Relearning system, relearning method, and program

Abstract

Abstract: A large-scale language model acquisition unit (101) of a relearning system (1) according to the present invention acquires a pre-trained large-scale language model. A training data acquisition unit (102) acquires training data for relearning of the large-scale language model, the training data having been created on the basis of activity data indicating activity of a user performed by a business support system that supports business of the user. A relearning unit (103) creates a business support model specific to the business support system, by executing relearning of the large-scale language model on the basis of the training data.

Description

Re-learning system, re-learning method, and program

[0001] The present disclosure relates to a relearning system, a relearning method, and a program.

[0002] Conventionally, business support systems that support user business operations have been known. For example, Patent Literature 1 describes, as an example of a business support system, groupware that supports communication between users belonging to an organization. The business support system of Patent Literature 1 inputs data indicating a user's usage pattern of the groupware to a trained model that estimates whether or not a no-input period will occur, during which the user does not perform any operation. The trained model estimates whether or not a no-input period will occur based on the data.

[0003] Japanese Patent Application Laid-Open No. 2023-091326

[0004] However, large-scale language models used in the field of natural language processing need to analyze a huge number of languages, unlike trained models that perform simple classification such as those in Patent Document 1. For this reason, conventional technologies such as those in Patent Document 1 have not been able to realize large-scale language models that are useful for business support systems.

[0005] One of the objectives of the present disclosure is to realize a large-scale language model that is useful for business support systems.

[0006] The re-learning system according to the present disclosure includes a large-scale language model acquisition unit that acquires a pre-trained large-scale language model, a training data acquisition unit that acquires training data for re-learning the large-scale language model, the training data being created based on activity data indicating a user's activity performed in a business support system that supports the user's business, and a re-learning unit that creates a business support model specific to the business support system by re-learning the large-scale language model based on the training data.

[0007] The present disclosure makes it possible to realize a large-scale language model that is useful for business support systems.

[0008] FIG. 1 is a diagram illustrating an example of the hardware configuration of a relearning system. FIG. 2 is a diagram illustrating an example of a screen displayed on a user terminal. FIG. 3 is a diagram illustrating an example of the relationship between a large-scale language model and a task support model. FIG. 4 is a diagram illustrating an example of functions realized in the relearning system. FIG. 5 is a diagram illustrating an example of a training database. FIG. 6 is a diagram illustrating an example of a user database. FIG. 7 is a diagram illustrating an example of processing executed in the relearning system. FIG. 8 is a diagram illustrating an example of multi-stage relearning. FIG. 9 is a diagram illustrating an example of functions realized in Modification 4.

[0009] [1. Hardware Configuration of the Re-learning System] An example of an embodiment of a relearning system, a relearning method, and a program according to the present disclosure will be described. FIG. 1 is a diagram showing an example of the hardware configuration of a relearning system. For example, the relearning system 1 includes a relearning terminal 10 and a business assistance system 2. The business assistance system 2 includes a business assistance server 20 and a user terminal 30. Each of the relearning terminal 10, the business assistance server 20, and the user terminal 30 is connected to a network N such as the Internet or a LAN.

[0010] The relearning terminal 10 is a computer that executes the relearning process described below. For example, the relearning terminal 10 is a personal computer, a server computer, or a tablet. For example, the relearning terminal 10 includes a control unit 11, a memory unit 12, a communication unit 13, an operation unit 14, and a display unit 15. For example, the control unit 11 includes at least one processor. The memory unit 12 includes at least one of a volatile memory such as RAM and a 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, a touch panel, or a keyboard. The display unit 15 is an LCD or organic EL display.

[0011] The business support server 20 is a server computer. For example, the business support server 20 includes a control unit 21, a storage unit 22, and a communication unit 23. The hardware configurations of the control unit 21, the storage unit 22, and the communication unit 23 may be similar to those of the control unit 11, the storage unit 12, and the communication unit 13, respectively.

[0012] The user terminal 30 is a user's computer. For example, the user terminal 30 is a personal computer, a tablet, 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.

[0013] The program stored in at least one of the memory units 12, 22, and 32 may be supplied to at least one of the relearning terminal 10, the business support server 20, and the user terminal 30 via the network N. Furthermore, the hardware configuration of at least one of the relearning terminal 10, the business support server 20, and the user terminal 30 is not limited to the example of FIG. 1 . For example, at least one of the relearning terminal 10, the business support server 20, and the user terminal 30 may include at least one of a reading unit (e.g., a memory card slot) that reads an information storage medium and an input / output unit (e.g., a USB terminal) that connects to the information storage medium. In this case, the program stored in the information storage medium may be supplied to at least one of the relearning terminal 10, the business support server 20, and the user terminal 30 via at least one of the reading unit and the input / output unit.

[0014] Furthermore, the computers included in the relearning system 1 are not limited to the example in FIG. 1. The relearning system 1 may include at least one computer. For example, the relearning system 1 may include only the relearning terminal 10. In this case, the business support system 2 exists outside the relearning system 1. For example, the relearning system 1 may include a computer not shown in FIG. 1.

[0015] [2. Overview of the Re-learning System] In this embodiment, the case where the business support system 2 is groupware is taken as an example. For example, the business support system 2 supports the business of a user who belongs to an organization such as a company or a government agency. The user uses the business support function of the business support system 2. The business support function is a function possessed by a program developed for business support (for example, a program stored in the business support server 20 or a program stored in the user terminal 30).

[0016] The business support function may be any of various known functions. For example, the business support function may be a thread function for users to communicate in threads, a schedule function for users to manage schedules, a database function for users to manage databases, a file management function for users to manage files, or an email management function for users to manage emails. The business support function may also be any other function implemented in known groupware.

[0017] The business support system 2 may be a system that is not classified as groupware. For example, the business support system 2 may be a system that supports business unrelated to an organization, a system that does not have a function for linking with other products, or any other system. The business support system 2 may be any system that supports some kind of business of the user.

[0018] 2 is a diagram showing an example of a screen displayed on the user terminal 30. For example, when a user operates the user terminal 30 to log in to the business support system 2, the user terminal 30 displays a business support screen SC on the display unit 35 to support the user's business. The user can use any business support function from the business support screen SC. The example of FIG. 2 shows the business support screen SC when the user uses the thread function. The user can post any message to the thread displayed on the business support screen SC.

[0019] In addition to the threads shown in FIG. 2 , various texts are registered in the business support system 2. If the business support system 2 could perform language analysis of the text registered therein, it would be possible to effectively support the user's business. For example, if the business support system 2 could analyze the text registered in a thread and create a summary of the thread, the user would be able to easily understand the topic of the thread. Similarly, language analysis of text would be useful for business support functions other than the thread function. However, it would be extremely time-consuming for the operating company of the business support system 2 to create a large-scale language model for language analysis from scratch.

[0020] Therefore, the retraining system 1 performs retraining of the pre-trained large-scale language model to create a business support model specific to the business support system 2. The pre-trained large-scale language model may be a known model used in the field of natural language processing. For example, the large-scale language model may be a recurrent neural network model, a long short-term memory network model, or a Transformer model.

[0021] In this embodiment, a case where the large-scale language model is a Transformer-type model is taken as an example. For example, the large-scale language model may be a Generative Pre-trained Transformer (GPT) or a Bidirectional Encoder Representations from Transformers (BERT) model. Pathways Language Model (PaLM) or Large Language Model Meta AI (LLaMA), which may be classified as a Transformer-type model, may be used as the large-scale language model. The large-scale language model may also be another model that utilizes a machine learning technique.

[0022] For example, when a text is input, a large-scale language model divides the text into multiple tokens. Tokens are elements that allow the large-scale language model to understand the text. For example, a token is a word, a morpheme, a phrase, or a character. The large-scale language model converts each of the multiple tokens into an embedding representation. The large-scale language model understands the context of the text input to it based on the arrangement of the embedding representations of each of the multiple tokens. The large-scale language model provides output according to its task based on the context of the text. The large-scale language model performs this series of processes based on parameters of each layer, such as the intermediate layer and the output layer. A large-scale language model may have hundreds of millions to hundreds of billions of parameters.

