JP7827686B2 - Business support system, business support method, and program - Google Patents

Description

This disclosure relates to a business support system, a business support method, and a program.

Technologies based on AI (Artificial Intelligence) to assist users in their work have been under consideration. For example, Patent Document 1 describes a large-scale language model such as BERT (Bidirectional Encoder Representations from Transformers), which is an example of such AI. Large-scale language models such as BERT are highly versatile and can process a variety of tasks based on input data representing user input.

Japanese Patent Application Laid-Open No. 2023-125311

However, because highly versatile AI such as that described in Patent Document 1 is not specialized for a specific task, the accuracy of task processing may not be sufficient. On the other hand, if a business support system were to prepare AI specialized for a specific task, it would need to prepare a large number of AIs to handle a variety of tasks. Users would have to select an AI specialized for a specific task from among the many AIs and have it process the task. For this reason, conventional technology does not sufficiently enhance user convenience. This point is not limited to AI, but can be said of all business support programs that support users' work.

One of the goals of this disclosure is to increase user convenience.

A business assistance system according to one aspect of the present disclosure includes an input data acquisition unit that acquires input data indicating user input; a selection unit that can select a business assistance program from among a plurality of business assistance programs specialized for a specific task in business assistance based on AI (artificial intelligence) that can select the business assistance program specialized for the task corresponding to the input data; and a task processing unit that, when the business assistance program is selected by the selection unit, processes the task corresponding to the input data based on the business assistance program.

This disclosure can improve user convenience.

FIG. 1 illustrates an example of a hardware configuration of a business support system. FIG. 10 is a diagram illustrating an example of a business support screen. FIG. 2 is a diagram illustrating an example of a relationship between a parent AI and a child AI. FIG. 2 is a diagram illustrating an example of functions realized by the business support system. FIG. 10 is a diagram illustrating an example of a training database. FIG. 2 is a diagram illustrating an example of processing executed in the business support system. FIG. 10 is a diagram illustrating an example of functions realized by a business support system according to a first modified example. FIG. 13 is a diagram showing an example of AI according to Modification Example 6.

[1. Hardware Configuration]
An example of an embodiment of a business support system, a business support method, and a program according to the present disclosure will be described. Fig. 1 is a diagram illustrating an example of the hardware configuration of the business support system. For example, the business support system 1 includes a learning terminal 10, a server 20, and a user terminal 30. Each of the learning terminal 10, the server 20, and the user terminal 30 is connected to a network N such as the Internet or a LAN.

The learning terminal 10 is a computer that performs AI (Artificial Intelligence) learning, which will be described later. For example, the learning terminal 10 is a personal computer, a tablet terminal, or a smartphone. For example, the learning terminal 10 includes a control unit 11, a memory unit 12, a communication unit 13, an operation unit 14, and a display unit 15. The control unit 11 includes at least one processor. The memory unit 12 includes at least one of volatile memory such as RAM and non-volatile memory such as flash memory. The communication unit 13 includes at least one of a communication interface for wired communication and a communication interface for wireless communication. The operation unit 14 is an input device such as a mouse or a touch panel. The display unit 15 is an LCD or organic EL display.

The server 20 is a server computer. For example, the server 20 includes a control unit 21, a memory unit 22, and a communication unit 23. The hardware configurations of the control unit 21, the memory unit 22, and the communication unit 23 may be similar to those of the control unit 11, the memory unit 12, and the communication unit 13, respectively.

The user terminal 30 is a user's computer. For example, the user terminal 30 is a personal computer, a tablet terminal, a smartphone, or a wearable terminal. For example, the user terminal 30 includes a control unit 31, a memory unit 32, a communication unit 33, an operation unit 34, and a display unit 35. The hardware configurations of the control unit 31, the memory unit 32, the communication unit 33, the operation unit 34, and the display unit 35 may be similar to those of the control unit 11, the memory unit 12, the communication unit 13, the operation unit 14, and the display unit 15, respectively.

The programs stored in the storage units 12, 22, and 32 may be supplied via the network N. The hardware configuration of each of the learning terminal 10, server 20, and user terminal 30 is not limited to the example shown in Figure 1. For example, at least one of the learning terminal 10, server 20, and user terminal 30 may include at least one of a reading unit (e.g., a memory card slot) that reads a computer-readable information storage medium and an input/output unit (e.g., a USB terminal) for direct connection to an external device. A program stored on an information storage medium may be supplied to at least one of the learning terminal 10, server 20, and user terminal 30 via at least one of the reading unit and the input/output unit.

Furthermore, the business support system 1 only needs to include at least one computer. The computers included in the business support system 1 are not limited to the example in Figure 1. For example, the business support system 1 may include only the learning terminal 10 and the server 20. In this case, the user terminal 30 exists outside the business support system 1. The business support system 1 may also include only the server 20. In this case, the learning terminal 10 and the user terminal 30 exist outside the business support system 1. The business support system 1 may also include the server 20 and other server computers.

[2. Overview of the business support system]
In this embodiment, the business support system 1 has a business support function that supports the user's business. For example, the business support function may be a communication function that enables the user to communicate with other users, a schedule management function that manages the user's schedule, a file management function that manages the user's files, or an email management function that manages the user's email. The business support function may also be any other known function.

For example, the business support system 1 may provide users with cloud-based or on-premise groupware. The business support system 1 may also provide users with services that support business operations, but are not classified as groupware. When a user logs in to the business support system 1, the user terminal 30 displays a business support screen on the display unit 35, allowing the user to use the business support functions. In this embodiment, an example is given in which the business support screen is displayed on the browser of the user terminal 30.

Figure 2 is a diagram showing an example of a business support screen. The example in Figure 2 shows a business support screen SC that allows a user to use a communication function, which is an example of a business support function. For example, the user terminal 30 displays a thread selected by the user from among multiple threads created in the business support system 1 on the business support screen SC. The business support screen SC lists posts made to the thread in chronological order. The user can post a new post to the thread or post a reply to an existing post.

As shown in the business support screen SC at the top of Figure 2, a user can input a new post into the input form F10. For example, a user can mention another user by specifying that user. A mention is a notification to another user. In the example of Figure 2, a user can mention another user by entering a specific symbol (e.g., @) followed by the other user's information (e.g., the other user's name). The mention mechanism may be similar to known mechanisms. A user can also mention an AI in the same way as mentioning other users.

AI is a program with artificial intelligence that assists users in their work. There are various definitions of AI, but the AI in this embodiment may be AI defined by various known definitions. The AI may be AI called generative AI or conversational AI. For example, the AI may be a large-scale language model, a machine learning model not classified as a large-scale language model, a program called a bot, or other programs. There are various definitions of machine learning, but the machine learning in this embodiment may be machine learning defined by various known definitions. The machine learning may be supervised learning, semi-supervised learning, or unsupervised learning.

In this embodiment, a case where a large-scale language model corresponds to AI is taken as an example. For example, a user can mention AI by entering a specific symbol (e.g., @) followed by a string indicating AI (e.g., bot) instead of information about another user. When a user enters a post, the user terminal 30 transmits input data indicating the post entered by the user to the server 20.

For example, when the server 20 receives input data from the user terminal 30, it determines whether or not a mention has been made to the AI based on the input data. If the server 20 determines that a mention has been made to the AI, it reflects the user's post in the thread. If the server 20 determines that a mention has been made to the AI, it causes the AI to process a task corresponding to the input data.

A task is the content of processing executed by a program, for example AI. A task can also be said to be the content of output from a program, for example AI. A task may be any task related to business support. For example, a task may be machine translation, summarizing, research (provision of knowledge), schedule adjustment, generation of text such as email or report, image creation, application creation, database configuration, input data classification, or other tasks. The business support system 1 is capable of processing multiple tasks, of which these are examples. Task processing is also an example of a business support function possessed by the business support system 1.

A task corresponding to input data is a task determined based on the input indicated by the input data (in the example of Figure 2, a post to a thread). A task corresponding to input data can also be said to be a task requested by the user who made the input indicated by the input data. For example, if the AI is generative AI, a task corresponding to input data is a task in which the AI generates content such as an image or text based on the input data. If the AI is interactive AI, a task corresponding to input data is a task in which the AI generates answer data indicating an answer based on the input indicated by the input data.

In this embodiment, multiple AIs are prepared that are specialized for specific tasks in business support. The AIs are an example of business support programs that support the user's business. As in the modified example described below, the business support program may be a program that is not classified as AI, but this embodiment takes as an example a case where the AI corresponds to the business support program. The server 20 has one of the multiple AIs prepared in advance that is specialized for a task corresponding to the input data process that task.

