Programs, analysis systems, and model building methods

Abstract

The goal is to enable analysts to easily identify the causes of errors that occur during the manufacturing process. [Solution] The analysis system (100) includes a reception unit (11) that receives input of the content of a target event, a first generation unit (253) that generates at least one candidate text indicating a candidate cause of the target event based on the content of the target event, and an output unit (12) that outputs the candidate text as options. The reception unit further receives input of a sentence indicating the causal event, which has been selected from the candidate texts or created based on the candidate texts.

Description

Technical Field

[0001] The present invention relates to a program, an analysis system, and a model construction method.

Background Art

[0002] As one of the methods for analyzing the causes of mistakes and troubles caused by human behavior in operations such as manufacturing processes and maintenance, for example, as described in Non-Patent Document 1 and Document 2, time-series factor analysis (Variation Tree Analysis: VTA), why-why analysis, etc. are known. In particular, the method called why-why analysis or root cause analysis is considered to be a simple and effective means because analysts logically describe sentences and pursue the causes.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in conventional analytical methods, the process of manually identifying factors and documenting them was extremely time-consuming. This is because these analytical methods require the analyst to refer to all information related to the occurrence of the error or trouble and list each factor one by one. As a result, problems arose where the identification of factors at a certain point in time was insufficient, leading to an insufficient identification of the factors that triggered those factors at a later point in time. Furthermore, the writing style of the analysis varied greatly depending on the analyst. For example, some analysts omitted the subject of their sentences in an attempt to write concisely. As a result, problems arose such as misinterpretations during the analysis, leading to analysis failures, or the analysis stalling due to a lack of understanding of factors based on experience, etc. In other words, conventional analytical methods also had the problem of producing inconsistent results even when analyzing the same error or trouble, depending on the analyst.

[0005] One aspect of the present invention aims to enable analysts to easily identify the causes of errors and troubles that occur in the manufacturing process. [Means for solving the problem]

[0006] The program according to aspect 1 of the present invention is configured to cause a computer to execute a first generation process that generates at least one candidate text indicating a candidate cause of the target event based on the content of the target event.

[0007] A program according to aspect 2 of the present invention may be configured such that, in the first generation process described in aspect 1 above, the computer is instructed to input the content of the target event into a classification prediction model that is constructed to output a classification to which the target event belongs when the content of the target event is input, and to execute a classification generation process that obtains at least one output from the classification prediction model as a candidate classification for the target event, and generates the candidate text based on a classification selected from the candidate classifications or created based on the candidate classifications.

[0008] The program according to aspect 3 of the present invention may be configured such that, in the first generation process described in aspect 2 above, after causing the computer to execute the category generation process, the selected or created category is input to a first concept prediction model which is constructed to output a concept that causes an event belonging to the category when a category is input, and the computer executes a concept generation process which obtains at least one output from the first concept prediction model as a candidate concept belonging to the selected or created category, and the candidate text is generated based on a concept selected from the candidate concept or created based on the candidate concept.

[0009] In the program according to aspect 4 of the present invention, in the concept generation process described in aspect 3 above, the first concept prediction model may be configured to obtain a concept example corresponding to the selected or created category from a list of concept examples as a candidate concept.

[0010] The program according to aspect 5 of the present invention may be configured such that, in the first generation process described in aspect 3 above, the computer is given an input of a sentence representing a concept to an event prediction model which is constructed to output text representing a causal event that is the cause of the target event included in the concept, and the computer is given an input of a sentence representing a concept to execute an event generation process which is obtained by acquiring at least one output from the event prediction model as the candidate text corresponding to the selected or created concept.

[0011] In the program according to aspect 6 of the present invention, in the event generation process described in aspect 5 above, the event prediction model may be configured to obtain, as candidate text, text indicating a past event corresponding to the selected or created concept from a database in which multiple texts indicating past events that have occurred in the past have been stored.

[0012] The program according to aspect 7 of the present invention may be configured such that, in aspect 5 above, after causing the computer to execute the first generation process, when a sentence indicating the causal event is input, the computer is further given a sentence indicating the causal event, selected from the candidate texts or created based on the candidate texts, to a second concept prediction model constructed to output another concept that is the concept of the cause of the causal event, and the computer is given at least one output from the second concept prediction model as a candidate for another concept corresponding to the sentence indicating the causal event.

[0013] The program according to embodiment 8 of the present invention may be configured such that, in embodiment 7 above, when one other concept is selected or created based on the candidate of the other concepts, the computer is made to execute the event generation process based on the one other concept, and to execute the second generation process based on the candidate text generated in correspondence with the one other concept.

[0014] The program according to aspect 9 of the present invention may be configured such that, in any of aspects 5 to 8 described above, if the execution of the event generation process is repeated, the computer further executes a first judgment process that determines whether the causal relationship between the content indicated by a sentence selected from the candidate texts generated in one event generation process, or created based on such candidate texts, and the content indicated by a sentence selected from the candidate texts generated in an event generation process executed before the first event generation process, or created based on such candidate texts, is logically correct.

