Information processing device, information processing method, and program

Abstract

To provide a technology that can generate an easily understandable explanation for the output results of a model. [Solution] An information processing device according to one aspect of the present disclosure has a first creation unit that creates model information including the output results of a model that predicts a target variable from explanatory variables, a second creation unit that creates a prompt that instructs the generation of an explanation for the output results based on the model information and domain knowledge of the task performed by the model, and a creation unit that generates information representing an explanation for the output results based on the prompt and a generation AI realized by a machine learning model including a large-scale language model.

Description

[Technical field]

[0001] The present disclosure relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] Specific tasks are executed using artificial intelligence (AI), machine learning, statistical methods, etc. For example, Patent Literature 1 discloses a technology for detecting anomalies, etc. in a monitored system using a model trained by machine learning.

[0003] On the other hand, in recent years, there are cases where an explanation is required for the output results of a model. For example, users may want to know how the output values ​​of a model that forecasts electricity demand are calculated, or the reasons why the forecast values ​​have increased or decreased from the most recent electricity demand results. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2024-41512 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, it is generally difficult to generate an explanation for the output of a model, and even if an explanation for the output of a model is generated, a user without specialized knowledge may not be able to understand the explanation.

[0006] The present disclosure has been made in consideration of the above points, and aims to provide a technology that can generate an easily understandable explanation for the output results of a model. [Means for solving the problem]

[0007] An information processing device according to one embodiment of the present disclosure has a first creation unit that creates model information including an output result of a model that predicts a target variable from explanatory variables, a second creation unit that creates a prompt that instructs the generation of an explanation for the output result based on the model information and domain knowledge of a task performed by the model, and a generation unit that generates information representing an explanation for the output result based on the prompt and a generation AI realized by a machine learning model including a large-scale language model. Effect of the Invention

[0008] It is possible to generate easy-to-understand explanations for the model output results. [Brief description of the drawings]

[0009] [Figure 1] FIG. 2 is a diagram illustrating an example of a hardware configuration of the model output result explanation device according to the present embodiment. [Diagram 2] 1 is a diagram illustrating an example of a functional configuration of a model output result explanation device according to an embodiment of the present invention. [Diagram 3] 13 is a flowchart illustrating an example of a model output result explanation process according to the embodiment. [Figure 4] FIG. 1 is a diagram showing an example (part 1) of domain knowledge. [Diagram 5] FIG. 13 is a diagram showing an example (part 2) of domain knowledge. [Figure 6] FIG. 13 illustrates an example of a prompt template. [Figure 7] FIG. 13 is a diagram showing an example (part 1) of a prompt. [Figure 8] FIG. 13 is a diagram showing an example (part 2) of a prompt. [Figure 9] FIG. 13 is a diagram showing an example (part 3) of a prompt. [Figure 10] FIG. 13 is a diagram showing an example (part 4) of a prompt. [Figure 11] FIG. 5 is a diagram showing an example of a prompt (part 5). [Figure 12] FIG. 13 is a diagram showing an example (part 1) of model output result explanation information. [Figure 13] FIG. 13 is a diagram showing an example (part 2) of model output result explanation information. [Figure 14] FIG. 13 is a diagram showing an example (part 3) of model output result explanation information. [Figure 15] FIG. 13 is a diagram showing an example (part 4) of model output result explanation information. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] An embodiment of the present invention will be described in detail below with reference to the drawings. In the following embodiment, a model output result explanation device 10 capable of generating an explanation that can be easily understood by a user for the output result of a model that executes some task (e.g., a classification task such as anomaly detection, a regression task such as prediction, etc.) will be described. Note that the task is not limited to a specific task, and the following embodiment can be applied to a model that executes any predetermined task.

[0011] Here, a model refers to a program or the like that has been learned, trained, or created to perform a certain task using artificial intelligence, machine learning, statistical methods, or the like. A model can generally be expressed as y = f(x) using some function f. x and y are generally expressed as x = (x1, , x N ), y=(y1, ,y M ), and each x n are explanatory variables, and each y m are called objective variables, etc. N is the number of explanatory variables, and M is the number of objective variables. Note that data collected from the target of a task (e.g., a plant, a facility, a piece of equipment, etc.) may contain various variables, but the explanatory variables and objective variables are defined or determined, for example, by a user or a model designer.

