System, method executed by system, and program

Abstract

To edit a learning data set so as to improve quality (accuracy) of a model when a learning data format is an original data format (other than a text data format).SOLUTION: A text data conversion unit 102 converts original data 132 having an original data format into original data correspondence text data 134. A measure determination unit 105 generates a measure 162 on the basis of an original data correspondence text data set 133 and a measure determination model 161. A measure execution unit 106 generates a text data set 141 for generation by editing the original data correspondence text data set 133 on the basis of the measure 162. An original data conversion unit 108 converts text data 142 for generation into generation data 144 having the original data format. The generation data 144 is learning data for training a determination model 165 for determining a class 191 of data 148 having the original data format by machine learning.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a technique for editing a learning dataset used for training a model by machine learning (for example, a technique for performing data augmentation or data generation).

Background Art

[0002] The quality (accuracy) of a model trained by machine learning may depend on the quality of the learning dataset used for training. For each of various properties that the learning data may have, it is considered useful for forming a high-quality learning dataset that the number of learning data having that property exists to a certain extent or more. However, for reasons such as the cost of collecting learning data, it may not be possible to collect a sufficient number of learning data to form a high-quality learning dataset. For example, in the collected learning dataset, there may be more than enough learning data having a certain property, an appropriate number of learning data having another property, and a shortage of the number of learning data having yet another property.

[0003] For the above reasons, conventionally, techniques for editing learning datasets have been known. Examples of techniques for editing learning datasets include data augmentation techniques and data generation techniques. The data augmentation technique is a technique for creating or adding new learning data by changing (editing) certain learning data included in the learning dataset. The data generation technique includes, in addition to the technique of generating learning data by data augmentation, a technique for generating new learning data not included in the learning dataset. As techniques for editing learning datasets (e.g., data augmentation techniques and data generation techniques), for example, there are the techniques disclosed in Non-Patent Document 1 and the techniques disclosed in Non-Patent Document 2. Non-Patent Document 1 discloses a technique for performing automatic data augmentation (AutoAugment) on learning data for realizing discrimination of classes of image data (image classification). Non-Patent Document 2 discloses a technique for performing automatic data augmentation (Text AutoAugment) on text data for learning data for realizing discrimination of classes of data in text data format (text classification).

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] A model trained by machine learning may handle data having combinations of a plurality of attribute values (including class values). For example, in a discrimination model (classification model, identification model) for discriminating classes indicating the types of objects included in image data (for example, classes indicating each of a car, a person, an animal (for example, a dog, a cat, a horse), and a non-living object (for example, a ball)), each of the image data handled by the discrimination model may have a combination of a plurality of attribute values (including class values). Specifically, the image data handled by the discrimination model may have a class value indicating the type of the object to be discriminated by the class (for example, a car, a person, an animal (for example, a dog, a cat, a horse), a non-living object (for example, a ball)), and may also have an attribute value that is not a class value. Here, the attribute value that is not a class value may be, for example, an attribute value indicating a detailed type, shape, size, or color of the object to be discriminated by the class (more detailed than the class), or an attribute value indicating the type (for example, a road, a garden, the sky), shape, size, or color of an object that is not the object to be discriminated by the class (for example, a background). For example, image data indicating "a black car running on a white road" has, as attribute values, a combination of the class value "car" and the attribute values other than the class value, "black as the color of the car", "road as the background", and "white as the color of the road". Also, for example, image data indicating "a black car running on a gray road" has, as attribute values, a combination of the class value "car" and the attribute values other than the class value, "black as the color of the car", "road as the background", and "gray as the color of the road". The above is an example of a discrimination model that handles image data, but the same may also apply to, for example, a discrimination model that handles audio data. For example, in the audio data handled by the discrimination model, the content indicated by the voice uttered by a person may be treated as a class value, and the rest (for example, the volume of the voice, the wavelength of the voice, the volume of the sound other than the voice (for example, music or noise), the wavelength of the sound other than the voice) may be treated as an attribute value other than the class value. Hereinafter, for example, a data format other than the text data format, such as image data or audio data, may be called the "original data format".

[0006] In order to improve the quality (accuracy) of a model that handles data in the original data format having combinations of a plurality of attribute values (including class values), in the learning dataset for the model, the number of learning data for each combination of a plurality of attribute values (including class values) should be present to a certain extent or more, and it is useful to balance the number of learning data among the combinations. Alternatively, even if there is an imbalance in the number of learning data among the combinations of a plurality of attribute values (including class values) in the learning dataset, it is useful to balance to some extent the quality (accuracy) provided by the model (trained with the learning data) among the combinations. That is, it is useful that for data having a specific combination among the combinations of a plurality of attribute values (including class values), the quality (accuracy) of the model does not become unacceptably low. For example, in a learning dataset for training a discrimination model that discriminates the class of image data, merely having sufficient learning data for each class value (for example, a value indicating any one of a car, a person, an animal (e.g., dog, cat, horse), and an inanimate object (e.g., ball)) may not result in sufficient discrimination accuracy of the class by the discrimination model after training. For example, assume that the discrimination model is trained so that it can discriminate the class of "car" when image data including a car is input. In this assumption, it is useful to evenly prepare learning data having various attribute values (not class values) such as the color of the car, what is in the background (e.g., road, garden, sky), and the color of the background for the learning data that is image data including a car. If the preparation is insufficient, the discrimination accuracy of the class by the discrimination model may be low for combinations of attribute values (including class values) with insufficient number of learning data. For example, when image data showing "a white car running on a black road" is input to the discrimination model, it may correctly discriminate the class of "car", while when image data showing "a black car running on a white road" is input to the discrimination model, it may not correctly discriminate the class of "car".

[0007] However, when the data handled by the model and the learning data for training the model by machine learning are data in the form of original data having a combination of a plurality of attribute values (including class values), such as image data or voice data, the technique of editing the conventionally known learning data set is not suitable. For example, the technique disclosed in Non-Patent Document 1 generates new image data, which is new learning data, by editing (data augmentation) the image data that is the learning data. The technique disclosed in Non-Patent Document 1 attempts to improve the accuracy of discriminating (classifying) the class of the image data by using automatic data augmentation (AutoAugment). However, in the technique disclosed in Non-Patent Document 1, there is little consideration for attribute values other than the class values that the image data has. That is, in the technique disclosed in Non-Patent Document 1, for each combination of a plurality of attribute values (including class values), the number of learning data is made to exist to a certain extent or more, and there is little consideration for balancing the number of learning data between the combinations or for balancing to a certain extent the quality (accuracy) provided by the model (trained by the learning data) between the combinations. In addition, the technique disclosed in Non-Patent Document 2 generates new learning data in the form of new text data by editing (data augmentation) the data in the form of text data, which is the learning data. In the technique disclosed in Non-Patent Document 2, the data handled by the model and the learning data for training the model by machine learning are always data in the form of text data. That is, the technique disclosed in Non-Patent Document 2 does not handle data in the original data format (which is a format other than the text data format), such as image data or audio data. Further, when performing automatic data augmentation (Text AutoAugment) in text data, the technique disclosed in Non-Patent Document 2 lacks consideration for ensuring that there are a certain number or more of learning data for each combination of a plurality of attribute values (including class values), and for balancing the number of learning data between the combinations, or for balancing to a certain extent the quality (accuracy) provided by the model (trained with the learning data) between the combinations.

[0008] Based on the above, when the format of the data handled by the model and the format of the learning data for training the model by machine learning are in the original data format (which is a format other than the text data format), it may be one of the objectives of the present disclosure to edit the learning data set so as to improve the quality (accuracy) of the model.

Means for Solving the Problem

[0009] In order to achieve at least one of the above objectives, the features that the present disclosure may include are, for example, as follows. One aspect of the present disclosure is a system. The system includes a text data conversion unit, a policy determination unit, a policy execution unit, and an original data conversion unit. The text data conversion unit converts each piece of original data in the original data format included in the original data set into each piece of original data corresponding text data in the text data format. The policy determination unit generates a policy to be applied to the original data corresponding text data set based on the original data corresponding text data set composed of each piece of original data corresponding text data and a policy determination model. The policy execution unit edits the original data corresponding text data set based on the policy to generate a generated text data set that is a set of generated text data. The original data conversion unit converts each piece of generated text data included in the generated text data set into each piece of generated data in the original data format. The generated data is learning data for training a discrimination model that discriminates classes of data in the original data format by machine learning.

Advantages of the Invention

[0010] When generating generated data (set) in the original data format, which is learning data for training a model by machine learning, the present disclosure converts the original data in the original data format into original data corresponding text data (set) in the text data format, and then performs text data editing on the original data corresponding text data (set). Furthermore, the present disclosure converts the generated text data (set) in the text data format into the generated data (set) in the original data format. That is, the present disclosure performs the substantial editing process of the learning data not in the original data format as it is, but after converting it into the text data format. Here, when editing information on combinations of attribute values (including class values) that the data has, it is easier to achieve intended editing, flexible editing, or accurate editing when the editing process is performed after converting the data into text data format rather than performing the editing process while the data remains in its original data format. That is, even if the present disclosure is learning data for a model that handles data in the original data format, it is possible to perform intended editing, flexible editing, or accurate editing on the information on combinations of attribute values (including class values) that the learning data has.

[0011] In addition, the present disclosure generates a strategy for a text data set corresponding to original data obtained by converting an original data set into text data format. Then, the present disclosure performs text data editing on the text data set corresponding to the original data based on the strategy. Therefore, the present disclosure can generate an appropriate strategy according to the situation of the original data set (text data set corresponding to the original data) (for example, the absolute number of original data (text data corresponding to the original data) for each combination of attribute values (including class values), the balance situation of the number of original data (text data corresponding to the original data) between the combinations, or the balance situation of the quality (accuracy) of the model between the combinations, etc.), and generate a text data set for generation (generated data set) that improves the quality (accuracy) of the model based on the strategy.

[0012] As described above, when the format of the data handled by the model and the format of the learning data for training the model by machine learning are in the original data format (a format other than the text data format), the present disclosure can edit the learning data set so as to improve the quality (accuracy) of the model.

[0013] A method or program that realizes the same thing as the processing realized by the above system can also obtain the same operational effects as the above system. In the case of a program aspect, costs are often reduced. In a program, it is also easy to make design changes regarding the processing. Features that the present disclosure may have other than those described above, and the effects corresponding to those features are disclosed in this specification, the claims, or the drawings.

Brief Description of the Drawings

[0014]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

Figure 20

Embodiments for Carrying Out the Invention

[0015] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. Note that the embodiments described below do not limit the disclosure according to the claims, and not all of the elements and combinations thereof described in the embodiments are essential for the solution means of the present disclosure. The following description and drawings are examples for explaining the present disclosure, and for the sake of clarity of explanation, appropriate omissions and simplifications have been made. The present disclosure can be implemented in various other forms. Unless otherwise particularly limited, each component may be in a single or plural number. The positions, sizes, shapes, ranges, etc. of the components shown in the drawings may not represent the actual positions, sizes, shapes, ranges, etc. in order to facilitate the understanding of the invention. For this reason, the present disclosure is not necessarily limited to the positions, sizes, shapes, ranges, etc. disclosed in the drawings. Each of the systems, devices, or functional units of the present disclosure may be integrated into one hardware-wise, or may be divided into a plurality of parts and the parts may cooperate to play a role. Some systems, devices, or functional units may be integrated hardware-wise. Each of the systems, devices, or functional units may be realized by causing a computer to execute software (program) (as shown in FIG. 3). A part of the functions of the systems, devices, or functional units may be realized by hardware (for example, hard-wired logic or field programmable gate array (FPGA)), and the remaining functions may be realized by executing software (program). All of the functions of each of the systems, devices, or functional units may be realized hardware-wise. Some or all of the steps shown in the flowcharts and the like described in the present disclosure may be realized hardware-wise. One or more of the hardware resources may implement each of the systems, apparatuses, or functional units of the present disclosure. For this purpose, each of the systems, apparatuses, or functional units of the present disclosure may be virtually implemented. For example, virtual machine or container techniques may be used. The program of the present disclosure is included in the concept of general software that, when software and hardware resources cooperate, constructs a specific information processing system (system) or its operation method according to the purpose of use. That is, the program of the present disclosure is not limited to a specific type or form of program. Also, the program may initially be recorded in a compressed format. Elements using the same reference numeral in multiple drawings are similar to each other. In the drawings showing flowcharts, rectangular boxes indicate processing steps, and hexagonal boxes indicate conditional branch steps. In the drawings showing flowcharts, "step" is abbreviated as "S". Also, the display or output modes shown in the drawings are merely examples. Within the scope of the gist of the present disclosure, the display or output modes may be different from those shown in the drawings.