[0023] For example, pre-training of a large-scale language model is performed by another computer not included in the relearning system 1. The operating company of the business support system 2 acquires, for a fee or free of charge, a large-scale language model that has been pre-trained by another computer. Pre-training of a large-scale language model can support a user's business to some extent. However, because pre-training of a large-scale language model is performed based on general-purpose data unrelated to business support, it is not capable of performing language analysis specific to business support. Therefore, the relearning system 1 re-trains a large-scale language model based on data specific to business support.

[0024] Retraining is a process for changing at least a part of a large-scale language model. For example, retraining is adjusting the parameters of a large-scale language model. Retraining may also be replacing some layers included in a large-scale language model. In this embodiment, an example is given in which fine-tuning corresponds to retraining. For example, fine-tuning is adjusting the parameters of a large-scale language model so that they are suited to a predetermined task. The meaning of fine-tuning may be any known meaning, with various theories. Fine-tuning is not limited to the example in this embodiment.

[0025] Note that relearning is not limited to fine tuning. Re-learning may be any process for changing at least a part of a large-scale language model. For example, relearning may be transfer learning or distillation. There are various theories about the meaning of transfer learning or distillation, but transfer learning or distillation may have any of the various well-known meanings. Re-learning may be a process other than fine tuning, transfer learning, and distillation. Hereinafter, the large-scale language model after relearning will be referred to as a business support model. Where simply referred to as a large-scale language model, it means a pre-trained large-scale language model (a large-scale language model before relearning).

[0026] FIG. 3 is a diagram showing an example of the relationship between a large-scale language model and a business support model. For example, pre-learning of the large-scale language model M1 is performed so that the large-scale language model M1 can process a specific task. The task of this embodiment is different from a general task and is a task for business support. Hereinafter, the task of this embodiment will be referred to as a business support task. In the example of FIG. 3, creating a summary of a thread corresponds to the business support task.

[0027] The business support task is not limited to creating a summary of a thread. For example, the business support task may be creating a summary of text other than the message of a thread (for example, the body of an email, a comment posted using a comment function of a database, or text on a website such as a help page), automatically creating some kind of text, machine translation, answering a user question, extracting data registered in the business support system 2, classifying data registered in the business support system 2, analyzing user behavior, predicting user behavior, or other tasks.

[0028] For example, suppose that thread data indicating a thread posted to the business support system 2 is input to the large-scale language model M1. The large-scale language model M1 analyzes the text of the thread indicated by the thread data based on parameters adjusted by pre-training. The large-scale language model M1 outputs summary data indicating a summary of the thread based on the analysis result of the text. Because the pre-training of the large-scale language model M1 is performed based on general-purpose data, the large-scale language model M1 can create summaries with a certain degree of accuracy, but may not be able to create summaries specific to business support.

[0029] For example, suppose that in a typical summary, the beginning and end of the text to be summarized are important. In this case, pre-training of the large-scale language model M1 is performed so that the large-scale language model M1 creates a summary based on the beginning and end of the text. The large-scale language model M1 creates a summary of the thread based on the beginning and end of the thread indicated by the thread data input to the large-scale language model M1.

[0030] On the other hand, suppose that in a summary for business support, the part including the date and time of the meeting and the text entered by the high-ranking user are more important than the beginning and end of the text. In this case, if these contents are not included at the beginning and end of the text, the summary created by the large-scale language model M1 may not be very useful. For example, the summary created by the large-scale language model M1 may not include the date and time of the meeting and may not include the text entered by the high-ranking user.

[0031] Therefore, the relearning system 1 of this embodiment performs relearning of the large-scale language model M1 based on activity data indicating user activity in the business support system 2. The activity is an input from the user to the business support system 2. Since the activity data includes information useful for business support, a business support model M2 useful for business support is created by relearning. In the example of FIG. 3 , thread data corresponds to the activity data. Examples of activity data other than thread data will be described later.

[0032] For example, the business support model M2 has learned training data that indicates the relationship between a thread prepared for relearning and a correct summary created from the thread. In this embodiment, an example is taken of a case where the operating company of the business support system 2 creates a correct summary. For example, the training data is created so that, among messages posted to a thread, a message including the date and time of a meeting and a message from a user with a high position are included in the correct summary. By performing relearning based on such training data, the business support model M2 becomes able to create summaries specific to business support.

[0033] As described above, the relearning system 1 creates a business assistance model M2 specific to business assistance by relearning the large-scale language model M1 that has been pre-trained based on general-purpose data. For example, it is not realistic for the operating company of the business assistance system 2 to train the business assistance model M2 from scratch, but the relearning system 1 can realize a business assistance model M2 that is useful for the business assistance system 2 by relearning the large-scale language model M1. Details of the relearning system 1 will be described below.

[0034] 3. Functions Realized by the Re-learning System FIG. 4 is a diagram showing an example of functions realized by the re-learning system 1. As shown in FIG.

[0035] [3-1. Functions Realized in the Re-learning Terminal] For example, the re-learning terminal 10 includes a data storage unit 100, a large-scale language model acquisition unit 101, a training data acquisition unit 102, and a re-learning unit 103. The data storage unit 100 is realized by the storage unit 12. Each of the large-scale language model acquisition unit 101, the training data acquisition unit 102, and the re-learning unit 103 is realized by the control unit 11.

[0036] [Data Storage Unit] The data storage unit 100 stores data necessary for relearning. For example, the data storage unit 100 stores a large-scale language model M1. The large-scale language model M1 includes parameters adjusted by pre-learning (parameters before relearning) and a program for processing such as calculation of embedded representations.

[0037] The parameters of the large-scale language model M1 are referenced by the program of the large-scale language model M1. For example, the parameters are weighting factors and biases. The parameters may be various well-known parameters. The parameters are not limited to weighting factors and biases. For example, the parameters may be matrices referenced when calculating an embedding representation, positional encodings referenced in encoding token positions, or other parameters.

[0038] The program of the large-scale language model M1 indicates the internal processing of the large-scale language model M1. For example, the program of the large-scale language model M1 indicates the processing of an encoder that calculates an embedded representation, a decoder that creates an output specific to a business support task from the embedded representation, an output layer that produces a final output, and other layers. Parameters are referenced in these processes. There may be parameters specific to a particular business support task. The program of the large-scale language model M1 may be a known program.

[0039] For example, when the re-learning unit 103 completes re-learning, the data storage unit 100 stores the re-learned business support model M2. In this embodiment, an example is given in which the large-scale language model M1 and the re-learned business support model M2 are separate data. Instead of these being separate data, the large-scale language model M1 may be overwritten by the business support model M2. The data storage unit 100 stores not only the large-scale language model M1 and the business support model M2, but also the training database DB1.

[0040] 5 is a diagram showing an example of the training database DB1. The training database DB1 is a database that stores training data necessary for relearning the large-scale language model M1. For example, the training data includes an input portion that is input to the large-scale language model M1 during relearning and an output portion that becomes the correct answer during relearning. The combination of the input portion and the output portion may be any combination that corresponds to the business support task. The training data may be prepared by manual input or by using a known tool.

[0041] The input portion of the training data may be the activity data itself or other data created based on the activity data. Because the large-scale language model M1 performs language analysis, the activity data represents text entered by users. The activity data may also represent other activities besides text. For example, the activity may represent user reactions (e.g., likes or emoji reactions), the number of views, the number of replies, the date and time the text was posted, or other content.

[0042] The output part of the training data indicates content according to the business support task. For example, if the business support task is to output some kind of text such as a summary, the output part of the training data is the correct text. If the business support task is to output some kind of classification (label), the output part of the training data is the correct classification. If the business support task is to extract some kind of data, the output part of the training data is the correct data (data that should be extracted from the input part of the training data). If the business support task is to sort some kind of data, the output part of the training data is the correct order.

