 (C203: yes), the control unit 2b compresses the lower character string class Ci to the upper character string class C (i-1) based on the longest match method (C204). The lower character string class Ci is regarded as the upper character string class C (i-1) and stored in the index word candidate table Tb14A (C205). In other words, "upper character string class C (i-1)" as an index word candidate of the (lower) character string class Ci, along with the position information L of the (lower) character string class Ci, the field of the index word candidate and the position information L Are stored in the index word candidate table Tb14A consisting of these fields. Therefore, at the time of generating the vector space described later, even if the lower character string class is included in the natural language sentence data of the corresponding field, the lower character string class is regarded as the upper character string class, and the upper character string class. A character string class is extracted as an element constituting a vector space. When ending C205, the control unit 2b returns to C202 via C206 and C207, and performs processing for the next “(i + 1) th character string class”. C202 to C207 are repeated until i becomes equal to the total number of suffixes J.

  Here, the index word candidate group generation processing of C200 will be described with reference to FIG. 14B. This figure conceptually shows how index word candidates are generated. First, the threshold T1 is set to “T1 = 5”, and the threshold T2 is set to “T2 = 2.0”. For example, when comparing the character string classes C1 “13 Rambus Heating” and C2 “13 Rambus”, these are inclusive relations and are calculated according to the above Equation 1, then “Gap (C1, C2) = | log8 / 8 | = Since “0 <threshold T2”, “13 Rambus” is compressed into “13 Rambus heating”, in other words, “13 Rambus heating” is considered and registered in the index word candidate table Tb14A, that is, “13 Rambus” Is registered as “13 Rambus Heating”. For example, when comparing the character string class C2 “13 Rambus” and the character string class C3 “CL13 Rambus heating”, the character string class C3 is regarded as an independent class because C2 is not in an inclusive relationship including C3. Then, it is registered in the index word candidate table Tb14A. In addition, for example, when comparing the character string class C4 “Ranbus heating freeze” and C5 “Ranbus”, these are in an inclusive relationship, so when calculated by the above equation 1, “Gap (C20, C21) = | log8 / 1000 | = 2.096> threshold T2 ”, the character string class C5 is registered in the index word candidate table Tb14A as an independent class.

  When the index word candidate group generation process (C02) is completed, the control unit 2b proceeds to C03 in FIG. 5A. In C03, the control unit 2b functions as an index word extraction unit, and as an index word extraction process, for each significant field, an index that overlaps each group among index word candidates of each group (suffix group, prefix group). A process of extracting word candidates as index words is performed. This process is performed, for example, by creating an index word table (not shown, see index word candidate table Tb14A) having the same data structure as the index word candidate table Tb14A. This is because the index word table only needs to include an index word field including the generated index word and a position information L field including all position information L in which the index word appears.

  When finishing the index word extraction process (C03), the control unit 2b proceeds to C04 in FIG. 5A. In C04, the control unit 2b functions as a vector space table creation unit, and as a vector space table creation process, from the daily report DB (or knowledge source table Tb13A), for each significant field, an index word ( The index word generated by executing C03 or the index word candidate generated by executing C02 may be extracted as an important word, and a vector space having the important word as an element is generated for all the columns, A vector space comprising a field of position information L (that is, a record number of the daily report DB5 or a record number of the knowledge source table Tb13A) of all the columns and a vector space field (extracted important words are stored) Create table Tb9A. That is, in C04, the control unit 2b refers to the index word table (not shown, refer to the index word candidate group table Tb14A) and displays the column displayed in the position information L of the daily report DB (or knowledge source table Tb13A). Search all index words included, position information L as a key field, extract the index words included in all the columns of each significant field represented by the position information L as an important word as an element of the vector space, The vector space table Tb9A is created by sorting in the order of the position information L (record number).

When finishing the vector space table creation process of C04, the control unit 2b proceeds to C05 of FIG. 5A. In C05, the control unit 2b reads and executes the importance table creation program C500, thereby functioning as an importance table creation means.
For each significant field, the importance of the index term is
Swj = log {(Fj / Dj) +1} (Formula 2)
(Where Swj: importance, Fj: appearance frequency, Dj: co-occurrence frequency),
For each significant field, the index word field, the importance field, and the natural language sentence database of each column in which the index word appears (i.e., the daily report DB 5, here may be the knowledge source table Tb13A) A process of creating an importance table Tb9B having a field of position information L for specifying a record number is performed.
Here, “appearance frequency Fj” means the appearance frequency of the index word in each significant field. “Co-occurrence frequency Dj” means type information (number of differences) of key words that co-occur with key words in each significant field.

  In C501, the control unit 2b sets an initial value 0 to the counter i and sets the total number of index words for each significant field to the total number of data j. In C502, the control unit 2b obtains the co-occurrence frequency Ci. The control unit 2b refers to an index word table (not shown, see index word candidate table Tb14A) having an index word field and a position information field, obtains position information L where a certain index word appears, By referring to the spatial table Tb9A, the total number of types of other index words included in the column indicated by the position information L is counted and obtained as the co-occurrence frequency Ci. Therefore, any number of co-occurring patterns is counted as one.

  In C503, the control unit 2b obtains the appearance frequency Ti. The control unit 2b refers to an index word table (not shown, see index word candidate table Tb14A) having an index word field and a position information L field, and how many pieces of position information L appear in a certain index word. Is counted for each significant field (that is, for each significant field).

  In C504, the control unit 2b obtains the importance of each index word by the above equation 2. When the control unit 2b obtains the importance according to the above formula 2, in C505, an index word field having an index word as an element, and an importance field having the importance of the index word obtained by the above formula 2 as an element, The importance level table Tb9B including the fields of the position information L of all the columns in which the index word appears is created for each significant field.