[0016] 1. Data generation, etc. considering combinations of basic functional configurations and attribute values (FIG. 1) FIG. 1 shows the basic functional configuration 100 (and the information handled) of the system according to an embodiment of the present disclosure. Note that not all of the functional configurations shown in FIG. 1 are essential. Also, the existence of functional configurations other than those shown in FIG. 1 is not precluded. System 101 generates a generated data set 143 which is a set of generated data 144. The generated data 144 is learning data for training a discrimination model 165 by machine learning. The discrimination model 165 discriminates the class of data 148 in the original data format and outputs a discrimination result class 191 (class value). That is, the discrimination model 165 handles a single-class (single-label) discrimination problem (classification problem) or a multi-class (multi-label) discrimination problem (classification problem). The data 148 in the original data format is data other than the text data format, and may be, for example, image data or audio data. The data 148 in the original data format may have not only one or more class values to be discriminated but also one or more attribute values other than the class values. For example, the image data indicating "a black car running on a white road" may have, as attribute values, a combination of "car" as the class value and "black as the color of the car", "road as the background", and "white as the color of the road" as attribute values other than the class value. The generated data 144 is learning data in the original data format. System 101 uses an original data set 131 which is a set of original data 132 (with teacher information (label information indicating the correct class)) in the original data format to obtain a generated data set 143 which is a set of generated data 144 (with teacher information (label information indicating the correct class)) in the original data format. That is, System 101 edits the original data set 131 (for example, performs data augmentation or data generation processing) to generate the generated data set 143.

[0017] System 101 has, as functional units, a text data conversion unit 102, a policy determination unit 105, a policy execution unit 106, and an original data conversion unit 108. Each of these functional units may be realized by executing a program or may be implemented more in a hardware manner. The text data conversion unit 102 converts each piece of original data 132 in the original data format included in the original data set 131 (which is data other than the text data format and may be, for example, image data or audio data) into corresponding text data 134 in the text data format for the original data. For example, if the original data 132 is image data indicating "a black car running on a white road", the text data conversion unit 102 may use data in the text data format of "a black car running on a white road" as the corresponding text data 134 for the original data. Incidentally, the text data conversion unit 102 may use, for example, a text data conversion model trained by machine learning to convert data in the original data format into data in the text data format. As a result of training by machine learning, this text data conversion model may be capable of associating the latent variables of the data in the original data format with the text data explaining the data. The policy determination unit 105 generates a policy 162 to be applied to the original data corresponding text data set 133 based on the original data corresponding text data set 133 composed of each piece of original data corresponding text data 134 and the policy determination model 161. Here, the policy 162 indicates the policy when editing the original data corresponding text data set 133 in the text data format. The policy 162 may include one or more editing operations. Each of the editing operations may be, for example, the content shown in FIG. 8 described later. As shown in FIG. 7 described later, the policy determination unit 105 may input a vector data set 139 derived from some or all of the original data corresponding text data 134 included in the original data corresponding text data set 133 into the policy determination model 161. Since the policy decision unit 105 generates the policy 162 as described above, the policy 162 is for correcting imbalance or the like according to the situation of the original data set 131 (original data corresponding text data set 133) (for example, the absolute number of the original data 132 (original data corresponding text data 134) for each combination of attribute values (including class values), the balance situation of the number of the original data 132 (original data corresponding text data 134) between the combinations, or the balance situation of the quality (accuracy) of the discrimination model 165 between the combinations, etc.). Based on the policy 162, the policy execution unit 106 edits the original data corresponding text data set 133 to generate the generation use text data set 141. For example, when the policy 162 consists of one or more editing operations, the policy execution unit 106 applies each of the editing operations included in the policy 162 to the original data corresponding text data set 133 to perform editing (text data editing) in the text data format. In this way, in order to obtain the generated data set 143 (learning data set) from the original data set 131, the policy execution unit 106 performs editing (text data editing) in the text data format instead of performing editing in the original data format. Therefore, as editing of the combination of attribute values (including class values) of the learning data (such as changing class values and attribute values), it is easy to realize intended editing such as correcting imbalance, flexible editing, or accurate editing. The original data conversion unit 108 converts each of the generation text data 142 included in the generation text data set 141 into each of the generation data 144 in the original data format. For example, if the generation text data 142 is data in the text data format of "a white dog running on a black road", the original data conversion unit 108 may generate image data indicating "a white dog running on a black road" and use the image data as the generation data 144. Incidentally, the original data conversion unit 108 may use, for example, an original data conversion model trained by machine learning so as to generate data in the original data format based on data in the text data format. Further, as will be described later, when the content of the editing operation included in the strategy 162 instructs to generate a plurality of generation data 144 for a certain generation text data 142, the original data conversion unit 108 may generate a plurality of mutually different image data although they are the content indicated by the certain generation text data 142. The set of the generation data 144 generated as described above becomes the generation data set 143. Compared with the situation regarding the original data set 131, the situation regarding the generation data set 143 (for example, the absolute number of the generation data 144 in the original data format for each combination of attribute values (including class values), the balance situation of the number of the generation data 144 between the combinations, or the balance situation of the quality (accuracy) of the discrimination model 165 between the combinations) is improved.

[0018] FIG. 2 shows an aspect 200 when the generation data set 143 is generated by editing (for example, data expansion / data generation) the original data set 131. The editing takes into account the situation of the combination of a plurality of attribute values (including class values) that the data has. In the upper left part of FIG. 2, a distribution 201 of the number of original data included in the original data set 131 is shown. Here, each of the original data 132 has one of A, B, C, and D as the first attribute value (which may be a class value), and one of a, b, and c as the second attribute value (which may be a class value). (Note that each of the original data 132 may have a third attribute value, a fourth attribute value, and so on.) In the example of the distribution 201 of the number of original data shown in the upper left part of FIG. 2, for each combination of the first attribute value and the second attribute value, the number of original data having the combination is shown as a column graph. (Hereinafter, a set of data for each combination of attribute values (including class values) may be referred to as a "subset".) And in the example of the distribution 201 of the number of original data shown in the upper left part of FIG. 2, there is no original data 132 whose first attribute value is D and second attribute value is a, nor is there original data 132 whose first attribute value is D and second attribute value is c. On the other hand, the number of original data 132 whose first attribute value is A and second attribute value is a is relatively large. Thus, in the original data set 131, there may be an imbalance in the number of original data 132 for each subset. In the upper right part of FIG. 2, the discrimination accuracy 202 of the discrimination model 165 (or 166 (see FIG. 4)) after being trained by machine learning using the original data set 131 as learning data is shown. The discrimination accuracy of the discrimination model 165 (or 166) here may indicate the degree of coincidence between the class 191 of the data 148 in the original data format discriminated by the discrimination model 165 (or 166) and the correct class of the data 148 in the original data format (for example, the class indicated by the teacher information (label)). In the upper right part of FIG. 2, for each subset of the verification data 146 in the original data format included in the verification data set 145 described later, the discrimination accuracy when the discrimination model 165 (or 166) discriminates the class of the verification data 146 included in the subset is shown by a column graph. As shown in the upper left part of FIG. 2, when there is no original data 132 whose first attribute value is D and second attribute value is a, and no original data 132 whose first attribute value is D and second attribute value is c, as shown in the upper right part of FIG. 2, the discrimination accuracy in the subset where the first attribute value is D and the second attribute value is a and the discrimination accuracy in the subset where the first attribute value is D and the second attribute value is c tend to be low. In the lower left part of FIG. 2, the distribution 211 of the number of generated data included in the generated data set 143 is shown. Similar to the distribution 201 of the number of original data, the distribution 211 of the number of generated data is also represented by a column graph for each subset. It is shown that the imbalance in the number of generated data between subsets is resolved in the generated data set 143 by editing (for example, data augmentation and data generation). In the lower right part of FIG. 2, the discrimination accuracy 212 of the discrimination model 165 (or 166) after being trained by machine learning using the generated data set 143 as learning data is shown. Similar to the discrimination accuracy 202 in the upper right part of FIG. 2, the discrimination accuracy 212 in the lower right part of FIG. 2 is also represented by a column graph for each subset. In the example of FIG. 2, in response to the correction of the imbalance between subsets in the distribution 211 of the number of generated data, the imbalance between subsets in the discrimination accuracy 212 is also corrected. It can be said that the strategy 162 is generated so that such correction of the imbalance can be realized. Further, FIG. 2 shows a state in which both the correction of the imbalance between subsets in the distribution 211 of the number of generated data and the correction of the imbalance between subsets in the discrimination accuracy 212 are achieved. Depending on the data in the original data format handled by the discrimination model 165, the correction of the imbalance between subsets in the distribution 211 of the number of generated data may not be sufficient, but the correction of the imbalance between subsets in the discrimination accuracy 212 may be achieved.

[0019] Since the system 101 in the embodiment of the present disclosure has the functional configuration as described above, it can have the effects shown in the above [Effects of the Invention]. In addition, the system 101 can widen the range in which the processing up to the generation of the generated data 144 can be automated.

[0020] 2. Computer Architecture for Realizing Embodiment of the Present Disclosure (FIG. 3) FIG. 3 shows a computer architecture 300 for realizing the system 101 of the embodiment of the present disclosure. To realize the system 101, some or all of the information processing device 301, the storage device 302, the non-volatile recording medium (recording device) 303, the external recording medium drive 304, the input device 306, the display or output device 307, the communication device 308, the external input / output port 309, and the reading device 310 may be interconnected by the interconnecting section 311. (Note that a part or all of the interconnecting section 311 may be a network. In that case, the system 101 is realized by a plurality of devices via the network.) The information processing device 301 may be, for example, a processor. Examples of this processor include a CPU, an MPU, or a GPU. Alternatively, the processor referred to here may be another semiconductor device as long as it is a main body that executes predetermined processing. Also, the information processing device 301 may be one or more (micro)processors. The storage device 302 may be, for example, a memory. The non-volatile recording medium (recording device) 303 may be, for example, a non-volatile memory (e.g., flash memory) or a non-volatile disk device. The external recording medium drive 304 may be, for example, a disk drive. The input device 306 may be, for example, a mouse, a keyboard, an imaging device, a sensor, a touch panel, or a pointing device. The display or output device 307 may be, for example, a display, a printer, or a speaker. The communication device 308 may be, for example, a communication device for wired communication or a communication device for wireless communication. The communication device 308 may be a network interface device (NIC) that controls communication with other systems, devices, terminals, or servers according to a predetermined protocol. The interconnecting section 311 may be, for example, a bus or a crossbar switch. (As described above, part or all of the interconnecting section 311 may be a network.)