[0043] In this embodiment, an example is taken of a case where creating a summary of a thread corresponds to a business support task. Therefore, in the example of Fig. 5 , the input portion of the training data is thread data for relearning. The thread data used as the input portion of the training data may be thread data that actually exists in the business support system 2, or may be thread data that indicates the contents of a virtual thread prepared by the operating company of the business support system 2. In the example of Fig. 5 , the output portion of the training data is summary data that indicates a correct summary created from the thread data that is the input portion of the training data.

[0044] The training data is not limited to the example of this embodiment. The training data may indicate content corresponding to the business assistance task. The input portion of the training data may be activity data for learning or data created based on activity data for learning. The output portion of the training data may indicate a correct answer as a processing result for the input portion of the training data. The input and output portions of the training data may indicate content corresponding to the business assistance task. An example of training data corresponding to a business assistance task will be described below, but the training data is not limited to the example below.

[0045] For example, if the task is to create a summary of text other than thread messages (e.g., the body of an email, a comment posted through a database comment function, or text on a website such as a help page), the input portion of the training data indicates the text to be summarized, and the output portion of the training data indicates the correct summary created from that text.

[0046] For example, if the business support task is to automatically create some kind of text (e.g., a message posted to a thread, the body of an email, a comment posted using a comment function in a database, or text on a website such as a help page), the input portion of the training data indicates the conditions for the text to be created. In this case, the output portion of the training data indicates the correct text according to the conditions.

[0047] For example, if the business support task is machine translation, the input portion of the training data is the text before translation (text in the original language). In this case, the output portion of the training data is the translated text that is the correct answer. For example, if the business support task is an answer to a user's question, the input portion of the training data indicates a question prepared for relearning. In this case, the output portion of the training data indicates the answer that is the correct answer to the question.

[0048] For example, if the business support task is to extract data registered in the business support system 2 (e.g., extracting the address of the sender of an email, extracting the address of the recipient of an email, extracting the organization and name written in the signature of an email, extracting the organization and name from the image data of a business card), the input portion of the training data indicates the data to be extracted. In this case, the output portion of the training data indicates the correct content extracted from the data.

[0049] For example, if the business support task is the classification of data registered in the business support system 2 (e.g., classification of topics in a thread, classification of the content of emails, or classification of files registered in the business support system 2), the input portion of the training data indicates data for relearning. In this case, the output portion of the training data indicates the classification of the data. For example, if the business support task is the analysis or prediction of user behavior in the business support system 2, the input portion of the training data indicates the user behavior. In this case, the output portion of the training data indicates the analysis or prediction results of the behavior.

[0050] The data stored in the data storage unit 100 is not limited to the above example. The data storage unit 100 may store any data. For example, the data storage unit 100 may store a re-learning program that indicates a series of processes in re-learning the large-scale language model M1. When a business support model M2 is created for each user organization, the data storage unit 100 may store a business support model M2 for each organization.

[0051] [Large-scale language model acquisition unit] The large-scale language model acquisition unit 101 acquires a pre-trained large-scale language model M1. In this embodiment, the pre-trained large-scale language model M1 is stored in the data storage unit 100, so the large-scale language model acquisition unit 101 acquires the pre-trained large-scale language model M1 from the data storage unit 100. If the pre-trained large-scale language model M1 is stored in a computer or information storage medium other than the relearning terminal 10, the large-scale language model acquisition unit 101 acquires the pre-trained large-scale language model M1 from the other computer or information storage medium.

[0052] [Training Data Acquisition Unit] The training data acquisition unit 102 acquires training data for relearning the large-scale language model M1, which is created based on activity data indicating user activities performed in the business support system 2 that supports the user's business. In this embodiment, the training data is stored in the training database DB1, so the large-scale language model acquisition unit 101 acquires the training data from the training database DB1. If the training data is stored in a computer or information storage medium other than the relearning terminal 10, the training data acquisition unit 102 acquires the training data from the other computer or information storage medium.

[0053] [Re-learning Unit] The re-learning unit 103 re-learns the large-scale language model M1 based on the training data, thereby creating a task assistance model M2 specific to the task assistance system 2. In this embodiment, an example is given in which the re-learning unit 103 re-learns the large-scale language model M1 based on a supervised learning algorithm. The re-learning algorithm may be a known algorithm. For example, the re-learning unit 103 may re-learn the large-scale language model M1 based on a semi-supervised learning or unsupervised learning algorithm.

[0054] For example, the re-learning unit 103 performs re-learning of the large-scale language model M1 by adjusting parameters of the large-scale language model M1 based on a supervised learning algorithm so that when an input portion of training data is input to the large-scale language model M1, an output portion of the training data is output from the large-scale language model M1. The re-learning unit 103 calculates a loss indicating the error between the output from the large-scale language model M1 when the input portion of training data is input to the large-scale language model M1, and the output portion of the training data, based on a known loss function. The re-learning unit 103 performs re-learning of the large-scale language model M1 by adjusting parameters of the large-scale language model M1 so that the calculated loss becomes smaller. The re-learning unit 103 may repeatedly perform re-learning of the large-scale language model M1 until the loss becomes less than a threshold.

[0055] For example, during relearning, the large-scale language model M1 divides text represented by the input portion of the training data into multiple tokens. The large-scale language model M1 encodes each of the multiple tokens and converts it into a sequence of embedded representations. Based on the sequence of embedded representations, the large-scale language model M1 predicts the continuation as needed and calculates an output according to the business support task. The greater the difference between the output of the large-scale language model M1 and the output portion of the training data, the greater the loss. The smaller the difference between the output of the large-scale language model M1 and the output portion of the training data, the smaller the loss. This series of processes is executed based on the program of the large-scale language model M1 (e.g., the encoder, decoder, output layer, and other layers) and the current parameters of the large-scale language model M1.

[0056] For example, assume that the input portion of the training data indicates text corresponding to the activity data. Furthermore, assume that the output portion of the training data is text that is the correct answer corresponding to the text. In this case, the business support task is a task of creating another text based on some text. When the re-learning unit 103 inputs the input portion of the training data to the large-scale language model M1, the large-scale language model M1 outputs another text corresponding to the sequence of embedded expressions of the text indicated by the input portion. The re-learning unit 103 re-learns the large-scale language model M1 based on the difference between the other text and the output portion of the training data.

[0057] For example, assume that the input portion of the training data represents text based on activity data. Furthermore, assume that the output portion of the training data represents a correct classification of the text. In this case, the business support task is a task of classifying some text. When the re-learning unit 103 inputs the input portion of the training data to the large-scale language model M1, the large-scale language model M1 outputs an estimated classification result according to the sequence of embedded expressions of the text represented by the input portion. The re-learning unit 103 re-learns the large-scale language model M1 based on the difference between the estimated classification result and the output portion of the training data.

[0058] In this embodiment, since the case where creating a summary of a thread corresponds to a business support task is taken as an example, the re-learning unit 103 inputs thread data, which is the input portion of the training data, to the large-scale language model M1. The large-scale language model M1 outputs summary data indicating a summary according to the arrangement of embedded expressions of the thread indicated by the thread data. The re-learning unit 103 re-learns the large-scale language model M1 based on the difference between the output summary data and the output portion of the training data. Similarly, for the various examples of training data described above, the re-learning unit 103 inputs the input portion of the training data to the large-scale language model M1, and re-learns the large-scale language model M1 based on the difference between the output from the large-scale language model M1 and the output portion of the training data.

[0059] The algorithm used in the re-learning may be a known algorithm. For example, the re-learning unit 103 re-learns the large-scale language model M1 based on the backpropagation algorithm. The re-learning unit 103 may also re-learn the large-scale language model M1 based on an algorithm other than the backpropagation algorithm. For example, the re-learning unit 103 may also re-learn the large-scale language model M1 based on the gradient descent algorithm, the momentum algorithm, the quasi-Newton algorithm, the conjugate gradient algorithm, the local search algorithm, or another algorithm.

