Method for generating machine learning data, and device for generating machine learning data

Abstract

[Problem] To enable a sufficient amount of machine learning data to be prepared even when third-party rights are required. [Solution] The present invention adds, in advance to relevant data (21) on a server, permission-required portion information (23) indicating portions requiring permission (permission-required portions (24)) and replacement data for replacing the relevant data in a permission-required portion for which permission is not obtained. Permission information (30) indicating whether there is permission for the permission-required portions is also created in advance. When generating machine learning data, the relevant data and the permission information are loaded, and the relevant data is replaced with the replacement data for the permission-required portion for which permission is not obtained, to generate the machine learning data. Thus, because the machine learning data can be generated even from relevant data for which permission is not obtained, a sufficient amount of machine learning data can easily be prepared.

Description

Machine learning data generation method and machine learning data generation device

[0001] The present invention relates to a technique for generating machine learning data used to generate an estimation model by machine learning.

[0002] Recent technological advances in machine learning have been remarkable. One such remarkable advance is the practical application of technology that creates estimation models with high estimation accuracy by performing preliminary learning using a large amount of data on general content, followed by learning data on individual content corresponding to the field of use, purpose of use, etc. In the following, data on individual content corresponding to the field of use, purpose of use, etc. will be referred to as "individual data," and learning using individual data will be referred to as "individual learning." Furthermore, data on general content will be referred to as "general data," and preliminary learning using general data will be referred to as "preliminary learning."

[0003] Generally, to create an estimation model with high estimation accuracy, it is necessary to train a large amount of data, but it is difficult to prepare a large amount of individual data. Therefore, by performing preliminary training using a large amount of easily available general data and then performing individual training using a smaller amount of individual data than the general data, it becomes relatively easy to create an estimation model with high estimation accuracy.

[0004] Furthermore, even if an estimation model with high estimation accuracy is obtained, the estimation accuracy may gradually decrease due to subsequent changes in environmental conditions. Therefore, various techniques have been proposed to maintain the accuracy of an estimation model by continuously collecting new individual data and performing individual learning even after the estimation model has been obtained (for example, Patent Document 1).

[0005] Furthermore, continuous individual learning may actually result in a decrease in the accuracy of the estimation model. Therefore, a technology has been proposed in which data learned through continuous individual learning (i.e., individual data) is stored, and when the accuracy of the estimation model decreases, the learned individual data can be verified (Patent Document 2).

[0006] However, these individual data are more difficult to obtain than the general data used in preliminary training, and are more likely to contain content that is subject to third-party rights (e.g., copyright, privacy rights, trade secrets, etc.) than general data. If the individual data used for training contains content that is subject to third-party rights, it becomes necessary to obtain the third party's permission after the fact, and if permission cannot be obtained, the estimation model must be discarded. For this reason, permission must be obtained before training for individual data that contains content that is subject to third-party rights.

[0007] JP 2023-086053 A JP 2022-150778 A

[0008] However, for machine learning data that requires third-party permission, such as individual data, permission is not always obtained, which makes it difficult to secure a sufficient amount of machine learning data.If a sufficient amount of machine learning data cannot be secured, it becomes difficult to obtain an estimation model with sufficient accuracy.

[0009] This invention has been made to solve the above-mentioned problems of conventional technology, and aims to provide technology that makes it possible to easily secure a sufficient amount of data for machine learning, even for data that requires third-party rights.

[0010] In order to solve the above-mentioned problems, the method for generating data for machine learning of the present invention has the following configuration: That is, a method for generating data for machine learning by a computer to generate data for machine learning to be used to generate an inference model by machine learning, the method comprising: a target data reading step of reading target data stored in a server, a license information reading step of reading license information indicating whether or not there is license to use the target data for the machine learning, and a generating step of generating the data for machine learning based on the target data and the license information, wherein the target data reading step is a step of reading the target data to which license-required portion information indicating a portion in the target data where content requiring the license is stated or displayed, and replacement data to be substituted for the target data in the license-required portion if the license for the license-required portion has been obtained, and the generating step is a step of generating the data for machine learning by replacing the target data in the license-required portion with the replacement data in accordance with the license information.

[0011] The above-described method for generating machine-learning data of the present invention can also be understood as an apparatus for generating machine-learning data. That is, the machine-learning data generation apparatus of the present invention generates machine-learning data for generating an estimation model by machine learning, and includes: a target data reading unit that reads target data stored in a server; a license information reading unit that reads license information indicating whether or not permission is granted to use the target data for the machine learning; and a machine-learning data generation unit that generates the machine-learning data based on the target data and the license information, wherein the target data reading unit reads the target data to which license-required portion information indicating a license-required portion in the target data, which is a portion of the target data where content requiring permission is stated or displayed, and replacement data to be substituted for the target data in the license-required portion if the permission has not been obtained for the license-required portion is added, and the machine-learning data generation unit generates the machine-learning data by replacing the target data in the license-required portion with the replacement data in accordance with the license information.