Accordingly, the control unit 2b executes the processes of C01 to C05 of the knowledge collection program C00, thereby causing the event vector space table Tb9A1, the event importance table Tb9B1, the cause vector space table Tb9A2, and the cause illustrated in FIGS. 9A and 9B. The importance level table Tb9B2, the treatment vector space table Tb9A3, the treatment importance level table Tb9B3, the countermeasure vector space table Tb9A4, the countermeasure importance degree table Tb9B4, the occurrence point vector space table Tb9A5, and the occurrence point importance degree table Tb9B5 are created.
When the above knowledge collection process is completed, the control unit 2b returns to A03 process selection (see FIG. 12B) via A07.

  As described above, the knowledge collection process has been described. However, since the control unit 2b performs the knowledge collection process as described above, the automatic exercise system 1 does not depend on the content and acquires knowledge necessary for automatically creating a problem. Can be collected.

[Question making program D00]
The question creation program D00 includes question set completion determination processing (D02 to D03), ranking list creation processing (D06, D09 to D11c), QA pair creation processing (D06, D07, D12 to D15), incorrect answer creation processing (D07, D08, D16 to D20) and a procedure for causing the control unit 2b to execute the problem integration process (D04). These processes are collectively called problem creation processes (D01 to D20). When performing each process, the control unit 2b functions as a problem set completion determination unit, a ranking list creation unit, a QA pair creation unit, an incorrect answer creation unit, and a problem integration unit. Cases that function to perform each of these processes are collectively referred to as problem creation means. Each process will be described below.

[Question creation program D00-Question set completion judgment processing: D02 to D03]
When detecting that the question creation button B12B2 is pressed in A03 (see FIG. 12B) in FIG. 3, the control unit 2b executes A05, that is, proceeds to D01 in FIG. The control unit 2b functions as a problem creating unit by calling and executing the problem creating program D00 in D01, and proceeds to D02.

In D02, the control unit 2b discards the ranking list RL illustrated in FIG. 16 described later when it is acquired, and discards all the problem sets (for example, 1000 sets of 10 questions per set (for example, a total of 10000 Q), 500 sets of 30 questions per set (15,000 questions in total), etc.The number of questions and the number of sets in one set can be set on the management screen (description omitted) If not (D02: no), proceed to D03. In D03, the control unit 2b determines whether the number of question creation trials (total of all question sets) exceeds δ. Here, δ is a preset upper limit of the number of question creation trials. The reason why the control unit 2b performs D03 is that an unexpected situation (such as the number of problems set by the user and the number of sets set to illegal) occurs, and the problem creation process after D06 is continued indefinitely. This is to prevent it.
If the controller 2b determines in D03 that the number of question creation trials does not exceed δ (D03: no), the control unit 2b proceeds to D06 to create the next question. That is, the control unit 2b proceeds to D06 until all the problem sets (a predetermined number of problems consisting of a predetermined number of problems) are generated unless the number of problem creation trials exceeds δ.

…(式３)、
により求め、当該類似度順に質問フィールドの各カラムをランキングしたランキングリストRLを作成する処理を行う。図１６にランキングリストRLを例示する。尚、「類似度」とは、対比する自然言語文データが類似する程度をそのベクトル空間を用いて演算することにより数値によって表したものであり、例えば、上記式３により求めることができる。 [Problem creation program D00-Ranking list creation processing: D06, D09 to D11c]
In the ranking list creation process (D06, D09 to D11c), the control unit 2b functions as a ranking list creation unit, and extracts an index word included in the query Qy from the importance level table Tb9B as an important word as an element of the query vector space. Query Q using the query vector space and a vector space included in a vector space table of a significant field (hereinafter, also simply referred to as “question field”) selected as a question field for creating a question sentence from a plurality of significant fields. And the similarity between all columns in the question field,

... (Formula 3),
And processing for creating a ranking list RL ranking each column of the question field in the order of similarity. FIG. 16 illustrates a ranking list RL. The “similarity” is a numerical value obtained by calculating the degree of similarity of the contrasted natural language sentence data using the vector space, and can be obtained by, for example, Equation 3 above.

First, the control unit 2b determines that the ranking list RL is not acquired in D06 (D06: no) when creating a new question (that is, when starting the nth question creation), and proceeds to D09. . In D09, the control unit 2b displays the query designation screen G15 in FIG.
Here, “query Qy” refers to natural language sentence data input to the input box BX151 of the query specification screen G15 by the user when a question is automatically created. Specifically, the natural language sentence data is This is a question sentence close to the problem that the user wants to create or a word included in the question sentence. Query Qy can also be automatically generated.In this case, query Qy is randomly generated from a natural language sentence database (for example, daily report DB5, but in a broad sense includes knowledge source table Tb13A). Natural language sentence data to be extracted. Query Qy is used to create a problem. Incidentally, the problem consists of a question sentence Q, one correct answer A option, and a plurality (four) of incorrect answer B (B1 to B4) options.

  In addition, “Check Box CB151” specifies which of the fields constituting one record of daily report DB5, that is, the event field, the cause field, the action field, the countermeasure field, and the occurrence location field is the question field. It is a box to do. In other words, the “question field” means a field whose natural language sentence data is the question sentence Q of the problem, and a significant field selected as a question sentence creation field from a plurality of significant fields in the natural language sentence database. That is, the concept includes the same significant field of the knowledge source table Tb13A. In the present embodiment, any of the event field, cause field, action field, countermeasure field, and occurrence location field can be used as the question field. In the present embodiment, when pressing of the automatic query generation button B152 is detected, the event field is set as a question field, but may be selected at random.

  “Check box CB152” is a box that specifies which field is one of the fields that make up one record of daily report DB5, that is, the event field, cause field, action field, countermeasure field, and occurrence field. It is. In other words, the `` answer field '' means a field in which the natural language sentence data becomes a question answer option (including correct answer A and incorrect answer B), and is used for creating an answer from a plurality of significant fields in the natural language sentence database. A significant field selected as a field is a concept including the same significant field in the knowledge source table Tb13A. In the present embodiment, any of the event field, cause field, action field, countermeasure field, and occurrence location field can be used as the answer field. In the present embodiment, when pressing of the automatic query generation button B152 is detected, the cause field is set as an answer field, but may be selected at random.