[0060] In this embodiment, the relearning unit 103 executes a series of relearning processes by executing a relearning program stored in the data storage unit 100. The relearning unit 103 records the large-scale language model M1 after relearning as a business assistance model M2 in the data storage unit 100. The relearning unit 103 transmits the business assistance model M2 to the business assistance server 20. The business assistance model M2 transmitted to the business assistance server 20 is made available for use by the user.

[0061] 3-2. Functions Realized by the Business Support Server For example, the business support server 20 includes a data storage unit 200 and a business support unit 201. The data storage unit 200 is realized by the storage unit 22. The business support unit 201 is realized by the control unit 21.

[0062] [Data Storage Unit] The data storage unit 200 stores data necessary for supporting the user's work. For example, the data storage unit 200 stores a user database DB2.

[0063] 6 is a diagram showing an example of the user database DB2. The user database DB2 is a database that stores various data of users who use the business support system 2. For example, the user database DB2 stores the organization ID of the organization to which the user belongs, the user ID of the user, and activity data. Any data may be stored in the user database DB2. For example, the user database DB2 may store other data such as the user's login password.

[0064] The organization ID is the ID of an organization that has signed a contract with the business support system 2. The user ID is the ID of a user that belongs to the organization. For example, the user ID is used as a login account for a user to log in to the business support system 2. At least one user ID is associated with one organization ID. The activity data is data that indicates a user's activity. For example, the activity data indicates the date and time when the activity was performed and the specific content of the activity.

[0065] In this embodiment, posting a message to a thread corresponds to an activity, and therefore the activity data indicates the text of the message. The activity data may include the date and time when a user posted a message to a thread, a reaction to the message, or other information. Every time a user of an organization posts a message to a thread, the business support unit 201 creates activity data, which is associated with the organization ID of the organization and the user ID of the user, and stores the activity data in the user database DB2.

[0066] The data stored in the data storage unit 200 is not limited to the user database DB2. The data storage unit 200 may store any data. For example, the data storage unit 200 may store a re-trained business assistance model M2. The data storage unit 200 may store various data, such as the contents of threads registered in the business assistance system 2, in addition to the activity data.

[0067] [Business Support Unit] The business support unit 201 executes various processes to support the user's business. For example, the business support unit 201 creates activity data based on operation data indicating the user's operation content acquired from the user terminal 30. The business support unit 201 updates the user database DB2 so that the user ID of the user and the created activity data are associated with each other. The business support unit 201 transmits the activity data stored in the user database DB2 to the relearning terminal 10. Training data is created based on the activity data.

[0068] For example, the business support unit 201 supports a user's business based on the re-trained business support model M2. In this embodiment, since the case where creating a summary of a thread corresponds to the business support task is taken as an example, the business support unit 201 inputs thread data of a thread for which the user wishes to create a summary to the business support model M2. The re-trained business support model M2 calculates an embedded expression of the thread indicated by the thread data. The re-trained business support model M2 creates summary data indicating a summary according to the calculated embedded expression. The business support unit 201 transmits the summary data created by the re-trained business support model M2 to the user terminal 30. The business support unit 201 similarly inputs data indicating some kind of text to the business support model M2 and supports the user's business based on the data output from the business support model M2 for other business support tasks.

[0069] [3-3. Functions Realized by User Terminal] For example, the user terminal 30 includes a data storage unit 300, a display control unit 301, and an operation reception unit 302. The data storage unit 300 is realized by the storage unit 32. The display control unit 301 and the operation reception unit 302 are realized by the control unit 31.

[0070] [Data Storage Unit] The data storage unit 300 stores data necessary for a user to use the business support system 2. For example, the data storage unit 300 stores a browser. For example, the data storage unit 300 stores a program dedicated to the business support system 2.

[0071] [Display Control Unit] The display control unit 301 displays various screens for business support on the display unit 35. For example, the display control unit 301 displays a screen in the business support system 2 on a browser. The display control unit 301 displays a screen in the business support system 2 on a program dedicated to the business support system 2. After the business support model M2 is created, the user can use the business support model M2 from the business support screen SC. In the example of FIG. 2, when the user selects "Create summary," summary data of the currently displayed thread is created by the business support model M2.

[0072] [Operation Receiving Unit] The operation receiving unit 302 receives operations for each screen displayed on the display unit 35 by the display control unit 301. The operation receiving unit 302 transmits operation data indicating the content of the operation performed by the user to the business support server 20.

[0073] 7 is a diagram showing an example of processing executed in the relearning system 1. The control units 11, 21, and 31 execute programs stored in the storage units 12, 22, and 32, respectively, to execute the processing in FIG.

[0074] 7 , the business support server 20 creates activity data based on operation data received from the user terminal 30 and stores the activity data in the user database DB2 (S1). The business support server 20 transmits the activity data stored in the user database DB2 to the relearning terminal 10 (S2). The relearning terminal 10 receives the activity data from the business support server 20 (S3). The relearning terminal 10 stores training data created based on the activity data in the training database DB1 (S4).

[0075] The relearning terminal 10 acquires the pre-trained large-scale language model M1 stored in the storage unit 12 (S5). The relearning terminal 10 acquires an arbitrary number of training data from the training database DB1 (S6). The relearning terminal 10 creates a business assistance model M2 by re-learning the pre-trained large-scale language model M1 based on the training data acquired in S6 (S7). The relearning terminal 10 transmits the business assistance model M2 created in S7 to the business assistance server 20 (S8). The business assistance server 20 receives the business assistance model M2 from the relearning terminal 10 (S9). The business assistance server 20 executes a process with the user terminal 30 to allow the user to use the business assistance model M2 (S10), and this process ends.

[0076] [5. Summary of the Embodiment] The relearning system 1 of this embodiment acquires a pre-trained large-scale language model M1. The relearning system 1 acquires training data for relearning the large-scale language model M1, which is created based on activity data. The relearning system 1 performs relearning of the large-scale language model M1 based on the training data. For example, it is not realistic for an operating company of a business assistance system 2 to create a business assistance model M2 from scratch. However, the relearning system 1 can realize a business assistance model M2 useful for the business assistance system 2 by relearning the pre-trained large-scale language model M1. For example, compared to a general-purpose large-scale language model M1, the relearning system 1 can effectively support a user's work based on the business assistance model M2 that has learned the user's activities in the business assistance system 2. Because the large-scale language model M1 has basic language analysis capabilities, the relearning system 1 can create a highly accurate business assistance model M2 even with a small amount of training data.

[0077] [6. Modifications] The present disclosure is not limited to the examples of the embodiments, and can be modified within the scope of the present disclosure.

[0078] [6-1. Modification 1] In the embodiment, training data including thread data for relearning and summary data that will be the correct answer during relearning has been described as an example. The relearning system 1 can provide more effective business support by creating multiple business support models M2 based on training data created from various perspectives, rather than performing relearning using training data created from a single perspective. Therefore, the relearning system 1 of Modification 1 performs multi-stage relearning based on training data created from various perspectives.

[0079] The training data acquisition unit 102 of Modification 1 acquires multiple pieces of training data created from different perspectives. A perspective is a criterion for creating the training data. In other words, a perspective is a type of activity data used in creating the training data. In Modification 1, as examples of multiple pieces of training data created from different perspectives, training data created for each business support task and training data created for each organization will be described. Multi-stage re-learning is performed based on the training data created from these perspectives.

[0080] 8 is a diagram showing an example of multi-stage re-learning. The re-learning unit 103 of Modification 1 creates multiple task support models M2 by performing multi-stage re-learning of the large-scale language model M1 based on multiple training data acquired by the training data acquisition unit 102 of Modification 1. For example, the re-learning unit 103 creates common task support models M2A1 and M2A2, which are first-stage task support models M2, by performing first-stage re-learning based on training data created from a certain perspective. Hereinafter, when the common task support models M2A1 and M2A2 are not distinguished from each other, they will be simply referred to as common task support models M2A. There may be only one common task support model M2A, or three or more common task support models M2A.

