Associative learning method, learning device, management device, learning program, management program, and associative learning system

Abstract

To achieve federated learning regardless of differences in data structures between locations. [Solution] A federated learning method that performs federated learning using multiple computers having a learning model that learns using training data which is the value of each item in an item group includes: a first generation process that obtains the item group from each of the multiple computers and generates a first common item group to be used by the multiple computers based on the multiple item groups; and a first application process that applies the first common item group generated by the first generation process to the item group in each of the multiple computers.

Description

Technical Field

[0001] The present invention relates to a federated learning method, a learning device, a management device, a learning program, a management program, and a federated learning system that execute federated learning.

Background Art

[0002] Patent Document 1 discloses a learning system that performs learning of a model using data collected in a plurality of databases managed by arbitrary names. In this learning system, the base database is provided for each of a plurality of bases, and stores measurement data measured from target devices provided in the bases in association with a name representing the type of the measurement data. The name database stores, in association with each other, names assigned to the same type of measurement data among the names used in the base databases for each base. The name specifying unit reads out a second name associated with a first name, which is a known name related to data to be used as a learning dataset for machine learning, from the name database. The learning unit performs learning processing of a learning model using a learning dataset generated from first measurement data, which is measurement data associated with the first name, and second measurement data, which is measurement data associated with the second name, in a plurality of base databases.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] When the data structures are different between bases, it is necessary to pre-design a predetermined rule for converting training data into specific features. When this pre-design is difficult, federated learning cannot be realized. The same applies when the training data possessed by each of the base groups includes unstructured data.

[0005] The present invention aims to achieve federated learning regardless of differences in data structures between locations. [Means for solving the problem]

[0006] The first aspect of the invention disclosed in this application is a federated learning method that performs federated learning using a plurality of computers having a learning model that learns on a training dataset which includes training data which is the value of each item in an item group, and includes a first generation process which obtains the item group from each of the plurality of computers and generates a common item group to be used by the plurality of computers based on the plurality of item groups, and a first application process which applies the common item group generated by the first generation process to the item group in each of the plurality of computers.

[0007] A learning device representing a second aspect of the invention disclosed in this application is a learning device among a plurality of learning devices having a learning model that is learned with training data which are the values ​​of each item in an item group, and comprises a processor that executes a program, a storage device that stores the program, and a communication interface that communicates with a management device that manages the plurality of learning devices, wherein the processor performs a first transmission process that transmits the item group to the management device, a first reception process that receives from the management device a common item group used by the plurality of learning devices which is generated by the management device based on a plurality of item groups of the plurality of learning devices, and a first application process that applies the common item group received by the first reception process to the item group.

[0008] A third aspect of the invention disclosed in this application is a management device capable of communicating with a plurality of learning devices having a learning model that learns using training data which are the values ​​of each item in an item group, and comprising: a processor that executes a program; a storage device that stores the program; and a communication interface that communicates with the plurality of learning devices, wherein the processor performs a first receiving process that receives the item group from each of the plurality of learning devices; a first generating process that generates a common item group to be used by the plurality of computers based on the plurality of item groups received by the first receiving process; and a first transmitting process that transmits the common item group generated by the first generating process to the plurality of learning devices.

[0009] A learning program, which is a fourth aspect of the invention disclosed in this application, is characterized in that it causes the processor of any of the learning devices among a plurality of learning devices having a learning model that learns with training data which are the values ​​of each item in an item group, to execute a first transmission process for transmitting the item group to a management device; a first reception process for receiving a common item group for use in the plurality of computers, which is generated based on the plurality of item groups of the plurality of learning devices, from the management device; and a first application process for applying the common item group received by the first reception process to the item group.

[0010] A fifth aspect of the invention disclosed in this application is a management program characterized in that it causes the processor of a management device, which can communicate with a plurality of learning devices having a learning model that learns using training data which are the values ​​of each item in an item group, to execute a first receiving process that receives the item group from each of the plurality of learning devices; a first generating process that generates a common item group to be used by the plurality of computers based on the plurality of item groups received by the first receiving process; and a first transmitting process that transmits the common item group generated by the first generating process to the plurality of learning devices.

[0011] A sixth aspect of the invention disclosed in this application is a federated learning system that performs federated learning using a plurality of learning devices and a management device that manages the plurality of learning devices, wherein each of the plurality of learning devices has a learning model that is learned using training data which is the value of each item in an item group, and each of the plurality of learning devices performs a first transmission process to transmit the item group to the management device, the management device performs a first reception process to receive the item group from each of the plurality of learning devices, a first generation process to generate a common item group to be used by the plurality of computers based on the plurality of item groups received by the first reception process, and a second transmission process to transmit the common item group generated by the first generation process to each of the plurality of learning devices, and each of the plurality of learning devices performs a second reception process to receive the common item group and a first application process to apply the common item group received by the second reception process to the item group. [Effects of the Invention]

[0012] According to a typical embodiment of the present invention, federated learning can be achieved regardless of differences in data structures between locations. Other issues, configurations, and effects will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0013] [Figure 1] Figure 1 is an explanatory diagram showing an example of the system configuration of an associative learning system. [Figure 2] Figure 2 is a block diagram showing an example of a computer hardware configuration. [Figure 3] Figure 3 is an explanatory diagram showing an example of a dataset to be analyzed. [Figure 4] Figure 4 is an explanatory diagram showing an example of an in-house dictionary. [Figure 5] Figure 5 is a sequence diagram illustrating the operation of the associative learning system. [Figure 6] Figure 6 is a flowchart showing a detailed example of the metadata extraction process (step S501) shown in Figure 5. [Figure 7] FIG. 7 is an explanatory diagram showing an example of metadata acquisition. [Figure 8] FIG. 8 is an explanatory diagram showing an example of acquisition results. [Figure 9] FIG. 9 is an explanatory diagram showing an example of filtering application settings in metadata extraction processing (step S501). [Figure 10] FIG. 10 is an explanatory diagram showing an example of common metadata generation in common metadata generation processing (step S503). [Figure 11] FIG. 11 is an explanatory diagram showing an example of application of common metadata M in common metadata application processing (step S505). [Figure 12] FIG. 12 is an explanatory diagram showing an example of application results of common metadata M in common metadata application processing (step S505). [Figure 13] FIG. 13 is a sequence diagram showing overall optimization in a federated learning system. [Figure 14] FIG. 14 is an explanatory diagram showing an example of metadata performance evaluation results. [Figure 15] FIG. 15 is a flowchart showing an example of detailed processing procedures of common metadata update processing (step S1303).

MODE FOR CARRYING OUT THE INVENTION

[0014] <FIG. 1 System Configuration Example> FIG. 1 is an explanatory diagram showing a system configuration example of a federated learning system. The federated learning system 100 is composed of a server 101 and a plurality of client terminals C1 to Cn (n is an integer of 2 or more). The server 101 and each client terminal Ci (i is an integer satisfying 1 ≦ i ≦ n) are communicably connected via a network 102 such as the Internet, a LAN (Local Area Network), or a WAN (Wide Area Network).

[0015] The server 101 and the plurality of client terminals C1 to Cn may be installed at the same location or at different locations.

[0016] Server 101 is a management device that manages multiple client terminals C1 to Cn and has the functions to generate common metadata M and common learning parameters P. Specifically, for example, Server 101 obtains metadata M1 to Mn extracted from the analysis target data L1 to Ln of multiple client terminals C1 to Cn, and generates common metadata M that is applied to multiple client terminals C1 to Cn based on the metadata M1 to Mn (common metadata generation function). Also, Server 101 obtains learning parameters P1 to Pn of the learning models LM1 to LMn of multiple client terminals C1 to Cn, performs federated learning, and generates common learning parameters P that are applied to multiple client terminals C1 to Cn (common learning parameter generation function).