[0012] In the machine learning data generation method and machine learning data generation device of the present invention, permission-required portion information and replacement data are added to target data in advance. Here, the permission-required portion information is information indicating a permission-required portion, which is a portion of the target data where content requiring permission (such as wording, text, figures, or graphs) is written or displayed. Furthermore, the replacement data is data that is substituted for the target data in the permission-required portion when permission for the permission-required portion has not been obtained. When generating machine learning data, the target data to which the permission-required portion information and replacement data have been added and the permission information are read. Here, the permission information is information indicating whether permission is available to use the target data for machine learning. Then, based on the permission information, it is determined whether permission has been obtained for the permission-required portion of the permission-required portion information. If permission has not been obtained, the target data in the permission-required portion is replaced with the replacement data, thereby generating machine learning data.

[0013] In this way, even if permission has not been obtained for the licensed portion, replacement data that has been modified to not infringe third-party rights or to allow permission can be prepared, and machine learning data can be generated by replacing the target data of the licensed portion with the replacement data. Furthermore, even if the target data contains licensed portions for which permission has been obtained and licensed portions for which permission has not been obtained, machine learning data can be generated even from target data that requires third-party rights by replacing the target data of the licensed portions for which permission has not been obtained with replacement data. Furthermore, instead of concealing the target data of the licensed portions for which permission has not been obtained by blacking them out, the target data can be replaced with appropriate replacement data prepared in advance. This makes it possible to generate machine learning data that is sufficient for machine learning even from target data that includes licensed portions for which permission has not been obtained. As a result, it becomes possible to easily secure a sufficient amount of machine learning data.

[0014] In addition, in the above-mentioned method for generating data for machine learning and device for generating data for machine learning of the present invention, the information on the portion requiring permission may be added to the target data by describing the information on the portion requiring permission as a separate layer of the target data.

[0015] This makes it easy to identify the parts of the target data that require third-party permission, making it easy to add permission-required part information to the target data. Additionally, because the parts that require third-party permission can be clearly identified without ambiguity, it also makes it easy to create replacement data. Furthermore, because there is no need to modify the target data itself, there are no copyright issues associated with modifications.

[0016] In addition, in the above-mentioned method for generating data for machine learning and device for generating data for machine learning of the present invention, when generating data for machine learning, target data to which replacement data has been added for each portion requiring permission and permission information indicating whether or not permission has been granted for each portion requiring permission may be read.

[0017] In this way, it becomes possible to generate appropriate machine learning data even from target data that has multiple licensed portions.

[0018] Furthermore, in the above-described machine learning data generation method and machine learning data generation device of the present invention, when generating machine learning data, license information including a partial license and target data to which replacement data corresponding to the partial license has been added may be read. Here, a partial license is a license that imposes restrictions on the content of the target data in the license-required portion, the party receiving the license, the use of the estimation model generated by machine learning, etc. Then, for the license-required portion for which partial license has been obtained, machine learning data may be generated by replacing the target data with replacement data for the partial license.

[0019] In this way, even if unlimited permission is not available, it is possible to generate data for machine learning from target data for which partial permission is available.

[0020] Furthermore, in the above-described method for generating machine learning data and device for generating machine learning data of the present invention, when or after generating the machine learning data, replacement information including information that can identify the target data, information on parts requiring permission, and replacement data replaced for each part requiring permission may be generated, and the replacement information 40 may be stored in a distributed ledger in blockchain format.

[0021] This makes it easy to prove at a later date that the machine learning data was generated with the permission of a third party, thus avoiding a situation in which an inference model generated by machine learning would have to be discarded because it is suspected that the model learned machine learning data without the permission of the third party.

