Information processing program, information processing method, and information processing device

Abstract

This information processing device: acquires a designated task; causes a language model to execute a planning process on sub-tasks obtained by dividing the acquired designated task so that the content of the designated task is described in natural language documents; determines, on the basis of the result of executing the planning process, whether or not the natural language documents generated in the sub-tasks satisfy a prescribed condition for the designated task; and outputs the sub-tasks when the determination result is that the prescribed condition for the designated task is satisfied.

Description

Information processing program, information processing method, and information processing apparatus

[0001] The present invention relates to an information processing program and the like.

[0002] When a task is complex, there are systems that complete the original task by dividing the original task into a plurality of subtasks and executing each of them. In this system, an LLM (Large Language Model) or the like is used. In the following description, the original task before being divided into a plurality of subtasks is referred to as the original task.

[0003] FIG. 14 is a diagram for explaining a conventional technique. As shown in FIG. 14, when the conventional system 1 receives an input of an original task 2 from a user U, the conventional system 1 divides the original task 2 into a plurality of subtasks. FIG. 14 shows an example in which the original task 2 is divided into subtasks 2-1, 2-2, ..., 2-N. N is a predetermined natural number.

[0004] The conventional system 1 obtains a final result 3-N by sequentially executing subtasks 2-1 to 2-N. For example, the conventional system 1 executes subtask 2-2 using a result 3-1 obtained by executing subtask 2-1, and similarly executes subsequent other subtasks using a result 3-2 obtained by executing subtask 2-2. In the conventional system 1, the result of the last subtask 2-N is obtained as the final result 3-N.

[0005] Here, the conventional system 1 uses an LLM to determine whether or not the final result 3-N of subtasks 2-1 to 2-N is a result that conforms to the original task 2. When the LLM determines that the final result 3-N is a result that conforms to the original task 2, the conventional system 1 presents the final result 3-N to the user U.

[0006] Japanese Unexamined Patent Application Publication No. 2023-011071

[0007] However, with the above-described conventional technology, it is not possible to accurately determine whether an execution result obtained by a plurality of subtasks is an execution result that conforms to the original task, and therefore, a task specified by a user cannot be accurately executed.

[0008] For example, in the conventional system 1 explained in Figure 14, the LLM determines whether the final result 3-N of subtasks 2-1 to 2-N is consistent with the original task 2. However, the determination result by the LLM can change probabilistically, so there is room for improvement.

[0009] In one aspect, the present invention aims to provide an information processing program, an information processing method, and an information processing device that can accurately perform a specified task.

[0010] In the first proposal, the computer is instructed to perform the following processes: The computer obtains a specified task, divides the obtained specified task into subtasks, and then uses a language model to perform a planning process to describe the content of the specified task in a natural language document. Based on the results of the planning process, the computer determines whether the natural language document generated in the subtask satisfies the predetermined conditions of the specified task, and if the determination is correct and the predetermined conditions of the specified task are met, the computer outputs the subtask.

[0011] It can perform the specified task with high accuracy.

[0012] Figure 1 is a flowchart showing the processing procedure of a conventional system. Figure 2 is a diagram showing an example of the data structure of sales data. Figure 3 is a diagram for supplementary explanation of the prior art. Figure 4 is a diagram showing an example of code generated by a conventional subtask. Figure 5 is a diagram showing an example of a graph generated by a conventional system. Figure 6 is a diagram for explaining the processing of the information processing device according to this embodiment. Figure 7 is a diagram (1) showing an example of a prompt. Figure 8 is a diagram showing an example of code generated by a subtask of this embodiment. Figure 9 is a diagram showing an example of a graph generated by the information processing device according to this embodiment. Figure 10 is a diagram (2) showing an example of a prompt. Figure 11 is a functional block diagram showing the configuration of the information processing device according to this embodiment. Figure 12 is a flowchart showing the processing procedure of the information processing device according to this embodiment. Figure 13 is a diagram showing an example of a computer hardware configuration that realizes the same functions as the information processing device of this embodiment. Figure 14 is a diagram for explaining the prior art.

[0013] The following describes in detail, with reference to the drawings, embodiments of the information processing program, information processing method, and information processing apparatus disclosed in this application. However, this invention is not limited to these embodiments.

