Information processing method, information processing apparatus, and information processing program

Abstract

To enable flexible evaluation of a model.SOLUTION: An information processing method to be executed by a computer includes: an obtaining process of inputting each of multiple pieces of input data to be subjected to inference processing to a model and obtaining multiple output values output by the model and representing inference results respectively corresponding to the multiple pieces of input data, and criterion information indicating a criterion for evaluating the model; and a processing process of selecting an evaluation data group to be used in the evaluation of the model, from among the multiple output values, by using a threshold determined on the basis of the criterion indicated by the criterion information obtained by the obtaining process and calculating an index value representing an evaluation of the model by using the selected evaluation data group.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an information processing method, an information processing device, and an information processing program. [Background technology]

[0002] In recent years, a technology has been proposed that generates models by having various models such as neural networks, such as DNNs (Deep Neural Networks), learn the features of training data. The trained models are then used for various inference processes such as prediction and classification. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-168042 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the above-mentioned technology, it is difficult to say that the evaluation of the trained model is sufficiently considered. For example, in the above-mentioned technology, the trained model is re-trained without being evaluated, and depending on the application of the model, the result of the re-training may result in a lower accuracy than the original model. Therefore, when training a model, it is desirable to perform flexible evaluation according to the application of the model. Thus, the above-mentioned technology has room for improvement in terms of the evaluation of the trained model.

[0005] The present application has been made in consideration of the above, and has an object to provide an information processing method, an information processing device, and an information processing program that enable flexible evaluation of a model. [Means for solving the problem]

[0006] The information processing method according to the present application is an information processing method executed by a computer, and includes an acquisition step of inputting each of a plurality of input data to be subjected to inference processing into a model, and acquiring a plurality of output values ​​indicating inference results corresponding to each of the plurality of input data output by the model, and reference information indicating a criterion for evaluating the model; and a processing step of selecting, from the plurality of output values, a group of evaluation data to be used for evaluating the model, using a threshold value determined based on the criterion indicated by the reference information acquired by the acquisition step, and calculating an index value indicating the evaluation of the model, using the selected group of evaluation data. Effect of the Invention

[0007] According to one aspect of the embodiment, an effect is achieved in that it is possible to enable flexible evaluation of the model. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating an example of information processing according to an embodiment. [Diagram 2] FIG. 1 is a diagram illustrating an example of a configuration of an information processing system according to an embodiment. [Diagram 3] FIG. 1 is a diagram illustrating an example of a configuration of an information processing device according to an embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of a model information storage unit according to the embodiment. [Diagram 5] 10 is a flowchart illustrating an example of information processing according to the embodiment. [Figure 6] FIG. 13 is a diagram showing an example of an experimental result. [Figure 7] FIG. 13 is a diagram showing an example of an experimental result. [Figure 8] FIG. 2 illustrates an example of a hardware configuration. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, the information processing method, the information processing device, and the information processing program according to the present application will be described in detail with reference to the drawings. Note that the information processing method, the information processing device, and the information processing program according to the present application are not limited to the embodiments. In addition, the same parts in the following embodiments are given the same reference numerals, and the duplicated description will be omitted.

[0010] (Embodiment) 1. Embodiment An example of information processing according to the embodiment will be described with reference to FIG. 1. FIG. 1 is a diagram showing an example of information processing according to the embodiment. For example, an information processing device 100 (see FIG. 3) executes the information processing shown in FIG. 1. In the following, a model (also called a "violation posting determination model") used to determine whether or not a posting (text information) such as a question and answer on the Internet constitutes a violation will be described as an example of a model to be evaluated. Note that the violation posting determination model is merely an example of a model to be evaluated in the information processing described below, and is not limited to the violation posting determination model. Any model can be adopted as the model to be evaluated as long as it is a model to which the information processing described below can be applied.

[0011] [1-1. Example of information processing] From here, an example of information processing executed by the information processing device 100 will be described with reference to FIG. 1. First, prior to the description of the processing shown below, a brief description of the premise will be given. For example, the violating post determination model to be evaluated in the example shown in FIG. 1 is a model that outputs an output value (also called a "score") indicating the possibility (degree) that character information such as a post corresponds to a violation when the character information is input. For example, the violating post determination model is a model that outputs a score (value) between 0 and 1, and the higher the possibility that the input character information corresponds to a violation when the character information is input, the higher the score (value) that is output. Note that the fact that a post corresponds to a violation here may be, for example, a violation of the guidelines of the service (Internet posting service, etc.) to which the post is posted.

[0012] 1, the information processing device 100 acquires a data group DT11 including a plurality of output values ​​(scores) corresponding to each of a plurality of pieces of character information output by a violation posting judgment model to which each of a plurality of pieces of character information subject to violation judgment (inference processing) has been input. For example, the information processing device 100 may input each of the pieces of character information subject to violation judgment to the violation posting judgment model, and generate a data group DT11 including a plurality of output values ​​(scores) indicating an inference result corresponding to each of the pieces of character information output by the violation posting judgment model.