In D09, the controller 2b waits for user input, and as shown in the query specification screen G15, the natural language sentence data is input as the query Qy, and the question field and the answer field are specified (default state). When the check box CB151 is checked due to an event and the check box CB152 is a cause), when the pressing of the designation button B151 is detected, the process proceeds to D11a.
On the other hand, when the control unit 2b detects that the automatic query generation button B152 is pressed in D09, the process proceeds to D10. In D10, the control unit 2b sets the event field as the question field and the answer field as the cause field among the significant fields (event / cause / action / measure / occurrence location), and randomly selects the columns from the question field. The selected natural language sentence data is selected as the query Qy, and the process proceeds to D11a.

The control unit 2b is in D11a.
(1) A query vector space is generated by extracting an index word included in the query Qy as an important word from the importance table TB9b illustrated in FIG. 9B, and the process proceeds to D11b.
(2) Query vector space and question vector space of all columns illustrated in FIG. 9A (any of event question vector space, cause question vector space, action question vector space, countermeasure question vector space, occurrence point question vector space) Is used to determine the similarity between the query Qy and all the columns in the question field using Equation 3 above, and proceed to D11c.
(3) A process of creating a ranking list RL (see FIG. 16) in which the contents of all the columns of the question field are sorted in the order of similarity is performed. Then, the control unit 2b returns to D06. When the control unit 2b detects that the designation button B151 is pressed without the query Qy input on the query designation screen G15, it displays an error message (not shown) and displays the query designation screen G15 again. Then, it waits for user input (D09).

[Question creation program D00-QA pair creation processing: D06, D07, D12 to D15]
When creating the ranking list RL as described above, the control unit 2b determines in D06 that the ranking list RL has been acquired (D06: yes), and proceeds to D07. When D07 is executed for the first time immediately after the control unit 2b acquires the ranking list RL, the question sentence Q and the correct answer A are not yet created. It is determined that a pair of Q and correct answer A (the same applies hereinafter) has not been created (D07: no), and the process proceeds to D12.

  In D12 to D15, the control unit 2b functions as a QA pair creation unit. As a QA pair creation process, the question creation template TP illustrated in FIG. 18 shows a question field of similarity Rα equal to or higher than the threshold value α in the ranking list RL. Is applied to the question sentence Q, and the answer field corresponding to the similarity Rα is a significant field selected as a correct answer answer field from a plurality of significant fields (that is, the answer field described above). Process to make the correct answer option. In other words, in D12 to D15, the control unit 2b creates a question sentence Q based on the contents of the question field corresponding to the query Qy, and creates a correct answer A based on the contents of the answer field corresponding to the query Qy. To do. Here, as illustrated in FIG. 18, the “question creation template TP” is an embedded field in which the question field of the natural language sentence database and / or the contents of the field such as the event, cause, treatment, countermeasure, and occurrence location is embedded Refers to a template provided with

Hereinafter, these processes will be described. In D12, the control unit refers to the ranking list RL, selects the highest similarity highest data Q _max having the highest similarity, proceeds to D13, and in D13, the similarity of the highest similarity data Q _max Is greater than or equal to the threshold value α. Here, the “threshold value α” is used to select a QA pair that is sufficiently close to the query Qy. In other words, the “threshold α” is used to prevent the creation of a QA pair that is far from the user's intention. As the threshold α, an arbitrary numerical value greater than 0 and equal to or less than 1 is set, preferably 0.9 or more, and more preferably 0.95 or more. If the control unit 2b determines that “the similarity of the highest data Q _max is equal to or greater than the threshold α” (D13: yes), the process proceeds to D14. On the other hand, when determining that “the similarity is less than α” (D13: no), the control unit 2b discards the selected highest similarity data Q _max and returns to D02. Note that the data selected in D12 is not limited to the highest similarity data, and can be arbitrarily set such as in a specified order or in a specified range.

In D14, the control unit 2b creates the question sentence Q by applying the question field data of the highest similarity data Q _max (and other significant field data as necessary) to the question creation template TP. That is, in D14, the control unit 2b determines which significant field is selected as the question field and the answer field, and determines which one to use in the question creation template TP. For example, the question creation template TPa (or TPb) is applied to “when an event field is specified as a question field and a cause field is specified as an answer field” as illustrated in the query specification screen G15.
FIG. 19 shows an application example of the problem creation template TPa. This is because the problem is caused by reading “Stop at empty DP chain break” from the question field (event question field) of record number 100001 of the highest similarity data Q _max and “Sub 1F device 5A” from the occurrence field. An example of embedding in [Embedded event field] and [Occurred location embedded field] of the created template TPa is shown. In this way, the control unit 2b creates the question sentence Q and stores it in the storage unit 2a (D14).

Control unit 2b in the D15, as the correct creation processing, response field that corresponds to the similarity maximum data Q _max, i.e., "daily DB5 corresponding to the record number of the similarity maximum data Q _max (or knowledge source table Tb13A) Natural language sentence data of “Answer field of record number” is extracted as correct answer A, and stored in the storage unit 2a as one of answer options. For example, as illustrated in FIG. 19, the control unit 2b, as illustrated in FIG. 19, responds to the record field 100001 of the daily report DB5 (or knowledge source table Tb13A) corresponding to the record number 100001 of the highest similarity data Q _max (ie, the check box CB152). The “side cam follower breakage” is extracted as the correct answer A from the cause field checked in step S1 and is stored in the storage unit 2a as one of the answer options. The control unit 2b returns to D07.