[0017] The client terminal Ci is a learning device installed at site Bi, possessing the analysis target dataset DSi and the site-internal dictionary DCi, and performing local learning. The analysis target dataset DSi is a combination of analysis target data Li and ground truth data Ai. There is one or more combinations of analysis target data Li and ground truth data Ai. Combinations of analysis target data Li and ground truth data Ai may be added over time. Ground truth data Ai may also be a reference to analysis target data Li itself.

[0018] The client terminal Ci extracts metadata Mi from the data Li to be analyzed. Metadata Mi is the information that is added when the data Li to be analyzed is summarized. For example, if the data Li to be analyzed is document data, the metadata Mi is the information that is added when the document data is summarized. The client terminal Ci may also use the local dictionary DCi when extracting metadata Mi from the data Li to be analyzed. The client terminal Ci sends the metadata Mi to the server 101.

[0019] The client terminal Ci updates metadata Mi with common metadata M. The client terminal Ci summarizes the data to be analyzed Li into structured information using the updated metadata Mi, and generates standardized data Si as a result of the summarization. Standardized data Si consists of words that become the values ​​of metadata Mi and is extracted from the data to be analyzed Li. Alternatively, the client terminal Ci may summarize the data to be analyzed Li into unstructured information using the updated metadata Mi, and generate standardized data Si as a result of the summarization. In this case, standardized data Si is represented, for example, as document data in which the correspondence between metadata Mi and the information that becomes the values ​​of metadata Mi is written in natural language. Standardized data Si is training data paired with ground truth data Ai. The combination of standardized data Si and ground truth data Ai is called the training dataset TSi. If ground truth data Ai is a reference to the data to be analyzed Li itself, ground truth data Ai may also be a reference to standardized data Si.

[0020] The client terminal Ci updates the learning model LMi using the training dataset TSi. The learning model LMi is, for example, a neural network with learning parameters Pi set. The learning parameters Pi are the set of weights for each neuron in the neural network. The client terminal Ci sends the learning parameters Pi to the server 101.

[0021] Server 101 and client terminal Ci may have a generating AI (Artificial Intelligence) 102. This allows Server 101 and client terminal Ci to input a query called a prompt to the generating AI 103 and obtain a response from the generating AI 103.

[0022] Furthermore, the generated AI 103 may be implemented on an external server 104 that is different from server 101 and client terminal Ci. The external server 104 is connected to server 101 and client terminal Ci via network 102 so as to be able to communicate with them.

[0023] The Generative AI 103 outputs a response upon receiving an inquiry. The Generative AI 103 includes a language model trained using natural language processing with a dataset, and generates sentences using this language model.

[0024] Furthermore, a language model is a type of probabilistic model used in natural language processing, and it is a model that probabilistically predicts how likely a given word or phrase or sentence is to occur in natural language. Specifically, a language model is a mathematical model used in the field of natural language processing to learn language patterns and grammatical rules, and to generate and understand natural language. For example, the Generative AI 103 uses a language model to calculate the probability of occurrence of a given word or phrase or sentence, and to compare the probability of occurrence of multiple words or phrases, in order to automatically generate the most likely word or phrase or sentence based on the context when predicting the next word or phrase or sentence.

[0025] In this way, when the Generative AI 103 receives a query called a prompt, it uses a language model trained on a vast amount of data to output an answer to that query.

[0026] Furthermore, the Generative AI 103 performs embedding, which converts the input text into numerical vectors that are easier for language models to process. Specifically, for example, the Generative AI 103 generates numerical vectors by using distributed representations such as word2vec, GloVe, BERT, and the text-embedding-ada-002 model.

[0027] The functions of server 101 may also be implemented on any of the client terminals C1 to Cn, specifically on client terminal Ci. In this case, server 101 becomes unnecessary.

[0028] <Figure 2: Example of hardware configuration of a computer (server 101, client terminal Ci, external server 104)> Figure 2 is a block diagram showing an example of the hardware configuration of a computer. Computer 200 includes a processor 201, a memory device 202, an input device 203, an output device 204, and a communication interface (communication IF) 205. The processor 201, memory device 202, input device 203, output device 204, and communication IF 205 are connected by a bus 206. The processor 201 controls computer 200. The memory device 202 serves as the work area for the processor 201. The memory device 202 is a non-temporary or temporary recording medium that stores various programs and data. Examples of memory devices 202 include ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), and flash memory. The input device 203 takes data in. Examples of input devices 203 include a keyboard, mouse, touch panel, numeric keypad, scanner, microphone, and sensor. The output device 204 outputs data. Output devices 204 include, for example, displays, printers, and speakers. The communication IF 205 connects to the network 102 and sends and receives data.

[0029] <Figure 3: Dataset DSi to be analyzed> Figure 3 is an explanatory diagram showing an example of the DSi dataset to be analyzed. The data Li to be analyzed is, for example, document data written in natural language. In this example, the data Li to be analyzed is, for example, document data representing the minutes of a meeting held at site Bi.

[0030] The data to be analyzed, Li, can be either unstructured data or structured data. Furthermore, the data to be analyzed, Li, may be semi-structured data with a granularity between unstructured and structured data. It may also be a combination of multiple data with different structures. For example, it may be a combination of document data and structured data containing different types of information related to that document data. In other words, the data to be analyzed, Li, is applicable regardless of the type of structure, the number of structures, and the granularity of the structure. Even when it is difficult to pre-design predetermined rules for converting unstructured data to semi-structured or structured data, or semi-structured data to structured data, or data combining multiple types of unstructured, semi-structured, and structured data to structured data, federative learning can be achieved. As a result, the client terminal Ci can stably update the highly accurate learning model LMi.

[0031] The ground truth data Ai is data that shows results related to the data being analyzed Li. For example, if the data being analyzed Li is the minutes of a meeting to increase sales at location Bi, then the ground truth data Ai would be the sales after that meeting. It could also be profit instead of sales. Alternatively, the ground truth data Ai could be the result indicating whether or not the sales reached the target amount. Furthermore, the ground truth data Ai could be the data being analyzed Li itself, for example, the ground truth data used to tune a language model that uses the data being analyzed Li as training data.

[0032] <Figure 4: In-house dictionary DCi> Figure 4 is an explanatory diagram showing an example of a site-internal dictionary DCi. The site-internal dictionary DCi has the following fields: classification 401, general name 402, and site-internal terminology 403. Classification 401 defines the affiliation of site-internal terminology 403. General name 402 is a general name for site-internal terminology 403. Unlike general name 402, site-internal terminology 403 is a specific or specialized term used within that site Bi. That is, classification 401 is a higher-level term for general name 402, and general name 402 is a higher-level term for site-internal terminology 403.

[0033] For example, the internal term 403 "DD0001" has a general name 402 of "storage device" and its classification 401 is "product code".

[0034] Furthermore, the internal term 403 "Sapporo First Branch" has a general name 402 of "Sales Branch" and its classification 401 is "Business Office".

[0035] <Figure 5 Sequence of the Associative Learning System 100> Figure 5 is a sequence diagram showing the operation of the federated learning system 100. The federated learning system 100 performs data commonization processing (step S500) and federated learning processing (step S510) using the server 101 and client terminals C1 to Cn.

[0036] (Step S500: Data standardization process) The data standardization process (step S500) is a process in which the server 101 generates common metadata M based on the metadata M1 to Mn of client terminals C1 to Cn, and client terminals C1 to Cn apply the common metadata M to generate standardized data S1 to Sn. The data standardization process (step S500) is executed repeatedly, either periodically or irregularly.

[0037] Furthermore, during repeated execution, the data to be analyzed Li and the ground truth data Ai may be added, modified, or deleted. Each client terminal Ci may input the newly added standardized data Si of the data to be analyzed Li into the learning model LMi, obtain prediction / classification results from the learning model LMi, and use them as new ground truth data Ai. The internal processing is described below.

