Training data deletion device, training data deletion method, and training data deletion program

Abstract

To provide a training data deletion device, a training data deletion method, and a training data deletion program for deleting training data overlapping with each other and deleting training data similar to each other.SOLUTION: The training data generation unit 40 of the training data deletion device 10 that uses the training data deletion method and executes the training data deletion program acquires a plurality of training data candidates, and compares a value related to a change in the data size of a training data set in a case where the training data candidates are added to the training data set T with a threshold value to determine whether to add or delete the training data candidates to or from the training data set T.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a training data deletion device, a training data deletion method, and a training data deletion program. [Background technology]

[0002] As described in Non-Patent Document 1, a method is known in the past in which overlapping training data is deleted using a full-text search that searches for an arbitrary character string from a plurality of document data. [Prior art documents] [Patent documents]

[0003] [Non-Patent Document 1] Lee et al., Deduplicating Training Data Makes Language Models Better, in Proc. of ACL2022 Summary of the Invention [Problem to be solved by the invention]

[0004] In the method described in Non-Patent Document 1, when character strings are completely identical to each other, the overlapping training data is deleted. Therefore, training data in which the description of the document data is similar to the character string or in which part of the document data is different from the character string is not deleted.

[0005] An object of the present disclosure is to provide a training data deletion device, a training data deletion method, and a training data deletion program that deletes overlapping training data and similar training data. [Means for solving the problem]

[0006] The invention described in claim 1 is a training data deletion device comprising an acquisition unit that acquires a plurality of training data candidates (c, ci) and a deletion unit that either adds or deletes the training data candidates from a training data set (T), which is a set of training data, and the deletion unit either adds or deletes the training data candidates from the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv).

[0007] The invention described in claim 7 is a training data deletion method that includes obtaining a plurality of training data candidates (c, ci) and either adding or deleting the training data candidates from a training data set (T), which is a set of training data, and that either adds or deletes the training data candidates from the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv).

[0008] Furthermore, the invention described in claim 8 is a training data deletion program that causes a training data deletion device to function as an acquisition unit that acquires a plurality of training data candidates (c, ci) and a deletion unit that either adds or deletes the training data candidates from a training data set (T), which is a set of training data, and the deletion unit either adds or deletes the training data candidates from the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv).

[0009] The change in the data size of the training data set when a training data candidate is added to the training data set corresponds to whether the training data candidate and the training data overlap or are similar to each other. Therefore, a training data deletion device compares a value relating to the change in the data size of the training data set when a training data candidate is added to the training data set with a threshold, and deletes training data that overlaps with each other and deletes training data that is similar to each other.

[0010] The reference symbols in parentheses attached to each component indicate an example of the correspondence between the component and the specific components described in the embodiments described below. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a configuration diagram of a training data deletion device according to a first embodiment. [Figure 2] 10 is a flowchart showing the processing of a condition acquisition unit of the training data deletion device. [Figure 3] 10 is a flowchart showing the processing of a training data generation unit of the training data deletion device. [Figure 4] 10 is a flowchart showing the processing of a training data generation unit in the training data deletion device of the second embodiment. [Figure 5] 10 is a flowchart showing the processing of a training data generation unit. [Figure 6] 11 is a flowchart showing the processing of a training data generation unit in the training data deletion device of the third embodiment. [Figure 7] 10 is a flowchart showing the processing of a training data generation unit in the training data deletion device of the fourth embodiment. [Figure 8] 13 is a flowchart showing the processing of a training data generation unit in the training data deletion device of the fifth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, embodiments will be described with reference to the drawings. In the following embodiments, identical or equivalent parts will be denoted by the same reference numerals, and description thereof will be omitted.

[0013] (First embodiment) A training data deletion device that uses the training data deletion method of this embodiment and executes a training data deletion program deletes overlapping training data and similar training data.

[0014] Specifically, the training data deletion device is mainly composed of a microcomputer, etc., and is equipped with a CPU, ROM, flash memory, RAM, I / O, and bus lines connecting these components, etc. As shown in Fig. 1, the training data deletion device 10 has a condition acquisition unit 20, a training data candidate storage unit 30, a training data generation unit 40, and a training data storage unit 50 as functional blocks.

