Educational material creation support device, educational material creation support method, and educational material creation support program
The teaching material creation support device addresses temporal leakage and consistency issues by separating records and using a forbidden word set for deterministic matching, ensuring accurate and leak-free educational materials are generated.
Patent Information
- Application Number
- JP2026161773
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-11-19
- Estimated Expiration
- 2046-07-10
AI Technical Summary
Existing systems face challenges in creating educational materials that prevent temporal leakage and ensure consistency between teacher-only and learner-only information, while maintaining numerical accuracy, by accurately referencing records up to a specified reference point in time.
A teaching material creation support device that separates records into pre- and post-reference time groups, generates materials using only pre-reference records, and employs a forbidden word set to detect and prevent temporal leaks through deterministic matching, ensuring consistency and numerical accuracy.
Efficient generation of educational materials that are free from post-reference information, maintaining temporal integrity and consistency, while mechanically preventing leaks and ensuring accurate numerical conditions.
Smart Images

Figure 0007953963000001_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a learning material creation support device, a learning material creation support method, and a learning material creation support program for creating learning materials used for education and training of medical personnel and the like. More specifically, the present invention relates to a technology for referring to actual records of a subject recorded in chronological order, and generating learning material documents using a large language model based on information up to any arbitrary point in time on the time series.
Background Art
[0002] In medical and nursing education, an educational method is known to be effective, in which a learner is presented with only information up to any arbitrary point in the course of a certain subject (patient), and made to perform exercises such as drafting a nursing care plan from the state at that point in time. For such exercises, it is necessary to create learning materials (hereinafter sometimes simply referred to as "learning material documents") such as exercise worksheets, evaluation rubrics, and exercise guides, based on actually recorded chronological information.
[0003] The present applicant has proposed a system for accumulating multiple types of chronological records related to patient conditions, examinations, nursing records, etc. through operations of a faculty terminal and a learner terminal, and allowing the records to be viewed on the learner terminal, in an educational electronic medical record system (Patent Document 1). In this system, for a certain subject, a plurality of records following the course from a starting point such as hospitalization to the present are accumulated in a state where they can be input or viewed for educational purposes.
[0004] There is a demand that a faculty member specifies an arbitrary point in time (hereinafter referred to as "reference point in time") for a subject with accumulated chronological records in this manner, and creates a learning material document based only on the information obtained up to the reference point in time. That is, in a usage mode, when records corresponding to 14 days after hospitalization are accumulated for a certain subject, the faculty member specifies "Day 3" as the reference point in time, and creates a learning material document based only on the records existing up to that point in time.
Prior Art Documents
Patent Documents
[0005] [Patent Document 1] Patent No. 7726558 [Overview of the Initiative] [Problems that the invention aims to solve]
[0006] However, as mentioned above, there were the following challenges in creating teaching materials from actual records by specifying a reference point.
[0007] Firstly, there is the problem of temporal leakage. If information about events that occurred after the reference point (for example, test results that become known later, or actions taken later) is included in the generated instructional documents, learners will be given advance information that they should not have known at that point in time, rendering the exercise ineffective. To prevent this temporal leakage, one could consider limiting the input (prompt) to the large-scale language model to only records up to the reference point, but this method has the following technical difficulties. That is, if records after the reference point are excluded in advance and not referenced by the large-scale language model at all, the large-scale language model cannot understand what events occur after the reference point, and therefore, even if the description in the generated instructional documents happens to include content that coincidentally matches actual events after the reference point, it will not be able to detect and exclude this. On the other hand, if records after the reference point are also referenced by the large-scale language model, it becomes possible to make the large-scale language model recognize that content based on those later records should be excluded, but then the risk of information after the reference point being directly reflected in the generated product (causing temporal leakage) increases. In other words, there is a technical challenge in that it is difficult to simultaneously prevent temporal leaks and generate appropriate educational materials based on an understanding of events that could lead to temporal leaks, all within a single generation process.
[0008] Secondly, there is the problem of teacher-only information being mixed in with learner-only information. Some teaching materials contain information that only teachers should refer to, such as model answers, explanations, and evaluation points (hereinafter referred to as "teacher-only information"), while others contain only information such as questions that learners should view (hereinafter referred to as "learner-visible information"). If teaching materials containing teacher-only information and teaching materials that do not contain teacher-only information are generated separately, there is a risk of inconsistencies in the content between the two. On the other hand, in the process of deriving one from the other, there is a risk that the content of teacher-only information may be mixed into the learner-visible information.
[0009] Thirdly, if the numerical consistency conditions that the teaching materials should satisfy (for example, the sum of the points allocated in the evaluation rubric matching the predetermined maximum score, or the sum of the time allocations in the exercise guide matching the predetermined total exercise time) are left to the results of a probabilistic generation process, there is a problem that teaching materials that do not satisfy these consistency conditions are often generated.