[0012] Since a model uses multiple explanatory variables as input, even if the same output value is obtained, the reasons for this may be different. For example, in a model in which the predicted value of power demand is the objective variable, various variables such as temperature, time, day of the week, etc. become explanatory variables. Therefore, even if the same predicted value is obtained, there are various patterns, such as high power demand due to high temperature and demand for cooling, high power demand due to low temperature and demand for heating, high power demand due to factories operating during the daytime on weekdays, etc. The number of these patterns is enormous, so it is difficult to prepare explanations for all patterns.

[0013] In addition, for example, when the forecast value or forecast error suddenly increases or decreases compared to the most recent power demand record, it may be necessary to consider the reason. This is because the forecast value changes significantly or the forecast error increases due to some special reason (e.g., abnormal weather, occurrence of a local event, etc.). In such cases, it is common to consider the cause by comparing and examining it with a normal time point. However, even if a highly explanatory model (e.g., multiple regression model, decision tree, etc.) is used, the user cannot obtain a clear cause unless he or she has both the domain knowledge of the task and the knowledge about the model. In other words, in this case, the user is required to have highly specialized knowledge about both the domain knowledge of the task and the knowledge about the model. For this reason, even if a general explanation of the output result of the model (e.g., a graph visualizing the output result of the model, etc.) is obtained, a user who does not have highly specialized knowledge about both the domain knowledge of the task and the knowledge about the model cannot fully interpret the output result of the model. Note that domain knowledge refers to specialized knowledge in the task or the industry or field that the task targets.

[0014] Therefore, in the model output result explanation device 10 according to the present embodiment, an explanation for the output result of the model (hereinafter also referred to as "target model") is generated using domain knowledge of the task executed by the model and a generation AI. This provides an explanation that can be easily understood even by a user who does not have specialized knowledge of both the target model and the domain knowledge of the task executed by the target model. Hereinafter, the target model will be represented by a function f with input x and output y, and the target model will be represented by f. Note that the model output result explanation device 10 according to the present embodiment is realized, for example, by an information processing device (computer) or an information processing system (computer system) composed of one or more information processing devices. Specific examples of information processing devices include personal computers, general-purpose servers, smartphones, tablet terminals, etc.

[0015] Here, generative AI refers to a technology or mechanism that generates content such as natural language sentences or images that satisfy given instructions or questions (these instructions or questions are called "prompts") Generative AI is generally realized by machine learning models including, for example, large language models (LLMs).

[0016] <Hardware configuration example of the model output result explanation device 10> An example of the hardware configuration of a model output result explanation device 10 according to this embodiment is shown in Fig. 1. As shown in Fig. 1, the model output result explanation device 10 according to this embodiment has an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these pieces of hardware is connected to each other via a bus 109 so as to be able to communicate with each other.

[0017] The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the model output result explanation device 10 does not necessarily have to have at least one of the input device 101 and the display device 102, for example.

[0018] The external I / F 103 is an interface with an external device such as a recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

[0019] The communication I / F 104 is an interface for connecting to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory. The processor 108 is an arithmetic device such as a central processing unit (CPU) or a graphic processing unit (GPU).

[0020] 1 is merely an example, and the hardware configuration of the model output result explanation device 10 is not limited to this. For example, the model output result explanation device 10 may have multiple auxiliary storage devices 107 and multiple processors 108, may not have some of the hardware shown in the figure, or may have various types of hardware other than the hardware shown in the figure.

[0021] <Example of functional configuration of model output result explanation device 10> An example of a functional configuration of the model output result explanation device 10 according to this embodiment is shown in Fig. 2. As shown in Fig. 2, the model output result explanation device 10 according to this embodiment has a data acquisition unit 201, a model learning unit 202, a model output unit 203, a model information creation unit 204, a domain knowledge acquisition unit 205, a prompt creation unit 206, an explanation generation unit 207, an explanation output unit 208, a user input acquisition unit 209, a data storage unit 301, a model storage unit 302, a domain knowledge storage unit 303, and a template storage unit 304.