[0081] For example, the relearning unit 103 performs second-stage relearning based on training data created from another perspective, thereby creating specific task support models M2B1 and M2B2, which are the second-stage task support model M2. Hereinafter, when the specific task support models M2B1 and M2B2 are not distinguished from each other, they are simply referred to as the common task support model M2B. There may be only one common task support model M2B, or three or more common task support models M2B. The relearning unit 103 may perform three or more stages of relearning.

[0082] In the first modification, as in the embodiment, it is assumed that users belonging to each of a plurality of organizations use the business assistance system 2. Furthermore, it is assumed that the user database DB2 stores activity data of each user of the plurality of organizations. For example, the training data acquisition unit 102 acquires common training data, which is training data common to the plurality of organizations and created based on the activity data of each of the plurality of organizations, for each business assistance task performed by the business assistance model M2.

[0083] For example, suppose there are two task support tasks: thread summary creation and machine translation. In this case, the training data acquisition unit 102 acquires common training data created based on thread data (an example of activity data) of each of multiple organizations to create a common task support model M2A1 that performs thread summary creation. For example, the training data acquisition unit 102 acquires common training data created based on activity data indicating some text entered by users belonging to each of multiple organizations to create a common task support model M2A2 that performs machine translation.

[0084] For example, the re-learning unit 103 performs re-learning of the large-scale language model M1 based on common training data for each business support task, thereby creating a common business support model M2A, which is a business support model M2 common to multiple organizations. The common business support model M2A of variant 1 is a model specific to a certain business support task, but is not a model specific to a specific organization. The common business support model M2A can generally support the respective business operations of multiple organizations. The common business support model M2A is an intermediate model for creating the specific business support model M2B described below. The method of re-learning based on common training data is the same as re-learning based on training data described in the embodiment.

[0085] For example, suppose there are two task support tasks: thread summarization and machine translation. In this case, the re-learning unit 103 re-learns the large-scale language model M1 based on common training data created from thread data (an example of activity data) of multiple organizations, thereby creating a common task support model M2A1 that performs thread summarization. The common task support model M2A1 that performs thread summarization has higher summary creation accuracy than the large-scale language model M1, but it cannot create summaries specific to a specific organization.

[0086] For example, the re-learning unit 103 creates a common task support model M2A2 that performs machine translation by re-learning the large-scale language model M1 based on common training data created from activity data that indicates some text entered by users belonging to each of multiple organizations. The common task support model M2A2 that performs machine translation has higher accuracy than the large-scale language model M1, but it cannot perform machine translation that is specific to a specific organization.

[0087] For example, the training data acquisition unit 102 acquires, for each organization, specific training data that is training data specific to that organization and that is created based on the activity data of that organization. In the first variant example, a case is described in which the activity data used to create the specific training data and the activity data used to create the general training data are the same, but they may be different.

[0088] For example, suppose there are two task support tasks: thread summary creation and machine translation. In this case, the training data acquisition unit 102 acquires specific training data created based on the thread data of a first organization to create a specific task support model M2B1 that creates summaries specific to the first organization. The training data acquisition unit 102 acquires specific training data created based on the thread data of a second organization to create a specific task support model M2B2 that creates summaries specific to the second organization.

[0089] For example, the training data acquisition unit 102 acquires specific training data created based on activity data indicating some text input by a user belonging to a first organization in order to create a specific task support model M2B3 that performs machine translation specific to the first organization. The training data acquisition unit 102 acquires specific training data created based on activity data indicating some text input by a user belonging to the second organization in order to create a specific task support model M2B4 that performs machine translation specific to another second organization.

[0090] For example, the re-learning unit 103 re-learns the common task support model M2A for each organization based on the specific training data of that organization, thereby creating a specific task support model M2B that is a task support model M2 specific to that organization. The specific task support model M2B of variant 1 is a model specialized for a specific task support task of a specific organization. The specific task support model M2B can support a specific task of a specific organization. The specific task support model M2B is a final model provided to users belonging to a specific organization. The method of re-learning based on the specific training data is the same as the re-learning based on the training data described in the embodiment.

[0091] For example, suppose there are two task support tasks: thread summary creation and machine translation. In this case, the re-learning unit 103 re-learns the common task support model M2A1 for summary creation based on the specific training data created from the thread data of the first organization, thereby creating a specific task support model M2B1 for creating summaries specific to the first organization. The re-learning unit 103 re-learns the common task support model M2A1 for summary creation based on the specific training data created from the thread data of the second organization, thereby creating a specific task support model M2B2 for creating summaries specific to the second organization.

[0092] For example, the re-learning unit 103 creates a specific task support model M2B3 that performs machine translation specific to the first organization by re-learning the common task support model M2A2 for machine translation based on specific training data created from activity data indicating some text input by a user belonging to the first organization. The re-learning unit 103 creates a specific task support model M2B4 that performs machine translation specific to the second organization by re-learning the common task support model M2A2 for machine translation based on specific training data created from activity data indicating some text input by a user belonging to the second organization.

[0093] In the above description, an example has been given in which there are two business support tasks and two organizations, but any number of business support tasks and organizations may be used. For example, if there are three or more business support tasks, three or more common business support models M2A may be created. If there are three or more organizations, three or more specific business support models M2B may be created for each business support task.

[0094] For example, the re-learning unit 103 records the specific task support model M2B for each of the multiple organizations in the data storage unit 100. The re-learning unit 103 transmits the specific task support model M2B for each of the multiple organizations to the task support server 20. The task support server 20 associates the organization ID of each of the multiple organizations with the specific task support model M2B for that organization and records them in the data storage unit 200. When a user belonging to a certain organization logs in to the task support system 2, the task support unit 201 supports the user's task based on the specific task support model M2B associated with the organization ID of that organization.

[0095] When a user belonging to an organization for which the specific task support model M2B has not been created logs in to the task support system 2, the task support unit 201 may support the task of the user based on the common task support model M2A. In this case, the task support unit 201 may support the task of the user based on the specific task support model M2B associated with the organization ID of another organization (for example, an organization in the same industry or with the same number of employees) rather than the common task support model M2A.

[0096] The relearning system 1 of Modification 1 acquires multiple training data created from different perspectives. The relearning system 1 creates multiple business assistance models M2 by performing multi-stage relearning of the large-scale language model M1 based on the multiple training data. This allows the relearning system 1 to realize a more useful business assistance model M2. For example, the relearning system 1 can realize more accurate business assistance through multi-stage relearning.

[0097] Furthermore, the relearning system 1 creates a common task support model M2A for each task based on the common training data. The relearning system 1 creates a specific task support model M2B for each organization based on the specific training data of that organization. This allows the relearning system 1 to create a common task support model M2A based on a larger amount of common training data, and then create a specific task support model M2B that is specialized for the tendencies of each organization, thereby efficiently creating a task support model M2 that is specific to each organization. For example, the relearning system 1 can realize highly accurate task support tailored to a specific task and a specific organization.

[0098] [6-2. Modification 2] For example, the multi-stage relearning method described in Modification 1 is not limited to the multi-stage relearning method according to the organization described in Modification 1. In Modification 2, as another example of the multi-stage relearning method, multi-stage relearning according to the business support task and the business support function will be described. Note that when there are multiple business support systems 2, the business support function may not be a function of a single business support system 2, but may be the business support system 2 itself. For example, when there are multiple groupware, each groupware may correspond to a business support function.

[0099] For example, the training data acquisition unit 102 acquires common training data, which is training data common to multiple business support functions, created based on the activity data of each of the multiple business support functions possessed by the business support system 2 for each business support task performed by the business support model M2.

[0100] For example, suppose there are two business support tasks: summary creation and machine translation. The summary creation in variant example 2 is not limited to threads, but is assumed to be summary creation for the entire business support system 2. In this case, the training data acquisition unit 102 acquires common training data for summary creation created based on the activity data of each of the multiple business support functions to create a common business support model M2A1 that performs summary creation. For example, the training data acquisition unit 102 acquires common training data for machine translation created based on the activity data of each of the multiple business support functions to create a common business support model M2A2 that performs machine translation.