[0013] The information processing device 100 may acquire (receive) a data group DT11 including a plurality of output values ​​(scores) indicating inference results corresponding to each of the plurality of pieces of character information output by the violating post determination model from another computer (external device) such as the information providing device 50. The plurality of pieces of character information input to the violating post determination model may be any character information, and may include character information posted in the past, or may include character information generated (prepared) for evaluation calculation. In other words, the character information (input data) input to the violating post determination model is not limited to information that has actually been posted, and any information may be used.

[0014] 1, the information processing device 100 calculates an evaluation of a violating posting determination model using a plurality of output values ​​(scores) as shown in a data group DT11. Note that, like a data group DT14 described later, each output value (score) included in the data group DT11 in FIG. 1 is associated with a label (correct answer information) indicating whether or not character information corresponding to the output value corresponds to a violation, but is not shown in the figure.

[0015] First, the information processing device 100 sorts the multiple output values ​​included in the data group DT11 in descending order of value (step S1). In Fig. 1, the information processing device 100 arranges (sorts) the multiple output values ​​included in the data group DT11 in descending order of value, thereby generating a data group DT12 in which the multiple output values ​​are arranged in descending order of value.

[0016] Then, the information processing device 100 determines the threshold value based on the criteria for evaluating the model (step S2). In Fig. 1, the information processing device 100 determines the threshold value PT1 based on the criteria indicating the number of pieces of character information that can be confirmed when a predetermined person confirms the pieces of character information, as shown in the data group DT13. For example, the information processing device 100 acquires criteria information indicating the number of pieces of character information that can be confirmed when a person in charge of handling violating posts visually checks whether or not the pieces of character information correspond to a violation, and determines the threshold value PT1 based on the acquired criteria information.

[0017] For example, when the number of pieces of character information that can be confirmed when a predetermined person confirms them is the top X% (X is an arbitrary value such as 5 or 10) of the data group DT13 (data group DT12), the information processing device 100 determines the threshold value PT1 to be a value that is equal to or greater than the top X% of the data (score) in the data group DT13. For example, when the number of pieces of character information that can be confirmed when a predetermined person confirms them is the top 8% of the data group DT13, the information processing device 100 determines the threshold value PT1 to be a value (for example, 0.7) that is equal to or greater than the top 8% of the data (score) in the data group DT13.

[0018] The number of verifiable items indicated by the reference information is not limited to a value indicating a percentage of a data set that can be confirmed, such as a top few percent (e.g., 5% or 10%), and may be a specific number. In this case, for example, when the number of verifiable items is 1,000 when a plurality of pieces of character information are confirmed by a predetermined number of personnel, the information processing device 100 may determine, as the threshold value, a value that is equal to or greater than the value of the data (score) of the top 1,000 items of data in the data group DT13.

[0019] Then, the information processing device 100 sets all output values ​​that are equal to or greater than the threshold value among the plurality of output values ​​after sorting to a first value greater than the threshold value, and sets all output values ​​that are less than the threshold value among the plurality of output values ​​after sorting to a second value less than the threshold value (step S3). In Fig. 1, as shown in the data group DT14, the information processing device 100 sets all output values ​​that are equal to or greater than the threshold value PT1 among the data (scores) of the data group DT13 to a first value greater than the threshold value PT1, and sets all output values ​​that are less than the threshold value PT1 to a second value less than the threshold value PT1.

[0020] For example, the information processing device 100 sets all output values ​​of the data (scores) of the data group DT13 that are equal to or greater than a threshold PT1 (e.g., 0.7) to a first value "1" that is greater than the threshold PT1, and sets all output values ​​that are less than the threshold PT1 (e.g., 0.7) to a second value "0" that is less than the threshold PT1. That is, the information processing device 100 sets the scores of the data groups that are equal to or greater than the top X% of the data group DT13 to a first value "1," and sets the scores of the data groups that are less than the top X% of the data group DT13 to a second value "0."

[0021] As a result, the information processing device 100 generates a data group DT14 in which all output values ​​equal to or greater than the threshold value PT1 (e.g., 0.7) are set to a first value "1" greater than the threshold value PT1, and all output values ​​less than the threshold value PT1 (e.g., 0.7) are set to a second value "0" less than the threshold value PT1. Note that the above-mentioned first and second values ​​are merely examples, and the first value is not limited to "1" and may be any value as long as it is greater than the threshold value, and the second value is not limited to "0" and may be any value as long as it is less than the threshold value.

[0022] Then, the information processing device 100 calculates an index value indicating the evaluation of the model using the evaluation data group set to the first value (step S4). In Fig. 1, the information processing device 100 selects data whose score is set to the first value "1" from the data group DT14 as the evaluation data group TG1. Then, the information processing device 100 calculates an index value indicating the evaluation of the violating posting detection model using the determined evaluation data group TG1.