In the example of Figure 2, the input data represents a post such as, "I will be explaining popular programming languages to my department manager at next week's regular meeting. Please tell me what popular programming languages are." Because the user is requesting the AI to research popular programming, the task corresponding to the input data is a research task. The server 20 has an AI specialized in research tasks, out of multiple pre-prepared AIs, process the research task. The server 20 transmits data indicating the task processing results output by the AI to the user terminal 30. As shown in the business support screen SC at the bottom of Figure 2, the user terminal 30 displays the AI's response based on the data.

As described above, the server 20 selects an AI from multiple AIs that is specialized for a task corresponding to the input data to process that task. For example, a user may select an AI from multiple AIs that is specialized for a task corresponding to the input data, and the server 20 may then select the AI selected by the user to process the task. In this case, the user must select the AI, which is time-consuming for the user. Furthermore, the user may not necessarily select the appropriate AI, so there is a possibility that the task may not be processed appropriately.

In this embodiment, therefore, in addition to AI specialized in a specific task, an AI capable of selecting that AI is prepared in advance. Hereinafter, an AI that performs processing specialized for a specific task will be referred to as a child AI. A child AI can also be referred to as a specialized AI. An AI that can select a child AI will be referred to as a parent AI. A parent AI can also be referred to as a selection AI. In this embodiment, an example is given in which the parent AI only performs the task of selecting a child AI, but as in the modified example described below, the parent AI may also be a general-purpose AI that can process a variety of tasks. In other words, the parent AI may not only be specialized in the task of selecting a child AI, but may also be able to handle a variety of tasks.

Figure 3 is a diagram showing an example of the relationship between parent AI and child AI. In the example of Figure 3, child AI201A specialized in the task of machine translation, child AI201B specialized in the task of summarizing, child AI201C specialized in the task of research, and child AI201D specialized in the task of schedule adjustment are prepared. Hereinafter, when there is no particular distinction between child AI201A to AI201D, they will simply be referred to as child AI201. Furthermore, when there is no particular distinction between parent AI200 and child AI201, the reference numerals will be omitted and they will simply be referred to as AI.

For example, the server 20 inputs input data received from the user terminal 30 into the parent AI 200. The parent AI 200 is assumed to have undergone prior learning so that it can select an appropriate child AI 201. The learning method of the parent AI 200 will be described later. Depending on the content of the input data, the parent AI 200 may not need to select a child AI 201. In the example of input data shown in Figure 2, the parent AI 200 infers from the content of the input data that child AI 201C, which is specialized in research tasks, is appropriate, and selects child AI 201C.

For example, the server 20 inputs input data received from the user terminal 30 to the child AI 201C selected by the parent AI 200. The child AI 201C processes the research task based on the input data and outputs output data indicating the results of the research. In the example of Figure 2, the output data output from the child AI 201C indicates an answer such as, "Trends in programming languages are constantly changing, but as of 2023, several languages are attracting attention. For example, the following five languages are..."

In this embodiment, the child AI201 selected by the parent AI200 varies depending on the content of the input data. For example, suppose a user inputs a post containing a sentence written in Japanese and a sentence such as "Please translate the above sentence into English." In this case, the parent AI200 infers from the content of the input data that child AI201A, which specializes in machine translation tasks, is appropriate, and selects child AI201A. Based on the input data, the server 20 has the child AI201A selected by the parent AI200 process the machine translation task. Child AI201A processes the machine translation task by translating the Japanese sentence indicated by the input data into English.

For example, suppose a user inputs a post that includes a sentence of a certain length and a sentence such as "Please create a summary of the above sentence." In this case, parent AI200 infers from the content of the input data that child AI201B, which specializes in the task of creating a summary, is appropriate, and selects child AI201B. Based on the input data, server 20 has child AI201B selected by parent AI200 process the task of creating a summary. Child AI201B processes the task of creating a summary by creating a summary of the sentence indicated by the input data.

For example, suppose a user inputs a post containing candidate dates for a meeting and a sentence such as, "Please check if my schedule is free." In this case, parent AI 200 infers from the contents of the input data that child AI 201D, which specializes in schedule adjustment tasks, is appropriate, and selects child AI 201D. Based on the input data, server 20 has child AI 201D selected by parent AI 200 process the schedule adjustment task. Child AI 201D processes the schedule adjustment task by referencing the user's schedule data registered in business support system 1 and checking whether the candidate dates indicated by the input data are free.

As described above, in the business support system 1 of this embodiment, the parent AI 200 can select a child AI 201 that specializes in a task corresponding to the input data based on input data indicating a post entered by a user. When a child AI 201 is selected by the parent AI 200, the child AI 201 processes the task corresponding to the input data. This improves the user's business efficiency, and the business support system 1 can increase user convenience. Details of the business support system 1 will be explained below.

[3. Functions realized by the business support system]
FIG. 4 is a diagram showing an example of functions realized by the business support system 1. As shown in FIG.

[3-1. Functions realized on the learning device]
For example, the learning terminal 10 includes a data storage unit 1000 and a learning unit 1001. The data storage unit 1000 is realized by the storage unit 12. The learning unit 1001 is realized by the control unit 11.

[Data storage unit]
The data storage unit 1000 stores data necessary for training at least one of the parent AI 200 and the child AI 201. For example, the data storage unit 1000 stores the parent AI 200. In this embodiment, an example is given in which the parent AI 200 is a large-scale language model based on a transformer, such as a generative pre-trained transformer (GPT) or a bidirectional encoder representations from transformers (BERT). The parent AI 200 may also be a large-scale language model using a method other than a transformer. The parent AI 200 may also be a machine learning model that is not classified as a large-scale language model. For example, the parent AI 200 may be a generative adversarial network, a neural network, or a support vector machine.

Note that, since the selection of an appropriate child AI 201 can also be considered as classification (labeling) of input data, the parent AI 200 may be a machine learning model capable of classifying input data. Furthermore, the parent AI 200 may be a program that is not classified as a machine learning model. For example, the parent AI 200 may be a rule-based program that indicates the relationship between input data and an appropriate child AI 201. A rule-based program defines rules that are judged based on all or part of the input indicated by the input data. A rule-based program contains program code that indicates that, when a certain rule is satisfied, the child AI 201 associated with that rule is appropriate. When the parent AI 200 is a rule-based program, learning is not performed, so the business support system 1 does not need to include a learning terminal 10.

In this embodiment, the parent AI 200 is a machine learning model, with a Transformer-based large-scale language model as an example, and therefore includes parameters that are adjusted through learning and a program that indicates processes such as calculating embedded representations. Furthermore, this embodiment takes as an example a case where a business that provides the business support system 1 to users (e.g., a company that developed groupware) creates the parent AI 200 by re-learning (e.g., fine-tuning, transfer learning, or distillation) a pre-trained large-scale language model provided by another business.

The program and parameters of the parent AI200 may be various known programs and parameters. The parameters of the parent AI200 are referenced by the program of the parent AI200. For example, the parameters are at least one of a weighting coefficient and a bias. The program of the parent AI200 includes code that indicates the internal processing of the parent AI200. For example, the program of the parent AI200 includes an intermediate layer that calculates an embedded representation and generates data for the output layer, and an output layer that performs final output based on the data.

For example, if the parent AI200 is a Transformer-based large-scale language model, the program of the parent AI200 indicates a process of dividing a sentence indicated by input data into multiple tokens, a process of calculating embedded representations for each token, a process of making a prediction based on the arrangement of the embedded representations, and a process of producing an output based on this series of processes. If the parent AI200 is another machine learning model, the program of the parent AI200 may indicate a process adopted for that other machine learning model. The data storage unit 1000 stores a training database DB that stores training data that the learning unit 1001, described below, has the parent AI200 learn.

Figure 5 is a diagram showing an example of a training database (DB). In this embodiment, supervised learning is used as an example. For example, training data includes an input portion that is input to the AI during learning, and an output portion that is the correct answer during learning. In the example of Figure 5, the input portion of the training data is input data for training. The output portion of the training data is child AI identification data for identifying the child AI 201 that will be the correct answer (child AI 201 specialized in the task according to the training input data). When the parent AI 200 is created by relearning a large-scale language model that has been pre-trained, the training data is data for relearning. Learning in this embodiment also includes relearning.

The training input data indicates training input. For example, the training input data indicates training sentences (character strings). As in the example of Figure 2, when input data indicating thread posts is analyzed by the parent AI 200, the training input data is the post that the learning unit 1001 causes the parent AI 200 to learn. The training input data may indicate information other than sentences. For example, the training input data may indicate files, reactions, images displayed on the business support screen SC, or other information.