[0038] (Step S501) Each client terminal Ci executes metadata extraction processing. Metadata extraction processing is the process of extracting metadata Mi from the data Li to be analyzed. For example, if the data Li to be analyzed is meeting minutes data as shown in Figure 3, then words and phrases from the meeting minutes data will be extracted as metadata Mi. Details of the metadata extraction processing will be described later in Figures 6 to 9.

[0039] (Step S502) Each client terminal Ci sends the metadata Mi extracted by the metadata extraction process (step S501) to the server 101. The server 101 receives the metadata M1 to Mn from the client terminals C1 to Cn.

[0040] (Step S503) Server 101 executes a common metadata generation process using metadata M1 to Mn received from client terminals C1 to Cn, and generates common metadata M. For example, if there are words common to a predetermined number or more of the metadata M1 to Mn, those words are included in the common metadata M. Details of the common metadata generation process will be described later in Figure 10.

[0041] (Step S504) Server 101 distributes common metadata M to client terminals C1 to Cn. Client terminal Ci receives the common metadata M.

[0042] (Step S505) Each client terminal Ci performs a common metadata application process. Specifically, for example, the client terminal Ci applies the received common metadata M to summarize the data Li to be analyzed into structured information, and generates standardized data Si as a result of the summarization. Details of the common metadata application process will be described later in Figures 11 and 12.

[0043] (Step S506) Each client terminal Ci executes the standardized data registration process. The combination of standardized data Si obtained from the data to be analyzed Li and the ground truth data Ai is used to update the learning model LMi. This completes the data commonization process (step S500).

[0044] (Step S510: Associative Learning Process) The associative learning process (step S510) generates common learning parameters P based on the learning parameters P1 to Pn of client terminals C1 to Cn and reflects them in the learning models LM1 to LMn of client terminals C1 to Cn. The associative learning process (step S510) is executed repeatedly, either periodically or irregularly.

[0045] Furthermore, during repeated execution, the data to be analyzed Li and the ground truth data Ai may be added, modified, or deleted. Each client terminal Ci may input the newly added standardized data Si of the data to be analyzed Li into the learning model LMi, obtain prediction / classification results from the learning model LMi, and use them as new ground truth data Ai. The internal processing is described below.

[0046] (Step S511) Server 101 distributes common learning parameters P to client terminals C1 to Cn. Client terminal Ci receives the common learning parameters P. In the initial stage, pre-configured common learning parameters P are distributed. If the federated learning process (step S510) is repeated, the latest common learning parameters P are distributed.

[0047] (Step S512) Each client terminal Ci executes the first model update process. Specifically, for example, when each client terminal Ci receives a common learning parameter P, it updates the learning parameter Pi of the learning model LMi with the received common learning parameter P. Furthermore, each client terminal Ci inputs standardized data Si into the updated learning model LMi and obtains the output result from the updated learning model LMi. Based on the difference between the output result and the ground truth data Ai, each client terminal Ci updates the learning parameter Pi of the updated learning model LMi.

[0048] (Step S513) Each client terminal Ci sends the updated learning parameter Pi in step S512 to the server 101. The server 101 receives the learning parameter Pi.

[0049] (Step S514) Server 101 executes the model generation process. Specifically, for example, Server 101 generates common learning parameters P based on the learning parameters P1 to Pn received from client terminals C1 to Cn. Specifically, Server 101 generates common learning parameters P from the learning parameters P1 to Pn through statistical processing. For example, Server 101 generates common learning parameters P by calculating the average value of the learning parameters P1 to Pn.

[0050] (Step S515) Server 101 distributes the common learning parameters P generated in step S514 to client terminals C1 to Cn. Client terminal Ci receives the common learning parameters P.

[0051] (Step S516) Each client terminal Ci executes the second model update process. Specifically, for example, each client terminal Ci uses the common learning parameters P generated in step S514 to perform the same process as the first model update process. This completes the federated learning process (step S510).

[0052] <Figure 6 Metadata extraction process (step S501)> Figure 6 is a flowchart showing a detailed example of the metadata extraction process (step S501) shown in Figure 5.

[0053] (Step S601) The client terminal Ci uses the generated AI103 to obtain metadata from the data Li to be analyzed. The metadata obtained in step S601 is denoted as obtained metadata m1i.

[0054] (Step S602) The client terminal Ci calculates the similarity between the acquired metadata m1i and the local term 403 in the local dictionary DCi, and detects metadata containing the local term 403 based on the similarity. Specifically, for example, the client terminal Ci converts the acquired metadata m1i and the local term 403 in the local dictionary DCi into numerical vectors by embedding the generated AI 103. Then, the client terminal Ci calculates the similarity between the acquired metadata m1i and the local term 403 in the local dictionary DCi based on the distance between the vectors. The shorter the distance between the vectors, the more similar they are considered to be. The metadata detected in step S602 is denoted as detected metadata m2i.

[0055] (Step S603) The client terminal Ci uses the generation AI 103 or the site dictionary DCi to convert the detected metadata m2i into metadata containing only common terms. Specifically, for example, if the detected metadata m2i is a site term 403, the client terminal Ci outputs a prompt to the generation AI 103 to convert the detected metadata m2i into metadata containing only common terms, and obtains the common terms as the response from the generation AI 103. If the generation AI 103 is used, the client terminal Ci may register the common terms obtained from the generation AI 103 in the site dictionary DCi as a general name 402 corresponding to the site term 403 included in the detected metadata m2i.

[0056] When using the local dictionary DCi, the client terminal Ci may convert the detected metadata m2i into a general name 402 corresponding to the local term 403 contained in the detected metadata m2i. Alternatively, if the detected metadata m2i is a local term 403, the client terminal Ci may convert it into its higher-level concept, a general name 402 or a classification 401. Alternatively, if the detected metadata m2i is a general name 402, the client terminal Ci may convert it into its higher-level concept, a classification 401. The metadata converted from the detected metadata m2i in step S603 is denoted as converted metadata m3i.

[0057] (Step S604) The client terminal Ci aggregates the transformed metadata m3i using the generating AI 103. Specifically, for example, the client terminal Ci classifies groups of words within the transformed metadata m3i into similar groups. The words are converted into numerical vectors by embedding using the generating AI 103. Therefore, the client terminal Ci groups together multiple words whose inter-vector distance is below a threshold.

[0058] The client terminal Ci concatenates the classified phrase groups by inserting delimiters between them. For example, if similar phrase groups are "attendees" and "participants," a " / " is inserted as a delimiter between "attendees" and "participants" to concatenate them into "attendees / participants." The metadata aggregated from the converted metadata m3i in step S604 is referred to as the aggregated metadata m4i. For the sake of explanation, phrases in the converted metadata m3i that were not aggregated are also included in the aggregated metadata m4i.

[0059] (Step S605) The client terminal Ci aggregates the m4i aggregate metadata and creates a frequency distribution. The frequency distribution shows, for example, the number of times each word or phrase appears in the m4i aggregate metadata.

[0060] (Step S606) The client terminal Ci filters the aggregated metadata m4i based on the frequency distribution created in step S605. Specifically, for example, the client terminal Ci extracts from the aggregated metadata m4i words whose occurrence count is above a threshold. The group of words extracted from the aggregated metadata m4i is referred to as the filtered metadata m5i. This completes the metadata extraction process.

[0061] The client terminal Ci outputs one of the following as extracted metadata Mi: acquired metadata m1i, transformed metadata m3i, aggregated metadata m4i, or filtered metadata m5i. In other words, in the metadata extraction process (step S501), at least one of steps S602 and S603, step S604, and steps S605 and S606 is executed. Which of these steps to execute may be pre-configured on the client terminal Ci, or it may be selected by the user of the client terminal Ci.

[0062] Furthermore, the client terminal Ci may perform filtering to output only the extracted metadata Mi that can be publicly shared to the server 101. Specifically, for example, the client terminal Ci may have a dictionary that defines words that cannot be publicly shared, and delete the words that cannot be publicly shared from the extracted metadata Mi before sending it to the server 101.