[0022] 1 is an explanatory diagram showing an overview of the machine learning data generation device 10 of the present embodiment. FIG. 2 is an explanatory diagram showing an overview of machine learning that performs preliminary learning and individual learning. FIG. 3 is an explanatory diagram showing the data structure of a target dataset 20 that is read into the machine learning data generation device 10 of the present embodiment. FIG. 4 is an explanatory diagram conceptually showing the relationship between target data 21 in the target dataset 20 and permission-requiring portion information 22. FIG. 5 is an explanatory diagram conceptually showing replacement data in the target dataset 20. FIG. 6 is an explanatory diagram of permission information 30 that is read to generate machine learning data. FIG. 7 is a flowchart showing the first half of a process for generating machine learning data. FIG. 8 is a flowchart showing the second half of a process for generating machine learning data. FIG. 9 is a flowchart showing a process for saving replacement information 40 in a blockchain format. FIG. 10 is an explanatory diagram showing the data structure of replacement information 40. FIG. 11 is an explanatory diagram showing how replacement information 40 is saved in a blockchain format. FIG. 12 is an explanatory diagram showing why a sufficient amount of individual data can be easily prepared by using the machine learning data generation device 10 of the present embodiment.

[0023] FIG. 1 is an explanatory diagram showing an overview of a machine-learning data generation device 10 according to this embodiment. As shown in the figure, the machine-learning data generation device 10 according to this embodiment generates machine-learning data by reading target data stored in a server 50 and preset permission information and performing predetermined processing. Here, target data refers to data that is the subject of machine learning, and includes document data, image data, and data that combines these. As will be described later, the machine-learning data generation device 10 according to this embodiment reads the target data not alone, but with permission-required portion information and replacement data, which will be described later, added.

[0024] Furthermore, permission information is information indicating whether permission has been obtained to use the target data for machine learning. In other words, the target data may contain content that is subject to third-party rights (e.g., privacy rights or copyrights), and if the target data is trained without the third-party's permission, in the worst case scenario, the estimated model obtained through machine learning may have to be discarded. Therefore, information indicating whether the target data contains content that requires permission, and if such content is included, whether permission has been obtained, is compiled in advance as permission information. Permission information will be explained in more detail later.

[0025] The machine-learning data generation device 10 also includes a target data reading unit 11, a permission information reading unit 12, a machine-learning data generation unit 13, a replacement information storage unit 14, and the like. Note that these "units" are abstract concepts that represent functions that the machine-learning data generation device 10 has for generating machine-learning data, and do not indicate the existence of components corresponding to these "units." These "units" can be realized as software programs executed by a computer, or as hardware such as an LSI or IC. Furthermore, they may be realized by combining a software program and hardware.

[0026] The target data reading unit 11 reads target data and supplies it to the machine learning data generation unit 13. The license information reading unit 12 reads license information and outputs it to the machine learning data generation unit 13.

[0027] When the machine learning data generation unit 13 receives the target data and the permission information, it checks whether the necessary permission has been obtained by referring to the permission information. If permission has been obtained, the machine learning data is generated using the target data as is. On the other hand, if permission has not been obtained, the machine learning data is generated by replacing the part of the target data that requires permission with preset replacement data. The process of generating machine learning data will be described in detail later.

[0028] The replacement information storage unit 14 generates replacement information by obtaining information indicating the portions of the target data that have been replaced with replacement data and information indicating the replaced replacement data from the machine learning data generation unit 13, and stores the replacement information in a distributed ledger on the blockchain network 60. The blockchain network 60 is configured such that multiple computer nodes (hereinafter referred to as nodes) are communicatively connected to one another, and when data is stored in one of the multiple nodes, the same data is also stored in the other nodes, thereby forming a distributed ledger. The replacement information storage unit 14 generates replacement information in blockchain format and transmits it to one node n1 of the blockchain network 60, thereby storing the replacement information in a distributed ledger on the blockchain network 60. A method for generating replacement information in blockchain format will be described in detail later.

[0029] Here, the machine learning data generated by the above-described machine learning data generation device 10 is machine learning data (individual data) used in a learning stage called "individual learning" in machine learning. That is, in recent machine learning, it has become common to first perform preliminary learning followed by individual learning, and the machine learning data generated by the machine learning data generation device 10 of this embodiment is particularly suitable as machine learning data for individual learning. Therefore, as a preparation for the explanation, an overview of machine learning that performs preliminary learning and individual learning will be provided.

[0030] FIG. 2 is an explanatory diagram showing an overview of machine learning that performs preparatory learning and individual learning. In machine learning, which has become mainstream in recent years, a large amount of data related to general content (general data) is learned in the first learning stage, preparatory learning. Machine learning for generating so-called "large-scale language models" also falls under preparatory learning. For example, to generate a large-scale Japanese language model, a large amount of Japanese document data must be learned, but the content of the document data can be general. In this way, a large amount of data is learned by machine learning in preparatory learning, but because the data to be learned is data with general content, even large amounts of data are relatively easy to obtain.