The child AI identification data is data that enables the business support system 1 to identify the child AI 201. For example, the child AI identification data is an ID or name assigned to the child AI 201. The child AI identification data indicates the child AI 201 that is specialized for a task corresponding to the training input data. In other words, the child AI identification data indicates the child AI 201 that is the correct answer during learning. In this embodiment, an example is given in which one piece of child AI identification data is associated with one piece of training input data (i.e., one piece of child AI identification data is the correct answer for one piece of training input data), but multiple pieces of child AI identification data may be associated with one piece of training input data.

For example, if the training input data indicates content related to machine translation, the child AI identification data paired with the training input data indicates child AI201A specialized in the task of machine translation. If the training input data indicates content related to summarization, the child AI identification data paired with the training input data indicates child AI201B specialized in the task of summarization. If the training input data indicates content related to research, the child AI identification data paired with the training input data indicates child AI201C specialized in the task of research. If the training input data indicates content related to schedule adjustment, the child AI identification data paired with the training input data indicates child AI201D specialized in the task of schedule adjustment.

The training data may be created by a person in charge of learning the parent AI 200. For example, the person in charge creates training input data, which is the input portion of the training data, based on posts in existing threads or virtual posts. The person in charge determines the appropriate task for these posts and designates (annotates) the child AI identification data of the child AI 201 specialized in that task as the output portion of the training data. The learning terminal 10 generates pairs of these input and output portions as training data and stores them in the training database DB. The training data may be generated using known tools rather than being generated by manual designation by the person in charge.

For example, the data storage unit 1000 stores multiple child AIs 201. In this embodiment, the child AI 201 is, like the parent AI 200, a large-scale Transformer-based language model, as an example. The child AI 201 may also be a large-scale language model using a method other than Transformer. The child AI 201 may also be a machine learning model that is not classified as a large-scale language model. For example, the child AI 201 may be a generative adversarial network, a neural network, or a support vector machine. Furthermore, the child AI 201 may be a program that is not classified as a machine learning model.

In this embodiment, an example is given of a case where a business providing a business support system 1 to a user creates a child AI 201 by relearning a large-scale pre-trained language model provided by another business. For example, the child AI 201 includes parameters that are adjusted through learning and a program that indicates processing such as calculating embedded expressions. The program and parameters of the child AI 201 may be various well-known programs and parameters. This is as explained for the program and parameters of the parent AI 200. The program and parameters of the child AI 201 may be the same as those explained as examples of the program and parameters of the parent AI 200.

For example, the data storage unit 1000 stores training data for relearning a large-scale language model that has already been pre-trained. That is, the data storage unit 1000 stores training data for the child AI 201 separately from the training data for the parent AI 200. The training data for the child AI 201 may be publicly known training data, and is therefore not shown in the figure. The training data for the child AI 201 is prepared according to the task for which the child AI 201 specializes. For example, the training data for the child AI 201 includes input data for training and output data that indicates an output that is the correct answer according to the task for the child AI 201. The output data is determined according to the content of the task.

For example, training data for child AI201A specialized in the task of machine translation includes training input data indicating the sentence before translation and output data indicating the sentence after translation. Training data for child AI201B specialized in the task of summary creation includes training input data indicating the sentence to be summarized and output data indicating the correct summary. Training data for child AI201C specialized in the task of research includes training input data indicating the details of the research request and output data indicating the correct details of the research. Training data for child AI201D specialized in the task of schedule adjustment includes training input data indicating the schedule to be adjusted and output data indicating the correct adjustment result. Similarly, training data appropriate to the task may be prepared for other tasks.

The data stored in the data storage unit 1000 is not limited to the example in this embodiment. The data storage unit 1000 may store data necessary for learning at least one of the parent AI 200 and the child AI 201. For example, the data storage unit 1000 may store other data such as a learning program that indicates a series of processes during learning. The learning program also indicates the formula for the loss function calculated during learning. The data storage unit 1000 may store all of the parent AI 200 before learning (parent AI 200 with initial parameters), parent AI 200 after learning (parent AI 200 with parameters adjusted by learning), child AI 201 before learning (child AI 201 with parameters adjusted by pre-learning), and child AI 201 after learning (child AI 201 with parameters adjusted by relearning).

[Study Department]
The learning unit 1001 performs learning of at least one of the parent AI 200 and the child AI 201. For example, the learning unit 1001 performs learning of the parent AI 200 based on each of multiple training data stored in the training database DB. In this embodiment, an example is given in which the learning unit 1001 performs learning of the parent AI 200 based on a supervised learning algorithm. The learning algorithm may be a well-known algorithm. The learning unit 1001 may also perform learning of the parent AI 200 based on a semi-supervised learning or unsupervised learning algorithm.

In this embodiment, the learning unit 1001 performs a series of learning processes by executing a learning program stored in the data storage unit 1000. For example, the learning unit 1001 may perform learning of the parent AI 200 by adjusting the parameters of the parent AI 200 so that when an input portion of training data stored in the training database DB is input to the parent AI 200, an output portion of the training data is output from the parent AI 200. The learning unit 1001 calculates the loss of the parent AI 200 based on the loss function described in the learning program. The learning unit 1001 repeats learning until the loss becomes sufficiently small.

In this embodiment, an example is given in which a Transformer-based large-scale language model corresponds to the parent AI 200, and therefore the learning unit 1001 may perform learning of the parent AI 200 based on a known algorithm adopted in the Transformer-based large-scale language model. The learning unit 1001 may also perform learning of the parent AI 200 based on a known algorithm according to the type of machine learning model adopted as the parent AI 200. For example, the learning unit 1001 may perform learning of the parent AI 200 based on a known algorithm such as backpropagation or gradient descent. The learning unit 1001 may also perform learning of the parent AI 200 from scratch, rather than re-training a pre-trained large-scale language model. The learning unit 1001 may also perform learning of the parent AI 200 based on unannotated training data.

For example, when learning is complete, the learning unit 1001 records the learned parent AI 200 in the data storage unit 1000. The data storage unit 1000 may store both the parent AI 200 before learning and the learned parent AI 200. The learning unit 1001 transmits the learned parent AI 200 to the server 20. The learned parent AI 200 transmitted to the server 20 is made available for use by the user.

For example, the learning unit 1001 performs learning of the child AI 201 based on the training data of the child AI 201. A specific example of the learning method of the child AI 201 may be the same as the method described as an example of the learning method of the parent AI 200. The learning unit 1001 performs learning of the child AI 201 by adjusting the parameters of the child AI 201 so that when an input portion of the training data of the child AI 201 is input to the child AI 201, an output portion of the training data is output from the child AI 201. In this embodiment, the learning unit 1001 performs re-learning of the pre-trained child AI 201. The re-learning may be performed using a known algorithm. The learning unit 1001 may perform learning of the child AI 201 from scratch, rather than re-training a pre-trained large-scale language model. The learning unit 1001 may perform learning of the child AI 201 based on unannotated training data.

For example, when learning is complete, the learning unit 1001 records the learned child AI 201 in the data storage unit 1000. The data storage unit 1000 may store both the child AI 201 before learning and the learned child AI 201. The learning unit 1001 transmits the trained child AI 201 to the server 20. The trained child AI 201 transmitted to the server 20 is made available for use by the user.

[3-2. Functions implemented by the server]
For example, the server 20 includes a data storage unit 2000, an input data acquisition unit 2001, a selection unit 2002, a task processing unit 2003, and a providing unit 2004. The data storage unit 2000 is realized by the storage unit 22. The input data acquisition unit 2001, the selection unit 2002, the task processing unit 2003, and the providing unit 2004 are each realized by the control unit 21.

[Data storage unit]
The data storage unit 2000 stores data necessary for business support. For example, the data storage unit 2000 stores a parent AI 200 and multiple child AIs 201. In this embodiment, the learning terminal 10 performs learning on each of the parent AI 200 and the multiple child AIs 201. The server 20 acquires the trained parent AI 200 and the trained multiple child AIs 201 from the learning terminal 10 and records them in the data storage unit 2000.

As mentioned above, child AI 201 is an example of a business support program, so any reference to child AI 201 can be read as business support program. In this embodiment, an example is given in which child AI 201 corresponds to one specific task, but child AI 201 may also correspond to multiple tasks. In other words, child AI 201 may be a single-tasking AI or a multi-tasking AI.

Furthermore, at least one of the parent AI 200 and the child AI 201 may be stored in a computer other than the server 20 (e.g., another server computer). For example, if the parent AI 200 is stored in another computer, the selection unit 2002 described below requests the other computer to select the child AI 201 by sending input data to the other computer. The other computer can select the child AI 201 based on the input data received from the server 20 and the parent AI 200 stored therein. The other computer sends selection result data indicating the selection result of the child AI 201 to the server 20. The selection unit 2002 may identify the child AI 201 selected by the parent AI 200 based on the selection result data received from the other computer.