[0010] The present invention has been made in view of the above-mentioned problems, and its objective is to provide a teaching material creation support device, a teaching material creation support method, and a teaching material creation support program that can efficiently generate teaching material documents based on actual records up to a target reference point, while mechanically preventing information contained in actual records after the reference point from being mixed into the generated teaching material documents (temporal leak). [Means for solving the problem]
[0011] To solve the aforementioned problems, the present invention provides a teaching material creation support device comprising: a reception unit that receives the designation of a target record including a plurality of records in a time series and the designation of a reference point in the time series; a separation unit that separates the records included in the target record received by the reception unit into a first record group consisting of records belonging to a time before the reference point and a second record group consisting of records belonging to a time after the reference point; a forbidden word synthesis unit that synthesizes a forbidden word set, which is a set of words representing events that have not occurred or are unknown at the reference point, by referring to both the first record group and the second record group; a teaching material generation unit that generates a teaching material document by referring only to the first record group for the target record and without including the forbidden word set as input; and a matching unit that checks whether the generated teaching material document contains words belonging to the forbidden word set by deterministic partial matching and suppresses the output of the teaching material document if it contains such words. Here, the synthesis of the set of forbidden words by the forbidden word synthesis unit and the generation of the teaching material document by the teaching material generation unit may be based on a trained language model, on predetermined rules, or a combination thereof. Preferably, the matching unit, when it detects the inclusion of a word belonging to the set of forbidden words, causes the teaching material generation unit to regenerate the teaching material document, and interrupts the generation of the teaching material document if the inclusion is not resolved after a predetermined number of regenerations.
[0012] According to the above configuration, the teaching material creation support device separates the target records into a first record group (before the reference time) and a second record group (after the reference time) at a reference point, and then instructs the teaching material generation unit, which generates teaching material documents, to refer only to the first record group. On the other hand, the forbidden word synthesis unit, which synthesizes the forbidden word set, is instructed to refer to both the first and second record groups, thereby identifying words representing events that should not have occurred or been determined at the reference time as a forbidden word set, based on what events may occur after the reference time. Then, by retrospectively applying a deterministic partial match matching using this forbidden word set to the teaching material documents generated by the teaching material generation unit, the generation of the teaching material documents themselves does not access any information after the reference time, yet it is possible to mechanically detect if such information is accidentally mixed into the generated product. In other words, the present invention adopts an asymmetrical configuration in which (i) information after a reference point is completely excluded from the input of the process for generating educational materials, (ii) the excluded information is converted into the form of forbidden words through a process separate from the process for generating educational materials (composition of a set of forbidden words), and (iii) it is used only as input for a subsequent deterministic comparison of the generated educational materials. This achieves both the prevention of temporal leaks and the generation of educational materials while being aware of events that could become temporal leaks. Furthermore, since this set of forbidden words is not included in the input (prompt) to the educational material generation unit, the process for generating educational materials itself does not have the opportunity to encounter words that suggest events after a reference point, and the possibility of such words being recalled in the generated product can be structurally reduced.
[0013] In the aforementioned teaching material creation support device, it is preferable that the forbidden word synthesis unit and the teaching material generation unit each cause a trained language model to generate the forbidden word set and the teaching material document. With this configuration, a two-stage setup is realized in which the product of one trained language model (forbidden word set) is used as input for verification against the product of another trained language model (teaching material document).
[0014] In the aforementioned teaching material creation support device, it is preferable that the forbidden word synthesis unit removes from the synthesized set of forbidden words words that represent events already appearing in the first record group, and mutually contradictory words, based on deterministic rules. With this configuration, excessive output suppression (where even descriptions that are not inherently problematic as teaching materials are judged as unsuitable) due to words representing events already known before the reference time being mistakenly included in the set of forbidden words is suppressed.
[0015] In the aforementioned teaching material creation support device, it is preferable that the verification unit causes the teaching material generation unit to regenerate the teaching material document when the teaching material document contains a word belonging to the set of prohibited words. With this configuration, the possibility of obtaining a teaching material document that does not contain a time leak is increased by regenerating the teaching material document when a time leak is detected.
[0016] In the aforementioned teaching material creation support device, it is preferable that the verification unit interrupts the generation of the teaching material document if the inclusion of words belonging to the prohibited word set is not eliminated even after a predetermined number of regenerations. With this configuration, the situation in which a teaching material document containing a time leak is output to learners or teachers is prevented.
[0017] In the aforementioned teaching material creation support device, it is preferable that the receiving unit, when receiving the designation of the reference time, verifies the validity of the reference time based on the time-series range of the target record and does not accept designations outside that range. With this configuration, abnormalities in subsequent processing caused by designating a time point that does not exist in the target record as the reference time are prevented before the start of the teaching material document generation process.
[0018] Preferably, the training material creation support device further includes a numerical verification unit that mechanically verifies the numerical consistency conditions that the training material document must satisfy (including at least one of the following: the sum of the points allocated matches a predetermined maximum score, and the sum of the time allocation matches a predetermined total time), and, if a violation is detected, causes the training material generation unit to regenerate the training material document with information indicating the nature of the violation. Furthermore, it is preferable that the numerical verification unit interrupts the generation of the training material document if the violation of the numerical consistency conditions is not resolved even after a predetermined number of regenerations. With this configuration, the numerical consistency that the training material document must satisfy is mechanically guaranteed without being left to probabilistic generation results.