[0021] In the non-volatile recording medium (recording device) 303, various programs included in the program group 331 (for example, programs for realizing the functional configurations according to the present disclosure. For example, various programs for implementing each of the functional units realized in the system 101.), various data groups included in the learning-related data group 332, various data groups included in the internal handling data group 333, or various information 334 may be recorded. The program group 331 may include a text data conversion program 102P, a word / part-of-speech extraction program 103P, a vector conversion program 104P, a policy decision program 105P, a policy execution program 106P, a text data loss calculation program 107P, an original data conversion program 108P, a discrimination model training program 109P, an original data loss calculation program 110P, a policy decision model training program 111P, a discrimination model operation program 112P, a data number display output control program 113P, a discrimination accuracy display output program 114P, an original data display output control program 115P, a text data display output control program 116P, or an attribute value display output control program 117P, each for realizing a corresponding functional unit among the text data conversion unit 102, the word / part-of-speech extraction unit 103, the vector conversion unit 104, the policy decision unit 105, the policy execution unit 106, the text data loss calculation unit 107, the original data conversion unit 108, the discrimination model training unit 109, the original data loss calculation unit 110, the policy decision model training unit 111, the discrimination model operation unit 112, the data number display output control unit 113, the discrimination accuracy display output control unit 114, the original data display output control unit 115, the text data display output control unit 116, or the attribute value display output control unit 117. Note that some of the above programs may be integrated into one program. Also, any of the above programs may be divided into multiple programs. The learning-related data group 332 may include an original data set 131, an original data-corresponding text data set 133, a word / part-of-speech data group 135, a text data set for policy decision 137, a vector data set 139, a generated text data set 141, a generated data set 143, or a verification data set 145. The internal processing data group 333 may include decision-making model information 161i which is information representing the decision-making model 161 (which may include model parameters), a policy 162, text data evaluation model information 163i which is information representing the text data evaluation model 163 (which may include model parameters), text data loss 164, discrimination model information 165i which is information representing the discrimination model 165 (which may include model parameters), original data loss 167, or policy loss 168. Alternatively, some or all of the various programs included in the above-described program group 331, the various data groups included in the learning-related data group 332, the various data groups included in the internal processing data group 333, or the various information 334 may be acquired from outside the configuration shown in FIG. 3.

[0022] The external recording medium drive 304 can connect to the external recording medium 305. The external recording medium 305 may be, for example, a portable recording disk (such as a DVD), an IC card, an SD card, a non-volatile memory (such as a flash memory), or a portable hard disk. Incidentally, from this external recording medium 305, various programs included in the program group 331, various data included in the learning-related data group 332, various data included in the internal processing data group 333, or information similar to the information in the various information 334 may be transferred and stored in the non-volatile recording medium (recording device) 303 or the storage device 302. The external recording medium 305 may be used to record programs and data handled in the system 101. The external recording medium drive 304 and the external recording medium 305 may be connected to the system 101 shown in FIG. 3 via a network. The various programs included in the program group 331, the various data such as those included in the learning-related data group 332, the various data such as those included in the internal processing data group 333, or the information in the various information 334 may be provided via the communication device 308, the external input / output port 309, the input device 306, and the reading device 310 and recorded or stored in the non-volatile recording medium (recording device) 303 or the storage device 302.

[0023] For the architecture of FIG. 3 to function as the system 101, each functional unit within the system 101, or a part of each functional unit (executing one or a series of processes (steps)), each of the various programs included in the program group 331 may be loaded into the storage device 302 (e.g., from the non-volatile recording medium (recording device) 303). The program after loading is indicated as 321 in FIG. 3. Then, the information processing device 301 may execute the program 321 (using, as necessary, various data etc. included in the learning-related data group 332 present in the non-volatile recording medium (recording device) 303 etc., various data etc. included in the internal handling data group 333, or information of various information 334). By executing the program 321, the functions of the system 101, each functional unit within the system 101, or a part of each functional unit are realized (one or a series of processes (steps) are executed). Various buffers 323 temporarily formed in the storage device 302 at this time may also be appropriately used.

[0024] 3. Processes performed by the embodiments of the present disclosure The processes executed by the system 101 which is an embodiment of the present disclosure are described. Note that it is not essential to realize all of the functional configurations described below and perform all of the processes. Also, it is not precluded that functional configurations other than those and processes described below are realized and processes are executed. Hereinafter, mainly with reference to FIGS. 4 to 9, the processes executed by the system 101 when training the policy decision model 161 by machine learning are described. Also, mainly with reference to FIG. 10, the processes executed by the system 101 when operating the trained policy decision model 161 are described. In addition, with reference to FIGS. 11 to 20, the display output control processes that the system 101 can perform are described as the generated data set 143 is generated based on the original data set 131.

[0025] 3.1 Processes during policy decision model training (FIGS. 4 to 9) FIG. 4 shows a functional configuration (detailed functional configuration) 400 (and information handled) related to the processing executed by the system 101 when training the policy decision model 161 by machine learning. Compared with FIG. 1 showing an outline of the functional configuration when operating the trained policy decision model 161, FIG. 4 additionally shows functional configurations etc. (for example, text data loss calculation unit 107, text data evaluation model 163, text data loss 164, verification data set 145 (verification data 146), original data loss calculation unit 110, original data loss 167, policy decision model training unit 111, policy loss 168) used for training the policy decision model 161, and also shows functional configurations etc. (for example, word / part-of-speech extraction unit 103, vector conversion unit 104, word / part-of-speech data group 135, vector data set 139, discrimination model training unit 109) that were omitted in FIG. 1. In FIG. 4, those surrounded by a circle are connected to each other. FIGS. 5 and 6 show flowcharts of the processing executed by the system 101 when training the policy decision model 161 by machine learning. In FIGS. 5 and 6, those surrounded by a circle with an alphabet (E or F) are connected to each other. Hereinafter, each of the processing steps included in the flowcharts of FIGS. 5 and 6 will be described while appropriately referring to FIGS. 4, 7, 8, and 9.

[0026] In step 501 (text data conversion step) of FIG. 5, the text data conversion unit 102 converts each piece of original data 132 in the original data format included in the original data set 131 into the corresponding text data 134 in the text data format. The details of the text data conversion unit 102 are as already shown in the description regarding FIG. 1. A set of the corresponding text data 134 forms the corresponding text data set 133. Note that at the upper part of FIG. 7, for each piece of original data 132 (original data (1), original data (2),...) included in the original data set 131, a state where each piece of corresponding text data 134 (corresponding text data (1), corresponding text data (2),...) included in the corresponding text data set 133 is generated is shown. In step 502 of FIG. 5, the word / part-of-speech extraction unit 103 performs part-of-speech analysis on each piece of the corresponding text data 134 in the text data format, for example, by morphological analysis. The word / part-of-speech extraction unit 103 stores each of the decomposed parts of speech (for example, nouns, verbs, adjectives, particles, etc.) as a word / part-of-speech data group 135. Note that the word / part-of-speech data group 135 may be added with words / parts-of-speech obtained by performing part-of-speech analysis on text data other than the corresponding text data 134. From each of the corresponding text data 134, a word / part-of-speech data group 135 is formed, and by utilizing the word / part-of-speech data group 135 in the text data editing executed by the policy execution unit 106, the possibility that the content of the generated text data 142 obtained by the text data editing becomes appropriate can be increased.

[0027] In step 503 of FIG. 5, the vector conversion unit 104 uses all or a randomly sampled part of the original data-corresponding text data 134 included in the original data-corresponding text data set 133 as each of the text data for policy decision. A text data set for policy decision 137 is formed by a set of the text data for policy decision. Note that from the upper part to the middle part of FIG. 7, all or a randomly sampled part of each of the original data-corresponding text data (original data-corresponding text data (1), original data-corresponding text data (2),...) is shown as each of the text data for policy decision (text data for policy decision (i), text data for policy decision (ii),...). The text data set for policy decision 137 is used to form a vector data set 139 that is input to the policy decision model 161. When determining the policy 162 of text data editing applied by the policy decision model 161 to the original data-corresponding text data set 133, it is not always necessary for the policy decision model 161 to know all the information of the original data-corresponding text data set 133. There may be a case where it is sufficient for the policy decision model 161 to know a certain situation regarding the original data-corresponding text data set 133 (for example, the imbalance situation between subsets formed by combinations of (attribute values (including class values)) in the set of the original data-corresponding text data 134 included in the original data-corresponding text data set 133). Therefore, the text data set for policy decision 137 may be formed by the above-described random sampling. In this way, the amount of information that needs to be handled when implementing the policy decision model 161 can also be suppressed. In step 504 of FIG. 5, the vector conversion unit 104 converts each piece of the decision-making policy text data included in the decision-making policy text data set 137 into each piece of vector data that can be handled by the decision-making policy model 161. A vector data set 139 is formed by a set of vector data. From the middle to the bottom of FIG. 7, the state in which each piece of vector data (vector data (i), vector data (ii), etc.) is generated corresponding to each piece of the decision-making policy text data (decision-making policy text data (i), decision-making policy text data (ii), etc.) is shown.

[0028] In step 505 (decision-making step) of FIG. 5, the decision-making unit 105 inputs each piece of the vector data included in the vector data set 139 into the decision-making policy model 161. The decision-making unit 105 obtains, as the output of the decision-making policy model 161, a policy 162 for text data editing applied to the original data corresponding text data set 133. As shown in the lower part of FIG. 7, the decision-making policy model 161 is capable of receiving a plurality of pieces of vector data included in the vector data set 139. Since the decision-making policy model 161 is enabled to receive such an input, the decision-making policy model 161 can know the situation of the original data corresponding text data set 133 that is the basis of the vector data set 139 (for example, the situation of the imbalance between subsets (formed by combinations of attribute values (including class values)) in the set of the original data corresponding text data 134 included in the original data corresponding text data set 133)). That is, the decision-making policy model 161 (after sufficient training) can generate a policy 162 suitable for the situation of the original data corresponding text data set 133.

[0029] The policy 162 output by the policy decision model 161 may include one or more operations (editing operations) of text data editing to be applied to the original data-corresponding text data set 133. When a plurality of editing operations are included in the policy 162, each of the editing operations may be applied to the same original data-corresponding text data set 133. FIG. 8 shows an example of the content of the editing operation included in the policy 162. Each of the editing operations may have one or more of, for example, "type of text data editing", "specification of the original data-corresponding text data to be edited", "number of generated text data generated from one original data-corresponding text data", and "number of generated data generated from one generated text data". The format of the information output from the policy decision model 161 (the format of the policy 162) may be predetermined so that a policy 162 including an editing operation having the content shown in FIG. 8 can be generated.