[0101] For example, the re-learning unit 103 performs re-learning of the large-scale language model M1 for each business support task based on common training data, thereby creating a common business support model M2A, which is a business support model M2 common to multiple business support functions. The common business support model M2A of variant example 2 is a model specific to a certain business support task, but is not a model specific to a specific business support function. The common business support model M2A can generally support the work of users who use each of multiple business support functions. The common business support model M2A is an intermediate model for creating the specific business support model M2B described below. The method of re-learning based on common training data is the same as re-learning based on training data described in the embodiment.

[0102] For example, suppose there are two business support tasks: summarization and machine translation. In this case, the re-learning unit 103 re-learns the large-scale language model M1 based on common training data for summarization created from the activity data of each of the multiple business support functions, thereby creating a common business support model M2A1 that performs summarization. The common business support model M2A1 that performs summarization has higher accuracy in summary creation than the large-scale language model M1, but it cannot create summaries specific to a specific business support function.

[0103] For example, the re-learning unit 103 creates a common task support model M2A1 that performs machine translation by re-learning the large-scale language model M1 based on common training data for machine translation created from activity data of each of the multiple task support functions. The common task support model M2A2 that performs machine translation has higher accuracy in machine translation than the large-scale language model M1, but it cannot perform machine translation specific to a specific task support function.

[0104] For example, the training data acquisition unit 102 acquires, for each business support function, specific training data that is training data specific to that business support function and that is created based on the activity data of that business support function. In the second modification, as in the first modification, a case will be described in which the activity data used to create the specific training data and the activity data used to create the general training data are the same, but they may be different.

[0105] For example, suppose there are two business support tasks: summary creation and machine translation. Furthermore, suppose there are two business support functions: a thread function and an email management function. In this case, the training data acquisition unit 102 acquires specific training data created based on thread data to create a specific business support model M2B1 that creates summaries specific to the thread function. The training data acquisition unit 102 acquires specific training data created based on activity data indicating the text of emails managed by the email management function to create a specific business support model M2B2 that creates summaries specific to the email management function.

[0106] For example, the training data acquisition unit 102 acquires specific training data created based on thread data to create a specific task support model M2B3 that performs machine translation specific to the thread function. The training data acquisition unit 102 acquires specific training data created based on activity data that indicates the text of emails managed by the email management function to create a specific task support model M2B4 that performs machine translation specific to the email management function.

[0107] For example, the re-learning unit 103 performs re-learning of the common task support model M2A for each task support function based on the specific training data of that task support function, thereby creating a specific task support model M2B, which is a task support model M2 specific to that task support function. The specific task support model M2B of variant example 2 is a model specialized for a specific task support task of a specific task support function. The specific task support model M2B can support the task of a user who uses a specific task support function. The specific task support model M2B is the final model provided to a user who uses a specific task support function. The method of re-learning based on the specific training data is the same as the re-learning based on the training data described in the embodiment.

[0108] For example, suppose there are two business support tasks: summary creation and machine translation. Furthermore, suppose there are two business support functions: a thread function and an email management function. In this case, the re-learning unit 103 re-learns the common business support model M2A1 for summary creation based on the specific training data created from the thread data, thereby creating a specific business support model M2B1 that creates summaries specific to the thread function. The re-learning unit 103 re-learns the common business support model M2A1 for summary creation based on the specific training data created from the activity data of a user who uses the email management function, thereby creating a specific business support model M2B2 that creates summaries specific to the email management function.

[0109] For example, the re-learning unit 103 creates a specific task support model M2B3 that performs machine translation specific to the thread function by re-learning the common task support model M2A2 for machine translation based on specific training data created from thread data. The re-learning unit 103 creates a specific task support model M2B4 that performs machine translation specific to the email management function by re-learning the common task support model M2A2 for machine translation based on specific training data created from activity data of a user who uses the email management function.

[0110] In the above description, an example has been given in which there are two business support tasks and two business support functions, but any number of business support tasks and business support functions may be used. For example, if there are three or more business support tasks, three or more common business support models M2A may be created. If there are three or more business support functions, three or more specific business support models M2B may be created for each business support task.

[0111] For example, the re-learning unit 103 records the specific business support model M2B for each of the multiple business support functions in the data storage unit 100. The re-learning unit 103 transmits the specific business support model M2B for each of the multiple business support functions to the business support server 20. The business support server 20 associates each of the multiple business support functions with the specific business support model M2B for that business support function and records them in the data storage unit 200. When a user uses a certain business support function, the business support unit 201 supports the user's business based on the specific business support model M2B associated with that business support function.

[0112] When a user uses a business support function for which a specific business support model M2B has not been created, the business support unit 201 may support the user's business based on the common business support model M2A. In this case, the business support unit 201 may support the user's business based on a specific business support model M2B associated with another business support function, rather than the common business support model M2A.

[0113] The relearning system 1 of Modification 2 creates a common task support model M2A by relearning the large-scale language model M1 based on common training data for each task support task. The relearning system 1 creates a specific task support model M2B, which is a task support model M2 specific to each task support function, by relearning the common task support model M2A based on specific training data for that task support function. This allows the relearning system 1 to create the common task support model M2A based on a larger amount of common training data, and then create a specific task support model M2B that is specialized for the tendencies of each task support function, thereby efficiently creating a task support model M2 specific to each task support function. For example, the relearning system 1 can realize highly accurate task support tailored to a specific task support task and a specific task support function.

[0114] [6-3. Modification 3] For example, in Modification 3, multi-stage relearning according to business support functions and organizations will be described as another example of the multi-stage relearning described in Modification 1. The number of business support tasks in Modification 3 may be one.

[0115] For example, the training data acquisition unit 102 acquires common training data, which is training data common to multiple organizations, created based on the activity data of each of the multiple organizations for each business support function possessed by the business support system 2.

[0116] For example, suppose there are two task support functions: a thread function and an email management function. In this case, the training data acquisition unit 102 acquires common training data created based on activity data indicating text entered by users belonging to each of multiple organizations using the thread function to create a common task support model M2A1 for the thread function. For example, the training data acquisition unit 102 acquires common training data created based on activity data indicating text entered by users belonging to each of multiple organizations using the email management function to create a common task support model M2A2 for the email management function.

[0117] For example, the re-learning unit 103 creates a common task support model M2A common to multiple organizations by re-learning the large-scale language model M1 for each task support function based on common training data. The common task support model M2A of variant example 3 is a model specialized for the task support system 2, but is not a model specific to a particular organization. The common task support model M2A can generally support the tasks of each of multiple organizations. The common task support model M2A is an intermediate model for creating the specific task support model M2B described below. The method of re-learning based on common training data is the same as re-learning based on training data described in the embodiment.

[0118] For example, suppose there are two business support functions: a thread function and an email management function. In this case, the re-learning unit 103 re-learns the large-scale language model M1 based on common training data created from activity data (activity data indicating text entered by users using the thread function) of each of multiple organizations, thereby creating a common business support model M2A1 for the thread function. The common business support model M2A1 for the thread function is capable of providing business support more specific to the thread function than the large-scale language model M1, but is not capable of providing business support specific to a specific organization.

[0119] For example, the re-learning unit 103 creates a common task support model M2A2 for the email management function by re-learning the large-scale language model M1 based on common training data created from activity data (activity data showing text entered by users using the email management function) of each of multiple organizations. The common task support model M2A2 for the email management function is more capable of providing task support specific to the email management function than the large-scale language model M1, but is not capable of providing task support specific to a specific organization.

[0120] For example, the training data acquisition unit 102 acquires, for each organization, specific training data that is training data specific to that organization and created based on the activities of that organization. In Variation 3, a case is described in which the activity data used to create the specific training data and the activity data used to create the general training data are the same, but they may also be different.

[0121] For example, suppose there are two business support functions: a thread function and an email management function. In this case, the training data acquisition unit 102 acquires specific training data created based on activity data of a first organization to create a specific business support model M2B1 for the thread function of the first organization. The training data acquisition unit 102 acquires specific training data created based on activity data of the second organization to create a specific business support model M2B2 for the thread function of another second organization.