[Problem creation program D00-Incorrect answer creation processing: D07, D08, D16 to D20]
In the incorrect answer creation process (D07, D08, D16 to D20), the control unit 2b creates a question answer option based on the similarity between the related data. For the incorrect answer B in the answer options, the control unit 2b has four elements (or an arbitrary number of remaining (n-1)) in ascending order from the lowest similarity in the ranking list RL of similar relation data sets. Branch element) is selected, and natural language sentence data of the corresponding answer field is extracted. The control unit 2b enumerates the created question sentence Q and answer options (enumeration of answer options is performed in random order using a random number generation algorithm), and creates a 5-choice (n-choice) type problem. The control unit 2b repeats the above process to create a problem set. Hereinafter, these processes will be described.

  The control unit 2b creates a correct answer A (D07, D12 to D15), and when returning to D07, determines that “QA pair has been selected” (D07: yes), so proceeds to D08 and answers the remaining incorrect answers. D16 to D20 are repeated until the options are completed. For example, in the case of a 5-choice (n-choice) problem, the correct answer has already been created in D12 to D15, so the control unit 2b performs the processes of D17 to D20 so that there are 4 branches (n branches) of incorrect answer options. Repeat at least 4 times (n-1 times).

  In D16, the control unit 2b refers to the ranking list RL, the “question field similarity” is the lowest or the record of the “question field similarity” higher in the ranking than the previously selected data (that is, Then, the corresponding daily report DB5 or knowledge source table Tb13A record) is selected (see FIG. 20A). Therefore, when creating the first incorrect answer choice for the n-th question, the control unit 2b records the lowest “question field similarity” in the ranking list RL (that is, the daily report DB5 or knowledge source table Tb13A corresponding to this). Record), when creating incorrect answer choices for the second and subsequent nth questions, the record with the highest `` question field similarity '' (that is, the corresponding daily report DB5 or Knowledge record table Tb13A). Here, the “highest ranking” may be “any fixed or variable interval”, and there is no particular limitation on which ranking is selected.

  For example, according to the example of FIG. 20A, the control unit 2b selects the 0.053 record having the lowest similarity when creating the first incorrect answer option for the nth question, and sets the second question for the nth question. When creating correct answer choices, select the 0.059 record with the highest similarity. When the control unit 2b selects a record including an answer field that can be an incorrect answer option in D16, the process proceeds to D17.

  In D17, the control unit 2b determines whether the similarity of the question field of the selected record is less than the threshold value β. This is to cause the control unit 2b to select an incorrect answer option that is sufficiently far from the correct answer A of the problem. The reason why the similarity of the question field is compared is that the lower the similarity of the question field, the more the content of the corresponding answer field can be determined to be farther. The “threshold value β” is preferably 0.6 or less, and is usually set to 0.5. When the control section 2b finishes D17, it proceeds to D18.

  In D18, the control unit 2b stores the natural language sentence data in the answer field of the selected record in the storage unit 2a as an incorrect answer option. After finishing D18, the controller 2b proceeds to D19.

In D19, the control unit 2b
(1) At the time of creating the first incorrect answer choice for the nth question, the similarity between the answer field of the selected incorrect answer option and the answer field of the correct answer option is calculated, the vector space of the answer field of the incorrect answer option, and the correct answer option Using the vector space of the answer field and the above equation 3,
(2) When creating incorrect answer choices for the second nth question and later, in addition to the similarity between the answer field of the selected incorrect answer option and the answer field of the correct answer option, the answer of the selected incorrect answer option Similarities between the field and all other incorrect answer options, the vector field of the answer field of the selected incorrect answer option, the vector space of the answer field of the correct answer option, and the answer field of the other incorrect answer option It calculates | requires by said Formula 3 using a vector space.

  Then, the control unit 2b determines whether or not the obtained similarity is less than the threshold value γ. The “threshold γ” is normally set to 0.90, but can be arbitrarily set in the range of more than 0 and less than 1. The reason why the control unit 2b determines whether or not the similarity is less than the threshold value γ is to select a question that is far enough from the correct answer A and other incorrect answers B as incorrect answer options. The similarity of the answer fields is compared in order to exclude options that are directly similar as answers. According to the example of FIG. 20B, the control unit 2b selects an answer field corresponding to the question field similarity 0.076 as an incorrect answer option and determines whether or not to register. In this case, the similarity between the natural language sentence data “Chain Stretch” in the answer field and the natural language sentence data “Chain Nobori” in the answer field of other answer options is 0.954, which is larger than the threshold γ. The 0.076 record is excluded from the incorrect answer choices (D19: no). This is determined by the relationship between the correct answer option and all other incorrect answer options. In this way, similarities with correct answer options and similarities between incorrect answer options can be avoided. If the control unit 2b determines in D19 that the similarity is less than the threshold γ (D19: yes), the process proceeds to D20.

  In D20, the control unit 2b registers the natural language sentence data in the answer field of the selected record as an incorrect answer option. Then, the control unit 2b returns to D08, and thereafter performs the processes from D16 to D20 until all the remaining incorrect answer options are completed. That is, when the control unit 2b determines that the incorrect answer option is completed (D08: yes), one question is completed at this point. FIG. 21 illustrates one completed question. Then, the control unit 2b returns to the process of D02, executes the process described above until the problem set is completed, and determines that the problem set is completed in D02 (D02: yes), proceeds to D04.

[Problem creation program D00-Problem integration processing: D04]
In D04, for each question, the control unit 2b answers the question sentence Q, the correct answer A and the incorrect answer B1 to B4 answer options (if all answer options are n, the incorrect answer options are (n-1 ))) With a new question number, question number of question DB 10, question sentence Q, (difficulty), (total number of questions), (total number of correct answers), correct answer A, incorrect answers B1-B4 To each field, and the process ends.
Note that the score in the figure is set based on the similarity between the correct answer A and the incorrect answers B1 to B4.