[0014] Before describing this embodiment, we will provide a supplementary explanation of the prior art. Figure 1 is a flowchart showing the processing procedure of a conventional system. As with Figure 14, the system of the prior art will be referred to as Conventional System 1. In addition, the original task before it is divided into multiple subtasks may be referred to as the original task.

[0015] As shown in Figure 1, the conventional system 1 receives instructions for the original task from the user (step S10). The conventional system 1 divides the original task into multiple subtasks (step S11).

[0016] Conventional System 1 executes multiple subtasks (step S12). The LLM of Conventional System 1 determines whether the execution results of the multiple subtasks are consistent with the original task (step S13). If the results are not consistent with the original task (step S14, No), Conventional System 1 proceeds to step S11.

[0017] On the other hand, if the conventional system 1 produces a result consistent with the original task (step S14, Yes), it presents the task execution result to the user (step S15).

[0018] Next, we will explain an example of how the system would process a task given to it in the conventional system 1. As a prerequisite, we assume that a sales dataset exists at the cake shop and that marketing analysis is being performed.

[0019] The sales dataset contains multiple sales data entries. Figure 2 shows an example of the data structure of the sales data. As shown in Figure 2, the sales data includes the sales date, advertising medium, customer attributes, product, quantity, and price. The advertising medium is the advertising medium that prompted the customer to purchase the product. Possible values ​​for customer attributes include seniors, families, young people, etc.

[0020] Figure 3 is a diagram to supplement the explanation of the conventional technology. For example, the conventional system 1 receives input from the user for the original task 11. The original task 11 contains the following: "Objective: Find out which advertising medium is most effective for which customer attributes," and "Instructions: Visualize the sales of each advertising medium by customer attribute in a graph."

[0021] Conventional System 1 divides the original task 11 into subtasks 11-1, 11-2, and 11-3. For example, subtask 11-1 is to "generate Python code to visualize sales by customer attribute for each advertising medium using a bar graph." Subtask 11-2 is to "execute the generated Python code and plot the graph." Subtask 11-3 is to "return the file path of the plotted graph."

[0022] Conventional system 1 generates code 12 (Python code) by executing subtask 11-1. Figure 4 shows an example of code generated by a conventional subtask. As shown in Figure 4, for example, the code 12 generated by subtask 11-1 is text (a document in natural language) written to describe the operation of the program. Code 12 includes code fragments 12-1, 12-2, 12-3, 12-4, and 12-5.

[0023] Next, the conventional system 1 executes subtask 11-2. For example, executing subtask 11-2 means executing code 12 to plot the graph. The conventional system 1 then uses the result of executing subtask 11-2 to execute subtask 11-3.

[0024] Conventional system 1 generates the final result 13 by executing subtasks 11-1, 11-2, and 11-3. For example, the final result 13 is a graph that visualizes the sales dataset, as explained in Figure 5.

[0025] Figure 5 shows an example of a graph generated by a conventional system. In graph G1 shown in Figure 5, the vertical axis corresponds to sales, and the horizontal axis corresponds to advertising media. As shown in Figure 5, graph G1 shows SNS, apps, and direct mail as advertising media, and visualizes sales corresponding to customer attributes (seniors, families, young people).

[0026] Here, comparing graph G1 generated by the conventional system 1 with the contents of the original task 11, graph G1 does not include the information regarding "most efficient" as stated in the objective of the original task 11. This is because the information regarding "most efficient" as stated in the objective of the original task 11 has been lost from subtask 11-1.

[0027] In the conventional system 1, as explained in step S13 of Figure 1, the LLM determines whether the execution results of multiple subtasks are consistent with the original task, but the determination result by the LLM can change probabilistically. In other words, the conventional technology cannot accurately determine whether the execution results of multiple subtasks are consistent with the original task, and therefore cannot obtain results that meet the user's task objectives.

[0028] The above provides supplementary information regarding conventional technology.

[0029] Next, the information processing device according to this embodiment will be described. The information processing device according to this embodiment will be referred to as the information processing device 100. The information processing device 100 divides the original task into multiple subtasks, and if a subtask is a subtask that generates code, it instructs the LLM to add comments to the code fragments. In particular, the information processing device 100 instructs the LLM to write the purpose of the original task as is in the comment line for the code fragment that deals with the purpose of the original task.