For example, if the child AI 201 is stored in another computer, the task processing unit 2003 described below requests the other computer to process the task by sending input data to that computer. If the parent AI 200 is not stored in that other computer, the task processing unit 2003 may send selection result data indicating the child AI 201 selected by the parent AI 200 to that other computer. Based on this data, the other computer causes the child AI 201 specialized in the task corresponding to the input data to process the task. The other computer sends processing result data indicating the task processing results to the server 20. The task processing unit 2003 acquires the processing result data received from the other computer. In this way, the task processing unit 2003 may process a task by requesting another computer to actually process the task and acquiring the processing result data.

The data stored in the data storage unit 2000 is not limited to the above examples. The data storage unit 2000 can store any data. For example, the data storage unit 2000 may store the programs and data of each function of the business support system 1 (for example, in the case of a communication function, data posted to a thread). The data storage unit 2000 may also store a database in which various data of users who use the business support system 1 is stored. The data storage unit 2000 may also store data such as data for displaying the business support screen SC (for example, HTML data).

[Input data acquisition section]
The input data acquisition unit 2001 acquires input data indicating user input. User input can also be referred to as user operation. For example, the input data indicates a character string entered by the user. The character string can also be referred to as text, a message, or a sentence. In the example of FIG. 2, the input data indicates a user post. The input data may also indicate input other than a character string. For example, the input data may indicate a voice uttered by the user, a file selection result by the user, a selection result of information (e.g., an image or text) displayed on the business support screen SC, or other input.

For example, when a user makes an input on the business support screen SC, the user terminal 30 transmits input data indicating the input to the server 20. The input data acquisition unit 2001 acquires the input data from the user terminal 30. The input data acquisition unit 2001 may acquire the input data from the user terminal 30 via a computer other than the server 20 and the user terminal 30. In the example of Figure 2, when the user performs an operation to post (for example, selecting the "Write" button), the user terminal 30 transmits input data indicating the post entered in the input form F10 to the server 20. The input data acquisition unit 2001 acquires input data indicating the user's post from the user terminal 30.

The input data may be stored in the data storage unit 2000. For example, the input data acquisition unit 2001 records the input data acquired from the user terminal 30 in the data storage unit 2000. The input data acquisition unit 2001 can acquire the input data from the data storage unit 2000 at any time. In this case, the data storage unit 2000 may also store data for identifying the user who made the input indicated by the input data.

[Selection section]
The selection unit 2002 can select a business support program based on the parent AI 200, which can select a business support program specialized for a task corresponding to input data from among multiple business support programs specialized for specific tasks in business support. In this embodiment, an example is given in which the child AI 201 corresponds to a business support program, so the selection unit 2002 can select the child AI 201 specialized for a task corresponding to the input data based on the parent AI 200. The selection unit 2002 may select multiple child AIs 201.

The selection unit 2002 is not necessarily required to select a child AI 201, and may not select a child AI 201. For example, the selection unit 2002 may not select a child AI 201 if there is no task corresponding to the input data, or if there is a task corresponding to the input data but no child AI 201 specialized for it. The selection unit 2002 may attempt to select a child AI 201 each time input data is acquired. When certain input data is acquired, the selection unit 2002 may not select a child AI 201 specialized for a task corresponding to the input data, and when other input data is acquired, the selection unit 2002 may select a child AI 201 specialized for a task corresponding to the other input data.

For example, the selection unit 2002 inputs input data to the parent AI 200. The input data is used as a prompt for the parent AI 200. In addition to the input data, a default prompt prepared by the business support system 1 may be input to the parent AI 200. The parent AI 200 calculates an embedded expression based on the input data and outputs output data indicating an output corresponding to the embedded expression. If a child AI 201 corresponding to the embedded expression exists, the parent AI 200 outputs the child AI identification data of the child AI 201 as output data. If a child AI 201 corresponding to the embedded expression does not exist, the parent AI 200 outputs output data indicating that the child AI 201 has not been selected. The selection unit 2002 acquires the output data output from the parent AI 200. A series of processes by the parent AI 200 is executed based on parameters adjusted by learning.

In this embodiment, a Transformer-based large-scale language model corresponds to the parent AI200, so the parent AI200 divides the string indicated by the input data into multiple tokens and then calculates the embedded representation of each token. The parent AI200 predicts the continuation as necessary based on the sequence of the embedded representation of each token, and then outputs output data. If a child AI201 exists that corresponds to the sequence of the embedded representation of each token, the parent AI200 outputs the child AI identification data of that child AI201 as output data. If a child AI201 does not exist that corresponds to the sequence of the embedded representation of each token, the parent AI200 outputs output data indicating that the child AI201 was not selected.

Note that the processing performed by parent AI200 is not limited to the above example. Parent AI200 may output output data indicating the selection result of child AI201 based on input data. The relationship between these inputs and outputs may be defined in the program code of parent AI200. For example, if parent AI200 is a machine learning model other than Transformer, parent AI200 may calculate an embedded representation based on input data and output output data corresponding to the embedded representation, as described above. If parent AI200 is a program not classified as a machine learning model, output data may be output from parent AI200 based on the program code, variables, etc. written in the program.

[Task processing section]
When a business support program is selected by the selection unit 2002, the task processing unit 2003 processes a task corresponding to the input data based on the business support program. In this embodiment, an example is given in which the child AI 201 corresponds to the business support program, so the task processing unit 2003 processes a task based on the input data and the child AI 201 selected by the selection unit 2002. In other words, the task processing unit 2003 causes the child AI 201 selected by the selection unit 2002 to process a task corresponding to the input data.

For example, the task processing unit 2003 inputs input data to the child AI 201 selected by the selection unit 2002. The input data is used as a prompt for the child AI 201. In addition to the input data, a default prompt prepared by the business support system 1 may be input to the child AI 201. The child AI 201 calculates an embedded expression based on the input data and outputs output data corresponding to the embedded expression. The child AI 201 outputs output data corresponding to the embedded expression. The output of the output data corresponds to processing of a task corresponding to the input data. The selection unit 2002 acquires the output data output from the child AI 201. A series of processes by the child AI 201 is executed based on parameters adjusted by learning.

For example, if the selection unit 2002 selects child AI201A specialized in the task of machine translation, child AI201A outputs output data indicating the translation result according to the embedded expression. If the selection unit 2002 selects child AI201B specialized in the task of summarizing, child AI201B outputs output data indicating the summary according to the embedded expression. If the selection unit 2002 selects child AI201C specialized in the task of research, child AI201C outputs output data indicating information according to the embedded expression. If the selection unit 2002 selects child AI201D specialized in the task of schedule adjustment, child AI201D outputs output data indicating the adjustment result according to the embedded expression.

In this embodiment, a Transformer-based large-scale language model corresponds to the child AI 201, so the child AI 201 divides the string indicated by the input data into multiple tokens and then calculates the embedded representation of each token. The child AI 201 predicts the continuation as necessary based on the sequence of the embedded representation of each token, and then outputs the output data.

Note that the processing performed by child AI201 is not limited to the above example. Child AI201 may output output data indicating the selection result of child AI201 based on input data. The relationship between these inputs and outputs may be defined in the program code of child AI201. For example, if child AI201 is a machine learning model other than Transformer, child AI201 may calculate an embedded representation based on the input data and output output data corresponding to the embedded representation, as described above. If child AI201 is a program not classified as a machine learning model, output data may be output from child AI201 based on the program code, variables, etc. written in the program.

In this embodiment, if the selection unit 2002 does not select a child AI 201, the task processing unit 2003 does not process the task corresponding to the input data. In this case, the task processing unit 2003 transmits data to the user terminal 30 indicating that the task corresponding to the input data has not been processed. Based on the data, the user terminal 30 displays an error message on the display unit 35 indicating that the task corresponding to the input data has not been processed. If the selection unit 2002 does not select a child AI 201, the task processing unit 2003 may cause a general-purpose parent AI 200 to process the task, as in a modified example described below. As another example, if the selection unit 2002 does not select a child AI 201, the task processing unit 2003 may cause a general-purpose AI prepared separately from the parent AI 200 to process the task.

[Provider]
The providing unit 2004 provides the user with the processing results of the task processing unit 2003. In this embodiment, the providing unit 2004 is realized by the server 20, and therefore the providing unit 2004 provides the processing results to the user by transmitting processing result data indicating the processing results of the task processing unit 2003 to the user terminal 30. Based on the processing result data, the user terminal 30 displays information (e.g., an image or text) indicating the processing results of the task processing unit 2003 on the business support screen SC.