[0063] <Figure 7 Example of metadata retrieval> Figure 7 is an explanatory diagram showing an example of metadata acquisition. The acquisition screen 700 is displayed on the client terminal Ci in step S601. The acquisition screen 700 has a metadata acquisition tab 701 and an acquisition results tab 702. Figure 7 shows the state where the metadata acquisition tab 701 is selected.

[0064] The metadata acquisition tab 701 includes a data specification method display unit 711, a language model setting display unit 712, a metadata acquisition prompt display unit 713, and a training data display unit 714.

[0065] The data specification method display unit 711 displays the data specification method. The data specification method is a method for specifying the data Li to be analyzed, and there are methods such as specifying by document input and specifying by file / folder. In the example in Figure 7, the data Li to be analyzed is specified in the learning data display unit 714 by document input.

[0066] The language model setting display unit 712 displays the language model settings. The language model setting is a method for setting the language model included in the generated AI 103, and the language model can be selected from a pull-down menu. The temperature parameter that controls the selected language model can also be adjusted. The temperature parameter ranges from 0 to 1, and the higher the value, the more random the response tends to be generated, while the lower the value, the more confident the response tends to be generated.

[0067] The metadata acquisition prompt display unit 713 displays the metadata acquisition prompt 730. The metadata acquisition prompt 730 is stored as a template and is displayed on the metadata acquisition prompt display unit 713 when the acquisition screen 700 is displayed. The metadata acquisition prompt 730 is a query requesting the generation AI 103 to generate combinations of metadata and their respective phrases, by assigning metadata representing item names to the words in the analysis target data Li displayed on the learning data display unit 714. In the metadata acquisition prompt 730, "input text" refers to the analysis target data Li displayed on the learning data display unit 714.

[0068] When the send button 731 is pressed, the metadata acquisition prompt 730 and the data Li to be analyzed are sent to the generating AI 103. As a result, the client terminal Ci obtains a response from the generating AI 103.

[0069] <Figure 8 Example of Acquisition Results> Figure 8 is an explanatory diagram showing an example of acquisition results. In Figure 8, the acquisition results tab 702 is selected on the acquisition screen 700. The acquisition results tab 702 includes a summary data display unit 801, an acquisition metadata display unit 802, and a save button 803.

[0070] The summary data display unit 801 displays summary data 810, which is a summary of the data Li to be analyzed, as the response obtained from the generating AI 103. The summary data 810 includes the acquired metadata m1i obtained from the data Li to be analyzed as a key, and the string that is its value. In the summary data 810, the acquired metadata m1i is a phrase enclosed in quotation marks (date and time, location, participants, purpose of the meeting, minutes, results of the meeting). The phrase that is the value follows the acquired metadata m1i.

[0071] For example, the value of the acquired metadata m1i, "Date and Time," is "October 15, 2023, 14:00-15:30."

[0072] The acquired metadata display unit 802 displays the acquired metadata m1i as "date and time, location, participants, purpose of the meeting, minutes, and results of the meeting."

[0073] The save button 803 is a user interface for saving the summary data 810 and acquired metadata m1i to the storage device 202 when pressed. In this example, the summary data 810 and acquired metadata m1i are assumed to be saved in the directory specified by "C:\Data\2024\".

[0074] <Figure 9: Example of filtering settings in the metadata extraction process (step S501)> Figure 9 is an explanatory diagram showing an example of filtering application settings in the metadata extraction process (step S501). The metadata filtering screen 900 is displayed on the client terminal Ci in steps S602 to S606. The metadata filtering screen 900 includes an acquisition metadata specification method display unit 901, a language model setting display unit 712, a filtering application setting display unit 902, an in-site term detection display unit 903, a metadata aggregation display unit 904, a frequency determination display unit 905, an apply button 906, and a filtered metadata display unit 907.

[0075] [Method for specifying retrieved metadata] The Acquisition Metadata Specification Method Display Unit 901 displays the acquisition metadata specification method. The acquisition metadata specification method is a method for specifying the data Li to be analyzed, and there are methods such as specifying by document input and specifying by file / folder. In the example in Figure 8, the file / folder specification is selected, and the acquisition metadata m1i is specified from the directory "C:\Data\2024\".

[0076] [Filtering Application Settings] The filtering application settings display unit 902 displays the filtering application settings. The filtering application settings are the settings for applying filtering, and include site term detection, metadata conversion, metadata aggregation, and frequency determination. When site term detection is selected, the site term detection display unit 903 is displayed, and steps S602 and S603 are executed. When metadata aggregation is selected, the metadata aggregation display unit 904 is displayed, and step S604 is executed.

[0077] When frequency determination is selected, the frequency determination display unit 905 is displayed, and steps S605 and S606 are executed. In Figure 9, site term detection (steps S602 and S603), metadata aggregation (step S604), and frequency determination (steps S605 and S606) are selected.

[0078] [Detection of in-house terminology and conversion to general terminology] The in-site terminology detection and display unit 903 includes an in-site dictionary specification unit 931, an execution button 932, an in-site terminology detection result display unit 933, a general term conversion prompt display unit 934, and another general term conversion prompt display unit 934. The in-site dictionary specification unit 931 is a user interface for receiving and calling the in-site dictionary DCi.

[0079] The execution button 932 is a user interface for executing step S602 using the in-site dictionary DCi specified by the in-site dictionary specification unit 931 when pressed. The in-site term detection result display unit 933 is an area that displays the detection metadata m2i, which is the in-site term detection result, when step S602 is executed by pressing the execution button 932.

[0080] In Figure 9, for example, the detection metadata m2i is " (DDD) Development Costs Twelve instances of "Usage Status" have been detected. Of these, the underlined part is: (DDD) Development Costs This is entry 403 in the DCi internal dictionary, which is an internal term within the base.

[0081] The general term conversion prompt display unit 934 displays the general term conversion prompt 935. The general term conversion prompt 935 is stored as a template and is displayed on the general term conversion prompt display unit 934 when site-specific term detection is selected in the filtering application setting display unit 902.

[0082] The general term conversion prompt 935 is a query requesting that metadata (acquired metadata m1i or detected metadata m2i) be converted into more generalized terminology. In the general term conversion prompt 935, "extracted term" refers to the detected metadata m2i displayed on the in-site terminology detection display unit 903 if in-site terminology detection is selected, and to the acquired metadata m1i if in-site terminology detection is not selected.

[0083] When the send button 936 is pressed, the general term conversion prompt 935 and metadata (acquired metadata m1i or detected metadata m2i) are sent to the generating AI 103. As a result, the client terminal Ci obtains a response from the generating AI 103.

[0084] The general term conversion result display unit 937 is an area that displays the conversion metadata m3i, which is the generalized term conversion result, as the answer obtained from the generating AI 103. In Figure 9, (DDD) Development Costs "Usage status" is development cost It has been generalized to "usage status", CD7800 "Delivery plan" is product It has been generalized to "delivery plan," XX factory "Order status to" is External factory This is generally referred to as "the status of orders to [the company / organization]."

[0085] [Metadata Aggregation] The metadata aggregation display unit 904 displays the metadata aggregation prompt 941. The metadata aggregation prompt 941 is stored as a template and is displayed in the metadata aggregation display unit 904 when metadata aggregation is selected in the filtering application setting display unit 902.

[0086] The metadata aggregation prompt 941 is a query that requests the aggregation of metadata (acquired metadata m1i or transformed metadata m3i) into groups of words with similar meanings. In the metadata aggregation prompt 941, "extracted metadata" refers to transformed metadata m3i if site-specific terminology detection is selected, and acquired metadata m1i if site-specific terminology detection is not selected.

[0087] When the send button 943 is pressed, the metadata aggregation prompt 941 and metadata (acquired metadata m1i or transformed metadata m3i) are sent to the generating AI 103. As a result, the client terminal Ci obtains a response from the generating AI 103.