[Problem creation program D00-Modification]
In D16, as illustrated in FIG. 22, the control unit 2b refers to the ranking list RL, and the daily report DB5 corresponding to the value specified by the “similarity of the question field” or the similarity satisfying the condition is satisfied. Alternatively, the record of the knowledge source table Tb13A) may be selected. In this case, the control unit 2b, when creating the incorrect answer choices for the nth question first and the second question and later, is a data that is a similarity that satisfies the value specified by the “similarity of the question field” in the ranking list RL or the condition. The record of (daily report DB5 or knowledge source table Tb13A corresponding to) is selected.

  The problem creation processing has been described above. However, as described above, the question field having the similarity Rα equal to or higher than the threshold value α is applied to the problem creation template, so that a problem according to the user's request is created. In addition, since the answer field corresponding to the similarity Rβ less than the threshold value β is an incorrect answer option, an incorrect answer far enough from the correct answer can be obtained. Furthermore, since the answer field whose similarity γ with the correct answer option and / or other incorrect answer options is less than the threshold γ is set as an incorrect answer option, it is avoided that the answer option has the same contents as the correct answer option and other incorrect answer options. Is done.

[Practice program E00]
The exercise program E00 includes a procedure for causing the control unit 2b to execute a preparation process (E01 to E04), a question scoring process (E05 to E14), and a management process (E15 to E20). These processes are collectively called exercise processes (E1 to E20). The control unit 2b functions as a preparation unit, a question grading unit, and a management unit when performing each process. These means are collectively called exercise means. The exercise means reads the user's proficiency level and problem difficulty level from the user information DB 11, selects the optimal problem set sequentially from the problem DB 10 based on this, and presents the problems to the user randomly through the interface screen 6d (Fig. 23), the user's answer is scored, calculation of the score result and calculation are performed based on the score result, and the skill level information and problem difficulty level information of the user in the user information DB 11 are updated. Hereinafter, these processes will be described.

[Preparation processing: E01 to E04]
The preparation processing (E01 to E04) is processing for acquiring an appropriate problem for each user from the problem set registered in the problem DB 10 after performing initialization processing of the temporary storage area of the RAM 2a. When the control unit 2b detects that the exercise button B12B3 is pressed in A03 in FIG. 3, A05 is executed, that is, the process proceeds to E01 in FIG. The control unit 2b functions as exercise means by calling and executing the exercise program E00 in E01, and proceeds to E02.
In E02, the control unit 2b initializes the temporary storage area of the RAM 2a necessary for the exercise process, and then proceeds to E03. In E03, the control unit 2b refers to the user information DB 11, acquires the proficiency level of the corresponding user, and proceeds to E04. In E04, the control unit 2b refers to the problem DB 10, selects one problem set that has the highest effect of improving the proficiency level of the corresponding user based on the acquired proficiency level, and sets the correct answer A / The incorrect answer B / score and the like are temporarily stored in the RAM 2a as question data, the timer for counting the answer time limit for all questions in the question set is activated, and the process proceeds to E05. Here, “Proficiency” means past correct / incorrect answer records, questions / already-answer records, and the question set selection includes X questions (X%) of questions that the user has answered correctly in the past. It is possible to make a question based on the rule of including Y questions (Y%) for questions that were incorrect in the past and Z questions (Z%) for questions that have not been answered (any number of X, Y, and Z).

[Question scoring: E05-E14]
The question grading process (E05 to E14) is a process of displaying a problem on the interface screen 6d, scoring a user answer to the problem, displaying the result, and temporarily storing it in the RAM 2a. The processes of E05 to E14 are repeated until the questions of the acquired problem set are completed.
In E05, the control unit 2b refers to the question set to determine whether or not there is an “unquestioned question in this exercise in the question set”. If there is an unquestioned question (E05: yes), the process proceeds to E06. . In E06, the control unit 2b selects one of the unquestioned questions (hereinafter referred to as “questioned question E06”), and then proceeds to E07. As illustrated in the problem display screen G23 of FIG. The process of displaying the question E06 on the screen 6d is performed. Then, the control unit 2b proceeds to the process of E08.

  In E08, when there is a user answer input within the time limit (E08: yes), the control unit 2b proceeds to E09. In E09, the control unit 2b accepts a response input from the user and proceeds to E10. In E10, the control unit 2b determines whether or not the response input from the user is a normal input. When the controller 2b detects that the answer button B231 is pressed while the answer option CB231 on the question display screen G23 is marked with a mouse or the like in E10, the controller 2b determines that the input is normal (E10: yes) Proceed to E11. In E08, when the control unit 2b determines that the time limit has been reached (E08: no), the process also proceeds to E11.

In E11, as shown in the result display screen G24 in FIG. 24, the control unit 2b displays the user answer UA for the question E06 and refers to the problem set temporarily stored in the RAM 2a to determine which user answer UA Judges whether the choice (correct answer A / incorrect answer B) matches. When the user answer UA matches the correct answer A, the control unit 2b, in E12, together with the user answer UA as an example of the result display screen G24, as a scoring result SR,
(1) Correct / incorrect answer,
(2) Correct answer rate (number of correct questions / cumulative number of questions in question set), and
(3) The acquired score, accumulated score, etc. are displayed and temporarily stored in the RAM 2a.

On the other hand, when the user answer UA matches the incorrect answer B of the question set, the control unit 2b, as illustrated in the result display screen G25 in FIG.
(1) Correct / incorrect answer,
(2) Correct answer rate (number of correct answers / cumulative number of questions in the problem set), and
(3) Acquisition score / cumulative score,
(4) Correct answer information (correct answer information to which pointer information for calling the knowledge collection program C00 is assigned) and the like are displayed and temporarily stored in the RAM 2a.