For example, the providing unit 2004 includes an output providing unit 2004A. The output providing unit 2004A provides the user with the output from the business support program selected by the selecting unit 2002, which is obtained by the task processing unit 2003. In this embodiment, since the child AI 201 corresponds to the business support program, the output providing unit 2004A provides the user with the output from the child AI 201 obtained by the task processing unit 2003. In other words, the output providing unit 2004A provides the user with the output obtained by the task processing unit 2003 as is. The output data output from the child AI 201 is an example of processing result data. Instead of the output data becoming processing result data as is, processing result data may be generated based on the output data, as in a modified example described below.

The processing result data may be in any format. In this embodiment, the processing result data may be data necessary for displaying some information on the business support screen SC. For example, if the business support screen SC is displayed in a browser, the processing result data may be data in a markup language such as HTML. If the business support screen SC is displayed in an application dedicated to the business support system 1, the processing result data may be data (e.g., image data or text data) necessary for displaying some information in that application. In the example of Figure 2, the providing unit 2004 displays the answer indicated by the processing result data as a reply to the post made to the user. The user's post is also displayed on the business support screen SC by processing by the server 20.

[3-3. Functions implemented on user terminals]
For example, the user terminal 30 includes a data storage unit 3000, a display control unit 3001, and an input reception unit 3002. The data storage unit 3000 is realized by the storage unit 32. The display control unit 3001 and the input reception unit 3002 are each realized by the control unit 31.

[Data storage unit]
The data storage unit 3000 stores data for business support. For example, the data storage unit 3000 stores a browser for displaying various screens of the business support system 1. For example, the data storage unit 3000 stores a program dedicated to the business support system 1. The business support screen SC may be displayed on a program dedicated to the business support system 1.

[Display control unit]
The display control unit 3001 causes the display unit 35 to display various screens in the business support system 1. For example, the display control unit 3001 causes the display unit 35 to display various screens such as the business support screen SC based on data received from the server 20.

[Input reception section]
The input receiving unit 3002 receives input from the user. For example, the input receiving unit 3002 receives input for various screens such as the business support screen SC. Input data indicating the input received by the input receiving unit 3002 is transmitted to the server 20 as appropriate.

[4. Processing Executed by the Business Support System]
Fig. 6 is a diagram showing an example of processing executed in the business support system 1. The processing in Fig. 6 is executed by the control units 11, 21, and 31 executing programs stored in the storage units 12, 22, and 32, respectively. The processing in Fig. 6 is an example of processing included in a business support method.

For example, the learning terminal 10 performs learning of the parent AI 200 based on the training data of the parent AI 200 stored in the training database DB (S1). For each pre-trained child AI 201, the learning terminal 10 performs learning of the child AI 201 based on the training data of that child AI 201 (S2). The learning terminal 10 transmits the trained parent AI 200 and multiple trained child AIs 201 to the server 20 (S3). The server 20 acquires the parent AI 200 and multiple child AIs 201 from the learning terminal 10 (S4). Thereafter, the user becomes able to use the parent AI 200 and multiple child AIs 201.

The user terminal 30 executes a login process between the user terminal 30 and the server 20 to allow the user to log in to the business support system 1 (S5). The user terminal 30 accepts user input based on a detection signal from the operation unit 34 (S6). The user terminal 30 transmits input data indicating the input accepted in S to the server 20 (S7). The server 20 acquires the input data from the user terminal 30 (S8). In the example of Figure 2, the following process is executed when an AI, including parent AI 200 and child AI 201, is mentioned.

The server 20 inputs input data to the parent AI 200 (S9). The server 20 obtains output data indicating the output from the parent AI 200 (S10). The server 20 determines whether the child AI 201 selected by the parent AI 200 is indicated in the output data (S11). If it is determined in S11 that the child AI 201 is not indicated in the output data (S11: N), this processing ends. In this case, a message indicating that the task was not processed by the child AI 201 may be displayed on the user terminal 30.

If it is determined in S11 that child AI 201 is indicated in the output data (S11: Y), the server 20 causes the child AI 201 to process a task corresponding to the input data (S12). The server 20 acquires output data indicating the output from the child AI 201 (S13). The server 20 executes processing between the server 20 and the user terminal 30 to provide the output from the child AI 201 to the user (S14), and the processing ends. In the example of Figure 2, the processing of S14 results in the answer from the child AI 201 being provided to the user as output.

5. Summary of the embodiment
The business support system 1 of this embodiment can select a child AI 201 specialized for a task corresponding to input data based on the parent AI 200. When a child AI 201 is selected, the business support system 1 processes the task corresponding to the input data based on the selected child AI 201. This allows the business support system 1 to effectively support the user's business, thereby improving user convenience. For example, the user does not need to select a child AI 201 to process a task from multiple child AIs 201, so the business support system 1 can save the user effort. If the user selects the child AI 201 themselves, there is a possibility that the user will select a child AI 201 that is not appropriate for processing the task corresponding to the input data. However, the business support system 1 can select an appropriate child AI 201 using the parent AI 200 to appropriately process the task.

The business support system 1 also provides the user with the output from the child AI 201 selected by the selection unit 2002. By providing the user with the output from the child AI 201 as the processing result of the task according to the input data, the business support system 1 can effectively support the user's business. The user can check the exact content of the output from the child AI 201.

6. Modifications
The present disclosure is not limited to the above-described embodiments, and may be modified as appropriate without departing from the spirit of the present disclosure.

[6-1. Modification 1]
For example, in the embodiment, an example is given in which the output providing unit 2004A provides the output from the child AI 201, which is specialized for a task according to the input data, to the user as is. In variant example 1, an example is given in which the parent AI 200 processes the output from the child AI 201 into an appropriate form. For example, the parent AI 200 modifies the output from the child AI 201 into more general terms. The parent AI 200 may merge the output from the child AI 201 with the processing results of the task it processed. In variant example 1, the output after processing by the parent AI 200 is provided to the user.

Figure 7 is a diagram showing an example of functions implemented by the business support system 1 of variant 1. The business support system 1 of variant 1 includes a processing execution unit 2005. The processing execution unit 2005 is implemented by the control unit 11. The providing unit 2004 of variant 1 includes a processing result providing unit 2004B. The providing unit 2004 of variant 1 does not have to include the output providing unit 2004A. The business support system 1 may include both the output providing unit 2004A and the processing result providing unit 2004B, and these may be used depending on the situation.

The processing execution unit 2005 executes processing on the output obtained by the task processing unit 2003, which is the output from the business support program selected by the selection unit 2002, based on the parent AI 2000. As in the embodiment, variant example 2 also takes the case where the child AI 201 corresponds to the business support program. For example, the processing execution unit 2005 inputs output data from the child AI 201 to the parent AI 200. The parent AI 200 may modify all or part of the output indicated by the output data from the child AI 201, or may add all or part of the input indicated by the input data to the output indicated by the output data from the child AI 201.

In addition, the processing execution unit 2005 may input not only the output data from the child AI 201 but also at least one of input data and prompt data indicating a predetermined prompt to the parent AI 200. The processing execution unit 2005 may execute processing for the output indicated by the output data from the child AI 201 based at least on the output data. The prompt data may indicate the processing content for the output from the child AI 201.

For example, if parent AI 200 is to correct the output from child AI 201 into more general language, the prompt data may include a prompt such as, "Please correct this sentence into more general language." If parent AI 200 is to merge the output from child AI 201 with its own processing results, the prompt data may include a prompt such as, "Please merge this sentence with your processing results." Parent AI 200 may execute processing indicated by the prompt data on the output from the child AI.

The processing result providing unit 2004B provides the user with the processing results of the processing execution unit 2005. The processing result providing unit 2004B provides the user with the processing results by sending processing result data indicating the processing results of the processing execution unit 2005 to the user terminal 30. The user terminal 30 displays information (e.g., an image or text) indicating the processing results of the processing execution unit 2005 on the business support screen SC based on the processing result data. The processing result providing unit 2004B differs from the output providing unit 2004A in that it does not provide the output from the child AI 201 to the user as is, but the method of providing information itself is the same as the output providing unit 2004A.

The business support system 1 of variant example 1 executes processing on the output obtained by the task processing unit 2003, which is the output from the business support program selected by the selection unit 2002, based on the parent AI 200. The business support system 1 provides the user with the processing results by the parent AI 200. This allows the business support system 1 to provide more appropriate information to the user, thereby effectively improving user convenience. For example, if the output from the child AI 201 is unnatural as a sentence, the business support system 1 can have the parent AI 200 correct it to an appropriate sentence and then provide the corrected sentence to the user, thereby improving user convenience. If the business support system 1 merges the output from the child AI 201 and the task processing results by the parent AI 200 and provides them to the user, it can present more information to the user.