[0019] In the aforementioned teaching material creation support device, the teaching material generation unit generates a first teaching material document containing answer information, and the teaching material creation support device further comprises a derived output unit that deterministically derives a second teaching material document that does not contain answer information by excluding the field corresponding to the answer information from the first teaching material document. Furthermore, the matching unit preferably further checks, by partial match matching, whether the content of the field corresponding to the answer information in the first teaching material document is included in the field that is visible in the second teaching material document. With this configuration, a first teaching material document containing teacher-only information (teacher version) and a second teaching material document not containing it (learner version) are consistently derived from a single generation process, and the inclusion of teacher-only information between the two (answer leak) is checked again by the same matching unit that checks for temporal leaks. In other words, the matching unit of the present invention is characterized by applying the same unified mechanism of deterministic partial matching to both leaks with a reference time as the boundary (leaks in the time axis direction) and leaks with a boundary between teacher-only and learner-visible content (leaks in the reader attribute direction).
[0020] Preferably, in the teaching material creation support device, the teaching material generation unit generates the teaching material document as structured data conforming to a predetermined data schema, and does not cause the teaching material generation unit to generate predetermined fixed value items included in the target record, and the predetermined fixed value items are directly transcribed from the target record when the teaching material document is output. According to the above configuration, since definitive items such as information for identifying a target are reliably reflected in the teaching material document without going through the result of generation by a large language model, alteration of the relevant items or incorporation of errors is structurally excluded.
[0021] Preferably, in the teaching material creation support device, the reception unit receives designation of the target record by an identifier that does not include the main body of the target record, and the teaching material creation support device requests a record management system that manages the target record to resolve the identifier and verify usage authority at each of the time of receiving the designation and the time of executing the teaching material document generation processing, and supplies only the resolved target record to the separation unit. According to the above configuration, errors associated with the main body of the target record being directly input to the teaching material creation support device and access to the target record without authority are repeatedly prevented by independent resolution and authority verification on the record management system side that manages the target record.
[0022] In addition, the present invention can be grasped as an invention of a method in which the respective units execute the processing, and as an invention of a teaching material creation support program that causes a computer to execute the processing. [Advantageous Effects of the Invention]
[0023] According to the present invention, it is possible to efficiently generate a teaching material document based on actual records up to a reference time point of a target, while mechanically preventing information contained in actual records after the reference time point from being mixed into the generated teaching material document (temporal leakage). As a result, it becomes possible to efficiently supply a practice teaching material document with an arbitrary designated reference time point with the certainty required in education. [Brief Description of the Drawings]
[0024] [Figure 1] FIG. 1 is a conceptual diagram showing an example of the configuration of a system including a learning material creation support apparatus according to an embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram illustrating a hardware configuration of the learning material creation support apparatus according to the embodiment of the present invention. [Figure 3] FIG. 3 is a block diagram illustrating an example of a functional configuration of the learning material creation support apparatus according to the embodiment of the present invention. [Figure 4] FIG. 4 is a flowchart showing the overall flow of a learning material document generation process according to the embodiment of the present invention. [Figure 5] FIG. 5 is a conceptual diagram for explaining separation of a target record into a first record group and a second record group by a separation unit according to the embodiment of the present invention. [Figure 6] FIG. 6 is a flowchart showing the flow of a prohibited word set synthesis process by a prohibited word synthesis unit according to the embodiment of the present invention. [Figure 7] FIG. 7 is a diagram for explaining deterministic removal processing (lint) for a synthesized prohibited word set according to the embodiment of the present invention. [Figure 8] FIG. 8 is a conceptual diagram for explaining the relationship between the generation processing of a learning material document by a learning material generation unit and the range of input according to the embodiment of the present invention. [Figure 9] FIG. 9 is a flowchart showing the flow of temporal leak inspection and regeneration processing by a collation unit according to the embodiment of the present invention. [Figure 10] FIG. 10 is a flowchart showing the flow of numerical consistency condition verification and regeneration processing by a numerical value verification unit according to the embodiment of the present invention. [Figure 11] FIG. 11 is a diagram for explaining derivation from a first learning material document to a second learning material document by a derivation output unit and answer leak inspection by a collation unit according to the embodiment of the present invention. [Figure 12] FIG. 12 is a sequence diagram showing the flow of target record identifier resolution and usage authority verification by an identifier resolution request unit according to the embodiment of the present invention. [Figure 13] FIG. 13 is a diagram showing a specific example of the relationship between a reference time point and the first record group and the second record group according to the embodiment of the present invention. [Modes for carrying out the invention]
[0025] The teaching material creation support device, teaching material creation support method, and teaching material creation support program according to embodiments of the present invention will be described below with reference to the figures. In the following description, the teaching material creation support device may be simply referred to as the "support device," and the large-scale language model may be referred to as "LLM." [Overall System Configuration]
[0026] As shown in Figure 1, System 1 of this embodiment is configured such that a teaching material creation support device 10, a user terminal 30 used by teachers, a large-scale language model 40, and a record management system 50 for managing records that identify subjects are all interconnected via a network 20 and can communicate with each other. The teaching material creation support device 10 generates teaching material documents in response to specifications from the user terminal 30 and outputs them to the user terminal 30.
[0027] The record management system 50 is, for example, an educational electronic medical record system as described in Patent Document 1, in which multiple types of records related to the condition, examinations, orders, and nursing records are accumulated chronologically for each subject through operation of a teacher's terminal or a student's terminal. In this embodiment, "subject record" refers to a collection of multiple chronological records related to a particular subject that are accumulated in the record management system 50.