[0015] The condition acquiring unit 20 executes a program stored in the ROM of the training data deletion device 10. As a result, the condition acquiring unit 20 acquires a threshold θ and a required number K of training data set by a user of the training data deletion device 10 or an external system such as the learning system 60. Furthermore, the condition acquiring unit 20 outputs the acquired threshold θ and the required number K of training data set to the training data generating unit 40, which will be described later. Details of the processing by the condition acquiring unit 20 will be described later.

[0016] The training data candidate storage unit 30 stores a set of document data that are candidates for training data. Here, the document data that are candidates for training data are referred to as training data candidate c. The set of training data candidate c is referred to as training data candidate set C.

[0017] The training data generation unit 40 executes a program stored in the ROM of the training data deletion device 10. As a result, the training data generation unit 40 acquires a threshold θ and a required number K from the condition acquisition unit 20. The training data generation unit 40 also reads the training data candidate storage unit 30. As a result, the training data generation unit 40 acquires a training data candidate set C from the training data candidate storage unit 30. Furthermore, the training data generation unit 40 uses the acquired threshold θ, required number K, and training data candidate set C to either add or delete a training data candidate c from the training data set T. In this way, the training data generation unit 40 generates the training data set T. The training data generation unit 40 outputs the generated training data set T to the training data storage unit 50 or the learning system 60, which will be described later. The training data set T is a set of training data. Details of the generation of the training data set T by the training data generation unit 40 will be described later.

[0018] The training data storage unit 50 stores the training data set T generated by the training data generation unit 40. The learning system 60 acquires the training data set T by reading the training data storage unit 50. Alternatively, the learning system 60 acquires the training data set T from the training data generation unit 40. Furthermore, the learning system 60 creates a machine learning model using the acquired training data set T.

[0019] The training data deletion device 10 of the first embodiment is configured as described above. Next, the processing of the condition acquisition unit 20 as a result of the program execution of the condition acquisition unit 20 will be described with reference to the flowchart in Fig. 2. The program of the condition acquisition unit 20 is executed, for example, when started by a user of the training data deletion device 10.

[0020] In step S100, the condition acquiring unit 20 acquires condition information. Specifically, the condition acquiring unit 20 acquires a threshold value θ and a required number K of training data. As described above, the threshold value θ and the required number K are set by a user of the training data deletion device 10 or an external system such as the learning system 60.

[0021] Subsequently, in step S102, the condition acquiring unit 20 outputs the condition information acquired in step S100, here the threshold value θ and the required number K, to the training data generating unit 40. Thereafter, the processing of the condition acquiring unit 20 ends.

[0022] As described above, the condition obtaining unit 20 performs processing. Next, the generation of the training data set T by the execution of the program of the training data generating unit 40 will be described with reference to the flowchart of Fig. 3. The program of the training data generating unit 40 is executed, for example, when started by the user of the training data deletion device 10.

[0023] In step S200, the training data generation unit 40 acquires various information. Specifically, the training data generation unit 40 acquires the threshold θ and the required number K from the condition acquisition unit 20. Furthermore, the training data generation unit 40 reads the training data candidate storage unit 30. As a result, the training data generation unit 40 acquires the training data candidate set C from the training data candidate storage unit 30.

[0024] In step S202 following step S200, the training data generation unit 40 performs an initialization process. Specifically, the training data generation unit 40 sets the training data set T to an empty set.

[0025] Here, the number of training data candidates c included in the training data candidate set C acquired by the training data generation unit 40 in step S200 is defined as the number of candidates Nc.

[0026] Then, in loop L1 following step S202, the training data generation unit 40 sets a variable i and repeats the processes from step S204 to step S220. Note that the initial value of the variable i is set to 1, for example.

[0027] Specifically, in step S204 of loop L1, the training data generation unit 40 selects the i-th training data candidate c from the training data candidates c included in the training data candidate set C acquired in step S200. As a result, the training data generation unit 40 selects ci, which is the i-th training data candidate c.