[0122] For example, the training data acquisition unit 102 acquires specific training data created based on activity data of a first organization to create a specific task support model M2B3 for the email management function of the first organization. The training data acquisition unit 102 acquires specific training data created based on activity data of the second organization to create a specific task support model M2B4 for the email management function of the second organization.

[0123] For example, the re-learning unit 103 performs re-learning of the common task support model M2A for each organization based on the specific training data of that organization, thereby creating a specific task support model M2B that is a task support model M2 specific to that organization. The specific task support model M2B of variant example 3 is a model specialized for a specific organization. The specific task support model M2B is a model for users belonging to a specific organization to use specific task support functions. The specific task support model M2B is the final model provided to users belonging to a specific organization. The method of re-learning based on the specific training data is the same as the re-learning based on the training data described in the embodiment.

[0124] For example, suppose there are two task support functions: a thread function and an email management function. In this case, the re-learning unit 103 re-learns the common task support model M2A1 for the thread function based on the specific training data created from the activity data of the first organization, thereby creating a specific task support model M2B1 that can support tasks specific to the first organization. The re-learning unit 103 re-learns the common task support model M2A1 for the thread function based on the specific training data created from the activity data of the second organization, thereby creating a specific task support model M2B2 that can support tasks specific to the second organization.

[0125] For example, the re-learning unit 103 creates a specific task support model M2B3 capable of providing task support specific to the first organization by re-learning the common task support model M2A2 for the email management function based on the specific training data created from the activity data of the first organization. The re-learning unit 103 creates a specific task support model M2B4 capable of providing task support specific to the second organization by re-learning the common task support model M2A2 for the email management function based on the specific training data created from the activity data of the second organization.

[0126] In the above description, an example was given in which there are two business support functions and two organizations, but any number of business support functions and organizations may be used. For example, if there are three or more business support functions, three or more common business support models M2A may be created. If there are three or more organizations, three or more specific business support models M2B may be created for each business support function.

[0127] For example, the re-learning unit 103 records the specific business support model M2B for each of the multiple organizations in the data storage unit 100. The re-learning unit 103 transmits the specific business support model M2B for each of the multiple organizations to the business support server 20. The business support server 20 associates the organization ID of each of the multiple organizations with the specific business support model M2B for that organization and records the associated organization ID in the data storage unit 200. When a user belonging to a certain organization uses a certain business support function, the business support unit 201 supports the user's business based on the specific business support model M2B for that business support function that is associated with the organization ID of that organization.

[0128] When a user belonging to an organization for which the specific task support model M2B has not been created logs in to the task support system 2, the task support unit 201 may support the task of the user based on the common task support model M2A. In this case, the task support unit 201 may support the task of the user based on the specific task support model M2B associated with the organization ID of another organization (for example, an organization in the same industry or with the same number of employees) rather than the common task support model M2A.

[0129] The relearning system 1 of Modification 3 creates a common task support model M2A by relearning the large-scale language model M1 based on common training data. The relearning system 1 creates a specific task support model M2B for each organization by relearning the common task support model M2A based on the specific training data of that organization. This allows the relearning system 1 to create the common task support model M2A based on a larger amount of common training data, and then create a specific task support model M2B that is specialized for the tendencies of each organization to which a user who uses a specific task support function belongs, thereby efficiently creating a task support model M2 that is specific to each organization. For example, the relearning system 1 can realize highly accurate task support tailored to a specific task support function and a specific organization.

[0130] [6-4. Modification 4] For example, in Modifications 1 to 3, multiple business support models M2, such as a common business support model M2A and a specific business support model M2B, are created. Each of these multiple business support models M2 may be used for data augmentation of training data.

[0131] 9 is a diagram showing an example of functions realized in Modification 4. For example, the relearning terminal 10 of Modification 4 includes a training data creation unit 104. The training data creation unit 104 is realized by the control unit 11. The training data creation unit 104 creates new training data based on each of the plurality of training data and each of the plurality of business assistance models M2. For example, the training data creation unit 104 inputs an input portion of the training data used in relearning the business assistance model M2 to each of the plurality of business assistance models M2. The business assistance model M2 performs output according to the input portion.

[0132] For example, the training data creation unit 104 acquires the input portion of the training data input to the business assistance model M2 as the input portion of new training data. The training data creation unit 104 acquires the output from the business assistance model M2 as the output portion of new training data. The output of the new training data does not have to be the output from the business assistance model M2 itself. For example, data in which a part of the output from the business assistance model M2 has been changed may be the output portion of the new training data. The training data creation unit 104 creates a pair of the input portion and the output portion as new training data. The training data creation unit 104 stores the new training data in the training database DB1.

[0133] The re-learning unit 103 of the fourth modification executes further re-learning of at least one of the plurality of business assistance models M2 based on new training data. The re-learning unit 103 may execute re-learning of all business assistance models M2 based on the new training data, or may execute re-learning of only some of the business assistance models M2. The further re-learning differs from the re-learning of the embodiment and the first to third modifications in that new training data is used, but the re-learning method may be the same as the embodiment and the first to third modifications.

[0134] The relearning system 1 of Modification 4 creates new training data based on each of the plurality of training data and each of the plurality of business assistance models M2. The relearning system 1 performs further relearning of at least one of the plurality of business assistance models M2 based on the new training data. This allows the relearning system 1 to realize data expansion using each of the plurality of business assistance models M2, thereby achieving effective relearning.

[0135] [6-5. Modification 5] For example, when the relearning system 1 attempts to create a business support model M2 specific to a particular organization, the organization may not have a sufficient amount of activity data. For this reason, activity data from another organization may be used to create the business support model M2 specific to the organization. In Modification 5, the organization for which the business support model M2 is to be created is referred to as the first organization. The other organization whose activity data is used to create the business support model M2 specific to the first organization is referred to as the second organization.

[0136] The training data acquisition unit 102 of the fifth modification acquires first-organization training data, which is training data created based on the activity data of a first organization, and second-organization training data, which is training data created based on the activity data of a second organization related to the first organization. The second organization is another organization related to the first organization. For example, the second organization is another organization in the same industry or with the same number of employees as the first organization.

[0137] In the training database DB1 of the fifth modification, each piece of training data is associated with the organization ID of the organization to which the user who performed the activity indicated by the activity data used to create the training data belongs. Furthermore, the association between the first organization and the second organization is defined in the data storage unit 100. For example, if the second organization is another organization in the same industry as the first organization, the industry of each organization is defined in the data storage unit 100. If the second organization is another organization with the same employee size as the first organization, the employee size of each organization is defined in the data storage unit 100.

[0138] The re-learning unit 103 of Modification 5 creates a task support model M2 specific to the first organization based on the first organization training data and the second organization training data. For example, the re-learning unit 103 creates the task support model M2 specific to the first organization by re-learning the large-scale language model M1 based on the first organization training data and the second organization training data. When a common task support model M2A is created as in Modifications 1 to 3, the re-learning unit 103 creates a task support model M2B specific to the first organization by re-learning the common task support model M2A based on the first organization training data and the second organization training data.

[0139] The relearning system 1 of the fifth modification creates a business support model M2 specific to the first organization based on the first organization training data and the second organization training data. This allows the relearning system 1 to efficiently create the business support model M2 specific to the first organization. For example, even if the first organization is a newly registered organization in the business support system 2 and does not have a sufficient amount of activity data, the relearning system 1 uses the second organization training data based on the activity data of the second organization, thereby improving the accuracy of the business support model M2 specific to the first organization.

[0140] [6-6. Modification 6] For example, the second organization in Modification 5 may be another organization that has been using the business support system 2 for a longer period of time than the first organization. The data storage unit 100 in Modification 6 stores data indicating the period of use of each organization. The relearning terminal 10 can identify which organization corresponds to the second organization by referring to the data.