[0028] The large-scale language model 40 is a trained language model that generates text and structured data in response to instructions (prompts) in natural language. It may be implemented on an information processing device separate from the material creation support device 10, or on the same device as the material creation support device 10. In this embodiment, the large-scale language model 40 used has a structured output function that pre-defines the type of data to be output and strictly adheres to that type. [Hardware configuration]
[0029] The teaching material creation support device 10 is realized, for example, by executing a teaching material creation support program on a workstation, personal computer, or other computer device. As shown in Figure 2, the hardware configuration of the teaching material creation support device 10 includes at least a CPU (Central Processing Unit) 11, ROM (Read Only Memory) 12, RAM (Random Access Memory) 13, a communication control interface 14, a storage device 15, and an input operation unit 16.
[0030] The CPU 11 performs various calculations and other processing to control the entire material creation support device 10. More specifically, the CPU 11 reads the material creation support program from the ROM 12 or storage device 15 and executes it using the RAM 13 as a work area, thereby controlling the operation of each component of the material creation support device 10. The communication control interface 14 controls the transmission and reception of data with the outside via the network 20, and the CPU 11 makes requests for generation to the large-scale language model 40 and receives the results, as well as requests for identifier resolution and authorization verification to the record management system 50 via the communication control interface 14. The storage device 15 stores various data such as the material creation support program, the material schema described later, and the cache of target records. The input operation unit 16 is a keyboard, mouse, etc., and accepts input operations from the user.
[0031] The user terminal 30 is implemented by an information processing device such as a desktop or notebook personal computer, a tablet terminal, or a smartphone, and includes at least a CPU, ROM, RAM, display unit, input unit, and communication control interface. [Functional Configuration]
[0032] Next, the functional configuration of the teaching material creation support device 10 will be explained based on Figure 3. Functionally, the teaching material creation support device 10 comprises at least a control unit 101, a storage unit 102, and an interface unit 103. Functionally, the control unit 101 comprises at least a reception unit 111, an identifier resolution request unit 118, a separation unit 112, a forbidden word synthesis unit 113, a teaching material generation unit 114, a matching unit 115, a numerical verification unit 116, and a derived output unit 117. The storage unit 102 stores at least a teaching material schema 121 and a target record cache 122.
[0033] The reception unit 111 is a means for receiving from the user terminal 30 the designation of the target record to be used for generating the teaching material document (an identifier to identify the target) and the designation of the reference point in the time series.
[0034] The identifier resolution request unit 118 is a means of requesting the record management system 50 to resolve the target record corresponding to the identifier received by the reception unit 111 and verify the user's access rights, and acquiring the resolved target record and temporarily storing it in the target record cache 122.
[0035] The separation unit 112 is a means for separating each record included in the target record stored in the target record cache 122 into a first group of records consisting of records belonging to a reference time before the reference time received by the reception unit 111, and a second group of records consisting of records belonging to a time after the reference time.
[0036] The forbidden word synthesis unit 113 is a means for causing the large-scale language model 40 to generate a set of words (forbidden word set) that represent events that have not occurred or are unknown at the reference time, by referring to both the first record group and the second record group.
[0037] The material generation unit 114 is a means of causing the large-scale language model 40 to generate material documents by referring only to the first set of records and without including the set of forbidden words as input.
[0038] The matching unit 115 is a means of checking whether the teaching material document generated by the teaching material generation unit 114 contains words belonging to the set of prohibited words by deterministic partial matching, and is also a means of checking whether the answer information of the first teaching material document is mixed into the second teaching material document derived by the derived output unit 117, which will be described later, by similar partial matching.
[0039] The numerical verification unit 116 is a means for mechanically verifying the numerical consistency conditions that the teaching material document must satisfy. The derived output unit 117 is a means for deterministically deriving a second teaching material document by excluding the fields corresponding to the answer information from a first teaching material document containing the answer information. [Overall processing flow]
[0040] Figure 4 is a flowchart showing an example of the overall flow of the material creation support device 10's process for generating material documents. First, the reception unit 111 receives the identifier of the target record and the specification of the reference time from the user terminal 30 (step S1). At this time, the reception unit 111 verifies the validity of the specified reference time based on the time-series range of the target record (the range of already existing records), and if it is outside the range, it does not accept the specification and requests the user terminal 30 to re-enter the information.
[0041] Once the designation is accepted, the identifier resolution request unit 118 requests the record management system 50 to resolve the identifier and verify the usage rights, and acquires the resolved target record (step S2). Next, the separation unit 112 separates the acquired target record into a first record group and a second record group at the reference time (step S3).
[0042] Next, the forbidden word synthesis unit 113 synthesizes a set of forbidden words using both the first and second record groups as input to the large-scale language model 40 (step S4). The synthesized set of forbidden words is then subjected to a deterministic removal process, which will be described later. Next, the teaching material generation unit 114 generates a teaching material document using only the first record group as input to the large-scale language model 40, and without including the set of forbidden words in the input (step S5).
[0043] Once the training material document is generated, the numerical verification unit 116 verifies the numerical consistency conditions (step S6). If a violation is detected, the content is added to the input to the large-scale language model 40, and the process returns to step S5 for regeneration. If the numerical consistency conditions are met, the matching unit 115 checks whether the generated training material document contains any words belonging to the prohibited word set (step S7). If contamination is detected, the process returns to step S5 for regeneration of the training material document. In either case, if the violation or contamination is not resolved after a predetermined number of regenerations, the generation of the training material document is interrupted.