[0031] Once the preliminary learning is complete, the second learning stage, individual learning, is performed. In individual learning, machine learning is performed using data corresponding to the content of the estimation model to be generated. For example, if an estimation model to be used in the financial field is to be generated, machine learning is performed using literature on economics, finance, etc., including as much recent information as possible. Similarly, if an estimation model to be used in the security field is to be generated, machine learning is performed using literature on computers, communications, security, etc., including as much recent information as possible. Furthermore, if it is necessary to adapt the estimation model to the circumstances of the field where it will be used, machine learning is also performed using materials and data used in the field. In this way, individual learning generates an estimation model by machine learning data (individual data) of individual content corresponding to the field of use, purpose of use, etc. In order to generate an estimation model with high estimation accuracy, a large amount of individual data must be machine learned, but a smaller amount of data than general data is acceptable.

[0032] Furthermore, even after an estimation model is obtained, its accuracy may gradually decrease due to subsequent changes in environmental conditions. Therefore, estimation accuracy can be maintained by continually collecting new individual data and updating the estimation model through individual learning.

[0033] Here, the individual data used in individualized learning and continuous individualized learning is data with individual content tailored to the purpose and use of the estimation model, and therefore it is difficult to prepare a large amount of data like general data. Furthermore, because the individual data is data with individual content, it may contain content that is subject to third-party rights (such as copyright, privacy rights, and trade secrets). In addition, to obtain an estimation model with high estimation accuracy, it is necessary to train the model using individual data with as much of the most recent content as possible, but the most recent individual data is more likely to contain content that is subject to third-party rights. For these reasons, it is difficult to prepare a sufficient amount of individual data. In contrast, by using the machine-learning data generation device 10 of the present embodiment described above, it is possible to prepare a sufficient amount of individual data. The reasons for this are explained below.

[0034] 3 is an explanatory diagram showing the data structure of a target dataset 20 to be read into the machine-learning data generation device 10 of this embodiment. The target dataset 20 has a data structure in which permission-required portion information 22 is added after target data 21 that is the subject of machine learning, and a replacement record set 23 is added after the permission-required portion information 22. Here, the permission-required portion information 22 is information that indicates portions of the target data 21 where content requiring permission is described or displayed (hereinafter, "permission-required portions").

[0035] 4 is an explanatory diagram conceptually showing the relationship between the target data 21 and the permission-requiring portion information 22 of the target data set 20. As shown in the figure, the permission-requiring portion information 22 is added to the target data 21 as a separate layer, and is data that specifies permission-requiring portions 24 on the separate layer. As shown in the figure, by displaying the permission-requiring portions 24 on a separate layer of the target data 21, even if there are multiple descriptions or indications requiring permission in the target data 21, they can be easily specified and recognized. Note that each permission-requiring portion 24 is assigned a serial number called a "license number."

[0036] 5 is an explanatory diagram conceptually illustrating a replacement record set 23 included in the target dataset 20. As shown in the figure, the replacement record set 23 is a collection of multiple replacement data records 25, and each replacement data record 25 is formed by arranging a license number, an assumed license status, and replacement data in that order. Here, the assumed license status indicates the assumed license status for which the replacement data listed later was created. For example, the replacement data record 25 second from the top in FIG. 5 is a replacement data record 25 for the license-requiring portion 24 with license number 2, and indicates that this replacement data was created assuming a license status of "none" (i.e., a situation in which permission cannot be obtained).

[0037] 5, the replacement data record 25 displayed at the top has a blank column for replacement data. This is because the permission status for the license-required portion 24 with license number 1 is "yes" (i.e., permission has been obtained), so there is no need to create replacement data.

[0038] 5, the fourth and fifth replacement data records 25 from the top both have a license requirement number of 4. That is, two replacement data are set for the license requirement section 24 with license requirement number 4. The reason for this is that two situations are assumed for the content described in the license requirement section 24 with license requirement number 4: when no license has been obtained and when partial license has been obtained, and replacement data is required for both the case where no license has been obtained (assumed license status is [none]) and the case where partial license has been obtained (assumed license status is [partial]).

[0039] In the above example, the two options are "a situation in which permission is not obtained" and "a situation in which partial permission is obtained." In contrast, if the two options are "a situation in which permission is obtained" and "a situation in which partial permission is obtained," it is sufficient to set a replacement data record 25 with an assumed permission status of "Yes" and a replacement data record 25 with an assumed permission status of "Partial." Furthermore, as will be described in detail later, if a replacement data record 25 with the same permission-requiring number exists, it is not necessary to set a replacement data record 25 with an assumed permission status of "Yes." Therefore, if the two options are "a situation in which permission is obtained" and "a situation in which partial permission is obtained," it is sufficient to set only a replacement data record 25 with an assumed permission status of "Partial."