[0022] The data acquisition unit 201, the model learning unit 202, the model output unit 203, the model information creation unit 204, the domain knowledge acquisition unit 205, the prompt creation unit 206, the explanation generation unit 207, and the explanation output unit 208 are realized, for example, by a process in which one or more programs installed in the model output result explanation device 10 are executed by the processor 108 or the like. Note that at least one of these units may be realized, for example, by a process executed by an external cloud server or the like.

[0023] The data storage unit 301, the model storage unit 302, the domain knowledge storage unit 303, and the template storage unit 304 are realized, for example, by a storage area of ​​the auxiliary storage device 107. At least one of these units may be realized by a storage area of ​​a storage device (e.g., a storage device provided in a database server) communicably connected to the model output result explanation device 10.

[0024] When learning a target model f, the data acquiring unit 201 acquires data for model learning (hereinafter also referred to as "learning data") from the data storage unit 301. When making a prediction using the target model f, the data acquiring unit 201 acquires data to be predicted (hereinafter also referred to as "prediction target data") from the data storage unit 301.

[0025] The learning data is represented by, for example, a pair (x, y) of data x consisting of explanatory variables and data y consisting of objective variables. On the other hand, the prediction target data is represented by, for example, data x consisting of explanatory variables. The data x is data that is input to the prediction model f, and is therefore called input data, etc. The data y is data that represents the correct answer of the output result when the input data x is input to the prediction model f, and is therefore called correct answer data or teacher data, etc.

[0026] If necessary, the data acquisition unit 201 may process or convert the input data x into a format that can be input to the target model f by performing preprocessing on the input data x.

[0027] The model learning unit 202 uses the learning data to learn, train, or create a target model f. That is, the model learning unit 202 learns the relationship between explanatory variables and objective variables using, for example, a known supervised learning method, thereby learning a target model f that predicts the value of an objective function from explanatory variables.

[0028] In addition, the model learning unit 202 stores the learned, trained, or created target model f in the model storage unit 302.

[0029] Here, since it is necessary to have the generating AI explain the output result of the target model f, a white-box AI, also known as an explainable AI, is used as the target model f. Specific examples of white-box AI include multiple regression models, decision trees, logistic regression, rule bases, and other models or algorithms that can present prediction grounds.

[0030] However, a black-box AI may be used as the target model f by also using a method or algorithm that explains the basis for prediction, such as SHAP (SHapley Additive exPlanations). A specific example of a black-box AI is a neural network.

[0031] The model output unit 203 outputs a prediction result (i.e., the output result of the target model f) using the prediction target data and the target model f stored in the model storage unit 302. That is, the model output unit 203 inputs the prediction target data x to the target model f, and outputs the prediction result y=f(x) of the target model f as the output result of the target model f.

[0032] The model information creation unit 204 creates model information including the explanatory variable names, objective variable names, model name, parameters, and output results output by the model output unit 203 of the target model f. Note that the model information creation unit 204 may create model information that further includes, for example, a description of the algorithm itself that realizes the target model f. Here, the model name refers to the name of the type of the target model f or the type of algorithm that realizes the target model f.

[0033] The domain knowledge acquisition unit 205 acquires domain knowledge of a task executed by the target model f from the domain knowledge storage unit 303. The domain knowledge includes, for example, explanations in natural language sentences regarding explanatory variables and objective variables of the target model f. The domain knowledge is, for example, created in advance by a user or the like.

[0034] The prompt creation unit 206 creates a prompt that instructs the generation of an explanation for the output result of the target model f, using the model information created by the model information creation unit 204, the domain knowledge acquired by the domain knowledge acquisition unit 205, and the prompt template stored in the template storage unit 304. The prompt creation unit 206 creates a prompt that instructs the generation of an explanation for the output result of the target model f, for example, by setting the explanatory variable name, objective variable name, model name, parameters, and output result included in the model information, and the domain knowledge in the prompt template. A prompt template is a model (template) of a prompt. The prompt template is created in advance, for example, by a user or the like.