[0044] If the temporal leak check is passed, the derived output unit 117 derives a second teaching material document by removing the fields corresponding to the answer information from the generated teaching material document (first teaching material document) (step S8), and the verification unit 115 checks whether the answer information from the first teaching material document is mixed into the second teaching material document (step S9). The teaching material document that has passed each of the above checks is output to the user terminal 30 (step S10).
[0045] The following provides a more detailed explanation of each component. [Component (1): Separation of target records by reference time]
[0046] As shown in Figure 5, the separation unit 112 compares each of the multiple records included in the target record with the time point associated with that record (for example, the number of days elapsed since the start of the target) and a reference time point. If the record is before the reference time point, it is classified into the first record group; if it is after the reference time point, it is classified into the second record group. As will be described later, only the first record group is supplied to the material generation unit 114, and the second record group is not supplied to the material generation unit 114 at all. This asymmetrical supply relationship is one of the technical features of this embodiment.
[0047] Furthermore, the records stored in the record management system 50 are assigned an attribute that indicates the time point to which the record corresponds (for example, the number of days relative to the starting point, such as the date of hospitalization), and the separation unit 112 performs the separation deterministically without using the large-scale language model 40 by comparing the value of the attribute with the value at the reference time. [Component (2): Composition of forbidden word sets and deterministic elimination process]
[0048] As shown in Figure 6, the forbidden word synthesis unit 113 receives both the first and second record groups as input to the large-scale language model 40 and synthesizes a set of words representing events that have not occurred or are unknown at the reference time (for example, test results that are known after the reference time, measures that are taken after the reference time, etc.) as part of the accompanying information representing the situation at the reference time.
[0049] As shown in Figure 7, the teaching material creation support device 10 further applies a deterministic rule-based removal process (hereinafter referred to as "lint") to the synthesized set of forbidden words. The lint process includes removing words that have already appeared in the first record group (words that represent events already known before the reference time) from the set of forbidden words, and removing words that contradict each other in the set of forbidden words (for example, when one presupposes the existence of the other).
[0050] The aforementioned linting prevents excessive output suppression caused by mistakenly prohibiting events already known before the reference point, and also prevents erroneous non-conformity judgments due to internal inconsistencies in the prohibited word set itself. If the prohibited word set is still empty after linting, it is preferable to have the request for prohibited word synthesis to the large-scale language model 40 executed again, up to a predetermined upper limit. [Component (3): Generation of teaching materials and verification / regeneration of temporal leaks]
[0051] As shown in Figure 8, the material generation unit 114 provides only the first set of records as input to the large-scale language model 40, and does not include the set of forbidden words synthesized by the forbidden word synthesis unit 113 as input. In other words, the process of generating the material document is configured not to access information after the reference time, nor the set of forbidden words derived from that information.
[0052] The reason why the set of prohibited words is not included in the input to the material generation unit 114 is as follows: If words relating to information after the reference point were explicitly included in the input to the large-scale language model 40 as an instruction "do not use", the words themselves would still be included in the input (the context that the large-scale language model 40 can refer to), and the risk of those words being reflected in the generated product could not be eliminated. In this embodiment, the set of prohibited words is completely excluded from the input to the material generation unit 114, and instead, this problem is avoided by using it only for subsequent deterministic matching of the generated material document (processing by the matching unit 115, described later).
[0053] As shown in Figure 9, the matching unit 115 checks for partial matches between the description in the teaching material document generated by the teaching material generation unit 114 and each word included in the set of prohibited words. If a partial match is detected for any word, the matching unit 115 instructs the teaching material generation unit 114 to regenerate the teaching material document. Regeneration is limited to a predetermined number of times (e.g., once), and if partial matches are still detected after regeneration up to the limit, the process is interrupted without outputting the teaching material document. [Component (4): Machine verification of numerical consistency conditions]
[0054] As shown in Figure 10, the numerical verification unit 116 mechanically verifies the numerical consistency conditions that the teaching materials must satisfy. Numerical consistency conditions include, for example, that the sum of the points allocated to each evaluation level in the evaluation rubric matches a predetermined maximum score, and that the sum of the time allocations for each item in the exercise guide matches a predetermined total exercise time.
[0055] If a violation of the numerical consistency condition is detected, the numerical verification unit 116 appends a message indicating the nature of the violation to the end of the input to the large-scale language model 40 and causes the material generation unit 114 to regenerate the material document. Regeneration is limited to a predetermined number of times, and if the violation is not resolved even after regeneration up to the limit, the process is interrupted.
[0056] In this embodiment, the verification of numerical consistency conditions (component (4)) is performed prior to the matching of temporal leaks (component (3)). That is, the generated training material document is first verified for numerical consistency conditions, and only after these conditions are met is it checked for temporal leaks. [Component (5): Deriving the learner's version from the teacher's version and verifying leaked answers]
[0057] The teaching material document (first teaching material document) generated by the teaching material generation unit 114 includes answer information such as model answers, explanations, and evaluation points. As shown in Figure 11, the derived output unit 117 deterministically derives a second teaching material document that does not contain answer information by mechanically removing fields corresponding to the answer information from the first teaching material document. In other words, generation by the large-scale language model 40 is performed only once when the first teaching material document is generated, and the second teaching material document is not generated again independently, but is derived from the first teaching material document through deterministic processing.