[0040] 5, the replacement data is recorded as is in the second, fourth, and fifth replacement data records 25. However, as in the third replacement data record 25 from the top, it is also possible to set the URI (Uniform Resource Identifier) ​​where the replacement data is saved.

[0041] 6 is an explanatory diagram of the license information 30 that the machine-learning data generation device 10 of this embodiment reads to generate machine-learning data. As shown in the figure, the license information 30 is a collection of multiple license status records 31. Each license status record 31 is formed by arranging a license number and a license status in this order. Here, the license status indicates whether or not permission has been obtained for the license-requiring portion 24 corresponding to the license number.

[0042] The machine learning data generation device 10 of this embodiment reads the target dataset 20 described above using Figures 3 to 5 and the license information 30 described above using Figure 6, and generates machine learning data by executing the machine learning data generation process described below.

[0043] 7 and 8 are flowcharts of the machine learning data generation process. In the machine learning data generation process, a target dataset 20 created in advance and stored in the server 50 and license information 30 created in advance are read (STEP 10). Then, after initializing the license number to 1 (STEP 11), the license status corresponding to the current license number is obtained by referencing the license information 30 (STEP 12). For example, immediately after the license number is initialized in STEP 11, the license number is 1, and therefore the license status of "Yes" set in the license status record 31 for the license number 1 in the license information 30 shown in FIG. 6 is obtained.

[0044] Next, it is determined whether the license status acquired in STEP 12 is "licensed" (STEP 13). If the result is "licensed" (STEP 13: yes), there is no need to replace the content of the target data 21 with replacement data for the license-required portion 24 of the current license-required number. In this case, it is determined whether the license-required number has reached the last license-required number (STEP 18 in FIG. 8). If the last license-required number has not been reached (STEP 18: no), the license-required number is incremented by one (STEP 19), and the process returns to STEP 12 in FIG. 7 to acquire the license status for which the license-required number is set in the license status record 31 for the new license-required number from the license information 30. Then, it is determined whether the acquired permission status is "yes" or not (STEP 13), and if it is "yes" (STEP 13), it is determined again whether the last permission required number has been reached, and if it has not yet reached the last permission required number, the permission required number is incremented by one (STEP 18 and STEP 19 in Figure 8), and then the process returns to STEP 12 in Figure 7 to repeat the same operation.

[0045] On the other hand, if the license status acquired in STEP 12 was not "licensed" (STEP 13: no), the replacement record set 23 of the target dataset 20 is referenced to acquire the assumed license status corresponding to the license-required number (STEP 14). That is, as described above with reference to Figures 3 and 5, the replacement data record 25 in the replacement record set 23 has the assumed license status and replacement data set in association with the license-required number. Therefore, in STEP 14, the assumed license status set in association with the current license-required number is acquired. At this time, if multiple assumed license statuses are set for the same license-required number, all assumed license statuses are acquired (see Figure 5).

[0046] Then, it is determined whether or not there is an assumed permission situation among the assumed permission situations acquired in STEP 14 that matches the permission situation acquired in STEP 12. For example, for license number 4 in the license information 30 shown in FIG. 6, the permission situation "Partial," which indicates that partial permission has been obtained, is set. On the other hand, for license number 4 in the target data set 20 shown in FIG. 5, the assumed permission situation "None" and the assumed permission situation "Partial" are set. Therefore, the assumed permission situation "Partial" that matches the permission situation "Partial" is set. If it is determined that there is an assumed permission situation that matches the assumed situation (STEP 15: yes), the content of the target data 21 described in the permission-requiring portion 24 is changed to replacement data that corresponds to the assumed permission situation that matches the assumed situation (STEP 16).

[0047] As described above, once the content of the target data 21 has been replaced with replacement data for the license-required portion 24 of the license-required number, it is determined whether or not the last license-required number has been reached (STEP 18 in Figure 8). If the last license-required number has not been reached (STEP 18: no), the license-required number is incremented by one (STEP 19), and the process returns to STEP 12 in Figure 7 to start the series of operations described above.

[0048] On the other hand, if the determination in STEP 15 is "no," i.e., if there is no assumed permission status among the assumed permission statuses acquired in STEP 14 that matches the permission status acquired in STEP 12, a predetermined alarm is output (STEP 17), and the machine learning data generation process is terminated. For example, if the permission status acquired in STEP 12 is "None," indicating that permission has not been obtained, but there is no replacement data record 25 with an assumed permission status of "None" in the replacement record set 23 referenced in STEP 14, the content of the permission-requiring portion 24 in the target data 21 cannot be replaced. Therefore, an alarm indicating this is output, and the process is terminated without generating machine learning data.