[0035] Note that the prompt generator 206 may generate a prompt by using, for example, performance data stored in the data storage unit 301 in addition to the model information and domain knowledge. Performance data refers to data that indicates the operational performance of the target of the task (e.g., a plant, a facility, a device, etc.). Like the learning data, the performance data is represented by, for example, a pair (x, y) of data x consisting of explanatory variables and data y consisting of objective variables.

[0036] The explanation generating unit 207 generates model output result explanation information representing an explanation for the output result of the target model f, using the prompt created by the prompt creating unit 206 and the generating AI. That is, the explanation generating unit 207 generates model output result explanation information by the generating AI by inputting the prompt created by the prompt creating unit 206 to the generating AI. The explanation generating unit 207 may use a generating AI provided as an external service, may use a generating AI possessed by the model output result explaining device 10, or may use other generating AI. Note that since the large-scale language model realizing the generating AI also acquires various domain knowledge during learning, when the model output result explanation information is generated by the generating AI, it is expected that the model output result explanation information will be generated using domain knowledge that combines the domain knowledge acquired by the domain knowledge acquiring unit 205 and the domain knowledge acquired by the large-scale language model realizing the generating AI during learning.

[0037] Furthermore, the explanation generating unit 207 generates new model output result explanation information by the generation AI by inputting user input information, which will be described later, as a prompt to the generation AI. This new model output result explanation information is, for example, information that represents an additional explanation to the model output result explanation information previously generated by the generation AI, information that represents a summary of the model output result explanation information previously generated by the generation AI, information that represents a response to some question or instruction regarding the model output result explanation information previously generated by the generation AI, etc.

[0038] The explanation output unit 208 outputs the model output result explanation information generated by the explanation generation unit 207 to a predetermined output destination. Examples of the predetermined output destination include the display device 102 such as a display, a storage area such as the auxiliary storage device 107, other devices or equipment connected in a communicable manner, and the like.

[0039] The user input acquisition unit 209 inputs information such as a natural language sentence input by a user (hereinafter, also referred to as "user input information"). Examples of the user input information include information expressing a natural language sentence requesting additional explanation for the model output result explanation information, information expressing a natural language sentence requesting a summary of the model output result explanation information, information expressing a natural language sentence such as a standard question for the model output result explanation information, and information expressing a natural language sentence meaning some kind of question or instruction for other model output result explanation information (e.g., a question or instruction requesting a comparison between actual values ​​and predicted values ​​obtained at a later date, a question or instruction requesting an explanation of the reason for the error between the actual values ​​and predicted values, etc.).

[0040] The data storage unit 301 stores various data (learning data, prediction target data, performance data, etc.) Note that the various data are collected from targets of a task executed by the target model f.

[0041] The model storage unit 302 stores the target model f. That is, the model storage unit 302 stores, for example, learned parameters of the target model f.

[0042] The domain knowledge storage unit 303 stores domain knowledge of the tasks executed by the target model f.

[0043] The template storage unit 304 stores prompt templates that represent prompt templates that are input to the generation AI.

[0044] <Example of model output result explanation process> An example of the model output result explanation process according to this embodiment will be described with reference to Fig. 3. For example, steps S101 to S102 in Fig. 3 are executed in advance, steps S103 to S109 are executed each time model output result explanation information is needed, and steps S110 to S112 are repeatedly executed one or more times as needed after model output result explanation information is obtained.

[0045] The data acquiring unit 201 acquires learning data (x, y) from the data storage unit 301 (step S101). At this time, the data acquiring unit 201 acquires a plurality of pieces of learning data (x, y) from the data storage unit 301.

[0046] The model learning unit 202 learns, trains, or creates a target model f using the learning data (x, y) acquired in step S101 above (step S102). That is, the model learning unit 202 updates the parameters of the target model f by using, for example, a known supervised learning method or the like, so as to minimize the error between a prediction result f(x) when input data x is input to the target model f and the correct answer data y corresponding to the input data x. In this way, the target model f is learned, trained, or created. The learned, trained, or created target model f is stored in the model storage unit 302.