[0015] The program according to embodiment 10 of the present invention may be configured such that, in the classification generation process described in embodiment 2 above, the computer vectorizes the text describing the content of the target event and calculates the vector value of the text, while weighting it based on the importance of each word input into the classification prediction model.

[0016] The program according to aspect 11 of the present invention may be configured such that, in aspect 7 above, the computer is further instructed to perform a second judgment process to determine whether the candidate concepts of other concepts output by the second concept prediction model for the text indicating the causal event were valid, after the second generation process has been performed.

[0017] The program according to aspect 12 of the present invention may be configured in aspect 4 above to further cause the computer to perform an update process that evaluates the quality of the list of conceptual examples and updates the list based on the evaluation result.

[0018] The program according to aspect 13 of the present invention may be configured such that, in aspect 12 above, the quality is at least one of the following: comprehensiveness of conceptual examples, clarity of distinction between conceptual examples, and accuracy of expression for each conceptual example.

[0019] An analysis system according to aspect 14 of the present invention comprises a receiving unit that receives input of the content of a target event, a first generating unit that generates at least one candidate text indicating a candidate cause of the target event based on the content of the target event, and an output unit that outputs the candidate text as an option, wherein the receiving unit further receives input of a sentence indicating a causal event that caused the target event, which is selected from the candidate text or created based on the candidate text.

[0020] A model construction method according to aspect 15 of the present invention is a method that includes the step of constructing a classification prediction model that outputs a classification when a sentence describing the content of a target event is input, by learning the relationship between the content of a target event and the classification to which the target event belongs as training data.

[0021] A model construction method according to aspect 16 of the present invention is a method that includes the step of constructing a first concept prediction model that outputs a concept when a category is input, by learning using the relationship between a category and the concept of the cause of an event belonging to that category as learning data.

[0022] The model construction method according to Embodiment 17 of the present invention includes a step of constructing an event prediction model that outputs text indicating a causal event when text indicating a concept is input, by learning the relationship between a concept and a causal event that is the cause of an event included in the concept, using learning data.

[0023] The model construction method according to Embodiment 18 of the present invention includes a step of constructing a second concept prediction model that outputs another concept when text indicating a causal event is input, by learning the relationship between a causal event that is the cause of an event and another concept that is the concept of the cause of the causal event, using learning data.

[0024] The model construction method according to Embodiment 19 of the present invention further includes, in Embodiment 15, 17 or 18 described above, a step of vectorizing the text used for learning by weighting based on the importance of each word included in the text, and a step of constructing a model using the vectorized text. This may be a method.

[0025] The model construction method according to Embodiment 20 of the present invention further includes, in Embodiment 15, 17 or 18 described above, a step of generating fictional text indicating an event with a relatively low occurrence frequency among a plurality of events, and a step of constructing the first concept prediction model using the fictional text. This may be a method.

Advantages of the Invention

[0026] According to one aspect of the present invention, when analyzing the causes of mistakes and troubles that occur in the manufacturing process, the analyst can easily derive the factors. Also, according to one aspect of the present invention, it is possible to reduce the variation in the analysis results among different analysts for the same mistakes and troubles.

Brief Description of the Drawings

[0027] [Figure 1] It is a block diagram showing an example of the schematic configuration of an analysis system according to an embodiment of Disclosure 1. [Figure 2]This block diagram shows an example of the functional configuration of the terminals included in the system. [Figure 3] This block diagram shows an example of the functional configuration of the information processing device included in the system. [Figure 4] This figure shows an example of the information stored in the device. [Figure 5] This figure shows an example of the information stored in the device. [Figure 6A] This flowchart shows an example of the first half of the processing flow performed by the device. [Figure 6B] This flowchart shows an example of the latter half of the processing flow performed by the device. [Figure 7] This diagram illustrates some of the processes performed by the device. [Figure 8] This diagram illustrates some of the processes performed by the device. [Figure 9] This diagram illustrates some of the processes performed by the device. [Figure 10] This diagram illustrates some of the processes performed by the device. [Figure 11] This diagram illustrates an example of a prompt that may be input to the device. [Figure 12] This figure shows an example of the information output by the device. [Figure 13] This diagram illustrates an example of a prompt that may be input to the device. [Figure 14] This is a flowchart showing an example of the flow of the model construction method according to the embodiment of Disclosure 2. [Modes for carrying out the invention]

[0028] <Disclosure 1: Embodiment of the analysis system> The following describes in detail one embodiment of the analysis system related to Disclosure 1.

[0029] When an incident (mistake or trouble) occurs at a manufacturing site, the person in charge at the manufacturing site (hereinafter referred to as the user) analyzes the cause of the incident. The analysis system 100 is a system that assists the user in analyzing the cause. As shown in Figure 1, the analysis system 100 comprises one or more terminal devices 1 and an information processing device 2. The terminal devices 1 and the information processing device 2 communicate with each other via a communication network N.

[0030] [Terminal device 1] As shown in Figure 2, the terminal device 1 according to this embodiment includes a reception unit 11, an output unit 12, a terminal communication unit 13, and a terminal calculation unit 14. The terminal device 1 according to this embodiment is composed of, for example, a PC (Personal Computer), a smartphone, a tablet terminal, etc.