[0030] The information processing device 100 searches whether the purpose of the original task matches the text of a comment line in any of the code snippets generated by the subtasks. If the purpose of the original task does not match the text of a comment line in any of the code snippets generated by the subtasks, the information processing device 100 executes the process of splitting the original task into multiple tasks again. At this time, the information processing device 100 also notifies the LLM of the subtasks that have been generated and notifies the LLM that the purpose of the original task is not included in the code snippets generated by the subtasks.

[0031] The information processing device 100 presents the execution results of multiple subtasks to the user if the purpose of the original task matches a comment line in any of the code snippets generated by the subtasks.

[0032] As described above, the information processing device 100 determines whether to further divide the original task into subtasks based on whether the purpose of the original task matches the text of a comment line in any of the code snippets generated by the subtasks. Whether the purpose matches the comment line of the code snippet is determined by string matching and can be determined mechanically. Therefore, it is possible to reliably determine whether the purpose of the original task matches the text of a comment line in any of the code snippets generated by the subtasks, and to execute the task specified by the user with high accuracy.

[0033] Figure 6 is a diagram illustrating the processing of the information processing device according to this embodiment. For example, the information processing device 100 receives input of the original task 20 from the user. The original task 20 contains the following: "Objective: Find out which advertising medium is most effective for which customer attributes" and "Instructions: Visualize the sales of each advertising medium by customer attribute in a graph."

[0034] The information processing device 100 uses LLM to divide the original task 20 into multiple subtasks. For example, the information processing device 100 inputs the prompt 30 shown in Figure 7 into LLM.

[0035] Figure 7 is Figure (1) showing an example of a prompt. As shown in Figure 7, area 30-1 of prompt 30 contains the following text: "Divide the given task into multiple subtasks in a logical flow. Each subtask should be independently executable and ultimately accomplish the entire task."

[0036] Area 30-2 of prompt 30 contains the following instructions: "When a subtask generates code, please follow these rules: *Include comments in the generated code snippets that describe the processing content of the snippet. *For code snippets that deal with the purpose of the task, add the purpose of the task directly to the comment line." Area 30-3 of the prompt contains the contents of the original task 20. Area 30-4 contains the attached data that is the subject of the task. The attached data is a sales dataset, etc.

[0037] Let's return to the explanation of Figure 6. For example, LLM executes the process based on prompt 30 and divides the original task 20 into subtasks 20-1, 20-2, and 20-3. For example, subtask 20-1 is "Generates Python code to visualize sales by customer attribute for each advertising medium in a bar graph." Subtask 20-2 is "Executes the generated Python code to plot the graph." Subtask 20-3 is "Returns the file path of the plotted graph." Of subtasks 20-1 to 20-3, subtask 20-1 is the subtask that generates the code.

[0038] The information processing device 100 generates code 22 (Python code) by executing subtask 20-1. Figure 8 shows an example of code generated by the subtask in this embodiment. As shown in Figure 8, for example, the code 22 generated by subtask 20-1 is text (a document in natural language) written to describe the operation of the program. Code 22 includes code fragments 22-1, 22-2, 22-3, 22-4, 22-5, 22-6, and 22-7.

[0039] For example, in the comment line of the code fragment 22-5, the statement "find which advertising medium is most effective for which customer attribute" that was described in the objective of the original task 20 is stated as it is.

[0040] Subsequently, the information processing apparatus 100 executes the subtask 20-2. For example, executing the subtask 20-2 means executing the code 22 to plot a graph. The information processing apparatus 100 executes the subtask 20-3 using the execution result of the subtask 20-2.

[0041] The information processing apparatus 100 compares the code fragment of the code 22 with the objective of the original task 20, and determines whether the objective of the task is additionally written as it is in a comment line of the code fragment that handles the objective of the original task 20. If the objective of the task is additionally written as it is in a comment line of the code fragment that handles the objective of the original task 20, the information processing apparatus 100 presents the execution results of the subtasks 20-1 to 20-3 to the user.

[0042] For example, the character string "find which advertising medium is most effective for which customer attribute" is set in the comment line 22-51 of the code fragment 22-5 of the code 22 shown in FIG. 8, and matches the objective "find which advertising medium is most effective for which customer attribute" set in the original task 20. In this case, the information processing apparatus 100 presents the execution results of the subtasks 20-1 to 20-3 to the user.