[0047] The data acquisition unit 201 acquires prediction target data x from the data storage unit 301 (step S103).

[0048] The model output unit 203 uses the prediction target data x acquired in step S103 above and the target model f stored in the model storage unit 302 to output the prediction result y=f(x) of the target model f as the output result of the target model f (step S104).

[0049] The model information creation unit 204 creates model information using the output results output in the above step S104, etc. (step S105). That is, the model information creation unit 204 creates model information including explanatory variable names, objective variable names, model name, parameters, and output results output by the model output unit 203, etc., of the target model f. As an example, when the target model f is a multiple regression model, the model information includes explanatory variable names of the multiple regression model, objective variable names of the multiple regression model, model name "multiple regression model", multiple regression coefficients and y-intercepts, output results y=f(x), etc.

[0050] The domain knowledge acquisition unit 205 acquires domain knowledge of the task executed by the target model f from the domain knowledge storage unit 303 (step S106). Note that the domain knowledge acquisition unit 205 may acquire domain knowledge input by a user.

[0051] <Examples of domain knowledge> Example of domain knowledge (part 1) An example (part 1) of domain knowledge is shown in FIG. 4. The domain knowledge 1100 shown in FIG. 4 includes an implementation task and a data description. The implementation task is a description of the task executed by the target model f. The data description is a description of the data that is the input and output of the target model f. In the example shown in FIG. 4, the implementation task is "power demand forecast by a power company," and the data description describes the units, possible values, etc. of each variable (objective variable, explanatory variable) included in the data.

[0052] In the example shown in FIG. 4, the electricity demand is the objective variable, and the maximum temperature, minimum temperature, minimum humidity, amount of solar radiation, weekday, Saturday, Sunday, and public holiday are the explanatory variables.

[0053] Example of domain knowledge (part 2) An example (part 2) of domain knowledge is shown in Fig. 5. The domain knowledge 1200 shown in Fig. 5 includes an implementation task, a data explanation, and specialized knowledge. That is, the domain knowledge 1200 shown in Fig. 5 includes specialized knowledge in addition to the domain knowledge 1100 shown in Fig. 4. The specialized knowledge is specialized knowledge related to the implementation task (e.g., knowledge related to the characteristics of data used in the implementation task, etc.). In the example shown in Fig. 5, the specialized knowledge includes knowledge related to power demand forecasting, such as "If there is demand for heating and cooling, power demand will increase" and "Power demand decreases on weekends and holidays."

[0054] 4 and 5 are merely examples. In addition to the performed tasks and data descriptions, the domain knowledge includes various information written in natural language sentences according to the performed tasks.

[0055] Returning to the description of Fig. 3, following step S106, the prompt creation unit 206 creates a prompt that instructs the generation of an explanation for the output result of the target model f, using the model information created in step S105, the domain knowledge acquired in step S106, and the prompt template stored in the template storage unit 304 (step S107). Note that the prompt creation unit 206 may create a prompt using, for example, performance data stored in the data storage unit 301 in addition to the model information and domain knowledge. At this time, the prompt creation unit 206 may also calculate feature quantities (e.g., statistics such as the average and variance for each variable) of the performance data, and then create a prompt using these feature quantities.

[0056] <<Example of a prompt template>> An example of a prompt template is shown in Fig. 6. The prompt template 2000 shown in Fig. 6 is a prompt template for generating model output result explanation information that explains the validity of the prediction result by the target model f. The prompt template 2000 shown in Fig. 6 includes variable parts, and each piece of information contained in the model information, domain knowledge, performance data, etc. is set in each variable part. By setting each piece of information contained in the model information, domain knowledge, performance data, etc. in these variable parts, a prompt that instructs the generation of an explanation for the output result of the target model f is created.