[6-2. Modification 2]
For example, if a business support program is not selected by the selection unit 2002, the task processing unit 2003 may process a task corresponding to the input data based on the parent AI 200. In variant example 2, as in the embodiment, an example is given in which the child AI 201 corresponds to the business support program. Therefore, if the child AI 201 is not selected by the selection unit 2002, the task processing unit 2003 processes a task corresponding to the input data based on the parent AI 200.

The parent AI 200 of variant example 2 is not an AI specialized for a specific task, but a general-purpose AI capable of processing a variety of tasks. For example, in the embodiment, an example is given in which the parent AI 200 is created by relearning a pre-trained large-scale language model, but if the pre-trained large-scale language model is a Transformer-based model such as GPT, the parent AI 200 created by pre-training also has general-purpose properties. The parent AI 200 may also be a multi-tasking AI capable of handling multiple specific tasks.

For example, the task processing unit 2003 inputs input data to the parent AI 200 selected by the selection unit 2002. The parent AI 200 calculates an embedded expression based on the input data and outputs output data corresponding to the embedded expression. The parent AI 200 outputs output data corresponding to the embedded expression. The output of the output data corresponds to processing of a task corresponding to the input data. The selection unit 2002 acquires the output data output from the parent AI 200. A series of processes of the parent AI 200 may be executed based on parameters adjusted by learning. For example, when re-learning is performed to make the parent AI 200 more versatile, a series of processes of the parent AI 200 is executed based on the parameters adjusted by re-learning.

In variant 2, if the selection unit 2002 does not select a child AI 201, the business support system 1 processes a task according to the input data based on the parent AI 200. This prevents the business support system 1 from being unable to provide any information to the user when the selection unit 2002 does not select a business support program, and allows the business support system 1 to provide some kind of information to the user. For example, if a user requests a task that cannot be handled by the child AI 201, the business support system 1 can process the task based on the general-purpose parent AI 200 and provide the user with the processing results.

[6-3. Modification 3]
For example, in the embodiment, a case has been described in which each of the multiple business support programs is a child AI 201. However, the multiple business support programs may include a child AI 201 that is an AI different from the parent AI 200, and a non-AI program that is a non-AI. A non-AI program is a program that is not classified as an AI. A non-AI program receives all or part of input data. A non-AI program outputs output data based on all or part of the input data input to the non-AI program. The non-AI program is assumed to be stored in the data storage unit 2000, but the non-AI program may also be stored in another computer. The non-AI program may be a simple program such as a program that obtains the current date. Even such a simple program can process the task of obtaining the current date.

For example, a non-AI program may be a machine translation program not classified as AI, a summary creation program not classified as AI, a research program not classified as AI (for example, a search engine program not classified as AI), a schedule adjustment program not classified as AI, or other programs. These non-AI programs may be programs used in known groupware or known business support services not classified as groupware. A non-AI program may also be the rule-based program mentioned above.

When a child AI 201 is selected by the selection unit 2002, the task processing unit 2003 of variant example 3 processes a task corresponding to the input data based on the child AI 201. The processing of the task processing unit 2003 in this case is as described in the embodiment. When a non-AI program is selected by the selection unit 2002, the task processing unit 2003 processes a task corresponding to the input data based on the non-AI program. The task processing unit 2003 inputs all or part of the input data to the non-AI program. The processing of the non-AI program may be processing adopted in known groupware or known business support services that are not classified as groupware.

For example, a non-AI program executes processing on all or part of the input data entered into it, based on the program code contained in the non-AI program. The non-AI program outputs output data indicating the execution result of the program code. The task processing unit 2003 acquires the output data output from the non-AI program. If the non-AI program is stored in a computer other than the server 20, the task processing unit 2003 may send input data to the other computer to request processing of a task by the non-AI program. The task processing unit 2003 acquires processing result data indicating the processing result of the task by the non-AI program from the other computer.

When a child AI 201 is selected by the selection unit 2002, the business support system 1 of variant example 3 processes a task corresponding to the input data based on the child AI 201. When a non-AI program is selected by the selection unit 2002, the business support system 1 processes a task corresponding to the input data based on the non-AI program. This allows the business support system 1 to have the appropriate one of the child AI 201 and the non-AI program process the task corresponding to the input data.

[6-4. Modification 4]
For example, in variant example 3, there may be a child AI 201 and a non-AI program specialized in the same task. For example, not only child AI 201A specialized in machine translation, but also a non-AI program not classified as AI may be able to process the task of machine translation. Not only child AI 201B specialized in summarization, but also a non-AI program not classified as AI may be able to process the task of summarization. Not only child AI 201C specialized in research, but also a non-AI program not classified as AI may be able to process the task of research. Not only child AI 201D specialized in schedule adjustment, but also a non-AI program not classified as AI may be able to process the task of schedule adjustment.

The selection unit 2002 of variant 4 selects both the child AI 201 and the non-AI program if both are capable of processing tasks according to the input data. When the selection unit 2002 selects both the child AI 201 and the non-AI program, the task processing unit 2003 processes the task according to the input data based on both the child AI 201 and the non-AI program, and may select either the output from the child AI 201 or the output from the non-AI program. Task processing by each of the child AI 201 and the non-AI program is as described in the embodiment or variant 3.

The task processing unit 2003 selects either the output from the child AI 201 or the output from the non-AI program based on a predetermined selection method. The selection method may be any predetermined method. In variant example 4, an example of a selection method is to have the parent AI 200 select the appropriate one from the output from the child AI 201 or the output from the non-AI program. For example, the training data for the parent AI 200 indicates the relationship between two training outputs and the correct output. The learning unit 1001 executes learning of the parent AI 200 so that when the two training outputs indicated in the input portion of the training data are input to the parent AI 200, the parent AI 200 selects and outputs the correct output indicated in the output portion of the training data. The learning algorithm may be any of a variety of well-known algorithms, as in the embodiment.

For example, the task processing unit 2003 inputs the output from the child AI 201 and the output from the non-AI program to the parent AI 200. The parent AI 200 calculates embedded representations for each of the output from the child AI 201 and the output from the non-AI program, and selects the appropriate one based on these embedded representations. Depending on the embedded representations of each of the output from the child AI 201 and the output from the non-AI program, the parent AI 200 may select both of them, or may select only one of them. The parent AI 200 outputs output data indicating the selection result of at least one of the output from the child AI 201 and the output from the non-AI program.

For example, the task processing unit 2003 obtains output data output from the parent AI 200. The providing unit 2004 provides the processing results to the user based on the selection result indicated by the output data. For example, the providing unit 2004 provides the processing results to the user based on at least one of the output from the child AI 201 and the output from a non-AI program selected by the parent AI 200.

Note that the selection method of the task processing unit 2003 is not limited to the method using the parent AI 200. The task processing unit 2003 may select either the output from the child AI 201 or the output from the non-AI program based on other selection methods. For example, if the child AI 201 is capable of calculating a score indicating the accuracy of the output and the non-AI program is also capable of calculating a score indicating the accuracy of the output, the task processing unit 2003 may select the output from the child AI 201 or the output from the non-AI program, whichever has the higher score. As another example, the task processing unit 2003 may select the output from the child AI 201 or the output from the non-AI program, whichever has the longer sentence.

When both the child AI 201 and the non-AI program are selected by the selection unit 2002, the business support system 1 of variant example 4 processes a task according to the input data based on both the child AI 201 and the non-AI program, and selects either the output from the child AI 201 or the output from the non-AI program. This allows the business support system 1 to provide the user with more appropriate processing results.

[6-5. Modification 5]
For example, the multiple business support programs may include a public program that processes public data that can be made public, and a private program that processes private data that cannot be made public.In variant example 5, as in the embodiment, an example is given in which child AI 201 corresponds to the business support program.Child AI 201 that processes public data is an example of a public program.Child AI 201 that processes private data is an example of a private program.Child AI 201 that processes public data may exist on a computer other than server 20.Child AI 201 that processes private data completes information processing in a secure environment within server 20.

Public data is data that may be made public outside the business support system 1. For example, data that may remain as a log outside the business support system 1 corresponds to public data. Public data is a portion of the data registered in the business support system 1. For example, data other than confidential data of the organization to which the user belongs corresponds to public data. Data that shows information posted on general websites such as Internet encyclopedias or homepages corresponds to public data. Classification data that indicates which data is classified as public data is assumed to be stored in advance in the data storage unit 2000.

Private data is data that must not be made public outside the business support system 1. For example, data that must not remain as a log outside the business support system 1 corresponds to private data. Private data is data registered in the business support system 1 other than public data. For example, among schedule data registered in the schedule management function, data such as the names of people outside the organization to which the user belongs corresponds to private data. Another example is confidential data of the organization to which the user belongs. Classification data indicating which data is classified as private data is assumed to be stored in advance in the data storage unit 2000. The child AI 201 of variant 5 references at least one of public data and private data to process tasks.