[0049] Furthermore, if it is determined in STEP 18 of Figure 8 that the license-requiring number has reached the final number (STEP 18: yes), machine learning data is obtained by replacing the description in the target data 21 for which no license has been obtained with replacement data. Therefore, the data name of the obtained machine learning data is obtained (STEP 20). The data name of the machine learning data may be set when the target dataset 20 and license information 30 are read into the machine learning data generation device 10 in STEP 10 of Figure 7, or input of the data name may be requested in STEP 20, and the input data name may be read. The machine learning data is then output under the obtained data name (STEP 21).

[0050] After the machine learning data has been output in this manner, the following replacement information storage process (STEP 30) is initiated to store the replacement information in a distributed ledger on the blockchain network 60. FIG. 9 is a flowchart of the replacement information storage process (STEP 30). The replacement information 40 is information indicating the content replaced in the target data 21 to generate the machine learning data, and as shown in FIG. 10 , is information in which the URI (Uniform Resource Identifier) ​​of the used target data 21, the licensed portion information 22 in the target dataset 20, the replacement record set 23, and the license information 30 are added to the data name of the machine learning data.

[0051] As shown in FIG. 9 , the replacement information saving process generates replacement information 40 by adding the URI of the target data 21 used to generate the machine learning data, the licensed portion information 22 and replacement record set 23 in the target dataset 20, and the permission information 30 to the data name of the machine learning data (STEP 31). The reason why the URI of the target data 21 is used as the data to be added to the data name of the machine learning data, rather than the data itself, is because the target data 21 is large in size. Furthermore, since the target data 21 is not created specifically for generating machine learning data but is literature stored in another institution (often a public institution), knowing the URI is sufficient. In contrast, the licensed portion information 22 and replacement record set 23 in the target dataset 20, and the permission information 30 were created for generating machine learning data, and therefore the data itself must be used. When generating replacement information 40 in STEP 31, information indicating the date and time when the machine learning data was output in STEP 21 of Fig. 8 may be acquired, and replacement information 40 including the date and time information may be generated. Furthermore, if the data size of target data 21 is small, replacement information 40 may be generated using target data 21 itself (instead of a URI).

[0052] Furthermore, among third parties who hold rights associated with the target data 21, there may be rights holders who are willing to perform machine learning on the target data 21 but who do not want the contents of the target data 21 to be made public, at least for the time being. In this case, analyzing the estimated model obtained by machine learning does not pose a problem because the contents of the target data 21 cannot be determined. However, if the URI of the target data 21 (or the target data 21 itself) is included in the replacement information 40, the contents of the target data 21 may be potentially made public, and the rights holder may be reluctant to do so. In such cases, the URI of the target data 21 (or the target data 21 itself) may be encrypted or concealed using secure computation. Furthermore, the URI of the target data 21 may be displayed normally, but the target data 21 itself identified by the URI may be encrypted or concealed using secure computation. In this way, the contents of the target data 21 will not be made public, making it possible to obtain permission for machine learning on the target data 21 even from rights holders who are reluctant to have the contents of the target data 21 made public.

[0053] Next, block data 41 is created by adding the replacement information 40 obtained in STEP 31 to the stored hash value (STEP 32). Here, the stored hash value is a value obtained by applying a predetermined hash function to previously created block data 41. The hash function converts data of any size into a hash value of the same data length, and if even a portion of the pre-conversion data is different, the converted hash value will be an entirely different value. Furthermore, it is impossible to revert to the pre-conversion data from the hash value. Therefore, since a one-to-one relationship always exists between the hash value and the original data, the hash value can be considered to uniquely represent the pre-conversion data.

[0054] Once the block data 41 is obtained in this way, a new hash value is calculated by applying a hash function to the block data 41 (STEP 33), and the stored hash value is changed to the new hash value (STEP 34). This hash value will be used when creating the next block data 41.