  After displaying the result in E12, the control unit 2b proceeds to E13, and when “the user answer UA matches the correct answer A in E11”, the control unit 2b waits for the next button B243 to be pressed, If it matches the incorrect answer in E11 ”, it waits for the next button B253 or the correct information button B254 with pointer information to be pressed. When the control unit 2b detects pressing of the correct answer information button B254 (E13: yes), the control unit 2b proceeds to E14.

  In E14, the control unit 2b passes the natural language sentence data of the question field (or question embedding field) of the question question E06 to D09 as the query Qy, and after performing the processing of D09 to D11c, the degree of similarity is high The natural language sentence data of the question field of the daily report DB 5 (or knowledge source table) corresponding in order from the data is displayed on the interface screen 6d as exemplified in the review screen G26 of FIG.

  For example, in the case of the review screen G26, the control unit 2b passes the question field (event field) “stopped on bogie overrun” as a query Qy to D09. Then, the control unit 2b searches and extracts the question field importance level table Tb9B (event importance level table Tb9B1) using the index word included in the `` stop on bogie overrun '' in D11a as an important word, a query vector Create a space (eg, (cart, overrun, stop), etc.). In D11b, the control unit 2b uses the query vector space and the question vector space to obtain the similarity between the question sentence Q of the question question E06 and the data of all the columns of the question field by the above equation 3. Then, the control unit 2b creates a ranking list RL from those having a high degree of similarity in D11c. After performing these processes, the control unit 2b displays the natural language sentence data of the question field of the daily report DB 5 (or knowledge source table Tb13A) corresponding to the data in descending order of similarity on the interface screen 6d. Furthermore, the control unit 2b displays the contents of other fields according to the screen size in descending order (or descending order from the question field specified by the mouse etc.) from the highest ranking among those displayed on the screen. (See FIG. 26). Note that when the control unit 2b detects that the return button B261 is pressed, the control unit 2b returns to E13.

  When the control unit 2b detects that the next button B243 (or B253) is pressed in E13, the control unit 2b returns to E05. The control unit 2b repeats the processes of E05 to E14 until the predetermined number of questions (10 questions or the like) in the problem set is completed in E05. When the control unit 2b determines that a predetermined number of problems have been completed in the problem set, that is, there is no unquestioned problem (E05: no), the process proceeds to E15.

[Management processing: E15 to E20]
In E15, the control unit 2b refers to the problem set temporarily stored in the RAM 2a and the user answer UA as the aggregation process, and refers to or calculates the number of correct answers, the number of trials, the acquired score, the cumulative score, and the current level, E16 Proceed to

  In E16, the control unit 2b performs a process of writing to the user information DB 11 using the data referenced or calculated in E15 and updating it as a proficiency level update process. That is, in E16, the control unit 2b updates the number of trials and the number of correct answers to the result obtained by adding the current result, and “adds the current score to the cumulative score of the cumulative score 1 shown in FIG. 11”. The new cumulative score is set to 1, and the previous cumulative score number is incremented by 1 and updated. If the control unit 2b determines that the cumulative score has reached the prescribed upper limit, the control unit 2b deletes the tail data of the cumulative score and then increments the cumulative score number of the existing data by one to increase the cumulative score. to add. Thereafter, the control unit 2b proceeds to E17.

  In E17, as the problem difficulty level update process, the control unit 2b sets the difficulty level for each problem based on the aggregation result for the problem set regardless of the user or the exercise date and time, “difficulty level = total number of correct answers of the problem / The total number of questions is calculated, and the result is updated by writing the result in each field of the difficulty level of each question, the total number of questions, and the total number of correct answers in the problem DB 10. Then, the control unit 2b proceeds to E18.

  In E18, the control unit 2b displays the user name, current level, cumulative score, accuracy rate (number of correct answers / number of current questions), acquired score, etc. As illustrated in the exercise end screen G27, it is displayed on the interface screen 6d.

  In E19, the control unit 2b waits for user input, and when it detects that the re-challenge button B271 is pressed, returns to E02 to perform exercise processing, and when it detects that the review button B272 is pressed, all of the problem sets are left unanswered Return to E05 to perform exercise processing based on the same problem, and when pressing back button B273 is detected, call automatic exercise program A00, jump to A03, display user processing selection screen G12B, and wait for user input . Further, when the control unit 2b determines that the input is not normal in E10 in the above process (E10: no), the error message is displayed as illustrated in the answer error warning screen G28 in FIG. The process proceeds to (E21) and E22, and as an error process, the automatic exercise program A00 is called, the process jumps to A03, the user process selection screen G12B is displayed, and a user input wait state is entered.

[Practice program E00-Modification]
The calculation and update processing of problem difficulty and user comprehension can be implemented by the following mechanism.
Calculate the similarity sim (Q_i, D_i, j) between the question field data Q_i and the choice data D_i, j (where i = 1,2,3, ... n and j = 1,2 , 3, 4).
At this time, the user's answer score is expressed as the sum of sim (Q_i, D_i, j).
score (U_k) = Σi = 1 ... n {sim (Q_i, D_i, j)}
The user's learning progress is expressed by an answer score.
prog (U_k) = score (U_k)
The degree of understanding of the user can be expressed by a deviation value based on the distribution of average values of answer scores.
U-level (U_k) = zscore (U_k) * 10 + 50
zscore (U_k) = (score (U_k)-mean (score (U_k))} / σ (score (U_k))
mean (score (U_k)) = Σk = 1 ... m {score (U_k)} / m
σ (score (U_k)) = sqrt {(score (U_k)-mean (score (U_k))) ^ 2 / m}
Therefore, the user information DB 11 may store a deviation value field as the user understanding level of the user information DB 11 in FIG.