[0031] [Reception area 11] The reception unit 11 receives input regarding the content of the target event. The reception unit 11 may be, for example, a keyboard, a pointing device (such as a mouse), or a touch panel.

[0032] [Output section 12] The output unit 12 outputs candidate text as options. In this embodiment, the output unit 12 is configured as a display device. That is, the output unit 12 in this embodiment displays candidate text as options. The output unit 12 may also be configured as, for example, a speaker, terminals connected to other output devices, etc.

[0033] [Terminal communication unit 13] The terminal communication unit 13 communicates with the information processing device 2. In this embodiment, the terminal communication unit 13 is composed of a wireless communication module. That is, in this embodiment, the terminal communication unit 13 communicates wirelessly with the information processing device 2. However, the terminal communication unit 13 may also be configured to communicate with the information processing device 2 via a wired connection.

[0034] [Terminal processing unit 14] The terminal processing unit 14 controls the terminal communication unit 13 when the receiving unit 11 receives input from the user. This causes the terminal communication unit 13 to transmit the received input to the information processing device 2. Furthermore, the terminal processing unit 14 controls the output unit 12 when the terminal communication unit 13 receives information, or data from a web page containing that information, from the information processing device 2. This causes the output unit 12 to display the information received by the terminal communication unit 13, or the web page containing that information. In this embodiment, the terminal processing unit 14 consists of a processor and memory.

[0035] [Information Processing Device 2] As shown in Figure 3, the information processing device 2 comprises a first storage unit 22 and a device calculation unit 25. The information processing device 2 according to this embodiment further comprises a device communication unit 21 and a second storage unit 23.

[0036] [Device communication unit 21] The device communication unit 21 communicates with the terminal device 1. As described above, the terminal communication unit 13 of the terminal device 1 according to this embodiment is composed of a wireless communication module. Therefore, the device communication unit 21 according to this embodiment is also composed of a wireless communication module. In other words, the device communication unit 21 according to this embodiment communicates with the terminal device 1 wirelessly. The device communication unit 21 may also be configured to communicate with the terminal device 1 via a wired connection.

[0037] [First storage section 22] The first storage unit 22 stores the program 221. The program 221 describes the processing to be executed by the device calculation unit 25. In other words, the program 221 is a program 221 that causes the computer to function as an information processing device 2, and is a program 221 that causes the computer to function as each control block of the device. In this embodiment, the first storage unit 22 is composed of a semiconductor memory, a hard disk drive, etc.

[0038] [Second storage unit 23] The second memory unit 23 stores a list of conceptual examples, such as those shown in Figure 4. The second memory unit 23 also contains a database 231. The database 231 stores multiple texts representing past events, such as those shown in Figure 5. The database 231 also stores multiple texts written on nodes of a tree diagram (conceptual tree) representing analysis results generated in past analyses. Furthermore, the second memory unit 23 stores a dictionary of words (multiple words and their respective vector values). In this embodiment, the second memory unit 23 is composed of semiconductor memory, a hard disk drive, etc.

[0039] [Third memory section 24] The third storage unit 24 stores a classification prediction model 241, a first concept prediction model 242, an event prediction model 243, and a second concept prediction model 244. In this embodiment, the third storage unit 24 is composed of a semiconductor memory, a hard disk drive, or the like.

[0040] (Classification prediction model 241) The classification prediction model 241 is a model constructed to output at least one classification to which a target event belongs when the content of the target event is input. In this embodiment, the classification prediction model 241 receives the text vector value as the content of the target event. Furthermore, the classification prediction model 241 in this embodiment outputs multiple classifications along with the probability to which the input target event belongs.

[0041] (First Concept Prediction Model 242) The first concept prediction model 242 is a model constructed to output a concept that is the cause of an event belonging to a given category when a category is input. In this embodiment, the first concept prediction model 242 is constructed to acquire a list of concept examples stored in the second memory unit 23, which includes a category selected or created by the user or computer, as a candidate concept. In addition, the category is input to the first concept prediction model 242 in this embodiment in the form of a vector value. Furthermore, the first concept prediction model 242 in this embodiment outputs multiple concepts along with their probabilities of being included in the input category. Note that the first concept prediction model 242 may be a pre-trained model constructed to generate a candidate concept when a category is input.

[0042] (Event prediction model 243) The event prediction model 243 is a model constructed to output text indicating causal events included in a concept when a sentence indicating a concept is input. A causal event is an event that causes the target event to occur. The event prediction model 243 according to this embodiment is constructed to obtain text indicating past events included in a concept selected or created by the user or computer from the database 231 as candidate text and output it. In addition, concepts are input to the event prediction model 243 according to this embodiment in the form of vector values. The first concept prediction model 242 may be a trained model constructed to generate text indicating causal events when a sentence indicating a concept is input.

[0043] (Second Concept Prediction Model 244) The second concept prediction model 244 is a model constructed to output other concepts when a sentence describing a causal event is input. These other concepts are concepts representing the causes of the causal event. In this embodiment, the vector value of the sentence is input to the second concept prediction model 244 as the content of the causal event. Furthermore, the second concept prediction model 244 in this embodiment outputs multiple other concepts along with the probability of including the input causal event.