[0043] When the information processing apparatus 100 executes the subtasks 20-1 to 20-3, a graph visualizing a sales dataset as shown in FIG. 9 is generated. The description of the sales dataset is the same as that described above.

[0044] FIG. 9 is a diagram showing an example of a graph generated by the information processing apparatus according to the present embodiment. In the graph G2 shown in FIG. 9, the vertical axis is an axis corresponding to sales, and the horizontal axis is an axis corresponding to advertising media. As shown in FIG. 9, in the graph G1, SNS, application, and posting are shown as advertising media, and sales corresponding to customer attributes (seniors, families, young people) are visualized.

[0045] Here, when comparing the graph G2 generated by the information processing apparatus 100 and the content of the original task 20, the graph G2 includes information related to "whether it is the most efficient" described in the objective of the original task 11. For example, for each sales amount related to SNS in the graph G2, it is indicated that the sales amount among young people is the maximum. For each sales amount related to applications in the graph G2, it is indicated that the sales amount for family-oriented applications is the maximum. For each sales amount related to applications in the graph G2, it is indicated that the sales amount among seniors is the maximum.

[0046] On the other hand, if the objective of the original task is not added as-is to a comment line in the code snippet that handles the objective of the original task 20, the information processing apparatus 100 uses an LLM to divide the original task 20 into a plurality of subtasks again. For example, the information processing apparatus 100 inputs the prompt 31 shown in FIG. 10 to the LLM.

[0047] FIG. 10 is a diagram (2) showing an example of a prompt. As shown in FIG. 10, in an area 31-1 of the prompt 31, the statement "Please divide a given task into a plurality of subtasks in a logical flow. Configure the task such that each subtask can be executed independently and the entire task can ultimately be achieved" is described.

[0048] In an area 31-2 of the prompt, the statement "If a subtask generates code, please comply with the following rules: * Please add a comment representing the processing content of the code snippet to the code snippet of the generated code * For code snippets that handle the objective of the task, please add the objective of the task as-is to a comment line" is described.

[0049] In an area 31-3 of the prompt 31, an example of subtask division that does not comply with the rules specified in the area 31-2 is set. For example, in the area 31-2, the subtasks 20-1 to 20-3 described with reference to FIG. 6 are set.

[0050] In an area 31-4 of the prompt 31, the content of the original task 20 is described. Furthermore, in an area 31-5 of the prompt 31, attached data that is the target of the task is attached. The attached data is a sales dataset or the like.

[0051] The information processing device 100 repeatedly executes the above process until the purpose of the task is added as a comment line to the code snippet that handles the purpose of the original task 20.

[0052] Next, an example of the configuration of the information processing device 100 that performs the above-described processing will be explained. Figure 11 is a functional block diagram showing the configuration of the information processing device according to this embodiment. As shown in Figure 11, the information processing device 100 has a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0053] The communication unit 110 performs data communication with external devices such as servers that provide LLM services via the network. In the following description, the servers that provide LLM services will be referred to as LLM servers. The communication unit 110 may also receive sales data sets 141, etc., from other external devices.

[0054] The input unit 120 inputs various information to the control unit 150. The user may also input the original task 20, etc., by operating the input unit 120.

[0055] The display unit 130 displays information output from the control unit 150. For example, the display unit 130 displays information that represents the execution result of the original task 20.

[0056] The storage unit 140 has a sales data set 141. The storage unit 140 is a memory or the like. The sales data set 141 contains multiple sales data. The data structure of the sales data is the same as the sales data structure described in Figure 2.

[0057] The control unit 150 includes an acquisition unit 151, a planning processing unit 152, and a determination unit 153. The control unit 150 is a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc.

[0058] The acquisition unit 151 acquires the original task 20 from the user. The acquisition unit 151 outputs the original task 20 to the planning processing unit 152. If the acquisition unit 151 acquires the sales data set 141 from the communication unit 110 or the input unit 120, it stores the sales data set 141 in the storage unit 140.

[0059] The planning processing unit 152 uses LLM to divide the original task 20 into multiple subtasks. The processing performed by the planning processing unit 152 corresponds to the planning process. For example, the planning processing unit 152 generates the prompt 30 described in Figure 7 and sends the prompt 30 to the LLM server, thereby obtaining the subtasks 20-1 to 20-3, which are the divisions of the original task 20, from the LLM server.