[0028] In step S206 following step S204, the training data generation unit 40 determines whether or not the variable i is 1. As a result, the training data generation unit 40 determines whether or not it is immediately after the initialization process of step S202.

[0029] If the variable i is 1, this means that the training data generation unit 40 has just completed the initialization process in step S202. At this time, the denominator of relational expression (1), which will be described later, is zero, and therefore si, which will be described later, cannot be calculated. Therefore, the processing of the training data generation unit 40 proceeds to step S216. In step S216 following step S206, the training data generation unit 40 adds c i when the variable i is 1, i.e., c1, to the training data set T as training data. If the variable i is not 1, this means that the training data generation unit 40 has not just completed the initialization process in step S202. Therefore, the processing of the training data generation unit 40 proceeds to step S208.

[0030] In step S208 following step S206, the training data generation unit 40 determines whether D(ci) is zero. Note that D(ci) is the data size of ci. The data size may also be called the file size, data capacity, file capacity, etc.

[0031] When D(ci) is zero, the denominator of relational expression (1) described below is zero, and therefore si described below cannot be calculated. Furthermore, when D(ci) is zero, ci does not contain character information. Therefore, the processing of the training data generation unit 40 proceeds to step S220. In step S220 following step S208, the training data generation unit 40 deletes ci when D(ci) is zero, adds 1 to the variable i, and performs processing on the next training data candidate c. Furthermore, when D(ci) is not zero, the processing of the training data generation unit 40 proceeds to step S210.

[0032] Here, the difference in the change in data size when overlapping or similar training data is compressed is significantly different from the difference in the change in data size when non-overlapping or non-similar training data is compressed.

[0033] For this reason, in step S210 following step S208, the training data generation unit 40 performs information compression on ci selected in step S204 and the training data set T being generated. This makes it easier for the data size of the training data set T to change when ci is added to the training data set T.

[0034] In step S212 following step S210, the training data generation unit 40 calculates si for ci. si for ci is a value related to the change in data size of the training data set T when ci is added to the training data set T.

[0035] Furthermore, let Dc(ci) be the data size of ci after information compression. Let Dc(T) be the data size of training data set T after information compression. Let Tc be the union of ci and training data set T. Let Dc(Tc) be the data size of Tc after information compression.

[0036] Then, the training data generation unit 40 substitutes Dc(ci), Dc(T), and Dc(Tc) into the following relational expression (1). As a result, the training data generation unit 40 calculates si for ci. Note that max{Dc(T), Dc(ci)} in the following relational expression (1) is the data size of the largest one of Dc(T) and Dc(ci). min{Dc(T), Dc(ci)} is the data size of the smallest one of Dc(T) and Dc(ci). Furthermore, Score(T,ci) is equal to or greater than 0 and equal to or less than 1.

[0037]

number

[0038] In step S214 following step S212, the training data generation unit 40 compares si calculated in step S212 with the threshold θ acquired in step S200. As a result, the training data generation unit 40 either adds or deletes ci from the training data set T.

[0039] Here, when the training data included in the training data set T and ci overlap or are similar to each other, the change in data size when ci is added to the training data set T is small. Therefore, si becomes relatively small at this time. Furthermore, when the training data included in the training data set T and ci do not overlap or are not similar to each other, the change in data size when ci is added to the training data set T is large. Therefore, si becomes relatively large at this time.

[0040] Therefore, the training data generator 40 determines whether si calculated in step S212 is greater than the threshold value θ acquired in step S200.

[0041] When si is equal to or smaller than the threshold θ, si is relatively small, and therefore the training data included in the training data set T and ci overlap or are similar to each other. Therefore, at this time, the training data generation unit 40 deletes ci selected in step S204 and does not use it as training data. Thereafter, the processing of the training data generation unit 40 proceeds to step S220. On the other hand, when si is greater than the threshold θ, si is relatively large, and therefore the training data included in the training data set T and ci do not overlap or are similar to each other. At this time, the processing of the training data generation unit 40 proceeds to step S216.