[0044] [Device calculation section 25] The device calculation unit 25 controls each part of the information processing device 2. The device calculation unit 25 includes an acquisition unit 251, an output control unit 252, a first generation unit 253, a second generation unit 254, a first judgment unit 255, a second judgment unit 256, an update unit 257, and a registration unit 258. In this embodiment, the device calculation unit 25 is composed of a processor. That is, the information processing device 2 is composed of a computer. Therefore, each function of each control block 251 to 258 is realized by the device calculation unit 25 executing various processes B1 to B9 in the flow shown in Figures 6A and 6B according to the program 221 stored in the first storage unit 22. Of the various processes, the acquisition process B1 is executed each time the terminal device 1 receives input of various information (the terminal communication unit 13 transmits various information). Also, the output control process B4 is executed each time the device calculation unit 25 generates various information.

[0045] (Acquisition process B1) In the initial acquisition process B1, the acquisition unit 251 acquires the content of the target event received by the device communication unit 21. In this embodiment, the device calculation unit 25 acquires a sentence that represents the content of the target event. The device calculation unit 25 may be configured to acquire the content of the target event in a form other than a sentence (e.g., an image). Also, when the terminal device 1 receives input of a new word from the user (A0), in acquisition process B1, the acquisition unit 251 acquires the word pair received by the device communication unit 21.

[0046] (First generation process B2) After the acquisition process B1 is executed, the process moves to the first generation process B2. In the first generation process B2, the first generation unit 253 generates at least one candidate text based on the content of the target event. The candidate text is text that indicates a candidate cause for the occurrence of the target event. In the first generation process B2 according to this embodiment, the first generation unit 253 executes the classification generation process B21, the concept generation process B22, and the event generation process B23.

[0047] • Classification generation process B21 In the classification generation process B21, the first generation unit 253 inputs the content of the target event into the classification prediction model 241, as shown in Figure 7, and obtains at least one output from the classification prediction model 241 as a candidate classification for the target event. As described above, the classification prediction model 241 according to this embodiment receives the vector value of a sentence as input for the content of the target event. Therefore, the first generation unit 253 according to this embodiment decomposes the sentence representing the content of the target event into multiple words (morphological analysis) and vectorizes each word. Then, the first generation unit 253 converts the vector of each word into a sentence vector. One conversion method is to calculate the average of the vector values ​​of each word as the sentence vector value. When calculating the sentence vector value according to this embodiment, the first generation unit 253 performs weighting based on the importance of each word input into the classification prediction model 241 (for example, giving higher importance to words that are input relatively frequently).

[0048] In output control processing B4 after the execution of category generation processing B21, the device calculation unit 25 controls the device communication unit 21 according to the processing result of category generation processing B21. As a result, the device communication unit 21 sends data of candidate categories or data of a web page indicating candidate categories to the terminal device 1. Upon receiving the data, the terminal device 1 outputs (displays) the candidate categories (A2). The user, after confirming the candidate categories displayed on the terminal device 1, performs an operation to select one of the candidate categories or an operation to create a category based on the content of the candidate categories. As a result, the reception unit 11 of the terminal device 1 receives input of the selected or created category (A3), and the terminal communication unit 13 transmits it to the information processing device 2.

[0049] Furthermore, in the acquisition process B1 after the execution of the classification generation process B21, the acquisition unit 251 acquires the selected or created classification received by the device communication unit 21.

[0050] • Concept generation process B22 After the execution of the classification generation process B21 (after the acquisition unit 251 acquires the selected or created classification), the process moves to the concept generation process B22. In the concept generation process B22, the first generation unit 253 inputs the selected or created classification into the first concept prediction model 242, as shown in Figure 8, and acquires at least one output from the first concept prediction model 242 as a candidate concept belonging to the selected or created classification.

[0051] In output control processing B4 after the execution of concept generation processing B22, the device calculation unit 25 controls the device communication unit 21 according to the processing result of concept generation processing B22. As a result, the device communication unit 21 sends data of candidate concepts or data of a web page showing candidate concepts to the terminal device 1. Upon receiving the data, the terminal device 1 outputs (displays) the candidate concepts (A4). The user, after confirming the candidate concepts displayed on the terminal device 1, performs operations to select one of the candidate concepts or to create a concept based on the content of the candidate concepts. As a result, the reception unit 11 of the terminal device 1 receives input of the selected or created concept (A5), and the terminal communication unit 13 transmits it to the information processing device 2.

[0052] Furthermore, in the acquisition process B1 after the execution of the concept generation process B22, the acquisition unit 251 acquires the selected or created concept received by the device communication unit 21.