[0060] The planning processing unit 152 outputs the original task 20 and subtasks 20-1 to 20-3 to the determination unit 153 and requests it to determine whether the code 22 generated by subtask 20-1 satisfies predetermined conditions.

[0061] If the planning processing unit 152 determines, based on the determination unit 153, that the code 22 generated by subtasks 20-1 to 20-3 satisfies predetermined conditions, the planning processing unit 152 executes subtasks 20-1 to 20-3 and outputs the execution results to the display unit 130 for display. For example, the planning processing unit 152 displays a graph G2, which is the execution result of subtasks 20-1 to 20-3, on the display unit 130.

[0062] On the other hand, if the planning processing unit 152 determines that the code 22 generated by subtasks 20-1 to 20-3 does not meet predetermined conditions, it performs the following processing to divide the original task 20 into multiple subtasks again. For example, the planning processing unit 152 generates the prompt 31 described in Figure 10 and sends the generated prompt 31 to the LLM server, thereby obtaining the multiple subtasks obtained by dividing the original task 20 from the LLM server.

[0063] The planning processing unit 152 outputs the original task 20 and the multiple subtasks that have been recreated to the determination unit 153 and requests it to determine whether the code generated by the subtasks satisfies predetermined conditions.

[0064] The planning processing unit 152 repeatedly executes the above process until the determination unit 153 determines that the code generated by the subtask satisfies predetermined conditions.

[0065] The determination unit 153 determines whether the code 22 generated by subtask 20-1 satisfies predetermined conditions and outputs the determination result to the planning processing unit 152.

[0066] For example, the determination unit 153 determines that a predetermined condition is met if the purpose of the original task 20 matches (exactly matches) the text of a comment line in any of the code snippets generated by the subtasks. On the other hand, the determination unit 153 determines that a predetermined condition is met if the purpose of the original task 20 does not match the text of a comment line in any of the code snippets generated by the subtasks.

[0067] Further explanations regarding the acquisition unit 151, planning processing unit 152, and determination unit 153 described above are the same as the explanation of the information processing device 100 shown in Figure 6. In the above explanation, the case where the planning processing unit 152 communicates with the LLM server to utilize LLM was described, but the information processing device 100 may also have LLM functionality.

[0068] Next, an example of the processing procedure of the information processing device 100 according to this embodiment will be described. Figure 12 is a flowchart showing the processing procedure of the information processing device according to this embodiment. As shown in Figure 12, the acquisition unit 151 of the information processing device 100 acquires the original task from the user (step S101). The planning processing unit 152 of the information processing device 100 generates a prompt 30 (31) and divides the original task into multiple subtasks using LLM (step S102).

[0069] The determination unit 153 of the information processing device 100 determines whether the subtask satisfies predetermined conditions (step S103). If the determination unit 153 determines that the subtask does not satisfy predetermined conditions (step S104, No), it proceeds to step 102. On the other hand, if the determination unit 153 determines that the subtask satisfies predetermined conditions (step S104, Yes), it proceeds to step S105.

[0070] The planning processing unit 152 generates execution results by executing multiple subtasks (step S105). The planning processing unit 152 displays the execution results on the display unit 130 (step S106).

[0071] Next, the effects of the information processing device 100 according to this embodiment will be described. The information processing device 100 performs a planning process using LLM to describe the contents of the original task in a natural language document for the subtasks obtained by dividing the original task. Based on the results of the planning process, the information processing device 100 determines whether the code generated in the subtask satisfies predetermined conditions in the original task, and if the result of the determination satisfies the predetermined conditions in the original task, it outputs the subtask. The information processing device 100 displays the execution result of the output subtask on the display unit 130. This enables the user to execute the specified task with high accuracy.

[0072] For example, in the conventional system 1, when the original task 11 is specified, graph G1 as shown in Figure 5 is generated, which does not include information regarding "most efficient" as stated in the objective of the original task 11, and therefore the task specified by the user cannot be executed accurately. On the other hand, in the information processing device 100, when the original task 20 is specified, graph G2 as shown in Figure 9 is generated, which includes information regarding "most efficient" as stated in the objective of the original task 20, and therefore the task specified by the user can be executed accurately.