In variant example 5, the output portion of the training data learned by parent AI 200 is child AI identification data that indicates either child AI 201 that processes public data or child AI 201 that processes private data. For example, if it is sufficient to make public data public and not necessary to make private data public in order to process a task corresponding to certain training input data, the child AI identification data paired with the training input data indicates child AI 201 that processes public data. For example, if it is necessary to make private data public as well as public in order to process a task corresponding to certain training input data, the child AI identification data paired with the training input data indicates child AI 201 that processes private data.

The parent AI 200 of variant 5 has learned higher-level training data, and is therefore able to estimate what type of input data should result in the selection of a child AI 201 that processes public data or a child AI 201 that processes private data. For example, the selection unit 2002 inputs input data to the parent AI 200. The parent AI 200 calculates an embedded representation of the input data and, based on the embedded representation, outputs output data indicating a child AI 201 that processes public data or a child AI 201 that processes private data. As described in the embodiment, the parent AI 200 may not select any child AI 201.

When a public program is selected by the selection unit 2002, the task processing unit 2003 of variant 5 processes a task corresponding to the input data based on the public program and public data. Variation 5 takes the example of a case where the child AI 201 that processes the public data corresponds to the public program, so the task processing unit 2003 inputs the input data to the child AI 201 that processes the public data. The child AI 201 calculates an embedded representation of the input data and acquires the public data based on the embedded representation. Instead of the child AI 201 acquiring the public data, the parent AI 200 may acquire the public data, and the selection unit 2002 may input the public data to the child AI 201. The child AI 201 leaves the public data as a log outside the business support system 1, and then processes a task corresponding to the input data. While variant 5 differs from the embodiment and other variants in that the public data is left as a log outside the business support system 1, the flow of task processing by the child AI 201 is the same as in the embodiment and other variants.

When a private program is selected by the selection unit 2002, the task processing unit 2003 of variant example 5 processes a task corresponding to the input data based on the private program and private data. Variation example 5 takes the example of a child AI 201 that processes private data as a private program, so the task processing unit 2003 inputs the input data to the child AI 201 that processes the private data. The child AI 201 calculates an embedded representation of the input data and acquires the private data based on the embedded representation. Instead of the child AI 201 acquiring the private data, the parent AI 200 may acquire the private data, and the selection unit 2002 may input the private data to the child AI 201. The child AI 201 processes a task corresponding to the input data without leaving the private data as a log outside the business support system 1. While variant example 5 differs from the embodiment and other variant examples in that the private data is not left as a log outside the business support system 1, the flow of task processing by the child AI 201 is the same as in the embodiment and other variant examples.

When a public program is selected by the selection unit 2002, the business support system 1 of variant example 5 processes a task corresponding to the input data based on the public program and public data. When a private program is selected by the selection unit 2002, the business support system 1 processes a task corresponding to the input data based on the private program and private data. This allows the business support system 1 to reliably prevent confidential data from leaking to the outside. For example, the business support system 1 can use child AIs 201 that can process the same task, such as a child AI 201 that processes private data and a child AI 201 that processes public data.

[6-6. Modification 6]
For example, if there are many child AIs 201 specialized in specific tasks, it may be impossible to select an appropriate child AI 201 with only one parent AI 200. For this reason, three or more hierarchies may exist in the AI, such as parent, child, and grandchild. In variant 6, three hierarchies are given as an example, but the number of hierarchies may be four or more. The AI at the lowest hierarchical level is an AI specialized in a specific task. The AIs other than the lowest hierarchical level are AIs that can select the AI at the hierarchical level below them. The AIs other than the lowest hierarchical level may be general-purpose AIs as described in variant 2.

Figure 8 is a diagram showing an example of an AI in variant 6. In Figure 8, there is a parent AI 200, multiple child AIs 201, and multiple grandchild AIs 202. The parent AI 200 can select at least one child AI 201 based on input data. As with the embodiment, the training data for the parent AI 200 indicates the relationship between the training input data and the child AI identification data of the child AI 201 that is the correct answer. However, variant 6 differs from the embodiment in that the child AI 201 is not an AI specialized for a specific task, but is an AI that can select a grandchild AI 202. Because the parent AI 200 in variant 6 has learned such training data, it can select an appropriate child AI 201 according to the input data.

The child AI 201 of variant 6 can select at least one grandchild AI 202 based on input data. In variant 6, the child AI 201 has the same functions as the parent AI 200 described in the embodiment and variants 1 to 5. The training data for the child AI 201 is the same as the training data for the parent AI 200 described in the embodiment. However, the output portion of the training data is grandchild AI identification data for identifying the grandchild AI 202. The grandchild AI identification data differs from the child AI identification data in that it is data for identifying the grandchild AI 202, but is similar to the child AI identification data in other respects. Because the child AI 201 of variant 6 has learned such training data, it can select an appropriate child AI 201 according to the input data.

For example, grandchild AI202 has the same functions as child AI201 described in the embodiment and variants 1 to 5. The description of child AI201 in the embodiment and variants 1 to 5 can be read as grandchild AI202. In the example of Figure 8, there are multiple grandchild AI202 specialized in the same task. For example, grandchild AI202 specialized in the task of machine translation includes grandchild AI202A specialized in English translation and grandchild AI202B specialized in Chinese translation. Grandchild AI202A has learned English training data. Grandchild AI202B has learned Chinese training data. Child AI201A can estimate which of grandchild AI202A and 202B is appropriate based on the input data and select the appropriate one. Note that parent AI200 estimates that the processing of child AI201A is appropriate based on the input data and selects child AI201A. Child AI 201A learns training input data to be translated as training data.

For example, grandchild AI202 specialized in the task of creating summaries include grandchild AI202C specialized in creating summaries for threads and grandchild AI202D specialized in creating summaries for emails. Grandchild AI202C has learned training data indicating thread posts. Grandchild AI202D has learned training data indicating the text of emails. Child AI201B can estimate which of grandchild AI202C or 202D is appropriate based on the input data and select the appropriate one. Note that parent AI200 estimates that child AI201B's processing is appropriate based on the input data and selects child AI201B. Child AI201B has learned training input data for creating summaries as training data. Similarly for other tasks, the grandchild AI202 that will process the final task is selected by child AI201.

The selection unit 2002 in variant example 6 selects a child AI 201 from which a business support program can be selected. The selection unit 2002 selects a business support program based on the selected child AI 201. In the example of Figure 8, the selection unit 2002 may correspond to the parent AI 200. If the selection unit 2002 does not correspond to the parent AI 200, the selection unit 2002 may select the child AI 201 based on the parent AI 200. For example, the selection unit 2002 selects a business support program based on a child AI 201 selected by the parent AI 200, which is another AI that can select any child AI 201, from among multiple child AIs 201. In variant example 6, a case will also be described in which the grandchild AI 202 corresponding to the child AI 201 described in the embodiment corresponds to a business support program. At least one of the parent AI 200 and the child AI 201 in variant example 6 may be a versatile AI. Note that the selection of the child AI 201 may be performed not by an AI such as the parent AI 200, but by a program that is not classified as an AI.

For example, the selection unit 2002 inputs input data to the parent AI 200. The parent AI 200 calculates an embedded expression based on the input data and outputs output data corresponding to the embedded expression. If a child AI 201 corresponding to the embedded expression exists, the parent AI 200 outputs the child AI identification data of the child AI 201 as output data. If a child AI 201 corresponding to the embedded expression does not exist, the parent AI 200 outputs output data indicating that the child AI 201 has not been selected. The selection unit 2002 acquires the output data output from the parent AI 200. The series of processes of the parent AI 200 are executed based on parameters adjusted by learning.

For example, the selection unit 2002 inputs input data to the child AI 201 indicated by the output data. The parent AI 200 calculates an embedded expression based on the input data and outputs output data corresponding to the embedded expression. If a grandchild AI 202 corresponding to the embedded expression exists, the parent AI 200 outputs the child AI identification data of the grandchild AI 202 as output data. If a grandchild AI 202 corresponding to the embedded expression does not exist, the parent AI 200 outputs output data indicating that the grandchild AI 202 was not selected. The selection unit 2002 acquires the output data output from the parent AI 200. A series of processes by the child AI 201 is executed based on parameters adjusted by learning.

The processing of the task processing unit 2003 after the grandchild AI 202 is selected may be the same as in the embodiment. The task processing unit 2003 may cause the grandchild AI 202 to process the task that was described in the embodiment as being processed by the child AI 201. The task processing unit 2003 inputs input data to the grandchild AI 202 selected by the selection unit 2002. The task processing unit 2003 obtains output data from the grandchild AI 202.