[0088] The metadata aggregation result display unit 944 displays the aggregated metadata m4i, which is the metadata aggregation result, as the answer obtained from the generating AI 103. In Figure 9, for example, the transformed metadata m3i "Attendees" and "Participants" are aggregated into the aggregated metadata m4i "Attendees / Participants". Also, the transformed metadata m3i "Location" remains as "Location" because there is no similar transformed metadata m3i, but for convenience it is designated as aggregated metadata m4i.

[0089] [Frequency determination] The frequency determination display unit 905 displays the frequency determination result 951 of the metadata (acquired metadata m1i, converted metadata m3i, or aggregated metadata m4i). In Figure 9, the metadata subject to the frequency determination result 951 is the aggregated metadata m4i. In the frequency determination result 951, for example, the number of occurrences of each word in the metadata (acquired metadata m1i, converted metadata m3i, or aggregated metadata m4i) is normalized to the proportion of the total number of occurrences.

[0090] [Apply filtering] The Apply button 906 is a user interface for applying filtering of the metadata (acquired metadata m1i, converted metadata m3i, or aggregated metadata m4i) displayed on the frequency determination display unit 905 when pressed. Since the frequency determination display unit 905 specifies a percentage of 10% or more, words with a frequency of 10% or more are extracted as metadata Mi.

[0091] [Filtering results] The filtered metadata display unit 907 displays the metadata Mi after filtering, the result editing button 971, and the result submission button 972. The filtered metadata Mi is the metadata Mi extracted by pressing the apply button 906.

[0092] The filtered metadata Mi displayed on the filtered metadata display unit 907 can be edited by operating the input device 203. The result editing button 971 is a user interface for confirming the latest metadata Mi displayed on the filtered metadata display unit 907 when pressed. The result transmission button 972 is a user interface for sending the filtered metadata Mi displayed on the filtered metadata display unit 907 to the server 101 (step S502) when pressed.

[0093] <Figure 10: Example of common metadata generation in the common metadata generation process (step S503)> Figure 10 is an explanatory diagram showing an example of common metadata generation in the common metadata generation process (step S503). In Figure 10, n=2, and the example shows how server 101 generates common metadata based on the extracted metadata M1 from client terminal C1 and the extracted metadata M2 from client terminal C2.

[0094] Server 101 calls a common metadata generation prompt 1000 from storage device 202 and sends extracted metadata M1 and M2 to generation AI 103. The common metadata generation prompt 1000 is a query requesting that groups of words in multiple metadata Mi be aggregated and redefined into a single word.

[0095] The generating AI 103 returns common metadata M to the server 101 as a response. Specifically, for example, if the extracted metadata M1 contains "Date and Time" and the extracted metadata M2 contains "Date and Time of Meeting," "Date and Time" is added to the common metadata M because it is common. Similarly, for "Purpose of the Meeting" in extracted metadata M1 and "Purpose" in extracted metadata M2, "Meeting Content" in extracted metadata M1 and "Meeting Minutes" in extracted metadata M2, and "Results of the Meeting" in extracted metadata M1 and "Results" in extracted metadata M2, "Purpose," "Meeting Minutes," and "Results" are added to the common metadata M.

[0096] In the case of "Participants" in extracted metadata M1 and "Participants" in extracted metadata M2, they are completely identical, so "Participants" is added to common metadata M.

[0097] Since the phrase containing "location" in extracted metadata M1 is not present in extracted metadata M2, "location" is added to common metadata M to improve comprehensiveness. Similarly, "action item" in extracted metadata M2 is also added to common metadata M.

[0098] Server 101 may perform the common metadata generation process (step S503) without using the generation AI 103.

[0099] <Figure 11: Example of application of common metadata M in the common metadata application process (step S505)> Figure 11 is an explanatory diagram showing an example of applying common metadata M in the common metadata application process (step S505). The common metadata application screen 1100 is displayed on the client terminal Ci during the common metadata application process (step S505). The common metadata application screen 1100 has a common metadata application tab 1101 and an application results tab 1102. Figure 11 shows the state where the common metadata application tab 1101 is selected.

[0100] The common metadata application tab 1101 includes a data specification method display unit 711, a language model setting display unit 712, a common metadata setting unit 1110, a common metadata display unit 1111, and a learning data display unit 714.

[0101] The common metadata setting unit 1110 sets the common metadata M. Specifically, for example, the common metadata setting unit 1110 retrieves the common metadata M by specifying the directory where the common metadata M is stored through user operation. The retrieved common metadata M is displayed on the common metadata display unit 1111.

[0102] The common metadata display unit 1111 stores the common metadata M and the common metadata application prompt 1112 as templates, and when the common metadata M is displayed, it is displayed on the common metadata display unit 1111. The common metadata application prompt 1112 is a query requesting the generating AI 103 to summarize the data Li to be analyzed into semi-structured information using the common metadata M as the key. In the metadata acquisition prompt 730, "document" refers to the data Li to be analyzed displayed on the learning data display unit 714.

[0103] When the send button 1113 is pressed, the common metadata application prompt 1112, the common metadata M, and the data to be analyzed Li are sent to the generating AI 103. As a result, the client terminal Ci obtains a response from the generating AI 103.

[0104] <Figure 12: Example of the application result of common metadata M in the common metadata application process (step S505)> Figure 12 is an explanatory diagram showing an example of the application result of common metadata M in the common metadata application process (step S505). In Figure 12, the application result tab 1102 is selected on the common metadata application screen 1100. The application result tab 1102 displays the common metadata application result 1200 and the common metadata editing unit 1210.

[0105] The common metadata application result 1200 includes a standardized data display unit 1201, a non-standardized data display unit 1202, a result save button 1203, and a federated learning start button 1204.

[0106] The standardized data display unit 1201 displays standardized data Si, which is a summary of the data Li to be analyzed into structured information, as the answer obtained from the generating AI 103. The standardized data Si consists of phrases obtained as values ​​from the data Li to be analyzed, using the common metadata M as the key.

[0107] For example, for the common metadata M, "Attendees," three values ​​were obtained from the data Li being analyzed: "Sales Manager Taro Yamada," "Director Ichiro Tanaka, XX City Water Bureau," and "XX City Water Bureau."

[0108] The non-standard data display unit 1202 displays the non-standard data 1206. The non-standard data 1206 is structured data obtained by retrieving words that become values ​​from the data Li to be analyzed, using words other than common metadata M as keys.

[0109] For example, regarding the phrase "technical question," which is not a common metadata M, two values, "water pressure supported by the water purifier" and "installation conditions," are obtained from the data Li being analyzed.

[0110] The result save button 1203 is a user interface for saving standardized data Si and non-standardized data 1206 to the storage device 202 when pressed.

[0111] The federated learning start button 1204 is a user interface that, when pressed, initiates federated learning; specifically, it allows the client terminal Ci to begin the local learning necessary for federated learning.

[0112] The common metadata editing area 1210 is an area for editing common metadata M. The common metadata M is displayed in the common metadata editing area 1210. The common metadata M can be added, modified, or deleted by operating the input device 203. For example, the non-standardized data 1206 can be added by operating the input device 203.

[0113] The common metadata update button 1211 is a user interface for updating the common metadata M and saving it to the storage device 202 when pressed.

[0114] <Figure 13 Overall Optimization Sequence> Figure 13 is a sequence diagram showing the overall optimization in the federated learning system 100. Overall optimization is the process of optimizing the common metadata M after the data commonization process (step S500) and the federated learning process (step S510). The learning model LMi used by the client terminal Ci is an explainable learning model, also known as XAI, and the learning model LMi calculates the influence of the input features. In this example, the input features are numerical vectors embedded with standardized data Si. The embedding of standardized data Si may be performed, for example, by the learning model LMi, or it may be performed by the generating AI 103 at the request of the client terminal Ci.