[0053] • Event generation process B23 After the execution of concept generation process B22 (after the acquisition unit 251 acquires the selected or created category), the process moves to event generation process B23. In event generation process B23, the first generation unit 253 generates candidate text based on the concept selected from the candidate concept or the concept created based on the candidate concept. Specifically, as shown in Figure 9, the first generation unit 253 inputs a sentence indicating the selected or created concept into the event prediction model 243 stored in the third storage unit 24, and acquires at least one output from the event prediction model 243 as candidate text corresponding to the selected or created concept. In the output control process B4 following the execution of the event generation process B23, the device calculation unit 25 controls the device communication unit 21 according to the processing result of the event generation process B23. As a result, the device communication unit 21 sends candidate text data or data of a web page showing the candidate text to the terminal device 1. Upon receiving the data, the terminal device 1 outputs (displays) the candidate text (A6). The user, after confirming the candidate text displayed on the terminal device 1, performs operations to select one of the candidate texts or to create a sentence indicating the cause event based on the content of the candidate text. As a result, the reception unit 11 of the terminal device 1 further accepts the input of the sentence indicating the cause event, which was selected from the candidate texts or created based on the candidate texts (A7), and the terminal communication unit 13 transmits it to the information processing device 2.

[0054] Furthermore, in the acquisition process B1 after the execution of the event generation process B23, the acquisition unit 251 acquires the selected or created text received by the device communication unit 21.

[0055] (Second generation process B3) After the execution of the first generation process B2, the process moves to the second generation process B3. In the second generation process B3, the second generation unit 254 inputs sentences indicating causal events, selected by the user or computer from among the candidate texts, or created by the user or computer based on the candidate texts, into the second concept prediction model 244, as shown in Figure 10, and obtains at least one output from the second concept prediction model 244 as a candidate for other concepts corresponding to the sentence indicating the causal event.

[0056] In the output control process B4 following the execution of the second generation process B3, the device calculation unit 25 controls the device communication unit 21 according to the processing result of the second generation process B3. As a result, the device communication unit 21 sends data of candidate other concepts, or data of a web page showing candidate other concepts, to the terminal device 1. Upon receiving the data, the terminal device 1 outputs (displays) the candidate other concepts (A8). The user, after reviewing the candidate other concepts displayed on the terminal device 1, performs operations to select one of the candidate other concepts, or to create another concept based on the content of the candidate other concepts. As a result, the reception unit 11 of the terminal device 1 receives input of the selected or created other concept (A9), and the terminal communication unit 13 transmits it to the information processing device 2.

[0057] Furthermore, in the acquisition process B1 after the execution of the second generation process B3, the acquisition unit 251 acquires other selected or created concepts received by the device communication unit 21.

[0058] • Event generation process B23 If another concept is selected or created based on other candidate concepts, i.e., if the acquisition unit 251 acquires the selected or created other concept and the analysis is not yet complete (B5:NO), the process moves back to event generation process B23. The second event generation process B23 is based on the other concept. That is, in the second event generation process B23, the first generation unit 253 inputs a sentence representing the selected or created other concept into the event prediction model 243, and acquires at least one output from the event prediction model 243 as candidate text corresponding to the selected or created other concept.

[0059] In the output control process B4 after the execution of the event generation process B23 again, the device calculation unit 25 controls the device communication unit 21 according to the processing result of the event generation process B23 again. As a result, the device communication unit 21 sends data of candidate text generated in correspondence with another concept, or data of a web page showing the candidate text, to the terminal device 1. Upon receiving the data, the terminal device 1 outputs (displays) the candidate text (A6). The user, after confirming the candidate text displayed on the terminal device 1, performs an operation to select one of the candidate texts, or an operation to create a sentence indicating the cause event based on the content of the candidate text. As a result, the reception unit 11 of the terminal device 1 further accepts input of the sentence indicating the cause event that was selected from the candidate texts or created based on the candidate texts (A7), and the terminal communication unit 13 transmits it to the information processing device 2.

[0060] Furthermore, in the acquisition process B1 after the execution of the event generation process B23, the acquisition unit 251 acquires the selected or created text received by the device communication unit 21.

[0061] (Second generation process B3) After the execution of the event generation process B23 based on the first other concept, the process moves again to the second generation process B3. The second generation process B3 is based on candidate text generated in response to the first other concept. That is, in the second generation process B3, the second generation unit 254 inputs a sentence indicating the causal event, selected by the user or computer from among the candidate texts, or created by the user or computer based on the candidate texts, into the second concept prediction model 244, and obtains at least one output from the second concept prediction model 244 as a candidate for the other concept corresponding to the sentence indicating the causal event.

[0062] In the output control process B4 following the execution of the second generation process B3, the device calculation unit 25 controls the device communication unit 21 according to the processing result of the second generation process B3. As a result, the device communication unit 21 sends data of other concept candidates or data of a web page showing other concept candidates to the terminal device 1. Upon receiving the data, the terminal device 1 outputs (displays) the other concept candidates. The user, after confirming the other concept candidates displayed on the terminal device 1, performs an operation to select one of the other concept candidates, an operation to create another concept based on the content of the other concept candidates, or an operation to instruct the end of the analysis (A9). As a result, the reception unit 11 of the terminal device 1 receives the input of the selected or created other concept or the instruction to end the analysis, and the terminal communication unit 13 transmits it to the information processing device 2.