[0055] The obtained block data 41 is then transmitted to node n1 of the blockchain network 60 (STEP 35), whereupon the replacement information storage process of FIG. 9 is terminated, and the process returns to the machine-learning data generation process of FIGS. 7 and 8 , after which the machine-learning data generation process is terminated. Upon receiving the block data 41, node n1 also transmits the block data 41 to other nodes n2 to n4 of the blockchain network 60, resulting in the block data 41 being stored in the distributed ledger of the blockchain network 60. Note that in the above description, the replacement information storage process generates block data 41 by adding a hash value to the replacement information, and then transmits the block data 41 to node n1 of the blockchain network 60. However, once the replacement information is generated in the replacement information storage process, the replacement information may be transmitted to node n1 of the blockchain network 60 without adding a hash value. Then, node n1 may generate block data 41 by adding a hash value and transmit the block data 41 to the other nodes n2 to n4.

[0056] 11 is an explanatory diagram showing how block data 41 is stored in a distributed ledger of a blockchain network 60. Because each block data 41 contains a hash value indicating the previous block data 41, multiple block data 41 are stored in a format in which they are linked in a row (blockchain format). Storing block data in this manner in blockchain format makes it extremely difficult to tamper with the replacement information in one block data 41, since tampering with the replacement information in that block data 41 requires changing the hash values ​​of all block data 41 connected after that block data 41.

[0057] The above has described in detail a method for generating machine learning data using the machine learning data generation device 10 of this embodiment. Generating machine learning data using such a method provides several advantages, as described below.

[0058] First, the machine learning data (individual data) used in the learning stage of machine learning, known as individual learning, is typically difficult to obtain because there is less of it than data with general content. In addition, because it is data about individual content, third-party permission may be required for use in machine learning, and if permission cannot be obtained, the data cannot be used for machine learning. Furthermore, if permission from multiple people is required, the data cannot be used for machine learning unless permission from all parties is obtained. For this reason, in addition to the small amount of data to begin with, the need for third-party permission makes it even more difficult to obtain.

[0059] In contrast, the method of the present embodiment described above identifies portions (license-requiring portions 24) that require third-party permission within the data to be subjected to machine learning (target data 21), and permission can be obtained for each permission-requiring portion 24. The permission-requiring portions 24 for which permission cannot be obtained can be replaced with content (replacement data) that does not require third-party permission. Therefore, even if there are permission-requiring portions 24 for which permission cannot be obtained, the target data 21 can be used for machine learning. Furthermore, if permission can be obtained with minor modifications to the content, the target data 21 can be used for machine learning by replacing the data with replacement data that has undergone the necessary modifications.

[0060] Furthermore, because replacement data can be prepared for each permission-required section 24, even if the target data 21 for a permission-required section 24 for which permission has not been obtained is replaced with replacement data, the meaning of the surrounding text, including that section, can be prevented from changing significantly. If the target data 21 for a permission-required section 24 for which permission has not been obtained were concealed by blacking out the data or replacing it with meaningless symbols, the meaning of the surrounding text, including that section, would become unclear. Such learning data is noise that adversely affects machine learning and cannot be used as machine learning data. In contrast, with the method of the present embodiment described above, for permission-required sections 24 for which permission has not been obtained, the target data 21 for that section can be replaced with replacement data prepared in advance, making it possible to generate machine learning data that is fully usable for machine learning.

[0061] Furthermore, because the permission-requiring passage 24 can be specified on a separate layer of the target data 21 (see FIG. 4), the permission-requiring passage 24 in the target data 21 can be easily and unambiguously identified. In addition, because the permission-requiring passage 24 in the target data 21 can be precisely identified, replacement data for replacing that portion can also be easily prepared. In addition, because there is no need to modify the target data 21 itself, permission-requiring passage information 22 can be easily added to the target data 21.

[0062] Furthermore, even if there are multiple third parties who require permission, it is not necessarily the case that all of them have the rights to all of the permission-requiring sections 24. Therefore, even if it is difficult to obtain permission from some of the third parties, it is possible to use the target data 21 for machine learning as is for the permission-requiring sections 24 for which permission has been obtained.

[0063] Furthermore, even if permission from a third party cannot be obtained, permission may be obtained or may be obtained with a slight modification depending on the party from whom permission is sought or the purpose and use of the machine learning. The method of this embodiment described above allows the permission status for each permission-requiring portion 24 to be set in the permission information 30, making it possible to flexibly respond to such cases. This point will be described with reference to FIG. 12.