[0030] The information on the "type of text data editing" may indicate one or more of "deletion", "replacement", "insertion", and "substitution". "Deletion" may be to delete Na words (for example, nouns, adjectives, particles, etc.) of the original data-corresponding text data 134 to be edited. For example, when Na = 2 and the original data-corresponding text data 134 to be edited is "a white big cat in a green garden", the text data after text data editing (generated text data 142) may be "a cat in a green garden". That is, here, two words, "white" and "big", are deleted. In addition, words indicating classes (class values) included in the original data-corresponding text data 134 to be edited may be defined so as not to be targets of "deletion". For example, when the original data-corresponding text data 134 to be edited is "a white big cat in a green garden", even if "deletion" is specified as an editing operation, "cat" indicating the class value may not be deleted. In order to achieve this, as described later (for example, in FIG. 9), a prohibited policy penalty 941 for imposing a penalty on a policy 162 including content for deleting a word indicating a class may be included in the calculation of the policy loss 168. Alternatively, when the policy 162 output by the policy decision model 161 includes content for deleting a word indicating a class, a policy correction unit may be provided to correct the policy 162 so as not to include content for deleting a word indicating a class. Alternatively, when the policy execution unit 106 generates the generation text data 142, (for example, as a result of "random deletion" of randomly deleting words from the original data corresponding text data 134), the generation text data 142 in which a word indicating a class is deleted may be deleted from the generation text data set 141. Alternatively, when the policy execution unit 106 generates the generation text data 142, (for example, as a result of "random deletion"), if the generation text data 142 in which a word indicating a class is deleted is generated, the text data before the deletion is restored and may be used as the generation text data 142 again. As described above, by ensuring that the text data after text data editing (generation text data 142) does not have a word indicating a class (class value) deleted, it is possible to prevent the generation of inappropriate learning data (learning data that does not correspond to any class) for training the discrimination model 165 for discriminating classes by machine learning. "Replacement" may be to interchange any Nb words (for example, nouns, adjectives, particles, etc.) in the original data corresponding text data 134 to be edited. For example, when Nb = 1 and the original data corresponding text data 134 to be edited is "a white dog in a green garden", the text data after text data editing (generated text data 142) may be "a brown dog and a white cat in a green garden". That is, here, two words, "brown" and "white", are interchanged with each other once. Note that in order for the content of the text data after editing to be appropriate, words with the same part of speech may be interchanged. In "Replacement", the interchange of particles in Japanese text data may be allowed, and the interchange of prepositions in English text data may be allowed. "Insertion" may be to insert Nc words (for example, nouns, adjectives, particles, etc.) into the original data corresponding text data 134 to be edited. For example, when Nc = 4 and the original data corresponding text data 134 to be edited is "a white dog in a green garden", the text data after text data editing (generated text data 142) may be "a white dog, a brown cat, and a running person in a green garden". That is, here, four words, "brown", "cat", "running", and "person", are inserted. Note that the number of particles does not have to be included in the number of inserted words Nc in "Insertion". In the above example, the number of particles "and" is not included in Nc. "Substitution" may be to replace Nd words (for example, nouns, adjectives, particles, etc.) included in the original data corresponding text data 134 with other words. For example, when Nd = 2 and the original data corresponding text data 134 to be edited is "a black car running on a white road", the text data after text data editing (generated text data 142) may be "a black horse running on a blue road". That is, here, the word "white" is replaced with the word "blue" of the same part of speech, and the word "car" is replaced with the word "horse" of the same part of speech. Note that in this example, the word indicating the class value is changed from "car" to "horse". The words used for word replacement may be obtained from the word - part - of - speech data group 135. Also, the word after replacement may be selected from the word - part - of - speech data group 135 so that the part of speech remains the same before and after replacement. By using the word - part - of - speech data group 135 (derived from the original - data - corresponding text data set 133) as the information source for the words for "replacement", it is expected that the content of the text data after text data editing (generation - use text data 142) will be appropriate.

[0031] The information of "specification of the original - data - corresponding text data to be edited" may specify which of the original - data - corresponding text data 134 included in the original - data - corresponding text data set 133 is to be the target of text data editing. When all of the original - data - corresponding text data 134 included in the original - data - corresponding text data set 133 are selected as the text data for policy decision, the "specification of the original - data - corresponding text data to be edited" may include the specifier Ne. The specifier Ne may directly indicate the index of the original - data - corresponding text data 134 to be edited. For example, when Ne = 2, the original - data - corresponding text data 134 with index 2 may be the target of text data editing. Or, Ne may be a numerical value that indirectly specifies the original - data - corresponding text data 134 to be edited. For example, if the original - data - corresponding text data set 133 contains 100 pieces of original - data - corresponding text data 134 and Ne = 0.2, the original - data - corresponding text data 134 with index 20 (which is the product of 100 and 0.2) may be the target of text data editing. (Note that when the above product is not an integer, the decimal part may be appropriately processed such as rounding down, rounding up, or rounding off.) Thus, when all of the original - data - corresponding text data 134 included in the original - data - corresponding text data set 133 are selected as the text data for policy decision, the editing operation included in the policy 162 can appropriately specify a specific original - data - corresponding text data 134 to be the target of text data editing. Alternatively, instead of the above-described method of "designating the original data-corresponding text data to be edited", Policy 162 may include one editing operation for each of the original data-corresponding text data 134 included in the original data-corresponding text data set 133. The editing operation may include flag information indicating whether the original data-corresponding text data 134 associated with the editing operation is the data to be edited. When the text data for policy determination is extracted by random sampling from the original data-corresponding text data 134 included in the original data-corresponding text data set 133, the "designation of the original data-corresponding text data to be edited" may include designators K, k, and Nf. The combination of K and k designates a combination of attribute values (including class values). (If the number of attribute values included in the combination is three or more, the "designation of the original data-corresponding text data to be edited" may further include a designator for designating the attribute values.) Nf designates the number of original data-corresponding text data 134 having the designated combination of attribute values (including class values). For example, when K = B, k = c, and Nf = 2 are designated, two pieces of original data-corresponding text data 134 having the combination of attribute values B and c are targeted for text data editing. In this way, even when random sampling is utilized and Policy 162 is determined, the original data-corresponding text data 134 to be edited can be appropriately selected from the original data-corresponding text data set 133.

[0032] The information on "the number of generated text data generated from one piece of original data-corresponding text data" designates the number Ng of text data after text data editing (generated text data 142) generated by text data editing based on one piece of original data-corresponding text data 134 that is the target of text data editing. For example, when Ng = 2 and (Nd = 1 and) the original data-corresponding text data 134 to be edited is "a black car running on a white road", the text data after text data editing (generated text data 142) may be two, namely, "a green car running on a white road" and "a blue car running on a white road". (In this example, the text data "a black car running on a white road" which was the text data before text data editing does not have to be included in the generated text data set 141.) Also, for example, when Ng = 0 and the original data-corresponding text data 134 to be edited is "a black car running on a white road", the original data-corresponding text data 134 of "a black car running on a white road" existing in the original data-corresponding text data set 133 is made not to exist in the generated text data set 141, and also no alternative generated text data 142 may exist. Since the "number of generated text data generated from one piece of original data-corresponding text data" can be designated by an editing operation, in the text data after text data editing (generated text data 142), it is possible to adjust the absolute number of generated text data 142 (generated data 144) for each subset (based on the combination of attribute values (including class values)) and correct the imbalance in the number of generated text data 142 (generated data 144) between subsets. Also, since it is possible to adjust to reduce the above absolute number, it is possible to reduce the number of learning data (generated data 144) for training the discrimination model 165 by machine learning within the range allowable as the accuracy of the discrimination model 165. When the number of learning data is reduced, the resources (memory area, computing resources, computing time) required for training the model can also be reduced. When the "number of generated text data generated from one original data-corresponding text data" or the "number of generated data generated from one generated text data" described later is specified by an editing operation, the absolute number of generated text data 142 (generated data 144) for each subset (based on the combination of attribute values (including class values)) can be adjusted, and the imbalance in the number of generated text data 142 (generated data 144) between subsets can be corrected efficiently. Therefore, the machine learning training of the discrimination model 165 using the generated data 144 as learning data can be performed efficiently. As the training of the discrimination model 165 becomes efficient, the training of the policy decision model 161 also becomes efficient.

[0033] The information on the "number of generated data generated from one generated text data" specifies the number Nh of generated data generated by the original data conversion unit 108 from one generated text data 142 generated by text data editing indicated by the editing operation. For example, when Nh = 3 and the text data after text data editing (generated text data 142) is "a white dog running on a black road", the original data conversion unit 108 may generate three pieces of image data (generated data 144) indicating "a white dog running on a black road". Also, for example, when Nh = 0 and the text data after text data editing (generated text data 142) is "a white dog running on a black road", the original data conversion unit 108 does not have to generate image data (generated data 144) indicating "a white dog running on a black road". The effect of being able to specify the "number of generated text data generated from one original data-corresponding text data" by an editing operation is the same as the effect of being able to specify the "number of generated text data generated from one original data-corresponding text data" already described. Further, in one editing operation, only one of "the number of generated text data generated from one original data-corresponding text data" and "the number of generated data generated from one generated text data" may be specified, or both may be specified. Also, when neither is specified, it may be implicitly assumed that Ng = 1 and Nh = 1 are specified.

[0034] Further, in step 505 of FIG. 5, when the vector data set 139 derived from the original data-corresponding text data set 133 is input to the policy decision model 161, the word / part-of-speech data group 135 (or the vector data group obtained by vector-converting each of the word / part-of-speech data included in the word / part-of-speech data group 135) may also be input. By inputting the information of the word / part-of-speech data group 135 to the policy decision model 161, it can be expected that the validity of the content of the policy 162 will be improved.

[0035] In step 506 (policy execution step) of FIG. 5, the policy execution unit 106 generates the generated text data set 141 by editing the original data-corresponding text data set 133 based on the policy 162 (one or more editing operations included therein). The policy execution unit 106 has already been described with reference to FIG. 1. Additionally, the policy execution unit 106 may also use the word / part-of-speech data group 135. The word / part-of-speech data group 135 can be a source of words after substitution, for example, when performing "substitution" as text data editing.

[0036] Steps 501 to 506 of FIG. 5 (and step 509 of FIG. 5 and step 601 of FIG. 6) described above are executed in the same manner when training the policy decision model 161 by machine learning and when operating the trained policy decision model 161. (The flowchart of the process when operating the trained policy decision model 161 is shown in FIG. 10. And steps 1001 to 1006 of FIG. 10 have the same content as steps 501 to 506 of FIG. 5.) On the other hand, step 507 in FIG. 5, step 508 in FIG. 5, and steps 602 to 607 in FIG. 6, which will be described hereinafter, are executed to train the policy decision model 161 by machine learning.

[0037] In step 507 of FIG. 5, the text data loss calculation unit 107 inputs each piece of generation text data 142 included in the generation text data set 141 (after vector conversion and vectorization if necessary) into the text data evaluation model 163. The text data evaluation model 163 may output an evaluation value indicating the validity of the content of the input text data (for example, the appropriateness of the text data, the naturalness of the text data). For example, high evaluation values may be given to text data such as "a person walking", and low evaluation values may be given to text data such as "a car walking". Note that the text data evaluation model 163 may be a model trained by machine learning using, for example, a group of text data created by a person as learning data. In step 508 of FIG. 5, the text data loss calculation unit 107 calculates the text data loss 164 based on the evaluation value output by the text data evaluation model 163 for each piece of generation text data 142. Generally, as the evaluation value is higher (the validity as text data is higher), the text data loss 164 becomes lower. By calculating the text data loss 164 and reflecting the text data loss 164 in the policy loss 168 described later, in the training of the policy decision model 161, it is possible to reflect the aspect of the validity of the content of the generation text data 142 based on the policy 162. If the content of the generation text data 142 is appropriate, the generated data 144 is also likely to be appropriate. An example of the text data loss 164 calculated by the text data loss calculation unit 107 is shown at the upper part of FIG. 9. The text data loss calculation unit 107 may calculate the overall text data loss 911, the subset text data loss 912 for each subset (based on the combination of attribute values (including class values)), the maximum value 913 in the set of subset text data losses 912, and the dispersion value 914. The overall text data loss 911 may reflect the evaluation values of all the generation text data 142 included in the generation text data set 141. It is reasonable that when the evaluation value is high (the validity as text data is high), the text data loss is low. Therefore, the overall text data loss 911 (J_text_all) may be calculated by, for example, the following linear combination formula for the evaluation value (S_text(i)) of each of the generation text data 142 (let the index be i). J_text_all := Σ(-S_text(i)) (Σ is the sum for all i) Note that the formula for obtaining the overall text data loss 911 may be arbitrary as long as it reflects the evaluation values of all the generation text data 142 included in the generation text data set 141 and is such that when the evaluation value is high (the validity as text data is high), the text data loss is low. When the overall text data loss 911 (J_text_all) is calculated and the overall text data loss 911 (J_text_all) is reflected in the policy loss 168 (J_GT) described later, it is possible to reflect the aspect of the validity of the content of the entire set of generation text data 142 based on the policy 162 in the training of the policy decision model 161. That is, the policy decision model 161 can be trained so that the content of the generation text data 142 included in the generation text data set 141 is overall appropriate. The partial set text data loss 912 is calculated for each partial set based on the combination of attribute values (including class values). The partial set text data loss 912 for a certain partial set may reflect the evaluation value of the generation text data 142 included in the certain partial set. Here, it is reasonable that as the evaluation value is high (the validity as text data is high), the text data loss becomes low. Therefore, the partial set text data loss 912 (J_text_part(j)) of a certain partial set (with index j) may be calculated, for example, by the following linear combination formula for the evaluation values (S_text(i)) of each of the generation text data 142 (with index i) included in the certain partial set (j). J_text_part(j) := Σ(-S_text(i)) (Σ is the sum for i included in partial set j) Note that not limited to the above linear combination formula, as long as it reflects the evaluation value of the generation text data 142 included in the partial set (j) of the generation text data set 141, and is such that as the evaluation value is high (the validity as text data is high), the text data loss becomes low, the formula for obtaining the partial set text data loss 912 can be arbitrary. The maximum value 913 (max_J_text_part) of the set of partial set text data losses 912 (J_text_part(j)) may indicate the maximum value among the partial set text data losses 912 (J_text_part(j)) for each partial set (j). By calculating the maximum value 913 (max_J_text_part) and reflecting the maximum value 913 (max_J_text_part) in the policy loss 168 (J_GT) described later, the policy decision model 161 can be trained to prevent the generation text data 142 having a specific combination of attribute values (including class values) (included in a specific partial set) from becoming text data with particularly low content validity. The variance value 914 (var_J_text_part) of the set of partial set text data losses 912 (J_text_part(j)) may indicate the variance value of the set of partial set text data losses 912 (J_text_part(j)) for each partial set (j). By calculating the variance value 914 (var_J_text_part) and reflecting the variance value 914 (var_J_text_part) in the policy loss 168 (J_GT) described later, the policy decision model 161 can be trained so as to prevent the degree of validity of the content of the generation text data 142 for each combination (partial set) of attribute values (including class values) from varying excessively.