[0063] If terminal device 1 transmits another selected or created concept, in acquisition process B1 after execution of second generation process B3, the acquisition unit 251 acquires the other selected or created concept received by device communication unit 21. Then, event generation process B23 based on the other concept is executed again. On the other hand, if terminal device transmits an analysis termination instruction (A10:YES), device calculation unit 25 terminates processing.

[0064] (First decision processing B6) If the execution of event generation process B23 is repeated, the process moves to first decision process B6. In first decision process B6, the first decision unit 255 determines whether the causal relationship between the content of a sentence selected from candidate texts generated in one event generation process, or created based on such candidate texts, and the content of a sentence selected from candidate texts generated in an event generation process executed before the one event generation process, or created based on such candidate texts, is logically correct. Specifically, the first decision unit 255 generates a prompt, for example, as shown in Figure 11, in response to input received from the user, and inputs the generated prompt to a trained model (e.g., ChatGPT, Claude, etc.) that determines causal relationships. The first decision unit 255 then obtains the output from the trained model as the decision result. In the first decision process B6 according to this embodiment, the first decision unit 255 outputs the decision result three times for a single user input. This is because the trained model does not necessarily output the same result for the same input. The first determination unit 255 may be configured to output a determination result two or four or more times in response to a single user input.

[0065] In the output control process B4 following the execution of the first judgment process B6, the device calculation unit 25 controls the device communication unit 21 according to the processing result of the first judgment process B6. As a result, the device communication unit 21 transmits the judgment result data, or the data of a web page indicating the judgment result, to the terminal device 1. The terminal device 1, having received the data, outputs (displays) the judgment result (A11). As described above, in the first judgment process S6, the first judgment unit 255 outputs the judgment result three times for a single user input. Therefore, the terminal device 1 displays the three judgment results in a list, as shown in Figure 12. This allows the user to finally confirm whether there are any problems with the causal relationship, even if there is variation in the judgment results. After the user has confirmed the judgment result displayed on the terminal device 1, they perform an operation to correct the sentences they entered that were pointed out as having incorrect causal relationships, if necessary. As a result, the reception unit 11 of the terminal device 1 receives the corrected sentences, and the terminal communication unit 13 transmits them to the information processing device 2.

[0066] If terminal device 1 sends a modified text, in acquisition process B1 after the execution of first decision process B6, the acquisition unit 251 acquires the modified text received by the device communication unit 21. Then, the second generation process B3 based on the modified text is executed again.

[0067] (Second decision processing B7) The second decision process B7 is executed at the appropriate time (for example, when terminal device 1 receives a predetermined processing start instruction from the user). In the second decision process B7, the second decision unit 256 determines whether the candidate of other concepts output by the second concept prediction model 244 for the sentence indicating the causal event was valid. Specifically, the second decision unit 256 generates a prompt, for example, as shown in Figure 13, in response to the input received from the user, and inputs the generated prompt to a trained model that judges the validity of the candidates. The second decision unit 256 then obtains the output of the trained model as the decision result.

[0068] (Update process B8) Update process B8 is a process that is executed at the appropriate time (for example, when terminal device 1 receives input from the user for a predetermined process start instruction). In update process B8, update unit 257 evaluates the quality of the list of conceptual examples stored in second memory unit 23 and updates the list based on the evaluation result. Quality is at least one of the following: comprehensiveness of conceptual examples, clarity of distinction between conceptual examples, and accuracy of representation of each conceptual example. Specifically, update unit 257 generates a prompt in response to input received from the user and inputs the generated prompt into a trained model (at least one of a trained model that evaluates comprehensiveness, a trained model that evaluates clarity of distinction, and a trained model that evaluates accuracy of representation). Then, update unit 257 obtains the output from the trained model as the evaluation result.

[0069] (Registration process B9) Registration process B9 is executed as needed (for example, when terminal device 1 receives input of a new word from the user (A0)). In registration process B9, the registration unit 258 vectorizes the input word and registers it in the dictionary stored in the second storage unit 23. The registration unit 258 creates a group of words that are similar in meaning to the word to be newly registered from among the words that are already registered, and uses the average value of the vector values ​​of the words included in the group as the vector value of the word to be newly registered.

[0070] [Effects of the Information Processing Device 2 (Analysis System 100)] The information processing device 2 (analysis system 100), as described above, generates at least one candidate text indicating a possible cause of a target event when the user inputs the details of the target event, and outputs it as a choice. The user can proceed with the analysis by selecting the most likely cause from the output choices, or by inputting a cause based on the choices. Therefore, with the information processing device 2 (analysis system 100), when analyzing the causes of errors or troubles that occur in the manufacturing process, the analyst can easily deduce the contributing factors.

[0071] [Differentiation] The present invention is not limited to the embodiments described above, and various modifications and / or selections are possible within the scope of the claims.