[0115] (Step S1301) The client terminal Ci performs metadata performance evaluation processing. Specifically, for example, in the federated learning process (step S510), the first model update process (step S512) performed by the client terminal Ci obtains the influence of metadata Mi calculated by the learning model LMi as the metadata performance evaluation result.

[0116] (Step S1302) The client terminal Ci sends the metadata performance evaluation result Ri to the server 101. The server 101 receives the metadata performance evaluation result Ri.

[0117] (Step S1303) Server 101 executes a common metadata update process based on the metadata performance evaluation results R1 to Rn obtained from client terminals C1 to Cn.

[0118] (Step S1304) Server 101 distributes the latest common metadata M, updated by the common metadata update process (step S1303), to client terminals C1 to Cn as the overall optimization result. Client terminal Ci receives the latest common metadata M.

[0119] (Step S1305) The client terminal Ci registers the latest common metadata M, which is the overall optimization result. Specifically, for example, the client terminal Ci uses the registered common metadata M to perform the common metadata application process (step S505) and the standardized data registration process (step S506). As a result, if the federated learning process (step S510) is then performed, the client terminal Ci will use the latest standardized data Si to perform local learning and generate and update the learning parameters Pi.

[0120] <Figure 14 Metadata Performance Evaluation Results> Figure 14 is an explanatory diagram showing an example of metadata performance evaluation results. In Figure 14, n=2, and the metadata performance evaluation results R1 for client terminal C1 and R2 for client terminal C2 are shown.

[0121] The metadata performance evaluation results R1 and R2 are obtained in step S1301 and sent to the server 101 in step S1302. The metadata performance evaluation results R1 and R2 have metadata Mi and an impact score of 1400. The impact score of 1400 is a value between 0.0 and 1.0, for example, and a larger value means that the impact score of 1400 for that metadata Mi is higher.

[0122] <Figure 15 Common metadata update process (step S1303)> Figure 15 is a flowchart showing a detailed example of the common metadata update process (step S1303).

[0123] (Step S1501) Server 101 performs statistical processing on the metadata performance evaluation results R1 to Rn for sites B1 to Bn. Specifically, for example, Server 101 performs statistical processing on the influence score of 1400 for each metadata Mi of the same item. Here, statistical processing may be done by adding up the influence scores of 1400 for each metadata Mi of the same item, or by taking the average value obtained by dividing the sum of the influence scores of 1400 for each metadata Mi of the same item by n. Alternatively, statistical processing may be done by taking the maximum, minimum, median, or mode from the influence scores of 1400 for each metadata Mi of the same item.

[0124] (Step S1502) Server 101 updates the common metadata M based on the statistical processing results of step S1501. Specifically, for example, Server 101 updates the common metadata M with metadata Mi whose statistical processing results are above a threshold. Additionally, Server 101 updates the common metadata M with metadata Mi up to the top x-th (x is a pre-set integer of 1 or greater) of the statistical processing results.

[0125] This means that the higher the impact score of 1400, the more likely it is to be adopted as the latest common metadata M. Therefore, the client terminal Ci can locally learn using the latest common metadata M as metadata Mi, and update the learning parameters Pi to ones that improve prediction accuracy compared to the learning parameters Pi obtained with the metadata Mi before the update.

[0126] Furthermore, since the associative learning process (step S510) is executed with the updated learning parameters Pi in this way, the server 101 can provide optimized common learning parameters P to client terminals C1 to Cn.

[0127] As explained above, this embodiment enables federated learning by absorbing differences in data structures between locations without disclosing the data Bi itself. Therefore, for example, even if the data structure of the data Li to be analyzed differs between locations, standardized data Si (training data) corresponding to the common metadata M common to locations B1 to Bn can be generated using the common metadata M. Consequently, the high-precision learning model LMi can be stably updated at each client terminal Ci.

[0128] Furthermore, even if the data to be analyzed, Li, is unstructured, it can be treated as structured data by extracting metadata Mi. Therefore, federative learning can be achieved even when it is difficult to pre-design predetermined rules for transforming the data to be analyzed, Li, into specific features. As a result, the client terminal Ci can stably update the highly accurate learning model LMi.

[0129] Furthermore, the above-described embodiment can also be configured as shown in (1) to (13) below.

[0130] (1) A federated learning method in which multiple computers (client terminals C1 to Cn) have a learning model (LMi) that is trained on a training dataset (TSi) which includes at least the training data, from the training data (standardized data Si) which is the value of each item in the item group (extracted metadata Mi) and the ground truth data (Ai) corresponding to the training data, A first generation process (step S503) which obtains the item group from each of the multiple computers and generates a first common item group (common metadata M) to be used by the multiple computers based on the multiple item groups, In each of the aforementioned computers, a first application process (step S505) is performed to apply the first common item group generated by the first generation process to the item group, An associative learning method characterized by including

[0131] (2) The associative learning method described in (1) above, In the first generation process (step S503), if the first item (e.g., location) in the first item group among the multiple item groups matches the second item (e.g., location) in the second item group among the multiple item groups, the first item (e.g., location) is added to the first common item group. A characteristic associative learning method.

[0132] (3) The associative learning method described in (1) above, In the first generation process (step S503), if the first item (e.g., date and time) in the first item group among the multiple item groups is included in the second item (e.g., event date and time) in the second item group among the multiple item groups, the first item (e.g., date and time) is added to the first common item group. A characteristic associative learning method.

[0133] (4) The associative learning method described in (1) above, In the first generation process (step S503), if the first item (e.g., action item) in the first item group among the multiple item groups is not included in the second item group among the multiple item groups, the first item (e.g., action item) is added to the first common item group. A characteristic associative learning method.

[0134] (5) An associative learning method described in any one of (1) to (4) above, An extraction process (step S501) is performed in which the training data is extracted from the data to be analyzed (Li) in each of the aforementioned multiple computers, An associative learning method characterized by including

[0135] (6) The associative learning method described in (5) above, The data to be analyzed is document data, The aforementioned item is the first word, The value of the aforementioned item is the second word or phrase corresponding to the aforementioned item in the document data. A characteristic associative learning method.

[0136] (7) The associative learning method described in (6) above, The document data is either unstructured data, structured data, or semi-structured data whose granularity of structure is between that of the unstructured data and the structured data. A characteristic associative learning method.

[0137] (8) The associative learning method described in (6) above, In the extraction process (step S501), each of the multiple computers outputs a query to the generating AI (103) requesting the addition of metadata to the document data. As a result, the metadata added to the document data by the generating AI is obtained as the first phrase. A characteristic associative learning method.

[0138] (9) The associative learning method described in (6) above, In the extraction process (step S501), each of the multiple computers outputs a query to the generating AI (103) requesting a summary of the document data and the addition of metadata. As a result, the summaries of the document data are obtained from the generating AI, and the metadata added to the summaries of the document data is obtained from the generating AI as the first phrase. A characteristic associative learning method.

[0139] (10) An associative learning method described in any one of (6) to (8) above, Each of the aforementioned computers has dictionary data (internal dictionary DCi) that associates higher-level terms with lower-level terms. In the extraction process (step S501), for each computer, the first phrase containing the word in the dictionary data is detected based on the similarity between the word in the dictionary data and the first phrase. A characteristic associative learning method.

[0140] (11) The associative learning method of (10) above, In the extraction process (step S501), for each computer, if the word in the dictionary data is a word of a lower-level concept (local term 403), the first word containing the word in the dictionary data is converted into the first word containing the corresponding word of a higher-level concept (general name 402). A characteristic associative learning method.

[0141] (12) An associative learning method described in any one of (6) to (11) above, In the extraction process (step S501), for each computer, multiple similar first phrases are aggregated. A characteristic associative learning method.

[0142] (13) An associative learning method described in any one of (6) to (12) above, In the extraction process (step S501), for each of the first terms, the frequency of occurrence of the second terms corresponding to the first term is counted for each computer, and the second terms are narrowed down based on the counting results. A characteristic associative learning method.