[0042] In step S216, the training data included in the training data set T and the ci selected in step S204 are not overlapping or similar to each other. Therefore, the training data generation unit 40 uses the ci selected in step S204 as training data and adds it to the training data set T.

[0043] In step S218 following step S216, the training data generation unit 40 counts Nt. Furthermore, the training data generation unit 40 determines whether the counted Nt is less than the required number K obtained in step S200. Note that Nt is the number of training data candidates c added by the training data generation unit 40 to the training data set T. Furthermore, since the initial training data set T is an empty set here, Nt corresponds to the number of training data in the training data set T.

[0044] If Nt is less than the required number K, the number of training data is insufficient, and the processing of the training data generation unit 40 proceeds to step S220. Furthermore, if Nt is equal to or greater than the required number K, the number of training data is sufficient, and the processing of the training data generation unit 40 exits loop L1 and proceeds to step S222.

[0045] In step S220, the training data generation unit 40 adds 1 to the current variable i and sets the value as a new variable i.

[0046] In this way, the processes from step S204 to step S220 are repeated while changing the variable i by 1 until the variable i reaches the number of candidates Nc. This generates a training data set T. Then, when loop L1 ends, the process of the training data generation unit 40 proceeds to step S222.

[0047] In step S222, the training data generation unit 40 outputs the generated training data set T to the training data storage unit 50 or the learning system 60. The training data storage unit 50 stores the training data set T generated by the training data generation unit 40. The learning system 60 acquires the training data set T by reading the training data storage unit 50. Alternatively, the learning system 60 acquires the training data set T from the training data generation unit 40. The learning system 60 also creates a machine learning model using the acquired training data set T.

[0048] As described above, the training data generation unit 40 generates the training data set T. Next, it will be explained how the training data deletion device 10 of this embodiment deletes overlapping training data and similar training data.

[0049] Here, as described in Non-Patent Document 1, a method is known in which duplicate training data is deleted using a full-text search that searches for an arbitrary character string from multiple document data. In the method described in Non-Patent Document 1, when character strings exactly match each other, duplicate training data is deleted. Therefore, training data in which a description in document data is similar to the character string or in which part of the document data is different from the character string is not deleted.

[0050] In contrast, the training data generation unit 40 of the training data deletion device 10 of this embodiment serves as an acquisition unit that acquires multiple training data candidates c. Furthermore, the training data generation unit 40 serves as a deletion unit that either adds or deletes the training data candidate c to the training data set T. The training data generation unit 40 also compares si for ci with a threshold θ to either add or delete the training data candidate c to the training data set T. Note that si for ci corresponds to a value related to the change in the data size of the training data set T when the training data candidate c is added to the training data set T.

[0051] The change in the data size of the training data set T when a training data candidate c is added to the training data set T corresponds to whether the training data candidate c and the training data overlap or are similar to each other. Therefore, the training data deletion device 10, which compares si for ci with a threshold θ, deletes both overlapping training data and similar training data.

[0052] Furthermore, the training data deletion device 10 of the first embodiment also provides the following effects.

[0053] [1-1] As described in Non-Patent Document 2 below, a method is known in which a hash value of each paragraph is calculated using a hash function, and the calculated hash values ​​are used to delete duplicate training data. However, the method described in Non-Patent Document 2 requires calculation of hash values, which increases the amount and time of calculation required to delete duplicate training data. Non-Patent Document 2 is Wenzek et al., CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data, in Proc. of LREC2020. A hash function is a function that converts arbitrary data into a unique value of a fixed length. A hash value is a value converted by the hash function.

[0054] Furthermore, as described in Non-Patent Document 3 below, a method for calculating the similarity of training data using data compression is known. It is conceivable to use this similarity of training data to delete duplicated training data. However, in the method using similarity described in Non-Patent Document 3, data are compared with each other. Therefore, for example, if the number of training data candidates c is a and the number of training data is b, the data must be compared a × b times at most. Therefore, the amount of calculation and calculation time required to delete duplicated training data increases. Non-Patent Document 3 is Li, Ming et al., The similarity metric, IEEE transactions on Information Theory 50.12 (2004): 3250-3264.