The problem difficulty level can be expressed by a deviation value based on a distribution of user comprehension levels that answered the problem.
Q-level (Q_i) = zscore (Q_i, U-level (U_k)) * 10 + 50
Ul_k = U-level (U_k)
zscore (Q_i) = (score (Q_i, Ul_k)-mean (score (Q_i, Ul_k))} /
σ (score (Q_i, Ul_k))
mean (score (Q_i, Ul_k)) = Σi = 1 ... {score (Q_i, Ul_k)} / n
σ (score (Q_i, Ul_k)) = sqrt {(score (Q_i, Ul_k)-
mean (score (Q_i, Ul_k))) ^ 2 / n}
Therefore, a deviation value based on the user comprehension distribution may be stored as the difficulty level of the problem DB 10 in FIG.

The problem difficulty level is reflected as a weight on the answer score at the next question.
Specifically, 1-Q-level (Q_i) / 100 is calculated and added to the answer score.
The initial state is Q-level (Q_i) = 0.
Therefore, a score weighted according to the problem difficulty level may be stored as the score of each answer option in the problem DB 10 of FIG.

  Using the above calculation, the user comprehension level and problem difficulty level can be recalculated and updated at regular intervals (the update period is determined by the setting of the problem creator). Thereby, a user comprehension degree and a problem difficulty level can always be hold | maintained at the newest value.

  As described above, the exercise process has been described above. However, as described above, the automatic exercise system 1 can perform automatic scoring processing of user answers and automatic aggregation processing of exercise results in addition to automatic problem creation. By using the automatic exercise system 1, knowledge and know-how of daily work in factories and offices can be passed on without a dedicated dictionary.

Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment. As the calculation method of importance and similarity, a calculation method other than that used in the above embodiment can be adopted.
Furthermore, although the case where daily report information in a factory is used as a knowledge source has been exemplified in the above embodiment, the present invention can be applied to any factory or office provided with a database composed of natural language sentence data.
Furthermore, although the query Qy is first used to create a question sentence, the query Qy can be used first to create an answer and then create a question sentence.

  Since the automatic exercise system according to the present invention can support active learning for the purpose of passing on knowledge and know-how related to daily operations such as factories and offices, even in the absence of skilled workers due to retirement, etc. Skilled know-how / skilled knowledge transfer is possible, which can contribute to improving the quality of work contents such as factories and offices, which is extremely beneficial to the industry.

1… Automatic exercise system
2 ... Server (2a ... Storage unit, 2b ... Control unit, 2c ... Input unit, 2d ... Output unit)
6… Client (6d… Interface screen)
5 ... Daily report DB
Tb13A ... Knowledge source table
9 ... Knowledge DB
Tb9A ... Vector space table
Tb9B ... Importance table
10 ... Problem DB
11 ... User information DB
A00 ... Automatic exercise program
B00 ... User management program
C00 ... Knowledge collection program
D00 ... Question making program
E00: Exercise program

Claims (20)

From a natural language sentence database comprising a plurality of significant fields having a causal relationship, a vector space table comprising a column position information field and a vector space field for each significant field, and an index word field for each significant field And a knowledge collection means for constructing a knowledge database comprising an importance table including an importance field and an index word position information field,
An index word included in a query is extracted as an important word from the importance table as an element of a query vector space, and the query vector space and a vector space of all columns of the question field selected from the significant fields are used. A ranking list creating means for obtaining a similarity between the query and all the columns of the question field, and creating a ranking list ranking all the columns of the question field based on the similarity;
In the ranking list, a question sentence with a similarity Rα equal to or higher than a threshold value α is applied to a question creation template to create a question sentence, and an answer field selected from the significant field corresponding to the similarity Rα is a correct answer option. An automatic exercise system characterized by comprising QA pair creation means. Furthermore,
In the ranking list, the answer field and the correct answer option using the vector space of the answer field corresponding to the similarity Rβ less than the threshold value β, the vector space of the correct answer option, and / or the vector space of other incorrect answer options. And / or when the similarity Rγ with the other incorrect answer option is obtained and the similarity Rγ is less than the threshold value γ, an incorrect answer creating means using the answer field corresponding to the similarity Rβ as the incorrect answer option is provided. The automatic exercise system according to claim 1, further comprising:   The automatic exercise system according to claim 1, wherein the similarity Rα is the highest similarity in the ranking list.   4. The automatic exercise system according to claim 2, wherein the similarity Rβ is the lowest similarity (exceeding 0) among those not selected as incorrect answers in the ranking list. 5. Furthermore,
A preparation means for obtaining a question to be asked from the problem database based on the user information database;
A question scoring means for displaying the question on the interface screen, scoring a user answer to the question, and displaying the result on the interface screen together with correct answer information;
The automatic exercise system according to claim 1, further comprising a management unit that updates the user information database and the problem database based on the result. Furthermore,
An index word included in the correct answer information is extracted from the importance level table as an important word as an element of the correct answer information vector space, and the correct answer is obtained using the correct answer information vector space and the vector space of all the columns of the question field. It is provided with a reviewing means for obtaining similarity between information and all columns of the question field, creating a ranking list ranking all the columns of the question field based on the similarity, and displaying the contents. The automatic exercise system according to claim 5, wherein the system is an automatic exercise system. The knowledge collecting means includes
For each significant field, a suffix group and / or a prefix group is generated by sorting in lexicographic order as a character string class, cutting out a matching portion of the preceding and following character string classes, excluding a character string class whose appearance frequency is less than a threshold T1, String class group generation means for sorting so that the upper string class in an inclusive relation is prioritized over the lower string class, and integration of the same string classes
For each significant field, when there is no inclusion relationship with the previous character string class of the character string class, the character string class is stored as an index word candidate. If the frequency ratio is greater than or equal to the threshold T2, the lower character string class is stored as an index word candidate, while if the frequency ratio is less than the threshold T2, the lower character string class is regarded as the upper character string class and the index word Index word candidate group generation means for storing as candidates,
Index word extraction means for extracting, as index words, index word candidates that overlap among the index word candidates of each group,
For each significant field, the index words included in all the columns are extracted as important words, a vector space having the important words as elements is generated for all the columns, and the natural language sentences in all the columns are generated. A vector space table creating means for creating a vector space table including a position information field capable of specifying a record number of the database and a vector space field;
For each significant field, the importance level of the index word is calculated, and for each significant field, the index word field, the importance level field, and the record numbers of the natural language sentence database of all the columns in which the index word appears are recorded. An automatic exercise system according to any one of claims 1 to 6, further comprising importance level table creation means for creating an importance level table including a position information field that can be specified. The frequency ratio is
Gap (C (n-1), Cn) = | log (Ti (n-1) / Tin) | (Formula 1)
(Where C (n-1): upper string class inclusive relation with Cn, Cn: lower string class inclusive relation with C (n-1), Gap (C (n-1), Cn ): C (n-1), frequency ratio of Cn, Ti (n-1): appearance frequency of upper character string class, Tin: appearance frequency of lower character string class) Automatic exercise system as described in. The importance is
Swj = log {(Fj / Dj) +1} (Formula 2)
9. The automatic exercise system according to claim 7 or 8, wherein the automatic exercise system is obtained by (where Swj: importance, Fj: appearance frequency, Dj: co-occurrence frequency). The similarity is