[0143] (14) An associative learning method described in any one of (6) to (13) above, In the extraction process, for each computer, the first phrase is narrowed down to a specific first phrase that can be made publicly available. A characteristic associative learning method.

[0144] (15) An associative learning method described in any one of (5) to (14) above, In the first application process (step S505), each of the multiple computers obtains a summary of the data to be analyzed based on the applied item group to which the first common item group has been applied, and obtains the values ​​of each item in the applied item group from the summary of the data to be analyzed. A characteristic associative learning method.

[0145] (16) The associative learning method of (15) above, In the first application process (step S505), each of the multiple computers outputs a query to the generating AI (103) requesting a summary of the data to be analyzed based on the applied item group. As a result, the summaries of the data to be analyzed are obtained from the generating AI, and the values ​​of each item in the applied item group are obtained from the summaries of the data to be analyzed. A characteristic associative learning method.

[0146] (17) The associative learning method described in (15) or (16) above, In the first application process (step S505), each of the multiple computers obtains a summary of the data to be analyzed based on items other than those based on the applied item group (the "technical questions" of the non-standardized data 1206), obtains the values ​​of the other items from the summary of the data to be analyzed, and accepts the addition of the other items to the applied item group. A characteristic associative learning method.

[0147] (18) An associative learning method described in any one of (1) to (17) above, A first update process (step S512) is performed on each of the plurality of computers to update the learning model using the updated training dataset, which includes the updated training data, which is the value of each item in the applied item group, and the ground truth data. A second generation process (step S514) generates common learning parameters to be used by the multiple computers based on the multiple learning parameters of the multiple learning models updated in each of the multiple computers by the first update process, In each of the aforementioned computers, a second update process (step S517) is performed to update the learning parameters with the common learning parameters generated by the second generation process, An associative learning method characterized by including

[0148] (19) The associative learning method of (18) above, A third generation process (step S1303) generates a second common item group (common metadata M) based on the influence of each item in the item group from each of the multiple computers (influence of metadata performance evaluation result Ri = 1400), based on the multiple item groups. In each of the aforementioned computers, a second application process (step S1305) is performed to apply the second common item group generated by the third generation process to the item group, An associative learning method characterized by including

[0149] (20) One of the learning devices (client terminals C1 to Cn) having a learning model (LMi) that is trained on a training dataset (TSi) which includes at least the training data, which consists of the values ​​of each item in the item group (extracted metadata Mi) (standardized data Si) and the ground truth data (Ai) corresponding to the training data, It has a processor (201) that executes a program, a storage device (202) that stores the program, and a communication interface (205) that communicates with a management device (101) that manages the plurality of learning devices. The aforementioned processor, A first transmission process (step S502) that transmits the aforementioned group of items to the management device, The management device receives from the management device a first receiving process (step S504) which is a first common item group (common metadata M) used by the multiple learning devices that have generated multiple item groups from the multiple learning devices, A first application process (step S505) applies the first common item group received by the first reception process to the item group, A learning device characterized by performing the following actions.

[0150] (21) A management device (101) that can communicate with a plurality of learning devices (client terminals C1 to Cn), each having a learning model (LMi) that is trained on a training dataset (TSi) which includes at least the training data, from the training data (standardized data Si) which is the value of each item in the item group (extracted metadata Mi) and the correct answer data (Ai) corresponding to the training data, It comprises a processor (201) for executing a program, a storage device (202) for storing the program, and a communication interface (205) for communicating with the plurality of learning devices. The aforementioned processor, A first receiving process (step S502) which receives the item group from each of the plurality of learning devices, A first generation process (step 503) generates a first common item group (common metadata M) to be used by the multiple computers based on the multiple item groups received by the first reception process, A first transmission process (step S504) transmits the first common item group generated by the first generation process to the plurality of learning devices, A management device characterized by performing the following actions.

[0151] (22) The control device described in (21) above, The management device is a computer (server 101) different from the plurality of learning devices. A management device characterized by the following features.

[0152] (23) The control device described in (21) above, The management device is one of the learning devices (client terminal Ci) among the plurality of learning devices. A management device characterized by the following features.

[0153] (24) A learning model (LMi) is trained on a training dataset (TSi) which includes at least the training data, which consists of the training data (standardized data Si) which is the value of each item in the item group (extracted metadata Mi), and the correct answer data (Ai) corresponding to the training data. The processor (201) of one of the learning devices (client terminals C1 to Cn) has a learning model (LMi) which is trained on a training dataset (TSi) which includes the training data. A first transmission process (step S502) that transmits the aforementioned group of items to the management device (101), A first reception process (step S504) receives from the management device a first common item group (common metadata M) for use in the multiple computers, which is generated based on multiple item groups of the multiple learning devices, A first application process (step S505) applies the first common item group received by the first reception process to the item group, A learning program characterized by causing the execution of [a specific action].

[0154] (25) A learning model (LMi) is trained on a training dataset (TSi) which includes training data (standardized data Si) which are the values ​​of each item in the item group (extracted metadata Mi), and ground truth data (Ai) corresponding to the training data. The processor (201) of a management device (101) that can communicate with multiple learning devices (client terminals C1 to Cn) has the following: A first receiving process (step S502) which receives the item group from each of the plurality of learning devices, A first generation process (step 503) generates a first common item group (common metadata M) to be used by the multiple computers based on the multiple item groups received by the first reception process, A first transmission process (step S504) transmits the first common item group generated by the first generation process to the plurality of learning devices, A management program characterized by causing the execution of a command.

[0155] (26) A federated learning system (100) that performs federated learning using multiple learning devices (client terminals C1 to Cn) and a management device (101) that manages the multiple learning devices, Each of the aforementioned learning devices has a learning model (LMi) that is trained on a training dataset (TSi) which includes at least the training data, which consists of training data (standardized data Si) which are the values ​​of each item in the item group (extracted metadata Mi), and ground truth data (Ai) corresponding to the training data. Each of the aforementioned learning devices A first transmission process (step S502) is performed to transmit the aforementioned group of items to the management device. The aforementioned control device A first receiving process (step S502) which receives the item group from each of the plurality of learning devices, A first generation process (step S503) generates a first common item group (common metadata M) to be used by the multiple computers based on the multiple item groups received by the first reception process, The process then executes a second transmission process (step S504) which transmits the first common item group generated by the first generation process to each of the multiple learning devices, Each of the aforementioned learning devices The second reception process (step S504) receives the first common item group, The system performs a first application process (step S505) which applies the first common item group received by the second reception process to the item group, An associative learning system characterized by the following:

[0156] It should be noted that the present invention is not limited to the embodiments described above, but includes various modifications and equivalent configurations within the spirit of the attached claims. For example, the embodiments described above are described in detail to make the present invention easier to understand, and the present invention is not necessarily limited to having all of the described configurations. Furthermore, some of the configurations of one embodiment may be replaced with those of another embodiment. Furthermore, some of the configurations of one embodiment may be added to those of another embodiment. Furthermore, some of the configurations of each embodiment may be added, deleted, or replaced with other configurations.

[0157] Furthermore, each of the aforementioned configurations, functions, processing units, and processing means may be implemented in hardware, for example, by designing them as integrated circuits, or they may be implemented in software by having a processor interpret and execute programs that realize each function.

[0158] Information such as programs, tables, and files that implement each function can be stored in memory, hard disks, SSDs (Solid State Drives), or on recording media such as IC (Integrated Circuit) cards, SD cards, and DVDs (Digital Versatile Discs).