[0058] Furthermore, the matching unit 115 further checks, by partial matching, whether a portion of the content of the field corresponding to the answer information in the first teaching material document (for example, a continuous string of characters longer than a predetermined number) is included in a field visible to the learner in the second teaching material document. This check is performed to detect retrospectively if there was an error in the mechanical exclusion process by the derived output unit 117, or if the large-scale language model 40 duplicates the content of the answer information in fields visible to the learner.
[0059] Thus, the matching unit 115 in this embodiment is used for both checking for temporal leaks (component (3)) with a reference time as the boundary, and checking for answer leaks (component (5)) with a boundary between teacher-only and learner-visible information as the boundary. Both are implemented as a common process of deterministic partial matching between "a set of information that should not be referenced or disclosed in principle" and "the teaching material document to be checked." [Component (6): Bounded customization and direct transfer of definite values]
[0060] The material generation unit 114 generates material documents as structured data that strictly adheres to a data schema predefined for each type of material document (e.g., worksheets, evaluation rubrics, exercise guides, etc.). The data schema has fixed types and structures for each field, such as questions and answer information, and does not allow free-form descriptions not defined in the schema.
[0061] Furthermore, among the information contained in the target record, items that are predetermined, such as information that identifies the target (hereinafter referred to as "determined value items"), are not generated by the material generation unit 114, but are directly copied from the target record when the material document is output. In other words, the content related to the determined value items is reflected in the material document without going through the generation results of the large-scale language model 40. [Component (7): Resolution of target records and verification of access rights using identifiers]
[0062] As shown in Figure 12, the reception unit 111 accepts the designation of an identifier that identifies the target record, rather than the target record itself. The identifier resolution request unit 118 requests the record management system 50 to resolve the target record corresponding to the identifier and to verify whether the user who made the designation has the authority to use the target record, both at the time the reception unit 111 accepts the designation of the identifier and at the time the material document generation process is executed.
[0063] The identifier resolution request unit 118 receives the resolution result and the authorization verification result from the record management system 50 and supplies only the target records for which authorization has been confirmed to the separation unit 112. The teaching material creation support device 10 itself does not permanently retain the main body of the target record, and storage in the target record cache 122 is temporary. This two-stage resolution and authorization verification (at reception and execution) prevents access to target records for which the user does not have authorization, which could be overlooked with only one verification. [Verification of the validity of the reference period]
[0064] As described above, when the reception unit 111 accepts the designation of a reference time, it verifies whether the designated reference time falls within the time-series range of the target record resolved by the identifier resolution request unit 118 (the minimum and maximum values of the time points corresponding to the already existing record). If a reference time outside the range is designated, the reception unit 111 does not accept the designation and requests the user terminal 30 to re-enter the designation along with the acceptable range. This verification is performed after the target record is resolved by the identifier resolution request unit 118 and before the separation process by the separation unit 112. [Specific examples of actions]
[0065] The following describes an example of operation based on specific specifications. For example, suppose that from the user terminal 30, the identifier of a certain target stored in the record management system 50 is specified as the identifier of the target record, and the reference time is set to the third day from the starting point of that target.
[0066] The reception unit 111 confirms that the specified reference point (day 3) falls within the time series range (for example, from day 0 to day 14) of the target record resolved by the identifier resolution request unit 118. Since it is within the range, the specification is accepted.
[0067] The separation unit 112 separates each record included in the target record into those corresponding to the third day or earlier and those corresponding to those corresponding to the third day or later (fourth day or later). The forbidden word synthesis unit 113 refers to both sets of records and synthesizes a set of forbidden words that represent events that can only be determined from the fourth day onward (for example, the results of tests not performed on the third day, the content of measures scheduled after the third day, etc.). The teaching material generation unit 114 refers only to the set of records corresponding to the third day or earlier and generates a teaching material document (first teaching material document) based on the third day, without including the forbidden word set as input.
[0068] The numerical verification unit 116 verifies the numerical consistency conditions, and the matching unit 115 checks for partial matches with the set of prohibited words. Once it is confirmed that there are no problems, the derived output unit 117 derives the second teaching material document (learner version) from the first teaching material document, and the matching unit 115 checks for any leaked answers. The first and second teaching material documents, having passed both checks, are output to the user terminal 30. [Variation]
[0069] The present invention is not limited to the embodiments described above, and various modifications are possible without departing from the spirit of the invention.
[0070] For example, each functional unit of the teaching material creation support device 10 may be implemented on a single device, or it may be implemented in a distributed manner across multiple devices. The large-scale language model 40 may be configured integrally with the teaching material creation support device 10.
[0071] Furthermore, the synthesis of the forbidden word set by the forbidden word synthesis unit 113 and the generation of the teaching material document by the teaching material generation unit 114 are not limited to using the large-scale language model 40. That is, these processes may be (a) generated by the large-scale language model, (b) generated based on predetermined rules (rule-based generation logic), or (c) a combination of these. In any case, the deterministic partial match matching by the matching unit 115 and the mechanical verification of numerical consistency conditions by the numerical verification unit 116 on the generated teaching material document are equally valid. Therefore, the effects of the present invention do not depend on the type of means for generating the forbidden word set and the teaching material document.