[0038] In step 509 (raw data conversion step) of FIG. 5, the raw data conversion unit 108 converts each of the generation text data 142 included in the generation text data set 141 into each of the generation data 144 in the raw data format. The details of the raw data conversion unit 108 have already been described with reference to FIG. 1. Additionally, if the information on "the number of generation data generated from one generation text data" is included in the editing operation included in the policy 162, the information on "the number of generation data generated from one generation text data" is also input to the raw data conversion unit 108. For the generation text data 142 to which the information on "the number of generation data generated from one generation text data" input is applied, the raw data conversion unit 108 may generate the number of generation data 144 in the raw data format indicated by the information on "the number of generation data generated from one generation text data". The set of generation data 144 forms the generation data set 143. After step 509, the control transitions to step 601 of FIG. 6.

[0039] In step 601 of FIG. 6, the discrimination model training unit 109 trains the discrimination model 165 by machine learning that uses each of the generated data 144 included in the generated data set 143 as learning data. For example, when supervised learning is realized, the discrimination model training unit 109 uses the generated data 144 as an input to the discrimination model 165, and sets the output from the discrimination model 165 as the class 191 of the discrimination result (which may be information indicating the probability for each class in some cases). The discrimination model training unit 109 compares the information of the class 191 of the discrimination result, which is the output of the discrimination model 165, with the teacher information about the generated data 144 (label information indicating the correct class of the generated data 144), and obtains the loss related to discrimination (in one piece of generated data 144). The discrimination model training unit 109 performs training of the discrimination model 165 (for example, updates the model parameters of the discrimination model 165) so as to eliminate the loss (group) related to discrimination in one or more pieces of generated data 144. Here, the training result of the discrimination model 165 (for example, the content of the update of the model parameters of the discrimination model 165) may be recorded in the non-volatile recording medium (recording device) 303 (or the storage device 302) as the discrimination model information 165i shown in FIG. 3. When the training of the discrimination model 165 by machine learning using each of the generated data 144 included in the generated data set 143 is completed and the discrimination model 165 is trained (shown as the trained discrimination model 166 in FIG. 4), the control transitions to step 602 to examine the discrimination accuracy of the trained discrimination model 166.

[0040] In step 602 of FIG. 6, the original data loss calculation unit 110 uses each of the original data-form verification data 146 included in the verification data set 145 as an input to the trained discrimination model 166, and sets the output from the trained discrimination model 166 as the class 191 of the discrimination result for each of the verification data 146 (which may be information indicating the probability for each class in some cases). In step 603 of FIG. 6, the original data loss calculation unit 110 calculates the original data loss 167 based on the information of the class 191 of the discrimination result for each of the verification data 146, which is the output of the trained discrimination model 166. For example, the original data loss calculation unit 110 compares the information of the class 191 of the discrimination result with the teacher information (label information indicating the correct class of the verification data 146) for each of the verification data 146, obtains the loss related to discrimination (in one verification data 146), and calculates the original data loss 167 based on the loss related to the discrimination. When the original data loss 167 is calculated and the original data loss 167 is reflected in the policy loss 168 described later, it is possible to reflect the validity of the discrimination accuracy of the discrimination model 165 trained using the generated data 144 based on the policy 162 in the training of the policy decision model 161. An example of the original data loss 167 calculated by the original data loss calculation unit 110 is shown in the middle of FIG. 9. The original data loss calculation unit 110 may calculate the overall original data loss 921, the subset original data loss 922 for each subset (based on the combination of attribute values (including class values)), the maximum value 923 and the variance value 924 in the set of subset original data losses 922. The overall original data loss 921 may reflect the loss related to discrimination in all the verification data 146 included in the verification data set 145. The overall original data loss 921 (J_ori_all) may be calculated, for example, by the following linear combination formula for the loss related to discrimination (L_ori(m)) for each of the verification data 146 (with the index being m). J_ori_all := Σ(L_ori(m)) (Σ is the sum for all m) Note that the formula for obtaining the overall original data loss 921 may be arbitrary as long as it reflects the loss related to discrimination of all the verification data 146 included in the verification data set 145, not limited to the above linear combination formula. The overall original data loss 921 (J_ori_all) is calculated, and by reflecting the overall original data loss 921 (J_ori_all) in the policy loss 168 (J_GT) described later, the policy decision model 161 can be trained so that the loss regarding the discrimination of the entire set of verification data 146 included in the verification data set 145 is overall reduced. The subset original data loss 922 is calculated for each subset based on the combination of attribute values (including class values). The subset original data loss 922 for a certain subset may reflect the loss regarding the discrimination of the verification data 146 included in the certain subset. The subset original data loss 922 (J_ori_part(n)) for a certain subset (with index n) may be calculated, for example, by the following linear combination formula for the loss (L_ori(m)) regarding the discrimination of each of the verification data 146 (with index m) included in the certain subset (n). J_ori_part(n) := Σ(L_ori(m)) (Σ is the sum for m included in subset n) Note that the formula for obtaining the subset original data loss 922 may be arbitrary as long as it reflects the loss regarding the discrimination of the verification data 146 included in the subset (n) of the verification data set 145, not limited to the above linear combination formula. The maximum value 923 (max_J_ori_part) of the set of subset original data losses 922 (J_ori_part(n)) may be the maximum value among the subset original data losses 922 (J_ori_part(n)) for each subset (n). By calculating the maximum value 923 (max_J_ori_part) and reflecting the maximum value 923 (max_J_ori_part) in the policy loss 168 (J_GT) described later, the policy decision model 161 can be trained to prevent the class discrimination accuracy from specifically decreasing for the verification data 146 having a specific combination of attribute values (including class values) (included in a specific subset). The variance value 924 (var_J_ori_part) of the set of partial set element data losses 922 (J_ori_part(n)) may be the variance value of the set of partial set element data losses 922 (J_ori_part(n)) for each partial set (n). By calculating the variance value 924 (var_J_ori_part) and reflecting the variance value 924 (var_J_ori_part) in the policy loss 168 (J_GT) described later, the policy decision model 161 can be trained so as to prevent the degree of loss related to the discrimination of the verification data 146 from varying excessively for each combination (partial set) of attribute values (including class values).

[0041] In step 604 of FIG. 6, the policy decision model training unit 111 calculates the policy loss 168. The policy loss 168 is used to train the policy decision model 161. FIG. 9 shows an aspect 900 of an information source and a calculation method for calculating a policy loss 168. As shown in FIG. 9, the policy loss 168 (J_GT) includes an overall text data loss 911 (J_text_all), a subset text data loss 912 (J_text_part(j)) for each subset, a maximum value 913 (max_J_text_part) of the set of subset text data losses 912, a variance value 914 (var_J_text_part) of the set of subset text data losses 912, an overall original data loss 921 (J_ori_all), a subset original data loss 922 (J_ori_part(n)) for each subset, a maximum value 923 (max_J_ori_part) of the set of subset original data losses 922, a variance value 924 (var_J_ori_part) of the set of subset original data losses 922, the number of generated data 931 (Num_all) which is the number of generated data 144 included in the generated data set 143, and may be calculated by a function 951 that takes as arguments some or all of the prohibited policy penalties 941 (P_abnormal). The policy decision model training unit 111 may calculate the policy loss 168 (J_GT), for example, according to the following linear combination formula. In the following formula, each of t_1·t_2·t_3·t_4·t_5·t_6·t_7·t_8 is a coefficient (parameter) and can be adjusted arbitrarily. J_GT := t_1×J_text_all + t_2×max_J_text_part + t_3×var_J_text_part + t_4×J_ori_all + t_5×max_J_ori_part + t_6×var_J_ori_part + t_7×Num_all + t_8×P_abnormal Regarding the effects of reflecting the terms J_text_all·max_J_text_part·var_J_text_part·J_ori_all·max_J_ori_part·var_J_ori_part among the terms included in the above linear combination formula on the policy loss 168 (J_GT), they have already been described. By reflecting the item of the number of generated data 931 (Num_all) in the policy loss 168 (J_GT), the policy decision model 161 can be trained to reduce the number of generated data 931 (if the adverse impact on other items is minor). If the number of generated data 931 can be reduced, the resources (memory area, computing resources, computing time) for training the discrimination model 165 by machine learning can be lowered. The prohibited policy penalty 941 (P_abnormal) indicates a penalty when the policy 162 contains undesirable content. For example, when any of the editing operations included in the policy 162 generated by the policy decision unit 105 indicates "deletion" as the "type of text data editing" and contains content indicating deletion of a word indicating a class from the original data corresponding text data 134, the prohibited policy penalty 941 (P_abnormal) may be set to a high value. By reflecting the item of the prohibited policy penalty 941 (P_abnormal) in the policy loss 168 (J_GT), the policy decision model 161 can be trained so as not to easily output a policy 162 containing undesirable content. (Note that, for example, when adopting an aspect that can prevent the format itself of the policy 162 output from the policy decision model 161 from containing undesirable content, when adopting an aspect of providing the policy correction unit already described, or when the generated text data 142 reflecting the undesirable content of the policy 162 is deleted or corrected, this item of the prohibited policy penalty 941 (P_abnormal) may not be included.) In step 605 of FIG. 6, the policy decision model training unit 111 performs machine learning training of the policy decision model 161 based on the policy loss 168 (J_GT). The policy decision model training unit 111 performs training of the policy decision model 161 (for example, updating of the model parameters of the policy decision model 161) to eliminate the policy loss 168 (J_GT). Here, the training result of the policy decision model 161 (for example, the content of the update of the model parameters of the policy decision model 161) may be recorded in the non-volatile recording medium (recording device) 303 (or the storage device 302) as the policy decision model information 161i shown in FIG. 3. After a series of processes from step 501 in FIG. 5 to step 604 in FIG. 6 are executed multiple times (e.g., a predetermined number of times) using different original data sets 131, step 605 in FIG. 6 may be executed.