[0064] FIG. 12 is an explanatory diagram showing how multiple types of machine learning data are generated from target data 21. The illustrated example shows a case in which machine learning data to be used at University B, at a related company C, and at a general company D are generated from one target data 21. Even with the same target data 21, when used for research at University B, permission is likely to be obtained for many license-requiring sections 24, as long as the content does not include trade secrets or the like. Furthermore, when used for product development at related company C, permission is likely to be obtained even for license-requiring sections 24 that include trade secrets or the like. Furthermore, when the target data 21 is used at general company D, permission is likely to be obtained for fewer license-requiring sections 24. Therefore, permission information 30 for the university (license information b), permission information 30 for the related company (license information c), and permission information 30 for the general company (license information d) are created in advance. For University B, license information 30 for the university (license information b) is applied to the target dataset 20 of the target data 21 to generate machine learning data for the university (machine learning data Ab). For Company C, license information 30 for the related company (license information c) is applied to generate machine learning data for the related company (machine learning data Ac). For Company D, license information 30 for general companies (license information d) is applied to generate machine learning data for general companies (machine learning data Ad). In this way, multiple types of machine learning data can be generated from a single piece of target data 21.

[0065] For the reasons described above, by using the method of this embodiment, it is possible to easily prepare a sufficient amount of individual data, even for machine learning data (individual data) used in individual learning, and as a result, it is possible to obtain an estimation model with high estimation accuracy through machine learning.

[0066] Furthermore, each time machine learning data is generated, replacement information 40 for that machine learning data is stored in blockchain format in the distributed ledger of the blockchain network 60. As is well known, data stored in blockchain format in the distributed ledger of the blockchain network 60 is virtually impossible to tamper with, making it easy to prove at a later date that the machine learning data was generated with the permission of a third party. Furthermore, if the replacement information 40 includes information on the date and time the machine learning data was output, the date and time the machine learning data was generated can also be identified, making proof even easier. This makes it possible to avoid situations where an estimation model generated by machine learning must be discarded.

[0067] The above describes the machine learning data generation device 10 of this embodiment, but the present invention is not limited to the above embodiment and can be implemented in various forms within the scope of its gist.

[0068] 10...machine learning data generation device, 11...target data reading unit, 12...license information reading unit, 13...machine learning data generation unit, 14...replacement information storage unit, 20...target dataset, 21...target data, 22...license-requiring portion information, 23...replacement record set, 24...license-requiring portion, 25...replacement data record, 30...license information, 31...license status record, 40...replacement information, 41...block data, 50...server, 60...blockchain network, n1 to n4...nodes.

Claims

1. A method for generating data for machine learning, in which data for machine learning used to generate an inference model by machine learning is generated by a computer, comprising: a target data reading step of reading target data stored in a server; a permission information reading step of reading permission information indicating the presence or absence of permission to use the target data for the machine learning; and a generation step of generating the data for machine learning based on the target data and the permission information, wherein the target data reading step is a step of reading the target data to which permission-required portion information indicating a portion in the target data where content requiring permission is described or displayed, and replacement data to be substituted for the target data in the permission-required portion if the permission has not been obtained for the permission-required portion is added, and the generation step is a step of generating the data for machine learning by replacing the target data in the permission-required portion with the replacement data in accordance with the permission information.

2. A method for generating data for machine learning as described in claim 1, characterized in that the target data reading process is a process for reading the target data in which the information on the parts requiring permission is written on a separate layer from the target data.

3. A method for generating data for machine learning as described in claim 2, characterized in that the target data reading process is a process for reading the target data to which the replacement data has been added for each of the permission-requiring parts, and the permission information reading process is a process for reading the permission information indicating the presence or absence of permission for each of the permission-requiring parts.

4. A method for generating data for machine learning as described in any one of claims 1 to 3, characterized in that the permission information reading process is a process for reading the permission information including partial permission which is a partial permission, and the target data reading process is a process for reading the target data to which the replacement data for when partial permission has been obtained has been added, for the permission-requiring portion for which partial permission has been obtained.

5. A method for generating data for machine learning as claimed in any one of claims 1 to 3, comprising a storage step of generating replacement information including information capable of identifying the target data, information on the portion requiring permission, and the replacement data replaced for each portion requiring permission, when generating or after the data for machine learning is generated, and storing the replacement information in a distributed ledger in blockchain format.

6. A machine learning data generation device that generates machine learning data for generating an estimation model by machine learning, comprising: a target data reading unit that reads target data stored in a server; a permission information reading unit that reads permission information indicating the presence or absence of permission to use the target data for the machine learning; and a machine learning data generation unit that generates the machine learning data based on the target data and the permission information, wherein the target data reading unit reads the target data to which permission-required portion information indicating a permission-required portion, which is a portion in the target data where content requiring permission is described or displayed, and replacement data to be substituted for the target data in the permission-required portion when the permission has not been obtained for the permission-required portion are added, and the machine learning data generation unit generates the machine learning data by replacing the target data in the permission-required portion with the replacement data in accordance with the permission information.

Images