[0072] Furthermore, the set of forbidden words is an example of matching information that represents events that have not occurred or are unknown at the reference time, i.e., non-disclosure items that should not be included in the teaching material document, in a format that can be identified by the matching unit 115 during inspection. Each element of the matching information is not limited to being held as a word (string) representing an event, but may also be held as (a) a string, (b) a code or identifier associated with the event, (c) structured data (event object) including the type, target, and time of the event as elements, (d) a feature quantity deterministically calculated from the string or the data, or (e) a combination thereof. In this embodiment, when we refer to "forbidden words" and "set of forbidden words," we mean to include these forms. Regardless of the form, the aforementioned asymmetric configuration remains unchanged, in which the matching information is synthesized by referring to both the first record group and the second record group, is not included in the input to the teaching material generation unit 114, and is used only for subsequent deterministic matching of the generated teaching material document, and the effects of the present invention are achieved in the same way.
[0073] Furthermore, the check performed by the matching unit 115 is not limited to deterministic partial matching. That is, the check may be (a) a partial match that determines whether a word belonging to the set of forbidden words is included as a string in the description of the teaching material document, (b) an exact match, prefix match, or suffix match, (c) a match after normalizing variations in notation such as full-width and half-width characters, uppercase and lowercase letters, and okurigana according to predetermined rules, (d) a match with a regular expression or other predetermined pattern, (e) a match of the code, identifier, or data when the matching information is held as a code, identifier, or structured data, (f) a determination of whether the similarity between the description of the teaching material document and the matching information, obtained by a predetermined calculation method, is above a predetermined threshold, or (g) any combination thereof. However, in any embodiment, the check is implemented as a deterministic operation that always returns the same determination result for the same input, namely the same teaching material document and the same matching information, and does not use probabilistic determination by a trained language model. By not introducing probabilistic behavior into the verification stage, the reproducibility and auditability of the detection of temporal leaks are ensured, and as long as this determinism is maintained, the specific method for determining agreement can be any of the above.
[0074] Furthermore, the suppression of output of the teaching material document when the matching unit 115 detects the inclusion of a word belonging to the prohibited word set is not limited to regenerating the teaching material document or interrupting it after a predetermined number of regenerations. That is, the suppression of output may be (a) causing the teaching material generation unit 114 to regenerate the teaching material document, (b) interrupting the generation of the teaching material document if the inclusion is not resolved even after a predetermined number of regenerations, (c) masking the portion of the teaching material document in which the inclusion was detected so that it is illegible before outputting it, (d) displaying a warning that the inclusion has been detected on the user terminal 30 and not providing the teaching material document in its original form, (e) withholding the provision of the teaching material document until approval is obtained from a user with predetermined authority, such as a teacher, (f) switching to a form in which the teaching material document is not presented to learners but only to teachers, or (g) any combination of these. In any of these embodiments, the effect of the present invention, which is the mechanical prevention of temporal leakage, is similarly achieved in that the teaching material document in which the inclusion was detected is prevented from being presented to learners in its original form.
[0075] Furthermore, the specific items of the target records and reference time, the classification of scenes used in the synthesis of the forbidden word set, and the specific content of the numerical consistency conditions are determined as appropriate according to the type of educational material document and the specifications of the record management system, and are not limited to those exemplified in this embodiment. The record management system 50 is not limited to one for nursing education, but may be a general record management system used for the education and training of healthcare professionals. [Technical significance of the present invention]
[0076] Each functional unit of the present invention (reception unit 111, identifier resolution request unit 118, separation unit 112, forbidden word synthesis unit 113, teaching material generation unit 114, matching unit 115, numerical verification unit 116, and derived output unit 117, etc.) is specifically realized by the CPU 11 of the teaching material creation support device 10 executing a teaching material creation support program stored in a storage device 15, etc., using RAM 13 as a working area, and cooperating with hardware resources such as the communication control interface 14 and the storage device 15. In other words, the present invention performs specific information processing using hardware resources, such as a deterministic process of separating target records into a first record group and a second record group at a reference time, and a deterministic partial match matching between the generated teaching material document and the forbidden word set (a judgment operation that always returns the same result for the same input). Therefore, the present invention is not merely limited to the presentation of information, artificial arrangements, or mathematical methods themselves, but achieves the technical effect of mechanically preventing temporal leaks through these specific information processing. This technical effect is achieved by the information processing performed by the separation unit and the matching unit, regardless of the type of means used to generate the forbidden word set and the instructional document (whether by a large-scale language model or by rules). [Explanation of symbols]
[0077] 1 System 10. Educational Material Creation Support Device 11 CPU 12 ROM 13 RAM 14 Communication control interface 15 Storage device 16 Input Operation Section 20 Networks 30 User terminals 40 Large-scale language models 50 Record Management Systems 101 Control Unit 102 Storage section 103 Interface section 111 Reception Department 112 Separation section 113 Forbidden word synthesis section 114 Teaching material generation department 115 Verification section 116 Numerical Verification Department 117 Derived Output Section 118 Identifier Resolution Request Section 121. Teaching Material Schema 122 Target Record Cache
Claims
1. A support device for creating educational materials used for education and training of medical professionals, etc., A receiving unit that accepts the designation of target records, which include multiple records arranged in chronological order, and the designation of a reference point in the said chronological order. A separation unit that separates the records included in the target record into a first group of records consisting of records belonging to a reference time or earlier, and a second group of records consisting of records belonging to a reference time or later. A forbidden word synthesis unit provides both the first and second record sets as input to a trained language model that generates text or structured data in response to natural language instructions, and causes the trained language model to synthesize a set of forbidden words, which is a set of words representing events that have not occurred or are unknown at the reference time; Regarding the aforementioned target records, only the first group of records is provided as input to a trained language model that generates text or structured data in response to instructions in natural language, and the trained language model generates the aforementioned teaching material document without including the set of forbidden words in the input, and the teaching material generation unit A matching unit that checks whether the generated teaching material document contains words belonging to the set of prohibited words by deterministic partial matching, and suppresses the output of the teaching material document if it contains such words, Equipped with, The aforementioned deterministic partial match matching determines whether there is a partial match between the description in the instructional document and each word included in the set of forbidden words, using a determination operation that always returns the same result for the same input. A device to support the creation of educational materials.