[0055] In contrast to these, the training data deletion device 10 of this embodiment does not use a hash function or the like. Furthermore, the comparison of si for ci with the threshold θ by the training data deletion device 10 corresponds to the comparison of ci with the training data set T. Therefore, if the number of training data candidates c is a, ci is compared with the training data set T a maximum of a times. This reduces the number of comparisons compared to the method using similarity in Non-Patent Document 3. Therefore, an increase in the amount of calculation and calculation time required to delete overlapping or similar training data is suppressed.

[0056] [1-2] Furthermore, in the training data deletion device 10 of this embodiment, overlapping training data and similar training data are deleted, so the number of training data is relatively reduced and the training data is diversified. Therefore, when a machine learning model of the learning system 60 is created using the training data set T generated by the training data deletion device 10, an increase in the amount of calculation and calculation time required for creating the machine learning model is suppressed, and a decrease in the accuracy of the machine learning model is suppressed.

[0057] [1-3] The training data generation unit 40 performs information compression on ci and the training data set T. Furthermore, the training data deletion device 10 calculates si using the information-compressed ci and the information-compressed training data set T.

[0058] This makes it easier for the data size of the training data set T to change when ci is added to the training data set T. This makes it easier to compare si with the threshold θ. Therefore, it is possible to suppress a decrease in accuracy due to the deletion of overlapping or similar training data.

[0059] [1-4] When Nt is equal to or greater than the required number K, the training data generation unit 40 terminates adding or deleting c i to the training data set T. Note that Nt is the number of training data candidates c added to the training data set T by the training data generation unit 40, as described above.

[0060] This prevents the number of training data included in the training data set T from becoming excessive.

[0061] [1-5] When the training data set T is an empty set, the training data generation unit 40 adds ci to the training data set T. Alternatively, when the data size of ci is zero, the training data generation unit 40 deletes ci.

[0062] This prevents the inability to calculate si, which is a comparison target for removing overlapping or similar training data.

[0063] (Second embodiment) The second embodiment differs from the first embodiment in the processing of the training data generation unit 40. Other than this, the second embodiment is similar to the first embodiment.

[0064] Here, due to information compression, Dc(Tc)-max{Dc(T),Dc(ci)}, which is the numerator on the right side of the above relational expression (1), may become a negative value. When Dc(Tc)-max{Dc(T),Dc(ci)} is a negative value, it is unlikely that si has been calculated appropriately.

[0065] 4 and 5, in step S300 following step S212, the training data generation unit 40 determines whether Dc(Tc)-max{Dc(T),Dc(ci)} is a negative value, thereby determining whether it is unlikely that si has been calculated appropriately.

[0066] When Dc(Tc)-max{Dc(T),Dc(ci)} is not a negative value, that is, when Dc(Tc)-max{Dc(T),Dc(ci)} is zero or a positive value, it is highly likely that si has been calculated appropriately. Therefore, in this case, the processing of the training data generation unit 40 proceeds to step S214. From step S214 onwards, the same processing as in the first embodiment is performed. Furthermore, when Dc(Tc)-max{Dc(T),Dc(ci)} is a negative value, it is highly unlikely that si has been calculated appropriately. Therefore, in this case, the processing of the training data generation unit 40 proceeds to step S302.

[0067] In step S302 following step S300, the training data generation unit 40 calculates a random number R.

[0068] In step S304 following step S302, the training data generation unit 40 determines whether the random number R calculated in step S302 is equal to a predetermined number Rf. As a result, the training data generation unit 40 either adds or deletes c i to the training data set T. Although it is unlikely that c i has been calculated appropriately, this prevents the number of training data in the training data set T from becoming insufficient because c i is randomly set as training data. The predetermined number Rf is set arbitrarily by a user of the training data deletion device 10 or an external system such as the learning system 60.