[0042] In step 606 of FIG. 6, the policy decision model training unit 111 determines whether to perform further training on the policy decision model 161. For example, if there is no longer a set of learning data in the original data format that can be the original data set 131 and has not yet been used for training the policy decision model 161, the determination result in step 606 may be negative. Also, if the absolute value of the policy loss 168 becomes sufficiently small and it can be said that the training of the policy decision model 161 has been sufficiently performed, the determination result in step 606 may be negative. If the determination result in step 606 is positive, the control transitions to step 607. If the determination result in step 606 is negative, the series of processes for training the policy decision model 161 is completed. In step 607 of FIG. 6, the policy decision model training unit 111 determines a new original data set 131 for use in training the policy decision model 161. The policy decision model training unit 111 may, for example, determine as the new original data set 131 a set of learning data in the original data format that can be the original data set 131 and has not yet been used for training the policy decision model 161. After step 607, the control returns to step 501 in FIG. 5.

[0043] As described above, the training of the policy decision model 161 is performed using the degree of validity of the content of the generated text data 142 generated based on the policy 162 output by the policy decision model 161 and the loss related to discrimination in the discrimination model 165 trained by the generated data 144 generated based on the policy 162 output by the policy decision model 161. In this way, even when it is difficult to assume the loss of the policy decision model 161 itself (excluding the number of generated data 931 (Num_all) and the prohibited policy penalty 941 (P_abnormal)), it is possible to perform the training of the policy decision model 161.

[0044] 3. Processing during operation of the decision-making model (Fig. 10) Fig. 10 shows a flowchart of the processing executed by the system 101 when operating the trained decision-making model 161. As already explained, among steps 1001 to 1009 included in the flowchart of Fig. 10, steps 1001 to 1008 have the same content as any of the steps in the flowcharts of Figs. 5 and 6. Specifically, each of steps 1001, 1002, 1003, 1004, 1005, 1006, 1007, and 1008 has the same content as each of steps 501, 502, 503, 504, 505, 506, 509, and 601, respectively. By executing steps 1001 to 1008, the training of the discrimination model 165 is completed. In step 1009 of Fig. 10, the discrimination model operation unit 112 inputs the data 148 in the original data format (also referred to as production data. Unlike data in other original data formats, it is normal for production data not to be accompanied by teacher information (label information indicating the correct class)).) into the trained discrimination model 165 (or 166). The discrimination model operation unit 112 obtains the class 191 of the discrimination result for the production data as the output of the discrimination model 165 (or 166). Note that among the processing steps included in the flowchart of Fig. 10, steps 1001 to 1007 are a series of processes for generating the generated data set 143 based on the original data set 131. When the system 101 performs up to the generation of the generated data set 143 when operating the trained decision-making model 161, the system 101 does not necessarily execute steps 1008 and 1009.

[0045] 3.3. Processing for display output control With the generation of the generated data set 143 based on the original data set 131, there can be several display output controls that the system 101 can perform. In the following, while mainly referring to FIGS. 11 to 13, processing related to the display output control of the number of data included in the original data set 131 and the generated data set 143 will be described. Also, while mainly referring to FIGS. 14 to 17, processing related to the display output control of the discrimination accuracy of the discrimination model trained using the original data set 131 and the discrimination accuracy of the discrimination model trained using the generated data set 143 will be described. In addition, while mainly referring to FIGS. 18 to 20, processing related to the display output control of individual information (such as text data) corresponding to the original data included in the original data set 131 and the generated data included in the generated data set 143 will be described. Note that in the following, the display output control of the number of data, the discrimination accuracy, and the individual information (such as text data) will be described separately, but these display output controls may be performed substantially simultaneously. Also, the display or output of the number of data, the discrimination accuracy, and the individual information (such as text data) may be performed integrally. Among the display output shown in FIG. 13 described later, the display output shown in FIG. 17, and the display output shown in FIG. 20, some or all of them may be simultaneously displayed or output by the display or output device 307. These displays or outputs improve the explanatory nature regarding the content of the generation process of the generated data set 143 based on the original data set 131. Also, these displays or outputs improve the interpretability regarding the behavior of the discrimination model 165 and the policy decision model 161.

[0046] 3.1. Processing of the number of data display output control (FIGS. 11 to 13) FIG. 11 shows a functional configuration 1100 (and the information to be handled) related to the display output control of the number of data included in the original data set 131 and the generated data set 143. FIG. 12 shows a flowchart 1200 of the processing related to the display output control of the number of data. FIG. 13 shows the display or output 1300 of the number of data. In the following, while appropriately referring to FIGS. 11 and 13, it will be described in the order of the processing steps of the flowchart in FIG. 12. In step 1201 of FIG. 12, the data number display output control unit 113 implemented (realized) in the system 101 acquires information on the number (n) of original data for each subset included in the original data set 131 (before the policy is applied), based on the combination of attribute values (including class values). The upper part of FIG. 11 shows the number (n) 1112 of original data for each subset 1111 when each of the original data 132 has a combination of a first attribute value (which can also be a class value) among A, B, C, and D and a second attribute value (which can also be a class value) among a, b, and c. As shown in the upper part of FIG. 11, for example, the number of original data included in the subset of the original data 132 having the first attribute value of B and the second attribute value of c is denoted as n[B, c]. The data number display output control unit 113 may acquire information on the number (n) 1112 of original data for each subset (for example, n[A, a], n[B, a] ··· n[D, c] shown in the upper part of FIG. 11) by aggregating the original data corresponding text data set 133 obtained from the original data set 131. Alternatively, when the text data conversion unit 102 generates the original data corresponding text data set 133 from the original data set 131, it may perform the aggregation process on the original data corresponding text data set 133 and record the information on the number (n) 1112 of original data for each subset in the non-volatile recording medium (recording device) 303. Then, the data number display output control unit 113 may acquire the information on the number (n) 1112 of original data for each subset. In step 1202 of FIG. 12, the data number display output control unit 113 acquires information on the number (N) of generated data 144 for each subset included in the generated data set 143 (after the policy is applied). The middle part of FIG. 11 shows the number (N) 1122 of original data for each subset 1121 when each of the generated data 144 has a combination of a first attribute value (which can also be a class value, such as any of A, B, C, and D) and a second attribute value (which can also be a class value, such as any of a, b, and c). As shown in the middle part of FIG. 11, for example, the number of original data included in the subset of the generated data 144 having a first attribute value of B and a second attribute value of c is denoted as N[B,c]. The data number display output control unit 113 can acquire information on the number (N) 1122 of generated data for each subset (for example, N[A,a], N[B,a] ··· N[D,c] shown in the middle part of FIG. 11) by aggregating the generation text data set 141 used to generate the generated data set 143. (In the case where the policy 162 specifies the number of generated data 144 generated based on a certain piece of generation text data 142, the information on that number may also be used in the aggregation process.) Alternatively, when generating the generation text data set 141, the policy execution unit 106 may perform an aggregation process on the generation text data set 141 and record the information on the number (N) 1122 of generated data for each subset in the non-volatile recording medium (recording device) 303. Then, the data number display output control unit 113 may acquire the information on the number (N) 1122 of generated data for each subset.

[0047] In step 1203 of FIG. 12, the data number display output control unit 113 controls the display or output device 307 to display or output the number of data included in the original data set 131 and the generated data set 143. As shown in the lower part of FIG. 11, the display or output device 307 performs a display or output (X) 1132 regarding the number of data (original data, generated data) for each subset (based on the combination of attribute values (including class values)). Here, the display or output (X) 1132 regarding the number of data may include content regarding the increase or decrease in the number of generated data as viewed from the number of original data. FIG. 13 shows a mode 1300 of the display or output (X) regarding the number of data (original data, generated data) for each subset. In FIG. 13, for each subset, the number (n) 1112 of original data is represented by a solid-line cylindrical graph. Also in FIG. 13, for each subset, the state of the increase or decrease in the number (N) 1122 of generated data as viewed from the number (n) 1112 of original data is represented by a dotted-line cylindrical graph. FIG. 13 shows that for the combination of the first attribute value being C and the second attribute value being c, for example, the number (N[C, c]) of generated data is larger than the number (n[C, c]) of original data. FIG. 13 shows that for the combination of the first attribute value being A and the second attribute value being a, for example, the number (N[A, a]) of generated data is smaller than the number (n[A, a]) of original data. FIG. 13 shows that for the combination of the first attribute value being D and the second attribute value being c, for example, the number (n[D, c]) of original data was zero, but the number (N[D, c]) of generated data became non-zero. FIG. 13 shows that for the combination of the first attribute value being A and the second attribute value being c, for example, the number (n[A, c]) of original data and the number (N[A, c]) of generated data are the same. As shown in FIG. 13, as long as the display or output includes content related to the increase or decrease in the number of generated data as viewed from the number of original data, the person viewing the display or output can understand the manner of editing (such as data expansion and data generation) performed by the system 101 on the original data set 131. Furthermore, the person viewing the display or output can understand the degree of improvement in the absolute number of data for each subset and the degree of correction of the imbalance in the number of data between subsets. Note that the information on the number of data for each subset can be obtained by text-based aggregation processing on the original data corresponding text data set 133 and the text data set 141 for generation. Therefore, it is easy to implement the aggregation processing.

[0048] Processing for discriminant accuracy display output control of 3 of 3 (FIGS. 14 to 17) FIGS. 14 and 15 show the functional configurations 1400 and 1500 (and the information handled) related to the display output control of the discriminant accuracy of the discriminant model 165. FIG. 16 shows the flowchart 1600 of the processing related to the display output control of the discriminant accuracy. FIG. 17 shows the display or output 1700 of the discriminant accuracy. Hereinafter, the description will be given in the order of the processing steps of the flowchart in FIG. 16 while appropriately referring to FIGS. 14, 15, and 17. Note that the discriminant accuracy of the discriminant model 165 (or 166) handled in FIGS. 14 to 17 is the discriminant accuracy of the discriminant model 165 (or 166) trained using the original data 132 included in the original data set 131 (before the policy is applied) as learning data, and the discriminant accuracy of the discriminant model 165 (or 166) trained using the generated data 144 included in the generated data set 143 (after the policy is applied) as learning data. In the flowchart of FIG. 16, steps 1601 to 1604 use the original data 132 included in the original data set 131 (before the policy is applied) as learning data to train the discrimination model 165 and obtain information on the discrimination accuracy of the trained discrimination model 165 (or 166). In contrast, steps 1605 to 1608 use the generated data 144 included in the generated data set 143 (after the policy is applied) as learning data to train the discrimination model 165 and obtain information on the discrimination accuracy of the trained discrimination model 165 (or 166). That is, steps 1601 to 1604 and steps 1605 to 1608 perform the same processing as each other except for the difference in the learning data for the discrimination model 165. Hereinafter, the processing steps of the flowchart of FIG. 16 will be described in order.