The automatic exercise system according to any one of claims 1 to 9, wherein the automatic exercise system is obtained by (Equation 3). Computer to create problems automatically,
From a natural language sentence database comprising a plurality of significant fields having a causal relationship, a vector space table comprising a column position information field and a vector space field for each significant field, and an index word field for each significant field And a knowledge collecting means for constructing a knowledge database comprising an importance level table including an importance level field and an index word position information field,
An index word included in a query is extracted as an important word from the importance table as an element of a query vector space, and the query vector space and a vector space of all columns of the question field selected from the significant fields are used. A ranking list creating means for obtaining similarity between the query and all the columns of the question field, and creating a ranking list ranking all the columns of the question field based on the similarity;
In the ranking list, a question sentence with a similarity Rα equal to or higher than a threshold value α is applied to a question creation template to create a question sentence, and an answer field selected from the significant field corresponding to the similarity Rα is a correct answer option. Automatic exercise program to function as a QA pair creation means. Said computer further
In the ranking list, the answer field and the correct answer option using the vector space of the answer field corresponding to the similarity Rβ less than the threshold value β, the vector space of the correct answer option, and / or the vector space of other incorrect answer options. And / or as an incorrect answer creation means for obtaining a similarity Rγ with the other incorrect answer options and using the answer field corresponding to the similarity Rβ as an incorrect answer option when the similarity Rγ is less than the threshold γ. The automatic exercise program according to claim 11 for making it function.   The automatic exercise program according to claim 11 or 12, wherein the similarity Rα is the highest similarity in the ranking list.   The automatic exercise program according to claim 12 or 13, wherein the similarity Rβ is the lowest similarity (but more than 0) among unselected incorrect choices in the ranking list. The computer for problem exercises,
A preparation means for acquiring questions to be asked from the problem database based on the user information database,
A question scoring means for displaying the question on the interface screen, scoring a user answer to the question, and displaying the result together with correct information on the interface screen;
The automatic exercise program according to any one of claims 11 to 14, which functions as a management unit that updates the user information database and the problem database based on the result. Said computer further
An index word included in the correct answer information is extracted from the importance level table as an important word as an element of the correct answer information vector space, and the correct answer is obtained using the correct answer information vector space and the vector space of all the columns of the question field. For obtaining similarity between information and all columns of the question field, creating a ranking list ranking all the columns of the question field based on the similarity, and functioning as a reviewing means for displaying the contents The automatic exercise program according to claim 15. In order to function as the knowledge collection means, the computer is
For each significant field, a suffix group and / or a prefix group is generated by sorting in lexicographic order as a character string class, cutting out a matching portion of the preceding and following character string classes, excluding a character string class whose appearance frequency is less than a threshold T1, String class group generation means for sorting so that the upper string class in the same string class is prioritized over the lower string class.
For each significant field, when there is no inclusion relationship with the previous character string class of the character string class, the character string class is stored as an index word candidate. If the frequency ratio is greater than or equal to the threshold T2, the lower character string class is stored as an index word candidate, while if the frequency ratio is less than the threshold T2, the lower character string class is regarded as the upper character string class and the index word Index word candidate group generation means for storing as candidates,
Index word extraction means for extracting, as index words, index word candidates that overlap among each group of index word candidates,
For each significant field, the index words included in all the columns are extracted as important words, a vector space having the important words as elements is generated for all the columns, and the natural language sentences in all the columns are generated. A vector space table creating means for creating a vector space table having a position information field capable of specifying a record number of the database and a vector space field;
For each significant field, the importance level of the index word is calculated, and for each significant field, the index word field, the importance level field, and the record numbers of the natural language sentence database of all the columns in which the index word appears are recorded. The automatic exercise program according to any one of claims 11 to 16, which functions as importance level table creation means for creating an importance level table including a position information field that can be specified. The frequency ratio is
Gap (C (n-1), Cn) = | log (Ti (n-1) / Tin) | (Formula 1)
(Where C (n-1): upper string class inclusive relation with Cn, Cn: lower string class inclusive relation with C (n-1), Gap (C (n-1), Cn ): C (n-1), Cn frequency ratio, Ti (n-1): Appearance frequency of upper character string class, Tin: Appearance frequency of lower character string class) Automatic exercise program described in. The importance is
Swj = log {(Fj / Dj) +1} (Formula 2)
The automatic exercise program according to claim 17 or 18, wherein the automatic exercise program is obtained by (where Swj: importance, Fj: appearance frequency, Dj: co-occurrence frequency). The similarity is

The automatic exercise program according to claim 11, wherein the automatic exercise program is obtained by (Equation 3).

Images