[0069] Then, when the random number R is the predetermined number Rf, the processing of the training data generation unit 40 proceeds to step S216. In step S216 following step S304, the training data generation unit 40 sets c i selected in step S204 as training data and adds it to the training data set T. From step S218 following step S216, the same processing as in the first embodiment is performed. Furthermore, when the random number R is not the predetermined number Rf, the training data generation unit 40 deletes c i selected in step S204, so that it is not used as training data. Thereafter, the processing of the training data generation unit 40 proceeds to step S220. From step S220 following step S304, the same processing as in the first embodiment is performed.

[0070] As described above, the training data generating unit 40 of the training data deleting device 10 of the second embodiment performs the process. In the second embodiment, the same effects as in the first embodiment are achieved.

[0071] (Third embodiment) The third embodiment differs from the first embodiment in the processing of the training data generation unit 40. Other than this, the third embodiment is similar to the first embodiment.

[0072] Here, depending on the threshold value θ, Nt may not reach the required number K. For this reason, as shown in the flowchart of Fig. 6, in step S400 following step S212, the training data generation unit 40 corrects the threshold value θ acquired in step S200 based on Nt. Note that Nt is the number of training data candidates c added to the training data set T by the training data generation unit 40, as described above.

[0073] For example, when Nt is less than the number threshold Nt_th, the training data generation unit 40 reduces the threshold θ acquired in step S200. This makes it easier for si to be greater than the threshold θ, making it easier for ci to be added to the training data set T. Furthermore, when Nt is equal to or greater than the number threshold Nt_th, the training data generation unit 40 leaves the threshold θ acquired in step S200 unchanged. Note that the number threshold Nt_th is set by experiment, simulation, or the like so that Nt is a desired number.

[0074] Then, in step S214 following step S400, the training data generation unit 40 compares si calculated in step S212 with the threshold value θ corrected in step S400. As a result, the training data generation unit 40 either adds or deletes ci from the training data set T. From step S216 following step S214, the same processing as in the first embodiment is performed.

[0075] As described above, the training data generation unit 40 of the training data deletion device 10 of the third embodiment performs the process. The third embodiment also achieves the same effects as the first embodiment. The third embodiment also achieves the following effects.

[0076] [2] When Nt is less than the number threshold Nt_th, the training data generation unit 40 changes the threshold θ. Here, when Nt is less than the number threshold Nt_th, the training data generation unit 40 decreases the threshold θ.

[0077] This makes it easier to add ci to the training data set T, preventing Nt from not reaching the required number K.

[0078] (Fourth embodiment) The fourth embodiment differs from the first embodiment in the processing of the training data generation unit 40. The rest is the same as the first embodiment.

[0079] In the first embodiment, the training data generation unit 40 performs information compression in step S210. In contrast, in the fourth embodiment, as shown in the flowchart of FIG.

[0080] Therefore, in this case, in step S212 following step S208, the training data generation unit 40 calculates si for ci using the following relational expression (2). In the following relational expression (2), f(Tc) is a value related to the data size of Tc. As described above, Tc is the union of ci and the training data set T. f(T) is a value related to the data size of the training data set T.

[0081]

number

[0082] In step S214 following step S212, the training data generation unit 40 compares si calculated in step S212 with the threshold θ acquired in step S200. As a result, the training data generation unit 40 either adds or deletes ci from the training data set T.

[0083] Here, when the training data included in the training data set T and ci overlap or are similar to each other, the change in data size when ci is added to the training data set T is small. Therefore, si becomes relatively small at this time. Furthermore, when the training data included in the training data set T and ci do not overlap or are not similar to each other, the change in data size when ci is added to the training data set T is large. Therefore, si becomes relatively large at this time.

[0084] Therefore, the training data generator 40 determines whether si calculated in step S212 is greater than the threshold value θ acquired in step S200.

[0085] When si is equal to or smaller than the threshold θ, si is relatively small, and therefore the training data included in the training data set T and ci overlap or are similar to each other. Therefore, at this time, the training data generation unit 40 deletes ci selected in step S204 and does not include it in the training data. Thereafter, the processing of the training data generation unit 40 proceeds to step S220. From step S220 onwards following step S214, the same processing as in the first embodiment is performed. On the other hand, when si is greater than the threshold θ, si is relatively large, and therefore the training data included in the training data set T and ci do not overlap or are similar to each other. At this time, the processing of the training data generation unit 40 proceeds to step S216. From step S216 onwards following step S214, the same processing as in the first embodiment is performed.