[0072] For example, in the first generation process B2 according to the above embodiment, the category generation process B21, the concept generation process B22, and the event generation process B23 were executed. However, the concept generation process B22 does not need to be executed in the first generation process B2. In that case, in the event generation process B23, the first generation unit 253 will generate candidate text based on a category selected from the category candidates or a category created based on the category candidates. Specifically, for example, a selected or created category is input to a trained model that is constructed to output candidate text when a category is input, and at least one output from the trained model is obtained as candidate text. Alternatively, only the generation of categories and the generation of candidate text may be repeated (concept generation may not be performed).

[0073] Furthermore, in the first generation process B2, it is not necessary to execute the classification generation process B21 and the concept generation process B22. In this case, in the event generation process B23, the first generation unit 253 will generate at least one candidate text based on the content of the target event. Specifically, for example, a selected or created sentence is input to a trained model that is constructed to output a candidate text when a sentence describing the content of the target event is input, and at least one output from the trained model is obtained as the candidate text.

[0074] Furthermore, in the analysis system 100 according to the above embodiment, the terminal device 1 outputs various candidates generated by the information processing device 2, and the user selects a candidate or creates a document based on the candidates. However, the analysis system 100 may be configured to automatically perform the selection of various candidates or the creation of a document based on the various candidates.

[0075] Furthermore, the analysis system 100 according to the above embodiment was composed of multiple devices (terminal device 1 and information processing device 2). However, the analysis system 100 may be composed of a single device. Specifically, the information processing device 2 may have a configuration corresponding to the receiving unit and output unit of the terminal device 1. In this case, the information processing device 2 does not need to have a device communication unit 21.

[0076] Furthermore, at least two of the first storage unit 22, the second storage unit 23, and the third storage unit 24 may be configured as a single unit. Also, the database 231 and at least one of the various prediction models may be stored in a storage device provided separately from the information processing device 2. In that case, the information processing device 2 does not need to have the second storage unit 23 or the third storage unit 24.

[0077] Furthermore, the program 221 may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the device. In the latter case, the program 221 may be supplied to the device via any wired or wireless transmission medium.

[0078] Furthermore, some or all of the functions of each of the control blocks 251 to 258 described above can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the control blocks 251 to 258 described above are formed is also included in the scope of the present invention.

[0079] <Disclosure 2: Embodiment of Model Construction Method> Next, an embodiment of the model construction method related to Disclosure 2 will be described. For the sake of clarity, components having the same function as those described in the embodiment of Disclosure 1 will be denoted by the same reference numerals, and their descriptions will not be repeated.

[0080] [How to build it] As shown in Figure 14, the model building method includes a model building step C1. The model building method according to this embodiment further includes a vectorization step C2. If a first concept prediction model 242 or a second concept prediction model 244 is to be generated, a text generation step C3 is further included.

[0081] [Text generation step C3] In the text generation step C3, a fictional text is generated that represents an event that occurs relatively infrequently among several events. In the text generation step C3 according to this embodiment, a prompt is input to a trained model constructed to generate fictional texts, instructing it to refer to statistical data of past events or statistical data of concepts that include such events, and output a fictional text corresponding to a concept that is rarely used. The output of the trained model is then used as the fictional text. The means for generating the fictional text may be the information processing device 2 described above, or it may be another device.

[0082] [Vectorization Step C2] In the vectorization step C2, various texts used for training are vectorized. These texts may include fictional texts generated in the text generation step C3. The means for vectorizing the text may be the information processing device 2 or other devices. Vectorization allows for the removal of particles, auxiliary verbs, etc., from the text that are not useful for prediction and often negatively affect the model's performance and computational complexity. Furthermore, since words with similar meanings are converted to similar values, variations in spelling can be handled. In the vectorization step C2 according to this embodiment, various texts used for training are vectorized by weighting them based on the importance of each word contained in each of the texts. The importance of words can be calculated, for example, using the TF-IDF (Term Frequency-Inverse Document Frequency) method.

[0083] [Model Building Step C1] In model building step C1, at least one of the following models is constructed using vectorized text: the classification prediction model 241, the first concept prediction model 242, the event prediction model 243, and the second concept prediction model 244. The classification prediction model 241 is constructed by learning, using the relationship between the content (vector value) of a target event and the classification to which the target event belongs as training data. The event prediction model 243 is constructed by learning, using the relationship between a concept and the events included in that concept as training data. The first concept prediction model 242 is constructed by learning, using the relationship between a classification (vector value) and the concept that caused the event belonging to that classification as training data. The second concept prediction model 244 is constructed by learning, using the relationship between a causal event, which is the cause of an event, and another concept, which is the concept that caused the causal event to occur, as training data. In the case of the first concept prediction model 242 or the second concept prediction model 244, a fictional text and the concept to which the text belongs are further used in the construction. [Explanation of Symbols]

[0084] 100 Analysis Systems 1 Terminal device 11 Reception Department 12 Output section 13 Terminal Communication Unit 14 Terminal Processing Unit 2. Information Processing Device 21. Device Communication Unit 22 First memory section 221 Programs 23 Second memory section 231 Databases 24 Third Memory 241 Classification Prediction Models 242 First Concept Prediction Model 243 Event Prediction Models 244 Second Concept Prediction Model 25 Device calculation section 251 Acquisition Department 252 Output Control Unit 253 First generation part 254 Second generation part 255 First Judgment Department 256 Second Judgment Department 257 Update Department 258 Registration Department C1 Model Building Step C2 Vectorization Step C3 Text Generation Step

Claims

1. A program that causes a computer to perform a first generation process that generates at least one candidate text indicating a possible cause of a target event, based on the content of that event.