[0073] The information processing device 100 determines that a predetermined condition is met if the code generated in the subtask matches the content of the objective of the specified task. Whether the objective and the comment lines of the code match is determined by string matching and can be determined mechanically. Therefore, it is possible to reliably determine that the objective of the original task matches the text of the comment lines of the code generated by the subtask.

[0074] The information processing device 100 receives a request for a task related to data visualization and executes the above processing. This allows the device to accurately generate graphs and other data from the task specified by the user and present them to the user.

[0075] The information processing device 100 outputs a prompt 30 to the LLM server or the like, instructing the subtask to write a document relating to the purpose of the original task in the code fragments of the code generated by the subtask, thereby dividing the subtask. This makes it possible to set the subtask to write a document relating to the purpose of the original task in the comment lines of the code when it generates code.

[0076] The information processing device 100 repeatedly executes the process of dividing the original task into multiple subtasks until the code generated in the subtasks satisfies the condition that the content matches the objective of the specified task. This makes it possible to generate subtasks that can accurately execute the task specified by the user.

[0077] Furthermore, the information processing device 100 can also use LLM to have an AI agent generate content. For example, the AI ​​agent can display content on the screen by executing a subtask. Specifically, when a goal is given to the AI ​​agent, the AI ​​agent generates subtasks to achieve the goal. The AI ​​agent then uses code snippets to collect data necessary to execute the generated subtasks. Next, the AI ​​agent generates content such as graphs that visualize the data based on the collected data. Finally, the AI ​​agent displays the generated content on the screen.

[0078] LLMs are pre-trained models such as multimodal, large-scale, and small-scale language models. An LLM is, for example, a transformer-based model trained using a token set in which some tokens are masked from a set of multiple tokens. Specifically, a language model is a transformer-based model trained using a token set in which some tokens are masked from a set of multiple tokens. For example, the information processing device 100 trains a language model using a token set (unsupervised learning dataset).

[0079] Next, we will describe in order an example of a computer hardware configuration that realizes the same functions as the information processing device 100 shown in the above embodiment.

[0080] Figure 13 shows an example of a computer hardware configuration that realizes similar functions to the information processing device of this embodiment. As shown in Figure 13, the computer 200 has a CPU 201 that performs various calculations, an input device 202 that receives data input from the user, and a display 203. The computer 200 also has a communication device 204 and an interface device 205 that exchange data with external devices via a wired or wireless network. The computer 200 also has a RAM 206 for temporarily storing various information and a hard disk drive 207. Each of the devices 201 to 207 is connected to a bus 208.

[0081] The hard disk drive 207 includes an acquisition program 207a, a planning processing program 207b, and a determination program 207c. The CPU 201 reads each of the programs 207a to 207c and loads them into the RAM 206.

[0082] The acquisition program 207a functions as the acquisition process 206a. The planning processing program 207b functions as the planning processing process 206b. The determination program 207c functions as the determination process 206c.

[0083] The processing in acquisition process 206a corresponds to the processing in acquisition unit 151. The processing in planning processing process 206b corresponds to the processing in planning processing unit 152. The processing in determination process 206c corresponds to the processing in determination unit 153.

[0084] Furthermore, it is not necessary to store each program 207a to 207c on the hard disk drive 207 from the beginning. For example, each program 207a to 207c may be stored on a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card inserted into the computer 200. Then, the computer 200 may read and execute each program 207a to 207c.

[0085] 100 Information processing device 110 Communication unit 120 Input unit 130 Display unit 140 Storage unit 141 Sales data set 150 Control unit 151 Acquisition unit 152 Planning processing unit 153 Determination unit

Claims

1. An information processing program characterized by: obtaining a specified task; dividing the obtained specified task into subtasks and having a language model perform a planning process to describe the content of the specified task in a natural language document; determining, based on the results of the planning process, whether the natural language document generated in the subtask satisfies predetermined conditions for the specified task; and, if the result of the determination satisfies the predetermined conditions for the specified task, having the computer perform a process to output the subtask.

2. The information processing program according to claim 1, wherein the natural language document generated in the subtask is associated with a code fragment that defines the processing content of the specified task and a natural language document relating to the processing content of the code fragment, and the determination process determines whether the natural language document generated in the subtask matches the content of the specified task, and further causes the computer to execute a process to execute the processing of the code fragment based on the determination result.