[0086] As described above, the training data generating unit 40 of the training data deleting device 10 of the fourth embodiment performs the processing. In the fourth embodiment, the same effects as in the first embodiment are achieved.

[0087] (Fifth embodiment) The fifth embodiment differs from the first embodiment in the processing of the training data generation unit 40. Other than this, the fifth embodiment is similar to the first embodiment.

[0088] Specifically, in step S214, the training data generating unit 40 sets si for ci to 1 / Score(T,ci) instead of Score(T,ci) in the above relational expression (1).

[0089] 8, in step S214 following step S212, the training data generation unit 40 compares si calculated in step S212 with θinv. As a result, the training data generation unit 40 either adds or deletes ci from the training data set T. Note that θinv is the reciprocal of the threshold θ acquired in step S200, i.e., 1 / θ.

[0090] Here, when the training data included in the training data set T and ci overlap or are similar to each other, the change in data size when ci is added to the training data set T is small, and Score(T, ci) is small. Therefore, at this time, si is relatively large. Furthermore, when the training data included in the training data set T and ci do not overlap or are not similar to each other, the change in data size when ci is added to the training data set T is large, and Score(T, ci) is large. Therefore, at this time, si is relatively small.

[0091] Therefore, the training data generating unit 40 determines whether si calculated in step S212 is less than θinv.

[0092] When si is equal to or greater than θinv, si is relatively large and Score(T, ci) is small, so that the training data included in the training data set T and ci overlap or are similar to each other. Therefore, at this time, the training data generation unit 40 deletes ci selected in step S204 and does not include it in the training data. Thereafter, the processing of the training data generation unit 40 proceeds to step S220. From step S220 onwards following step S214, the same processing as in the first embodiment is performed. On the other hand, when si is less than θinv, si is relatively small and Score(T, ci) is large, so that the training data included in the training data set T and ci do not overlap or are similar to each other. At this time, the processing of the training data generation unit 40 proceeds to step S216. From step S216 onwards following step S214, the same processing as in the first embodiment is performed.

[0093] As described above, the training data generating unit 40 of the training data deleting device 10 of the fifth embodiment performs the process. In the fifth embodiment, the same effects as in the first embodiment are achieved.

[0094] (Other embodiments) The present disclosure is not limited to the above-described embodiments, and appropriate modifications can be made to the above-described embodiments. Furthermore, it goes without saying that the elements constituting the embodiments in the above-described embodiments are not necessarily essential unless they are specifically stated as essential or are considered to be clearly essential in principle.

[0095] The acquisition unit, deletion unit, and method described herein may be implemented by a special-purpose computer configured with a processor and memory programmed to perform one or more functions embodied in a computer program. Alternatively, the acquisition unit, deletion unit, and method described herein may be implemented by a special-purpose computer configured with a processor configured with one or more dedicated hardware logic circuits. Alternatively, the acquisition unit, deletion unit, and method described herein may be implemented by one or more special-purpose computers configured with a processor and memory programmed to perform one or more functions in combination with a processor configured with one or more hardware logic circuits. Furthermore, the computer program may be stored as instructions executed by a computer on a computer-readable non-transitory tangible storage medium.

[0096] In the first embodiment, si and Score(T, ci) are expressed as in the above relational expression (1). However, as the variable i increases, Dc(T)>Dc(ci). Therefore, si and Score(T, ci) may be expressed as in the following relational expression (3).

[0097]

number

[0098] In each of the above embodiments, the training data set T is set to an empty set in the initialization process of step S202 of the training data generation unit 40. However, the initialization process is not limited to setting the training data set T to an empty set. In the initialization process, the training data set T may include at least one training data. This at least one training data is set by, for example, a user of the training data deletion device 10 or an external system such as the learning system 60.

[0099] In each of the above embodiments, ci is selected in order starting from 1. However, ci is not limited to being selected in order starting from 1, and may be selected after being adjusted by a user of the training data deletion device 10 or an external system such as the learning system 60, or may be selected randomly.