2. The trained language model used by the forbidden word synthesis unit and the trained language model used by the material generation unit each have a structured output function that predefines the type of data to be output and strictly adheres to that type. The teaching material creation support device according to claim 1.
3. The forbidden word synthesis unit performs the following processes from the synthesized forbidden word set: removing words that have already appeared in the first record group, and removing words that are mutually contradictory to each other, based on a deterministic rule that always returns the same result for the same input. The teaching material creation support device according to claim 1 or 2.
4. The matching unit, when the teaching material document contains a word belonging to the set of prohibited words, causes the teaching material generation unit to regenerate the teaching material document. The teaching material creation support device according to claim 1 or 2.
5. The matching unit interrupts the generation of the instructional document if the inclusion of words belonging to the prohibited word set is not eliminated even after a predetermined number of regenerations. The teaching material creation support device according to claim 4.
6. The reception unit, when accepting the designation of the reference time, verifies the validity of the reference time based on the time-series range of the target record and does not accept designations outside the range. The teaching material creation support device according to claim 1 or 2.
7. The system further includes a numerical verification unit that mechanically verifies the numerical consistency conditions that the aforementioned teaching material document must satisfy (including at least one of the conditions that the sum of the points allocated matches the predetermined maximum score, and that the sum of the time allocation matches the predetermined total time), and, if a violation is detected, causes the teaching material generation unit to regenerate the teaching material document with information indicating the nature of the violation. The teaching material creation support device according to claim 1 or 2.
8. The numerical verification unit interrupts the generation of the training material document if the violation of the numerical consistency condition is not resolved even after a predetermined number of regenerations. The teaching material creation support device according to claim 7.
9. The aforementioned material generation unit generates a first material document containing answer information, The aforementioned teaching material creation support device further comprises a derived output unit that deterministically derives a second teaching material document that does not contain answer information by removing fields corresponding to answer information from the first teaching material document. The teaching material creation support device according to claim 1 or 2.
10. The matching unit further checks, by partial matching, whether the content of the field corresponding to the answer information in the first teaching material document is included in the field that is visible in the second teaching material document. The teaching material creation support device according to claim 9.
11. The aforementioned training material generation unit generates the training material document as structured data that conforms to a predetermined data schema, The predetermined fixed value items included in the aforementioned target record are not generated by the aforementioned teaching material generation unit, but are directly copied from the aforementioned target record when the teaching material document is output. The teaching material creation support device according to claim 1 or 2.
12. The receiving unit receives the designation of the target record by an identifier that does not include the body of the target record. The aforementioned teaching material creation support device requests the record management system that manages the target records to resolve the identifier and verify the usage rights at the time of receiving the designation and at the time of executing the process of generating the teaching material document, and supplies only the resolved target records to the separation unit. The teaching material creation support device according to claim 1 or 2.
13. A method for creating educational materials, which is executed by a computer, for creating educational materials documents used for education and training of medical professionals, etc. A step that accepts the designation of target records, which include multiple records arranged in chronological order, and the designation of a reference point in the said chronological order. The steps include separating the records included in the target record into a first group of records consisting of records belonging to a reference time or earlier, and a second group of records consisting of records belonging to a reference time or later. The steps include providing both the first and second record sets as input to a trained language model that generates text or structured data in response to natural language instructions, and causing the trained language model to synthesize a set of forbidden words, which is a set of words representing events that have not occurred or are unknown at the reference time; The steps include providing only the first group of records as input to a trained language model that generates text or structured data in response to natural language instructions, and causing the trained language model to generate the training document without including the set of forbidden words in the input, The steps include: checking whether the generated instructional document contains words belonging to the set of prohibited words by deterministic partial matching, and suppressing the output of the instructional document if it contains such words; Includes, The aforementioned deterministic partial match matching determines whether there is a partial match between the description in the instructional document and each word included in the set of forbidden words, using a determination operation that always returns the same result for the same input. Methods for supporting the creation of teaching materials.
14. A textbook creation support program that causes a computer to perform each step of the textbook creation support method described in claim 13.
Citation Information
Patent Citations
Educational electronic medical chart system and educational electronic medical chart program
JP2026056431A
Educational electronic medical record system and educational electronic medical record program
JP7726558B1