2. In the first generation process, The computer is instructed to input the content of the target event into a classification prediction model that is constructed to output the classification to which the target event belongs when the content of the target event is input, and to execute a classification generation process that obtains at least one output from the classification prediction model as a candidate classification for the target event. The candidate text is generated based on a category selected from the candidate categories or a category created based on the candidate categories. The program according to claim 1.

3. In the first generation process, After causing the computer to perform the classification generation process, the computer is then given the selected or created classification as input to a first concept prediction model, which is constructed to output a concept that causes an event belonging to that classification when a classification is input, and the computer is given a concept generation process to obtain at least one output from the first concept prediction model as a candidate concept belonging to the selected or created classification. Based on a concept selected from the candidate concepts or a concept created based on the candidate concepts, the candidate text is generated. The program according to claim 2.

4. In the concept generation process, the first concept prediction model obtains from the list of concept examples the concept examples included in the selected or created category as candidate concepts. The program according to claim 3.

5. In the first generation process, the computer is instructed to input the selected or created concept into an event prediction model, which is constructed to output text indicating a causal event that is the cause of the target event, when a sentence indicating a concept is input to the computer, and to execute an event generation process in which at least one output from the event prediction model is obtained as the candidate text corresponding to the selected or created concept. The program according to claim 3.

6. In the event generation process, the event prediction model obtains, as candidate text, text representing past events included in the selected or created concept from a database containing multiple texts representing past events that have occurred in the past. The program according to claim 5.

7. After the computer has performed the first generation process, if a sentence indicating the causal event is input, the computer is further instructed to input a sentence indicating the causal event, selected from the candidate texts or created based on the candidate texts, into a second concept prediction model, which is constructed to output other concepts that are concepts that are the cause of the causal event, and to obtain at least one output from the second concept prediction model as a candidate for another concept corresponding to the sentence indicating the causal event. The program according to claim 5.

8. If one other concept is selected or created based on the other candidate concepts mentioned above, the computer will: The event generation process based on the other concept mentioned above is executed, The second generation process is executed based on the candidate text generated in correspondence with the other concept of the first, The program according to claim 7.

9. If the execution of the event generation process is repeated, the computer is further instructed to perform a first judgment process that determines whether the causal relationship between the content of a sentence selected from the candidate texts generated in one event generation process, or created based on such candidate texts, and the content of a sentence selected from the candidate texts generated in an event generation process executed before the first event generation process, or created based on such candidate texts, is logically correct. The program according to claim 8.

10. In the aforementioned classification generation process, the computer, The text describing the content of the aforementioned event is vectorized, When calculating the vector values ​​of the aforementioned text, weighting is performed based on the importance of each word input into the classification prediction model. The program according to claim 2.

11. After having the computer perform the second generation process, it is further instructed to perform a second judgment process to determine whether the candidate concepts output by the second concept prediction model for the text representing the causal event were valid. The program according to claim 7.

12. The computer is further instructed to perform an update process that evaluates the quality of the list of conceptual examples and updates the list based on the evaluation results. The program according to claim 4.

13. The aforementioned quality is at least one of the following: comprehensiveness of the conceptual examples, clarity of distinction between each conceptual example, and accuracy of the expression of each conceptual example. The program according to claim 12.

14. A reception unit that accepts input of the details of the target event, A first generation unit generates at least one candidate text indicating a possible cause of the target event based on the content of the target event, An output unit that outputs the aforementioned candidate text as options, Equipped with, The reception unit further accepts input of a sentence that indicates the causal event causing the target event, which is selected from the candidate texts or created based on the candidate texts. Analysis system.

15. This includes the step of constructing a classification prediction model that outputs a classification when a sentence describing the content of a target event is input, by using the relationship between the content of the target event and the classification to which the target event belongs as training data. Model construction method.

16. This includes the step of constructing a first concept prediction model that outputs a concept when a category is input, by learning using the relationship between a category and the concept of the cause of the event belonging to that category as training data. Model construction method.

17. This includes the step of constructing an event prediction model that outputs text representing a causal event when a sentence representing a concept is input, by using the relationship between a concept and the causal event that is the cause of the events contained in that concept as training data. Model construction method.

18. This includes the step of constructing a second concept prediction model that outputs another concept when a sentence describing a causal event is input, by learning the relationship between a causal event, which is the cause of an event, and another concept, which is the concept of the cause of the causal event, as training data. Model construction method.

19. The process further includes the step of vectorizing the text used for learning by weighting each word in the text based on its importance, The step of constructing a model using the vectorized text is included, A model construction method according to any one of claims 15 to 17.

20. The process involves generating a fictional text that describes an event that occurs relatively infrequently among several events, and The further step includes constructing the first conceptual prediction model using the aforementioned fictional text, The model construction method according to claim 16.

Images