The business support system 1 of variant example 6 selects a child AI 201 that can select a grandchild AI 202. This allows the business support system 1 to appropriately process tasks according to input data even if the number of business support programs, including grandchild AI 202, increases.

[6-7. Modification 7]
For example, when multiple business support programs are selected by the selection unit 2002, the task processing unit 2003 may process a task corresponding to input data based on the multiple business support programs and merge the outputs from each of the multiple business support programs. In the seventh variant, as in the embodiment, an example is given in which the child AI 201 corresponds to the business support program. When the sixth variant and the seventh variant are combined, the grandchild AI 202 corresponds to the business support program. Merging the outputs may mean that the task processing unit 2003 simply combines multiple outputs, or may mean that the task processing unit 2003 generates a new output based on the multiple outputs.

For example, when multiple child AIs 201 are selected by the selection unit 2002, the task processing unit 2003 causes each of the multiple child AIs 201 to process a task according to the input data. Task processing by each child AI 201 is as described in the embodiment. The task processing unit 2003 combines the outputs from each of the multiple child AIs 201 to generate a single output. For example, the task processing unit 2003 may merge the outputs from each of the multiple child AIs 201 by ANDing them, or may merge the outputs from each of the multiple child AIs 201 by ORing them. The providing unit 2004 provides the processing result to the user based on the single output data generated by the task processing unit 2003.

For example, suppose the task corresponding to the input data is schedule adjustment. Furthermore, suppose the user's schedule is managed separately by multiple schedule management functions. The task processing unit 2003 merges the output from a child AI 201 that adjusts the schedule based on a schedule managed by one schedule management function with the output from a child AI 201 that adjusts the schedule based on a schedule managed by another schedule management function. The task processing unit 2003 can obtain an output that indicates the user's free time by merging these outputs to perform an AND operation on the user's free time indicated by these outputs. Merging can be performed similarly for other tasks according to the child AI 201.

When multiple child AIs 201 are selected by the selection unit 2002, the business support system 1 of variant 7 processes tasks according to the input data based on the multiple child AIs 201 and merges the output from each of the multiple child AIs 201. This allows the business support system 1 to merge processing results from more child AIs 201, thereby providing useful information to the user. For example, even if a user's schedule is managed separately by multiple schedule management functions, the business support system 1 can present the user with times when they are definitely free by having each of the multiple child AIs 201 check the user's availability managed by these schedule management functions and merging the output.

[6-8. Modification 8]
For example, when multiple business support programs are selected by the selection unit 2002, the task processing unit 2003 may process a task according to the input data by inputting the output from some of the multiple business support programs to other of the multiple business support programs. In the same manner as in the embodiment, the eighth variant also takes as an example a case in which the child AI 201 corresponds to the business support program. When the sixth variant and the eighth variant are combined, the grandchild AI 202 corresponds to the business support program.

For example, the task processing unit 2003 may input output data from child AI 201A specialized in the task of machine translation to child AI 201B specialized in the task of summarizing. In this case, child AI 201B creates a summary in a language other than the input language indicated by the input data. The task processing unit 2003 acquires the output data output by child AI 201B. As another example, the task processing unit 2003 may input output data from child AI 201B specialized in the task of summarizing to child AI 201C specialized in the task of research. In this case, child AI 201C performs research according to the summary of the input indicated by the input data. The task processing unit 2003 acquires the output data output by child AI 201C.

In addition, the question of which child AI 201 the task processing unit 2003 should input the output from which child AI 201 to which child AI 201 may be predetermined or may be determined by the parent AI 200. When the decision is made by the parent AI 200, the training data of the parent AI 200 includes training data regarding which child AI 201 the output from which child AI 201 should be input to. By having the parent AI 200 learn such training data, the parent AI 200 can estimate which child AI 201 the output from which child AI 201 should be input to. The task processing unit 2003 simply performs input and output in the flow estimated by the parent AI 200 to obtain the final output.

When multiple child AIs 201 are selected by the selection unit 2002, the business support system 1 of variant 8 processes a task according to the input data by inputting the output from some of the multiple child AIs 201 to other child AIs 201. This allows the business support system 1 to link multiple child AIs 201, thereby improving the accuracy of task processing.

[6-9. Other Modifications]
For example, two or more of Modifications 1 to 8 may be combined. For example, although the case where AI-based business support is provided when a user uses a communication function such as a thread has been exemplified, the business support system 1 may also provide AI-based business support when a user uses another business support function. In this case, the business support system 1 may provide business support by executing the same processes as those in the embodiment and Modifications 1 to 8.

For example, functions described as being implemented by the server 20 may be implemented by the user terminal 30. In this case, the functions may be implemented by a browser script or an application installed on the user terminal 30. For example, each function may be shared among multiple computers, or may be implemented by a single computer.

1 Business support system, 10 Learning terminal, 11, 21, 31 Control unit, 12, 22, 32 Memory unit, 13, 23, 33 Communication unit, 14, 34 Operation unit, 15, 35 Display unit, 20 Server, 30 User terminal, N Network, DB Training database, SC Business support screen, F10 Input form, 1000 Data memory unit, 1001 Learning unit, 2000 Data memory unit, 2001 Input data acquisition unit, 2002 Selection unit, 2003 Task processing unit, 2004 Provision unit, 2004A Output provision unit, 2004B Processing result provision unit, 2005 Processing execution unit, 3000 Data memory unit, 3001 Display control unit, 3002 Input reception unit, 200 Parent AI, 201, 201A, 201B, 201C, 201D Child AI, 202 Grandchild AI.

Claims (12)

an input data acquisition unit that acquires input data indicating a prompt input by a user;
a selection unit that can select a business assistance program based on the input data and AI (Artificial Intelligence) , which is a large-scale language model that can select a business assistance program specialized for a task corresponding to the input data from among a plurality of business assistance programs specialized for a specific task in business assistance;
a task processing unit that processes the task corresponding to the input data based on the input data and the business assistance program when the business assistance program is selected by the selection unit;
Business support system including. the task support system further includes an output providing unit that provides the user with an output from the task support program selected by the selection unit and acquired by the task processing unit.
The business support system according to claim 1 . The business support system includes:
a processing execution unit that executes, based on the AI, processing on the output from the task assistance program selected by the selection unit and acquired by the task processing unit;
a processing result providing unit that provides the user with a processing result of the processing execution unit;
The business support system according to claim 1 or 2, further comprising: When the business assistance program is not selected by the selection unit, the task processing unit processes the task according to the input data based on the AI.
The business support system according to claim 1 or 2. The plurality of business support programs include an AI different from the AI and a non-AI program that is a non-AI,
The task processing unit
When the different AI is selected by the selection unit, the task according to the input data is processed based on the different AI;
When the non-AI program is selected by the selection unit, the task according to the input data is processed based on the non-AI program.
The business support system according to claim 1 or 2. When both the different AI and the non-AI program are selected by the selection unit, the task processing unit processes the task according to the input data based on both the different AI and the non-AI program, and selects either an output from the different AI or an output from the non-AI program.
The business support system according to claim 5 . The plurality of business support programs include a public program that processes public data that can be made public, and a private program that processes private data that cannot be made public,
The task processing unit
When the public program is selected by the selection unit, the task corresponding to the input data is processed based on the public program and the public data;
When the non-public program is selected by the selection unit, the task corresponding to the input data is processed based on the non-public program and the non-public data.
The business support system according to claim 1 or 2. The selection unit selects the AI that can select the business assistance program.
The business support system according to claim 1 or 2. when the selection unit selects a plurality of the business assistance programs, the task processing unit processes the task corresponding to the input data based on the plurality of business assistance programs and merges outputs from the plurality of business assistance programs;
The business support system according to claim 1 or 2. when the selection unit selects a plurality of the business assistance programs, the task processing unit processes the task according to the input data by inputting an output from some of the business assistance programs to the other of the plurality of business assistance programs;
The business support system according to claim 1 or 2. Get input data indicating the prompts entered by the user;
The business assistance program can be selected based on the input data and AI (Artificial Intelligence) , which is a large-scale language model that can select a business assistance program specialized for a task corresponding to the input data from among a plurality of business assistance programs specialized for a specific task in business assistance;
When the business support program is selected, the task corresponding to the input data is processed based on the input data and the business support program .
Business support methods. an input data acquisition unit that acquires input data indicating a prompt entered by a user;
a selection unit capable of selecting a business assistance program based on the input data and AI (Artificial Intelligence) , which is a large-scale language model capable of selecting a business assistance program specialized for a task corresponding to the input data from among a plurality of business assistance programs specialized for a specific task in business assistance;
a task processing unit that processes the task corresponding to the input data based on the input data and the business assistance program when the business assistance program is selected by the selection unit;
A program that allows a computer to function as a