[0057] For example, the variable portion [Performance Task] included in the prompt template 2000 shown in FIG. 6 is set with the performance task included in the domain knowledge. Similarly, for example, the variable portion [Model Name] included in the prompt template 2000 shown in FIG. 6 is set with the model name included in the model information. Similarly, for example, the variable portion [Objective Variable] included in the prompt template 2000 shown in FIG. 6 is set with the objective variable name included in the model information. Similarly, for example, the variable portion [Explanatory Variable] included in the prompt template 2000 shown in FIG. 6 is set with the explanatory variable name included in the model information. Similarly, for example, the variable portion [Data Description] included in the prompt template 2000 shown in FIG. 6 is set with the data description included in the domain knowledge.

[0058] 6 is just an example. Various prompt templates can be used as long as they are prompt templates that instruct some kind of explanation for the output result of the target model f (that is, prompts that include a variable portion). Hereinafter, information set in the variable portion included in the prompt template will also be referred to as "setting information."

[0059] <<Example of a prompt>> Prompt example (part 1) An example (part 1) of a prompt is shown in Fig. 7. Prompt 3100 shown in Fig. 7 is a prompt created by prompt creation unit 206 using domain knowledge 1100 shown in Fig. 4, prompt template 2000 shown in Fig. 6, and model information in the case where target model f is a certain multiple regression model. Here, in prompt 3100 shown in Fig. 7, the parts written in bold and italics are setting information.

[0060] Prompt example (part 2) An example (part 2) of a prompt is shown in Fig. 8. Prompt 3200 shown in Fig. 8 is a prompt created by prompt creation unit 206 using domain knowledge 1100 shown in Fig. 4, a certain prompt template, model information in the case where target model f is a certain multiple regression model, and performance data. Here, in prompt 3200 shown in Fig. 8, the parts written in bold and italics are setting information.

[0061] Prompt example (part 3) An example (part 3) of a prompt is shown in Fig. 9. Prompt 3300 shown in Fig. 9 is a prompt created by prompt creation unit 206 using certain domain knowledge including the execution task "state prediction by sensor", a certain prompt template, and model information in which the target model f is a certain decision tree. Here, in prompt 3300 shown in Fig. 9, the parts written in bold and italics are setting information.

[0062] Prompt example (part 4) An example (No. 4) of a prompt is shown in Fig. 10. Prompt 3400 shown in Fig. 10 is a prompt created by prompt creation unit 206 using certain domain knowledge including the execution task "product quality prediction", a certain prompt template, and model information in the case where target model f is a certain black-box AI that also uses SHAP. Here, in prompt 3400 shown in Fig. 10, the parts written in bold and italics are setting information.

[0063] Prompt example (part 5) FIG. 11 shows an example of a prompt (part 5). The prompt 3500 shown in FIG. 11 is a prompt created by the prompt creation unit 206 using the domain knowledge 1100 shown in FIG. 4, a certain prompt template, model information in the case where the target model f is a certain multiple regression model, and performance data. Here, in the prompt 3500 shown in FIG. 11, the part written in bold and italics is the setting information. In the example shown in FIG. 11, it is written, "Please add your opinion based on general knowledge of power demand" so that the model output result information is generated using the domain knowledge previously acquired by the large-scale language model that realizes the generation AI. This makes it possible for the generation AI to take into account the specialized knowledge contained in the domain knowledge previously acquired by the large-scale language model that realizes the generation AI when generating the model output result explanation information. In addition, by using the domain knowledge 1200 shown in FIG. 5 instead of the domain knowledge 1100 shown in FIG. 4, it is expected that model output result information with more specialized content will be generated.

[0064] Returning to the description of Fig. 3, following step S107, the explanation generating unit 207 uses the prompt created in step S107 and the generating AI to generate model output result explanation information that represents an explanation for the output result of the target model f (step S108). That is, the explanation generating unit 207 inputs the prompt created in step S107 to the generating AI, thereby generating model output result explanation information by the generating AI.

[0065] The explanation output unit 208 outputs the model output result explanation information generated in the above step S108 to a predetermined output destination (step S109).