3. The information processing program according to claim 2, wherein the specified task is a task relating to data visualization, and the program repeatedly performs the process of performing the planning in the language model until the condition is met that a natural language document relating to the processing content in the code fragment matches a document relating to the purpose of the predetermined task, and further causes the computer to perform the process of generating data visualization content based on the subtask for which the planning process has been performed, and displaying the generated content on the screen.

4. The information processing program according to claim 3, characterized in that the computer further causes the language model to output a prompt instructing the code snippet of the natural language document generated for the subtask to describe a document relating to the purpose of the specified task.

5. The information processing program according to claim 3, characterized in that it determines whether a natural language document relating to the processing content in the code fragment matches a document relating to the purpose of the predetermined task, and if they do not match, causes the computer to execute the process of executing the planning process using the language model again.

6. The information processing program according to claim 3, characterized in that the AI ​​agent collects data necessary for executing the subtask, and the AI ​​agent displays the generated content on the screen based on the collected data.

7. An information processing method characterized by the following: obtaining a specified task; dividing the obtained specified task into subtasks and having a language model perform a planning process to describe the content of the specified task in a natural language document; determining, based on the results of the planning process, whether the natural language document generated in the subtask satisfies predetermined conditions for the specified task; and, if the result of the determination satisfies the predetermined conditions for the specified task, having a computer perform a process to output the subtask.

8. The information processing method according to claim 7, wherein the natural language document generated in the subtask is associated with a code fragment that defines the processing content of the designated task and a natural language document relating to the processing content of the code fragment, and the determination process determines whether the natural language document generated in the subtask matches the content of the designated task, and the computer further executes a process to execute the processing of the code fragment based on the determination result.

9. The information processing method according to 8, characterized in that the specified task is a task relating to data visualization, and the computer repeatedly performs the process of performing the planning in the language model until the condition is met that a natural language document relating to the processing content in the code fragment matches a document relating to the purpose of the predetermined task, and the computer further performs the process of generating data visualization content based on the subtask for which the planning process has been performed, and displaying the generated content on the screen.

10. The information processing method according to claim 9, characterized in that the computer further performs a process of outputting a prompt to the language model instructing it to write a document relating to the purpose of the specified task in the code fragment of the natural language document generated for the subtask.

11. The information processing method according to claim 9, characterized in that the computer determines whether a natural language document relating to the processing content in the code fragment matches a document relating to the purpose of the specified task, and if they do not match, the computer repeats the process of executing the planning process using the language model.

12. The information processing method according to claim 9, characterized in that the AI ​​agent collects data necessary for executing the subtask, and the AI ​​agent displays the generated content on the screen based on the collected data.

13. An information processing device having a control unit that obtains a specified task, divides the obtained specified task into subtasks and performs a planning process using a language model to describe the content of the specified task in a natural language document, determines whether the natural language document generated in the subtask satisfies predetermined conditions for the specified task based on the result of the planning process, and outputs the subtask if the result of the determination satisfies the predetermined conditions for the specified task.

14. The information processing apparatus according to 13, wherein the natural language document generated in the subtask is associated with a code fragment that defines the processing content of the designated task and a natural language document relating to the processing content of the code fragment, the determination process determines whether the natural language document generated in the subtask matches the content of the designated task, and the control unit further executes a process to execute the code fragment based on the determination result.

15. The information processing apparatus according to 13, wherein the specified task is a task relating to data visualization, the control unit repeatedly performs the process of performing the planning process in the language model until the condition is met that a natural language document relating to the processing content in the code fragment matches a document relating to the purpose of the specified task, and further performs the process of generating data visualization content based on the subtask on which the planning process has been performed, and displaying the generated content on the screen.

16. The information processing apparatus according to claim 15, wherein the control unit further performs a process that outputs a prompt to the language model instructing it to write a document relating to the purpose of the specified task in the code fragment of the natural language document generated for the subtask.

17. The information processing apparatus according to 15, wherein the control unit determines whether the natural language document relating to the processing content in the code fragment matches the document relating to the purpose of the specified task, and if they do not match, it repeats the process of executing the planning process in the language model.

18. The information processing device according to claim 15, characterized in that the AI ​​agent collects data necessary for the execution of the subtask, and the AI ​​agent displays the generated content on a screen based on the collected data.

Images