[0100] The above embodiments may be combined as appropriate.

[0101] (Aspects of the present disclosure) [Point 1] A training data deletion device, comprising: an acquisition unit that acquires multiple training data candidates (c, ci); a deletion unit that either adds or deletes the training data candidate from a training data set (T), which is a set of training data; Equipped with The deletion unit is a training data deletion device that either adds or deletes the training data candidate to the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv). [Point 2] The training data deletion device according to aspect 1, wherein the deletion unit performs information compression on the training data candidates and the training data set, and calculates a value relating to the change in data size using the training data candidates after information compression and the training data set after information compression. [Point 3] The training data deletion device according to aspect 1 or 2, wherein the deletion unit terminates either adding or deleting the training data candidate from the training data set when the number (Nt) of training data candidates added to the training data set is equal to or greater than the required number (K). [Point 4] 4. The training data deletion device according to any one of Aspects 1 to 3, wherein the deletion unit adds the training data candidate to the training data set when the training data set is an empty set. [Point 5] The training data deletion device according to any one of Aspects 1 to 4, wherein the deletion unit deletes the training data candidate when the data size of the training data candidate is zero. [Point 6] The training data deletion device according to any one of aspects 1 to 5, wherein the deletion unit changes the threshold when the number (Nt) of training data candidates added to the training data set is less than a number threshold (Nt_th). [Point 7] A training data deletion method, comprising: Obtaining multiple training data candidates (c, ci); adding or deleting the training data candidate to a training data set (T), which is a set of training data; Including, A training data deletion method that either adds or deletes the training data candidate to the training data set by comparing a value (si) related to the change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv). [Point 8] A training data deletion program, A training data deletion device an acquisition unit that acquires multiple training data candidates (c, ci); functioning as a deletion unit that either adds or deletes the training data candidate from a training data set (T), which is a set of training data; The deletion unit is a training data deletion program that either adds or deletes the training data candidate to the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv). [Explanation of symbols]

[0102] 10 Training data deletion device 20 Condition acquisition section 30 Training data candidate storage unit 40 Training data generation unit 50 Training data storage unit 60 Learning System

Claims

1. A training data deletion device, comprising: an acquisition unit that acquires a plurality of training data candidates (c, ci); a deletion unit that either adds or deletes the training data candidate from a training data set (T), which is a set of training data; Equipped with The deletion unit either adds or deletes the training data candidate to the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with a threshold (θ, θinv).

2. 2. The training data deletion device according to claim 1, wherein the deletion unit performs information compression on the training data candidates and the training data set, and calculates a value related to the change in data size using the training data candidates after information compression and the training data set after information compression.

3. 3. The training data deletion device according to claim 1, wherein the deletion unit terminates either adding or deleting the training data candidate to the training data set when the number (Nt) of training data candidates added to the training data set is equal to or greater than a required number (K).

4. 3. The training data deletion device according to claim 1, wherein the deletion unit adds the training data candidate to the training data set when the training data set is an empty set.

5. 3. The training data deletion device according to claim 1, wherein the deletion unit deletes the training data candidate when the data size of the training data candidate is zero.

6. 3. The training data deletion device according to claim 1, wherein the deletion unit changes the threshold when the number (Nt) of training data candidates added to the training data set is less than a number threshold (Nt_th).

7. A training data deletion method, comprising: Obtaining a plurality of training data candidates (c, ci); adding or deleting the training data candidate to a training data set (T), which is a set of training data; Including, A training data deletion method that either adds or deletes the training data candidate to the training data set by comparing a value (si) related to the change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv).

8. A training data deletion program, A training data deletion device an acquisition unit that acquires a plurality of training data candidates (c, ci); and functioning as a deletion unit that either adds or deletes the training data candidate from a training data set (T), which is a set of training data; The deletion unit either adds or deletes the training data candidate from the training data set by comparing a value (si) related to a change in data size of the training data set (Tc) when the training data candidate is added to the training data set with thresholds (θ, θinv).

Images