[0066] <Example of model output result explanation information> Example of model output result explanation information (part 1) An example (part 1) of the model output result explanation information is shown in Fig. 12. The model output result explanation information 4100 shown in Fig. 12 is model output result explanation information generated by inputting the prompt 3200 shown in Fig. 8 to the generation AI. By referring to the model output result explanation information 4100 shown in Fig. 12, the user can easily understand the explanation for the output result of the target model f, which is a multiple regression model, in a form compared with past performance data.

[0067] Example of model output result explanation information (part 2) An example (part 2) of the model output result explanation information is shown in Fig. 13. The model output result explanation information 4200 shown in Fig. 13 is model output result explanation information generated by inputting the prompt 3300 shown in Fig. 9 to the generation AI. By referring to the model output result explanation information 4200 shown in Fig. 13, the user can easily understand the explanation for the output result of the target model f, which is a decision tree.

[0068] Example of model output result explanation information (part 3) An example (part 3) of the model output result explanation information is shown in Fig. 14. The model output result explanation information 4300 shown in Fig. 14 is model output result explanation information generated by inputting the prompt 3400 shown in Fig. 10 to the generation AI. By referring to the model output result explanation information 4300 shown in Fig. 14, the user can easily understand the explanation for the output result of the target model f, which is a certain black-box type AI.

[0069] Example of model output result explanation information (part 4) An example (part 4) of the model output result explanation information is shown in Fig. 15. The model output result explanation information 4400 shown in Fig. 15 is model output result explanation information generated by inputting the prompt 3500 shown in Fig. 11 to the generation AI. By referring to the model output result explanation information 4400 shown in Fig. 15, the user can easily understand the explanation for the output result of the target model f, which is a multiple regression model.

[0070] Returning to the description of Fig. 3, following step S109 above, the user input acquisition unit 209 acquires user input information input by the user (step S110). Note that the user input acquisition unit 209 may, for example, acquire performance data stored in the data storage unit 301, and then acquire user input information including this performance data. At this time, the user input acquisition unit 209 may also calculate features of the performance data (e.g., statistics such as average and variance for each variable), and then acquire user input information including these features.

[0071] The explanation generating unit 207 generates new model output result explanation information using the user input information acquired in step S110 and the generating AI (step S111). Note that this new model output result explanation information may be, for example, information representing an additional explanation to the model output result explanation information previously generated by the generating AI, information representing a summary of the model output result explanation information previously generated by the generating AI, information representing a response to some question or instruction to the model output result explanation information previously generated by the generating AI, etc., depending on the user input information.

[0072] The explanation output unit 208 outputs the new model output result explanation information generated in the above step S111 to a predetermined output destination (step S112).

[0073] <Modification> The prompt and its template are not limited to natural language sentences, and may include content such as images, videos, sounds, etc. Furthermore, they may include mathematical expressions, program codes, etc.

[0074] In particular, when instructing the AI ​​to perform some kind of numerical calculation or processing, the prompt may contain the formula or program code used in the numerical calculation or processing, which makes it possible to reduce calculation errors when the AI ​​performs numerical calculations.

[0075] <Summary> As described above, the model output result explanation device 10 according to the present embodiment uses domain knowledge of the task executed by the target model f and the generation AI to generate and output an explanation for the output result of the target model f. As a result, the model output result explanation device 10 according to the present embodiment can achieve the following effects.

[0076] (1) Domain knowledge is input into the generative AI, so that an explanation for the output results of the target model f can be generated based on the domain knowledge.

[0077] (2) By storing domain knowledge for each target model f in the domain knowledge storage unit 303 or by the user inputting domain knowledge, it is possible to use the model output result explanation device 10 for a variety of tasks without changing the configuration of the model output result explanation device 10.

[0078] (3) Since model output result explanation information, which represents an explanation for the output result of the target model f, is generated by a generative AI realized by a large-scale language model, etc., it becomes possible to obtain an explanation that can be easily interpreted by the user.

[0079] (4) Even if a user is unable to understand the model output result explanation information, new model output result explanation information can be obtained by providing a question or other information as user input information, thereby enabling the user to deepen their understanding of the model output result explanation information or resolve their doubts.