[0049] In step 1601 of FIG. 16, the discrimination model training unit 109 trains the discrimination model 165 by machine learning using each of the original data 132 included in the original data set 131 (before the policy is applied) (with teacher information (label information indicating the correct class)) as learning data. The details of the training of the discrimination model 165 executed by the discrimination model training unit 109 are as already described. Note that in this step 1601, the original data 132 rather than the generated data 144 is used as learning data. In step 1602 of FIG. 16, the original data loss calculation unit 110 inputs each of the verification data 146 included in the verification data set 145 into the discrimination model 166 trained in step 1601. The original data loss calculation unit 110 obtains information on the class 191 of the discrimination result of each of the verification data 146 as the output from the trained discrimination model 166. In step 1603 of FIG. 16, the original data loss calculation unit 110 calculates a subset original data loss 922 (denoted as 922o in FIG. 14 for distinction) for each subset (based on the combination of attribute values (including class values)) of the verification data 146, based on the information of class 191 of the discrimination result obtained in step 1602 for each of the verification data 146 and the teacher information (label information indicating the correct class) for the verification data 146. The details of the process in which the original data loss calculation unit 110 calculates the subset original data loss 922 are as already described. Here, the larger the subset original data loss 922o of a certain subset, the lower the discrimination accuracy 1410 of the discrimination model 165 for the certain subset. Based on this relationship, either the original data loss calculation unit 110 or the discrimination accuracy display output control unit 114 may calculate the discrimination accuracy 1410 of the discrimination model 165 for each subset based on the subset original data loss 922o for each subset. In step 1604 of FIG. 16, the discrimination accuracy display output control unit 114 acquires information on the discrimination accuracy 1410 of the discrimination model 165 (or 166) trained using the original data 132 included in the original data set 131 (before the policy is applied) as learning data. This information on the discrimination accuracy 1410 is information for each subset. The form of this information on the discrimination accuracy 1410 is shown at the upper part of FIG. 15. The upper part of FIG. 15 shows the accuracy (p) 1512 of the discrimination model for each subset 1511 when each of the verification data 146 has a combination of a first attribute value (which can also be a class value) among A, B, C, and D and a second attribute value (which can also be a class value) among a, b, and c. As shown in the upper part of FIG. 15, for example, the accuracy of the discrimination model for the subset of the verification data 146 having the first attribute value of B and the second attribute value of c is denoted as p[B,c].

[0050] Steps 1605 to 1608 of FIG. 16 are the same as steps 1601 to 1604, except that the learning data for training the discrimination model 165 is the generated data 144 instead of the original data 132. Therefore, the description of steps 1605 to 1608 is generally omitted. However, the subset source data loss generated in step 1607 is denoted as "subset source data loss 922n" in FIG. 14. Also, the discrimination accuracy of the discrimination model for each subset handled in step 1608 is denoted as "discrimination accuracy 1420 by the discrimination model after training using the generated data set" in FIG. 14. Furthermore, in the middle part of FIG. 15, the aspect of the information of the discrimination accuracy 1420 is shown. The middle part of FIG. 15 shows the accuracy (P) 1522 of the discrimination model for each subset 1521 when each of the verification data 146 has a combination of a first attribute value (which may also be a class value, any one of A, B, C, and D) and a second attribute value (which may also be a class value, any one of a, b, and c). As shown in the middle part of FIG. 15, for example, the accuracy of the discrimination model for the subset of the verification data 146 having the first attribute value of B and the second attribute value of c is denoted as P[B,c].

[0051] In step 1609 of FIG. 16, the discrimination accuracy display output control unit 114 controls the display or output device 307 to display or output the discrimination accuracy of the discrimination model 165 (or 166) trained using the original data set 131 and the discrimination accuracy of the discrimination model 165 (or 166) trained using the generated data set 143. As shown in the lower part of FIG. 15, the display or output device 307 performs a display or output (Y) 1532 regarding the discrimination accuracy of the discrimination model for each subset (based on the combination of attribute values (including class values)). Here, the display or output (Y) 1532 regarding the discrimination accuracy may include the content regarding the increase or decrease in the discrimination accuracy of the discrimination model 165 (or 166) trained using the generated data set 143 as viewed from the discrimination accuracy of the discrimination model 165 (or 166) trained using the original data set 131. FIG. 17 shows a mode 1700 of display or output (Y) regarding the discrimination accuracy of the discrimination model 165 for each subset. In FIG. 17, for each subset, the discrimination accuracy (p) 1512 of the discrimination model trained using the original dataset 131 is represented by a solid-line cylindrical graph. Also in FIG. 17, for each subset, the increase and decrease of the discrimination accuracy (P) 1522 of the discrimination model trained using the generated dataset 143, as seen from the discrimination accuracy (p) 1512 of the discrimination model trained using the original dataset 131, is represented by a dotted-line cylindrical graph. FIG. 17 shows, for example, for the subset indicated by the combination of the first attribute value being C and the second attribute value being c, that the discrimination accuracy (P[C,c]) after training with the generated data is higher than the discrimination accuracy (p[C,c]) after training with the original data. FIG. 17 shows, for example, for the subset indicated by the combination of the first attribute value being A and the second attribute value being a, that the discrimination accuracy (P[A,a]) after training with the generated data is lower than the discrimination accuracy (p[A,a]) after training with the original data. FIG. 17 shows, for example, for the subset indicated by the combination of the first attribute value being D and the second attribute value being c, that the discrimination accuracy (p[D,c]) after training with the original data was extremely low, but the discrimination accuracy (P[D,c]) after training with the generated data became quite high and was improved. FIG. 17 shows, for example, for the subset indicated by the combination of the first attribute value being A and the second attribute value being c, that the discrimination accuracy (p[A,c]) after training with the original data and the discrimination accuracy (P[A,c]) after training with the generated data are the same. As shown in FIG. 17, as long as the display or output includes content regarding the increase and decrease of the discrimination accuracy of the discrimination model 165 (or 166) trained using the generated dataset 143 as seen from the discrimination accuracy of the discrimination model 165 (or 166) trained using the original dataset 131, the viewer of the display or output can grasp the effect of the improvement in the discrimination accuracy of the discrimination model due to the editing (such as data expansion and data generation) performed by the system 101 on the original dataset 131. Furthermore, the viewer of the display or output can grasp the degree of improvement in the discrimination accuracy of the discrimination model for each subset.

[0052] 3.3.3. Processing for Display Output Control of Individual Information (Text Data, etc.) (Figs. 18 to 20) Fig. 18 shows a functional configuration 1800 (and the information to be handled) regarding the display output control of individual information (e.g., data in the original data format, data in the text data format, attribute values (including class values)) related to the original data 132 included in the original data set 131 and the generated data 144 included in the generated data set 143. (In Fig. 18, those with alphabets (A and B) enclosed in parentheses are connected to each other. Fig. 19 shows a flowchart 1900 of the process regarding the display output control of the individual information. Fig. 20 shows the display or output 2000 of the individual information. Hereinafter, the explanation will be given in the order of the processing steps of the flowchart in Fig. 19 while appropriately referring to Figs. 18 and 20. In Fig. 18, the original data display output control unit 115, the text data display output control unit 116, or the attribute value display output control unit 117 is individually shown. Hereinafter, these display output control units may be collectively referred to as the "individual information display output control unit". As an implementation aspect in the system 101, the original data display output control unit 115, the text data display output control unit 116, or the attribute value display output control unit 117 may be separate functional units or substantially integrated functional units. In step 1901 of FIG. 19, the individual information display output control unit acquires information indicating changes, additions, and deletions between the original data set 131 (before the policy is applied) and the generated data set 143 (after the policy is applied). In the example of FIG. 18, when the generated data set 143 is generated from the original data set 131, the original data (1) is changed to the generated data (α), the generated data (β) is added using the information of the original data (3), and the original data (4) is deleted. To acquire information indicating such changes, additions, and deletions, for example, the individual information display output control unit acquires information for managing data included in the original data set 131 or the original data corresponding text data set 133, and acquires information for managing data included in the generated data set 143 or the generation text data set 141, and then may compare the information of both. Alternatively, one or both of the policy execution unit 106 and the original data conversion unit 108 may create information indicating changes, additions, and deletions when generating the generation text data set 141 and the generated data set 143, and record the information on the non-volatile recording medium (recording device) 303. Then, the individual information display output control unit may acquire the recorded information. In step 1902 of FIG. 19, the individual information display output control unit acquires data in the original data format, data in the text data format, and attribute values (including class values) regarding changes, additions, and deletions. Here, the data in the original data format is acquired from the original data set 131 or the generated data set 143. The data in the text data format is acquired from the original data corresponding text data set 133 or the generation text data set 141. The attribute values (including class values) may be acquired from the original data corresponding text data set 133 or the generation text data set 141. Alternatively, when the text data conversion unit 102 generates the original data corresponding text data set 133 or the policy execution unit 106 generates the generation text data set 141, information on attribute values (including class values) may be extracted from the generated text data, and the extracted information may be recorded on the non-volatile recording medium (recording device) 303. Then, the individual information display output control unit may acquire the recorded information. In the example of FIG. 18, individual information (data in the original data format, data in the text data format, attribute values (including class values)) regarding each of the original data (1), original data (4), generated data (α), and generated data (β) is acquired by the individual information display output control unit.

[0053] In step 1903 of FIG. 19, the individual information display output control unit controls the display or output device 307 to display or output individual information (data in the original data format, data in the text data format, attribute values (including class values)) regarding changes, additions, and deletions. FIG. 20 shows mode 2000 when displaying or outputting individual information (data in the original data format, data in the text data format, attribute values (including class values)) regarding changes, additions, and deletions when the generated data set 143 is generated from the original data set 131. In FIG. 20, when one original data 132 is changed to one generated data 144, the original data 132 may be denoted as "deletion" and the generated data 144 may be denoted as "addition". (Therefore, in the example of FIG. 20, there is no notation of "change".) As shown in FIG. 20, for each record (one row shown in FIG. 20), there are columns of "editing type", "original data format", "text data format", and "attribute values, etc.". (Note that some of these columns may not be displayed or output.) The "editing type" column indicates whether the record is a deletion or an addition. In the record of deletion, the "original data format" column indicates the deleted original data 132. On the other hand, in the record of addition, the "original data format" column indicates the added generated data 144. When the original data format is an image format, the image data itself may be displayed or output. When the original data format is an audio format, a hyperlink may be displayed, and when the hyperlink is clicked, audio data may flow from the speaker. The column of "text data format" in the record of the deletion part indicates the original data corresponding text data 134 corresponding to the deleted original data 132. The column of "text data format" in the record of the addition part indicates the generation text data 142 corresponding to the added generated data 144. The column of "attribute value etc." indicates the attribute value (including class value) of the deleted original data 132 (original data corresponding text data 134), or the attribute value (including class value) of the added generated data 144 (generation text data 142). Devises may be made to make it easier to distinguish between the records of the deletion part and the records of the addition part. In the example of FIG. 20, the records of the deletion part are shaded. Or, other devises are possible, such as using different background colors or different character fonts between the records of the deletion part and the records of the addition part. Also, the table (table) showing the records of the deletion part and the table (table) showing the addition part may be separated and displayed or output.

[0054] As described above, when the generated data set 143 is generated from the original data set 131, in addition to or instead of displaying or outputting data in the original data format (for example, image data or audio data) for the changed part, added part, and deleted part, information on data in the text data format or attribute values (including class values) is displayed or output, so that the person viewing the display or output can easily grasp the situation regarding the attribute values (including class values) etc. before and after editing (data expansion and data generation).

[0055] In addition, although the display or output regarding the changed part, added part, and deleted part is shown above, not limited to the changed part, added part, and deleted part, for the original data 132 included in the original data set 131 and the generated data 144 included in the generated data set 143, the display or output of the columns of "original data format", "text data format", "attribute value etc." as described above may be performed. In that case, the situation regarding the attribute values (including class values) in the original dataset 131 and the generated dataset 143 becomes easier to grasp. For example, it becomes easier for the person viewing the display or output to grasp that there is insufficient data regarding a specific attribute value or the like.

[0056] 4. Others (Modification Examples) The present disclosure is not limited to the above-described embodiments and includes various modification examples. Part of the configuration and processing of the embodiment may be replaced with the configuration and processing of other conceivable embodiments. The configuration and processing of other conceivable embodiments may be added to the configuration and processing of the embodiment. For example, in the present disclosure, there may be modification examples of the following embodiments.

[0057] (A) Policy decision-making without using a policy decision model In the above-described embodiment, the policy 162 for generating the generation text dataset 141 by editing the original data corresponding text data 134 was generated by the policy decision model 161 handled by the policy decision unit 105. And the policy decision model 161 was trained by machine learning. In the modification example, the policy decision unit 105 may generate the policy 162 without using the policy decision model 161. For example, the policy decision unit 105 may perform statistical processing on the set of the original data corresponding text data 134 included in the original data corresponding text dataset 133 to grasp the absolute number situation of the original data corresponding text data 134 for each subset based on the combination of attribute values (including class values), and the imbalance situation of the number of the original data corresponding text data 134 between the subsets. And the policy decision unit 105 may determine the policy 162 based on the grasped situation. Modification examples as described above eliminate the need for model construction of the policy decision model 161 and training of the policy decision model 161, so resources (human resources, financial resources, time resources) for developing the system 101 and resources (memory area, computing resources, computing time) for operating the system 101 can be saved.