[0159] Furthermore, the control lines and information lines shown are those deemed necessary for explanation purposes and do not necessarily represent all control lines and information lines required for implementation. In reality, it can be assumed that almost all components are interconnected. [Explanation of Symbols]

[0160] 100 Associative Learning Systems 101 Server (Management Device) 201 Processor 202 Storage Devices 205 Communication IF 1000 Common metadata generation prompt 1100 Common Metadata Application Screen 1400 Impact AI Correct Answer Data 103 Generation AI Ci client terminals DCi internal dictionary Li Data to be analyzed LMI Learning Model M Common Metadata Mi Extracted Metadata Si standardized data Pi learning parameters TSi training dataset

Claims

1. A federated learning method that involves federated learning by multiple computers, each having a learning model that learns using training data consisting of the values ​​of each item in a set of items, A first generation process that obtains the item group from each of the multiple computers and generates a first common item group to be used by the multiple computers based on the multiple item groups, Each of the aforementioned computers performs a first application process that applies the first common item group generated by the first generation process to the item group, An associative learning method characterized by including

2. The associative learning method according to claim 1, In the first generation process, if the first item in the first item group among the multiple item groups matches the second item in a second item group that is different from the first item group among the multiple item groups, the first item is added to the first common item group. A characteristic associative learning method.

3. The associative learning method according to claim 1, In the first generation process, if the first item in the first item group among the multiple item groups is included in the second item in a second item group that is different from the first item group among the multiple item groups, the first item is added to the first common item group. A characteristic associative learning method.

4. The associative learning method according to claim 1, In the first generation process, if the first item in the first item group among the multiple item groups is not included in a second item group different from the first item group among the multiple item groups, the first item is added to the first common item group. A characteristic associative learning method.

5. The associative learning method according to claim 1, Each of the aforementioned computers performs an extraction process to extract the training data from the data to be analyzed, An associative learning method characterized by including

6. The associated learning method according to claim 5, The data to be analyzed is document data, The aforementioned item is the first word, The value of the aforementioned item is the second word or phrase corresponding to the aforementioned item in the document data. A characteristic associative learning method.

7. The associated learning method according to claim 6, The document data is either unstructured data, structured data, or semi-structured data whose granularity of structure is between that of the unstructured data and the structured data. A characteristic associative learning method.

8. The associated learning method according to claim 6, In the extraction process, each of the multiple computers outputs a query to the generating AI requesting the addition of metadata to the document data, and as a result, the metadata added to the document data by the generating AI is obtained as the first phrase. A characteristic associative learning method.

9. The associated learning method according to claim 6, In the extraction process, each of the multiple computers outputs a query to the generating AI requesting a summary of the document data and the addition of metadata. As a result, the summaries of the document data are obtained from the generating AI, and the metadata added to the summaries of the document data is obtained from the generating AI as the first phrase. A characteristic associative learning method.

10. The associated learning method according to claim 6, Each of the aforementioned computers has dictionary data that associates higher-level terms with lower-level terms. In the extraction process, for each computer, the first phrase containing the word in the dictionary data is detected based on the similarity between the word in the dictionary data and the first phrase. A characteristic associative learning method.

11. The associative learning method according to claim 10, In the extraction process, for each computer, if a word in the dictionary data is a word of a subordinate concept, the first word containing the word in the dictionary data is converted into a first word containing the corresponding word of a higher concept. A characteristic associative learning method.

12. The associated learning method according to claim 6, In the extraction process, each computer aggregates multiple similar first phrases. A characteristic associative learning method.

13. The associated learning method according to claim 6, In the extraction process, for each of the first terms, the frequency of occurrence of the second terms corresponding to the first term is counted for each computer, and the second terms are narrowed down based on the counting results. A characteristic associative learning method.

14. The associated learning method according to claim 6, In the extraction process, for each computer, the first phrase is narrowed down to a specific first phrase that can be made publicly available. A characteristic associative learning method.

15. The associated learning method according to claim 5, In the first application process, each of the plurality of computers obtains a summary of the data to be analyzed based on the applied item group to which the first common item group has been applied, and obtains the values ​​of each item in the applied item group from the summary of the data to be analyzed. A characteristic associative learning method.

16. The associative learning method according to claim 15, In the first application process, each of the multiple computers outputs a query to the generating AI requesting a summary of the data to be analyzed based on the applied item group. As a result, the summaries of the data to be analyzed are obtained from the generating AI, and the values ​​of each item in the applied item group are obtained from the summaries of the data to be analyzed. A characteristic associative learning method.

17. The associative learning method according to claim 15, In the first application process, each of the multiple computers obtains a summary of the data to be analyzed based on items other than those based on the applied item group, obtains the values ​​of the other items from the summary of the data to be analyzed, and accepts the addition of the other items to the applied item group. A characteristic associative learning method.

18. The associative learning method according to claim 1, Each of the aforementioned computers performs a first update process to update the learning model using updated training data, which is the value of each item in the applied item group. A second generation process generates common learning parameters to be used by the multiple computers based on the multiple learning parameters of the multiple learning models updated in each of the multiple computers by the first update process, Each of the aforementioned computers performs a second update process that updates the learning parameters with the common learning parameters generated by the second generation process, An associative learning method characterized by including

19. The associated learning method according to claim 18, A third generation process generates a second common item group based on the multiple item groups, based on the degree of influence of each item in the item group from each of the multiple computers. Each of the aforementioned computers performs a second application process that applies the second common item group generated by the third generation process to the item group, An associative learning method characterized by including

20. One of several learning devices, each having a learning model that learns using training data which consists of the values ​​of each item in a set of items, It comprises a processor that executes a program, a storage device that stores the program, and a communication interface that communicates with a management device that manages the plurality of learning devices, The aforementioned processor, A first transmission process that transmits the aforementioned group of items to the management device, The management device receives from the management device a first receiving process which includes a first common item group used by the multiple learning devices, which is generated based on the multiple item groups of the multiple learning devices. A first application process that applies the first common item group received by the first reception process to the item group, A learning device characterized by performing the following actions.

21. A management device capable of communicating with multiple learning devices, each having a learning model that learns using training data which consists of the values ​​of each item in a set of items, It comprises a processor for executing a program, a storage device for storing the program, and a communication interface for communicating with the plurality of learning devices, The aforementioned processor, A first receiving process that receives the item group from each of the plurality of learning devices, A first generation process generates a first common item group to be used by the multiple computers based on the multiple item groups received by the first reception process, A first transmission process that transmits the first common item group generated by the first generation process to the plurality of learning devices, A management device characterized by performing the following actions.

22. A control device according to claim 21, The management device is a computer different from the plurality of learning devices. A control device characterized by the following features.

23. A control device according to claim 21, The management device is one of the learning devices among the plurality of learning devices. A control device characterized by the following features.

24. In the processor of one of the learning devices among several learning devices, which have a learning model that is trained with training data consisting of the values ​​of each item in the item set, A first transmission process that transmits the aforementioned group of items to a management device, A first receiving process receives from the management device a first common item group for use in the multiple computers, which is generated based on multiple item groups of the multiple learning devices, A first application process that applies the first common item group received by the first reception process to the item group, A learning program characterized by causing the execution of [a specific action].

25. A processor in a management device that can communicate with multiple learning devices, each having a learning model that learns using training data consisting of the values ​​of each item in a set of items, A first receiving process that receives the item group from each of the plurality of learning devices, A first generation process generates a first common item group to be used by the multiple computers based on the multiple item groups received by the first reception process, A first transmission process that transmits the first common item group generated by the first generation process to the plurality of learning devices, A management program characterized by causing the execution of a command.

26. A federated learning system that performs federated learning using multiple learning devices and a management device that manages the multiple learning devices, Each of the aforementioned learning devices has a learning model that is trained with training data which is the value of each item in the item group. Each of the aforementioned learning devices A first transmission process is performed to transmit the aforementioned group of items to the management device. The aforementioned control device A first receiving process that receives the item group from each of the plurality of learning devices, A first generation process generates a first common item group to be used by the multiple computers based on the multiple item groups received by the first reception process, The process includes a second transmission process that transmits the first common item group generated by the first generation process to each of the multiple learning devices, Each of the aforementioned learning devices A second receiving process that receives the first common items group, The system performs a first application process that applies the first common item group received by the second reception process to the item group, An associative learning system characterized by the following:

Images