[0080] (5) The model output result explanation information explains the basis for the output result of the target model f. Therefore, even a user who does not understand the algorithm of the target model f can use a system incorporating the target model f with confidence.

[0081] The present invention is not limited to the above-described embodiments specifically disclosed, and various modifications, changes, combinations with known technologies, etc. are possible without departing from the scope of the claims. [Explanation of symbols]

[0082] 10. Model output result explanation device 101 Input Device 102 Display device 103 External I / F 103a Recording media 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage device 108 processors 109 Bus 201 Data Acquisition Department 202 Model Learning Department 203 Model Output Section 204 Model Information Creation Department 205 Domain Knowledge Acquisition Department 206 Prompt Creation Department 207 Explanation Generation Unit 208 Description output section 209 User Input Acquisition Unit 301 Data storage unit 302 Model Memory Section 303 Domain Knowledge Storage Unit 304 Template storage section

Claims

1. a first creation unit that creates model information including an output result of a model that predicts a target variable from an explanatory variable; a second generator for generating a prompt for instructing generation of an explanation for the output result based on the model information and domain knowledge of a task performed by the model; A generation unit that generates information representing an explanation for the output result based on the prompt and a generation AI realized by a machine learning model including a large-scale language model; having The second creation unit is an information processing device that creates a prompt that instructs an explanation for a result of comparing the output result with a performance value represented by the performance data, based on the model information, the domain knowledge, and performance data acquired from a target of the task.

2. The second creation unit is The information processing apparatus according to claim 1 , wherein the prompt is created by setting each piece of information included in the model information, the domain knowledge, and the performance value for a template created in advance.

3. 3. The information processing device according to claim 2, wherein the template includes a natural language sentence instructing generation of an explanation for a result of comparing the output result with the performance value, and a variable portion in which each piece of information included in the model information, the domain knowledge, and the performance value are respectively set.

4. The information processing device according to claim 3 , wherein the template includes at least one of a formula used for a numerical calculation performed by the generating AI and a program code for a predetermined process performed by the generating AI.

5. 5. The information processing device according to claim 2, wherein the model information includes the output result, a name of the explanatory variable, a name of the target variable, and a name of the type of the model or a name of a type of algorithm that realizes the model.

6. an acquisition unit that acquires user input information representing an instruction for summarizing an explanation for the output result, or an instruction for additional explanation for the explanation for the output result, a question, or a predetermined instruction; The generation unit is The information processing device of claim 1 , which generates information representing a summary of an explanation for the output result, or information representing the additional explanation, information representing an answer to the question, or information satisfying the specified instructions, based on the user input information and the generation AI.

7. The information processing apparatus according to claim 1 , wherein the domain knowledge includes a natural language sentence that describes the task, and a natural language sentence that describes input / output data of the model.

8. The information processing apparatus according to claim 7 , wherein the domain knowledge includes a natural language sentence that describes a characteristic of the input / output data.

9. A first generation step of generating model information including an output result of a model predicting a target variable from an explanatory variable; a second generation step of generating a prompt for generating an explanation for the output result based on the model information and domain knowledge of the task performed by the model; A generation step of generating information representing an explanation for the output result based on the prompt and a generation AI realized by a machine learning model including a large-scale language model; The computer executes The second creation procedure includes: An information processing method, comprising: creating a prompt that instructs an explanation for a result of comparing the output result with a performance value represented by the performance data, based on the model information, the domain knowledge, and performance data obtained from a subject of the task.

10. A first generation step of generating model information including an output result of a model predicting a target variable from an explanatory variable; a second generation step of generating a prompt for generating an explanation for the output result based on the model information and domain knowledge of the task performed by the model; A generation step of generating information representing an explanation for the output result based on the prompt and a generation AI realized by a machine learning model including a large-scale language model; Run the following on your computer: The second creation procedure includes: a program that creates a prompt that instructs an explanation for a result of comparing the output result with a performance value represented by the performance data, based on the model information, the domain knowledge, and performance data obtained from a target of the task.

Images