[0141] The training data acquisition unit 102 of the sixth modification acquires first organization training data of a first organization newly registered in the business support system 2 and second organization training data of a second organization that has been using the business support system 2 for a longer period than the first organization. The training data acquisition unit 102 does not acquire training data for creating a business support model M2 specific to the first organization based on activity data of another organization that has been using the business support system 2 for a shorter period than the first organization.

[0142] The re-learning unit 103 of the sixth modification creates a business support model M2 specific to a first organization newly registered in the business support system 2. The re-learning unit 103 creates the business support model M2 specific to the first organization based on second-organization training data of a second organization that has been using the business support system 2 for a longer period than the first organization, without using training data based on activity data of another organization that has been using the business support system 2 for a shorter period than the first organization. Although the method of acquiring the second-organization training data is different from that of the fifth modification, the method of relearning based on the first-organization training data and the second-organization training data is the same as that of the fifth modification.

[0143] The relearning system 1 of the sixth modification creates a business support model M2 specific to the first organization newly registered in the business support system 2, based on first organization training data of the first organization newly registered in the business support system 2 and second organization training data of the second organization that has been using the business support system 2 for a longer period than the first organization. Since the longer the period of use, the more activity data there may be, the more efficiently the relearning system 1 can create a business support model M2 specific to the first organization newly registered in the business support system 2.

[0144] [6-7. Variation 7] For example, data other than the activity data to be processed may be input to the business support model M2. In Variation 7, as in the embodiment, an example is taken of a business support task being the creation of a summary of a thread. For example, a thread may contain keywords that the user wants to include in the summary. In this case, not only the thread data of the thread for which a summary is to be created, but also keywords specified by the user may be input to the business support model M2. In Variation 7, a business support model M2 that can handle such input is created.

[0145] The training data acquisition unit 102 of the seventh modification acquires training data representing the entire activity indicated by the activity data and a portion selected from the activity data. The portion selected from the activity data is a portion that particularly requires relearning. For example, the operating company of the business support system 2 selects a portion from the entire activity and performs annotation. The input portion of the training data is a pair of the entire activity indicated by the activity data and the annotated portion. The output portion of the training data may be the same as in the embodiment.

[0146] The re-learning unit 103 of the seventh modification creates a task support model M2 by re-learning the large-scale language model M1 based on training data indicating the whole and the part. Although the input part of the training data is different from that of the embodiment, the re-learning method is the same as that of the embodiment. The re-learning unit 103 re-learns the large-scale language model M1 so that when the whole and the part indicated by the input part of the training data are input to the large-scale language model M1, the large-scale language model M1 outputs the output part of the training data.

[0147] For example, when a user uses the business support model M2, the user specifies a portion corresponding to the portion. When the business support task is to create a summary of a thread, the business support unit 201 inputs the thread data of the thread for which a summary is to be created and keywords specified by the user from within the thread into the business support model M2. The business support model M2 outputs summary data based on the embedded representation of these texts. The summary indicated by the summary data includes the keywords specified by the user.

[0148] Note that the portion selected from the activity data is not limited to the above examples. The portion may be a portion corresponding to the business support task. For example, if the business support task is creating an answer to a question, the portion may be the content of the question. If the business support task is creating text, the portion may be a keyword to be included in the text. Alternatively, the portion may be any portion that has some correlation with the output portion of the training data.

[0149] The relearning system 1 of Variation 7 acquires training data representing the entire activity represented by the activity data and a portion selected from the activity data. The relearning system 1 creates a business support model M2 by relearning the large-scale language model M1 based on the training data representing the entire activity and the portion. This allows the relearning system 1 to create a more flexible business support model M2.

[0150] [6-8. Other Modifications] For example, the above modifications may be combined.

[0151] For example, the functions described as being realized by the relearning terminal 10 may be realized by the business support server 20. The functions described as being realized by the relearning terminal 10 may be shared among multiple computers.

Claims

1. A re-learning system including: a large-scale language model acquisition unit that acquires a pre-trained large-scale language model; a training data acquisition unit that acquires training data for re-learning the large-scale language model, the training data being created based on activity data indicating a user's activity performed in a business support system that supports the user's business; and a re-learning unit that creates a business support model specific to the business support system by performing re-learning of the large-scale language model based on the training data.

2. The re-learning system of claim 1, wherein the training data acquisition unit acquires a plurality of pieces of training data created from different perspectives, and the re-learning unit creates a plurality of the business support models by performing multi-stage re-learning of the large-scale language model based on the plurality of training data.

3. The re-learning system described in claim 2, wherein the training data acquisition unit acquires, for each business support task performed by the business support model, common training data, which is the training data common to the multiple organizations, created based on the activity data of each of the multiple organizations; the re-learning unit creates a common business support model, which is the business support model common to the multiple organizations, by performing re-learning of the large-scale language model based on the common training data, for each business support task; the training data acquisition unit acquires, for each organization, specific training data, which is the training data specific to the organization, created based on the activity data of the organization; and the re-learning unit creates a specific business support model, which is the business support model specific to the organization, by performing re-learning of the common business support model based on the specific training data of the organization, for each organization.

4. The re-learning system described in claim 2, wherein the training data acquisition unit acquires, for each business support task performed by the business support model, common training data, which is the training data common to the multiple business support functions possessed by the business support system, created based on the activity data of each of the multiple business support functions; the re-learning unit creates a common business support model, which is the business support model common to the multiple business support functions, by executing re-learning of the large-scale language model based on the common training data, for each business support task; the training data acquisition unit acquires, for each business support function, specific training data, which is the training data specific to the business support function, created based on the activity data of the business support function; and the re-learning unit creates a specific business support model, which is the business support model specific to the business support function, by executing re-learning of the common business support model based on the specific training data of the business support function, for each business support function.

5. The re-learning system of claim 2, comprising: the training data acquisition unit acquires, for each business support function possessed by the business support system, common training data, which is the training data common to the multiple organizations and created based on the activity data of each of the multiple organizations; the re-learning unit, for each business support function, performs re-learning of the large-scale language model based on the common training data, thereby creating a common business support model common to the multiple organizations; the training data acquisition unit acquires, for each organization, specific training data, which is the training data specific to the organization and created based on the activity of the organization; and the re-learning unit, for each organization, performs re-learning of the common business support model based on the specific training data of the organization, thereby creating a specific business support model, which is the business support model specific to the organization.

6. The re-learning system according to any one of claims 2 to 5, further comprising: a training data creation unit that creates new training data based on each of the plurality of training data and each of the plurality of business support models; and the re-learning unit that performs further re-learning of at least one of the plurality of business support models based on the new training data.

7. A re-learning system as described in any of claims 1 to 5, wherein the training data acquisition unit acquires first organization training data, which is the training data created based on the activity data of a first organization, and second organization training data, which is the training data created based on the activity data of a second organization related to the first organization, and the re-learning unit creates the business support model specific to the first organization based on the first organization training data and the second organization training data.

8. The re-learning system described in claim 7, wherein the training data acquisition unit acquires the first organization training data of the first organization newly registered in the business support system and the second organization training data of the second organization that has been using the business support system for a longer period than the first organization, and the re-learning unit creates the business support model specific to the first organization newly registered in the business support system.

9. A re-learning system as described in any of claims 1 to 5, wherein the training data acquisition unit acquires training data indicating the entire activity indicated by the activity data and a portion selected from the activity data, and the re-learning unit creates the business support model by re-learning the large-scale language model based on the training data indicating the entire activity and the portion.

10. A re-learning method comprising: acquiring a pre-trained large-scale language model; acquiring training data for re-learning the large-scale language model, the training data being created based on activity data indicating a user's activity performed in a business support system that supports the user's business; and performing re-learning of the large-scale language model based on the training data, thereby creating a business support model specific to the business support system.

11. A program for causing a computer to function as: a large-scale language model acquisition unit that acquires a pre-trained large-scale language model; a training data acquisition unit that acquires training data for re-learning the large-scale language model, the training data being created based on activity data indicating a user's activity performed in a business support system that supports the user's business; and a re-learning unit that creates a business support model specific to the business support system by performing re-learning of the large-scale language model based on the training data.

Images