[0058] (B) Discriminant model trained by unsupervised learning In the above embodiment, the discriminant model 165 was trained by supervised learning. In a modification, the discriminant model 165 may be trained by unsupervised learning. For example, the discriminant model 165 may be trained by a clustering method with a set of generation data 144 in the original data format input thereto. The original data loss 167 at this time may be based on a loss function suitable for the clustering method. Modifications as described above can broaden the types of the discriminant model 165 used in the present disclosure. Also, since it is unsupervised learning, there is no need to associate teacher information (label information indicating the correct class) with the original data 132 or the generation data 144.

[0059] (C) Variation of subsets In the above embodiment, for example, when data (original data 132, generation data 144, verification data 146) has a first attribute value (which may be a class value) that is any one of A, B, C, and D and a second attribute value (which may be a class value) that is any one of a, b, and c, the subset formed by the data was defined by the combination of the first attribute value and the second attribute value. That is, the finest-grained subset that can be formed by the combination of the attribute values (including class values) that the data has was handled. In a modification, the subset formed by the data (original data 132, generation data 144, verification data 146) does not necessarily have to be the finest-grained. For example, when data (original data 132, generation data 144, verification data 146) has a first attribute value that is any one of A, B, C, and D and a second attribute value that is any one of a, b, and c, the subset may be defined based only on the first attribute value. Or, a combination of some of the finest-grained subsets that can be formed by the combination of the attribute values (including class values) that the data has may be redefined as a subset. As described above, the modification examples enable flexible definition of subsets, allowing for processing suitable for the discrimination problem (classification problem) handled by the discrimination model 165. Further, when the number of attribute values (including class values) of the data (original data 132, generated data 144, verification data 146) is large, the number of subsets with the finest granularity that can be formed by combinations of the attribute values (including class values) of the data also becomes large. However, in the modification examples, the number of subsets can be adjusted to a reasonable value.

[0060] (D) Variation of attribute values In the description of the above embodiment, it was explained as if the attribute values of the data (original data 132, generated data 144, verification data 146) were discrete values. In the modification example, there may be attribute values that are continuous values. For example, when color information is used as an attribute value, the attribute value can be regarded as a continuous value represented by the wavelength of light. In this case, a range of the attribute value may be specified, and subsets as described above may be formed. The modification examples as described above can broaden the types of attribute values that can be handled in the present disclosure.

[0061] (E) Calculating policy loss without using text data loss In the above embodiment, text data loss 164 is used in calculating the policy loss 168 for training the policy decision model 161. In the modification example, it is not necessary to use text data loss 164 in calculating the policy loss 168. Although the modification examples as described above make it difficult to utilize the validity degree of the content of the generated text data 142 for training the policy decision model 161, they can save the resources (storage area, computing resources, computing time) required for training the policy decision model 161.

[0062] (F) Not using the word / part-of-speech extraction unit (utilizing the word / part-of-speech database) In the above embodiment, a word / part-of-speech data group 135 is generated based on the original data corresponding text data 134. In a modification, instead of generating the word and part-of-speech data group 135 based on the original data-corresponding text data 134, a database of words and parts of speech may be available. The database of words and parts of speech serves the role of the word and part-of-speech data group 135. Modifications such as the above can save resources (computing resources and computing time) related to the extraction of words and parts of speech.

[0063] (G) Variations in the display or output of the number of data and discrimination accuracy In the above-described embodiment, as shown in FIGS. 13 and 17, for each subset based on the combination of attribute values (including class values), the increase or decrease in the number of data and discrimination accuracy before and after applying the strategy is displayed or output in a graph (for example, a column graph). A modification may display or output either a graph of the number of data and discrimination accuracy for each subset before applying the strategy (original data set 131) or a graph of the number of data and discrimination accuracy for each subset after applying the strategy (generated data set 143). Even in such a modification, for those who view the display or output, it has the effect of making it easier to grasp the number of learning data and the discrimination accuracy of the discrimination model 165 for each subset based on the combination of attribute values (including class values).

[0064] (H) Generation of verification data In the above-described embodiment, the generated data 144 generated by the system 101 was learning data for training the discrimination model 165 by machine learning. In a modification, a part of the generated data 144 generated by the system 101 may be used as a supplement to the verification data 146. Even if a modification initially has a verification data set 145 with an imbalance in the number of verification data for each subset based on the combination of attribute values (including class values) or an imbalance in the number of verification data between subsets, by using a part of the generated data 144 as a supplement to the verification data 146, a verification data set 145 with the imbalance resolved can be formed.

[0065] (I) Utilization of the Policy Decision Model for Generating Learning Data for Training the Text Data Discrimination Model In the above embodiment, the discrimination model 165 was for discriminating the class of the data 148 in the original data format (for example, image data or audio data). In a modified example, instead of the discrimination model 165 for the original data format, it may be a text data discrimination model for discriminating the class of data in the text data format. Further, the policy decision model 161 may be for generating a policy for text data editing when generating learning text data in the text data format for training the text data discrimination model by machine learning. In this modified example, the learning text data before text data editing is used instead of the original data-corresponding text data 134, the learning text data after text data editing is used instead of the generation-use text data 142, and the learning text data after text data editing is used as learning data for training the text data discrimination model. Also in the modified example, some or all of the learning text data (vector-converted) included in the learning text data set before text data editing is input to the policy decision model 161. Therefore, the policy decision model 161 can output a policy for text data editing for the learning text data set before text data editing, taking into account the situation of the learning text data set before text data editing (such as the absolute number of learning text data before text data editing for each subset based on the combination of attribute values (including class values), and the situation of the imbalance in the number between subsets).

[0066] The technical matters shown in each of the above-described embodiments and modified examples of the present disclosure can be appropriately combined as long as no technical contradiction occurs.

Claims

1. A system comprising: a text data conversion unit that converts each piece of original data in the original data format included in the original data set into each piece of original data corresponding text data in the text data format; an original data corresponding text data set composed of each of the original data corresponding text data, and a policy decision unit that generates a policy to be applied to the original data corresponding text data set based on a policy decision model; a policy execution unit that edits the original data corresponding text data set based on the policy to generate a generated text data set that is a set of generated text data; a system having an original data conversion unit that converts each piece of the generated text data included in the generated text data set into each piece of generated data in the original data format, wherein the generated data is learning data for training a discrimination model that discriminates classes of data in the original data format by machine learning.

2. The system according to claim 1, further comprising: a text data loss calculation unit that obtains a text data loss based on the generated text data set and a text data evaluation model; a discrimination model training unit that trains the discrimination model by machine learning using each of the generated data as learning data; an original data loss calculation unit that obtains an original data loss based on losses related to discrimination of each class of the original data format verification data included in a verification data set in the trained discrimination model by the discrimination model training unit; a system having a policy decision model training unit that determines a policy loss that is a loss of the policy decision model based on the text data loss and the original data loss, and trains the policy decision model by machine learning based on the policy loss.

3. The system according to claim 2, further comprising: The policy loss is based on any one of the overall text data loss based on all of the generation text data included in the generation text data set, the maximum value of the subset text data loss based on each subset of the generation text data included in the generation text data set, the variance value of the subset text data loss, the overall original data loss based on all of the verification data included in the verification data set, the maximum value of the subset original data loss based on each subset of the verification data included in the verification data set, or the variance value of the subset original data loss. System.

4. The system according to claim 3, wherein the subset of the generation text data or the subset of the verification data is formed based on one or more of the classes associated with each of the generation text data or the verification data, or a combination of class values, attribute values, or class values or attribute values associated with one or more attributes associated with each of the generation text data or the verification data. System.

5. The system according to claim 3, wherein the policy loss is based on a value obtained by adding together values obtained by multiplying coefficients to the overall text data loss, the maximum value of the subset text data loss, the variance value of the subset text data loss, the overall original data loss, the maximum value of the subset original data loss, and the variance value of the subset original data loss, respectively. System.

6. The system according to claim 1, wherein the policy determination unit generates the policy that does not include content for deleting words indicating the class included in the original data corresponding text data. System.

7. The system according to claim 1, wherein the policy determination unit generates a policy applied to the original data corresponding text data set based on a part or all of the original data corresponding text data included in the original data corresponding text data set and the policy determination model. System.

8. The system according to claim 2, wherein the policy loss is based on the number of the generated data generated based on the policy. System.

9. The system according to claim 1, The system, wherein the policy decision-making unit generates a policy including information specifying which of the original data corresponding text data included in the original data corresponding text data set is to be the target of text data editing performed by the policy execution unit.

10. The system according to claim 1, wherein the policy decision-making unit generates a policy including information specifying the number of generated data generated by the original data conversion unit based on the generated text data generated by the policy execution unit by text data editing.

11. The system according to claim 1, having a data number display / output control unit that controls to display or output the number of the original data or the generated data for each subset of the original data or the generated data formed based on one or more of the classes associated with each of the original data or the generated data, or class values, attribute values, or combinations of class values or attribute values related to one or more of the attributes associated with each of the original data or the generated data, wherein the data number display / output control unit controls to display or output, for each subset, content indicating an increase or decrease in the number of data in the number of generated data included in the subset after the policy is applied as viewed from the number of the original data included in the subset before the policy is applied.

12. The system according to claim 2, having a discrimination accuracy display / output control unit that controls to display or output information indicating the discrimination accuracy of the discrimination model with respect to each subset of the verification data formed based on one or more of the classes associated with each of the verification data, or class values, attribute values, or combinations of class values or attribute values related to one or more of the attributes associated with each of the verification data. The discrimination accuracy display output control unit controls to display or output, for each subset, the content indicating the increase or decrease in the discrimination accuracy of the discrimination model trained by machine learning using each of the original data included in the original data set before the policy is applied as learning data, in terms of the discrimination accuracy of the discrimination model trained by machine learning using each of the generated data included in the generated data set after the policy is applied as learning data.

13. The system according to claim 1, A system having a text data display output control unit that controls to display or output part or all of the generated text data or the original data corresponding text data corresponding to the change, addition, or deletion of the data in the original data format in the set of generated data included in the generated data set after the policy is applied, as viewed from the set of original data included in the original data set before the policy is applied.

14. A method executed by a system, A text data conversion step of converting each of the original data in the original data format included in the original data set into each of the original data corresponding text data in the text data format, A policy determination step of generating a policy to be applied to the original data corresponding text data set based on the original data corresponding text data set composed of each of the original data corresponding text data and a policy determination model, A policy execution step of editing the original data corresponding text data set based on the policy to generate a generated text data set that is a set of generated text data, A method having an original data conversion step of converting each of the generated text data included in the generated text data set into each of the generated data in the original data format, wherein the generated data is learning data for training a discrimination model that discriminates the class of the data in the original data format by machine learning.

15. A program, to the system, A text data conversion step of converting each of the original data in the original data format included in the original data set into each of the original data corresponding text data in the text data format, A metadata-corresponding text data set consisting of each of the metadata-corresponding text data, and a policy determination step of generating a policy applied to the metadata-corresponding text data set based on a policy determination model; A policy execution step of editing the metadata-corresponding text data set based on the policy to generate a generation text data set that is a set of generation text data; A metadata conversion step of converting each of the generation text data included in the generation text data set into each of the generation data in the metadata format, wherein the generation data is a program for executing the metadata conversion step, which is learning data for training a discrimination model for discriminating classes of data in the metadata format by machine learning.

Images