[0023] 1, the information processing device 100 calculates an index value VL1 indicating the evaluation of the violating post determination model, as shown in new evaluation index information NM1. For example, the information processing device 100 calculates an index value indicating the evaluation of the violating post determination model using information on labels (correct answer information) associated with each score of the evaluation data group TG1, which is a data group that is equal to or greater than a threshold value PT1, i.e., equal to or greater than the top X%, of the data group DT14. For example, character information corresponding to a score associated with a label "1" indicates that the character information (post) corresponds to a violation, and character information corresponding to a score associated with a label "0" indicates that the character information (post) does not correspond to a violation.

[0024] For example, the information processing device 100 calculates a value indicating the proportion of data corresponding to a violation included in the evaluation data group TG1 as an index value VL1 indicating the evaluation of the violating posting determination model. In Fig. 1, the information processing device 100 calculates the index value VL1 indicating the evaluation of the violating posting determination model based on the proportion of data associated with the label "1" indicating a violation in the evaluation data group TG1. For example, the information processing device 100 calculates the index value VL1 by dividing the number of data associated with the label "1" among the data (scores) included in the evaluation data group TG1 by the total number of data (scores) included in the evaluation data group TG1.

[0025] As described above, the information processing device 100 calculates the evaluation of the model using only the data corresponding to the number that can be confirmed when a predetermined number of people check the data group. This allows the information processing device 100 to calculate the evaluation of the model using only the data corresponding to the number that can be targeted by humans. Therefore, the information processing device 100 can calculate the evaluation according to the state of use of the model, rather than simply using the entire data to calculate the evaluation, and therefore allows flexible evaluation of the model.

[0026] In addition, any model to which the above-mentioned information processing can be applied can be adopted as the model to be evaluated, not limited to the violating post judgment model, and the model to be evaluated may be various models such as a violating product judgment model, a violating image judgment model, etc.

[0027] Furthermore, the information processing system 1 may use the evaluation of the model calculated by the information processing device 100 for various processes. The information processing system 1 may provide information indicating the evaluation of the model calculated by the information processing device 100 to a service provider (such as a user) that provides a service using the model. In this case, the information processing system 1 transmits information indicating the evaluation of the model calculated by the information processing device 100 to a computer (such as the terminal device 10) used by the service provider that provides a service using the model. Furthermore, for example, the information processing system 1 may perform a process related to optimization of the model using the evaluation of the model calculated by the information processing device 100. For example, the information processing system 1 may repeat an update process for updating the model (parameters, etc.) using the evaluation of the model calculated by the information processing device 100.

[0028] For example, the information processing system 1 may evaluate multiple models, calculate the evaluation of each model, select a model with a high evaluation from among the multiple models, and perform an update process to update the selected model to improve the accuracy of the model. For example, the information processing system 1 may learn multiple models by selecting a model with the highest calculated evaluation from among the multiple models and performing an update process to update the selected model. Then, the information processing system 1 may select a model with the highest calculated evaluation from among the multiple learned models and perform an update process.

[0029] In this way, the information processing system 1 may repeatedly update (improve) the model using the evaluation of the model calculated by the information processing device 100. For example, the update process for updating the model may be performed by any device of the information processing system 1. Note that the above-mentioned process is merely an example, and the information processing system 1 may perform the process related to optimization of the model in any manner as long as it uses the evaluation of the model calculated by the information processing device 100.

[0030] For example, the information processing device 100 may perform an update process to update the model. Also, for example, the information providing device 50 may perform an update process to update the model. In this case, the information processing device 100 may transmit information indicating the calculated evaluation of the model to the information providing device 50, and perform an update process to update the model based on the evaluation of the model received by the information providing device 50 from the information processing device 100. Then, the information processing device 100 may receive information on the model after the update process from the information providing device 50, and calculate the evaluation of the model using the received information. For example, the information processing device 100 may receive information on the model after the update process and a data group generated using the model from the information providing device 50, and calculate the evaluation of the model using the received information.

[0031] As described above, the information processing device 100 may use information on the calculated evaluation of the model (e.g., index value) to improve the quality of various information processing, services, and the like. A specific example of this point is described below. Below, a case where the calculated evaluation information of the model is used to improve the quality of an Internet posting service (also simply called a "posting service") is described as an example. Note that the process of applying the calculated evaluation information of the model is not limited to the process described below, and may be used to improve the quality of various information processing, services, and the like.

[0032] For example, calculating an index value indicating the evaluation of a model using only data corresponding to the number of posts that can be confirmed by personnel who check violating posts at the posting service out of a data group in the posting service leads to obtaining the most appropriate model for that posting service. That is, calculating an index value of a model by the above-mentioned process enables the information processing device 100 to use, for example, the deletion rate of violating posts, which is a KPI (Key Performance Indicator) pursued by the posting service, as an evaluation index of the model, and to optimize the model, such as by selecting a model.

[0033] In other words, a high index value of a model calculated by the information processing device 100 indicates a high deletion rate of violating posts by that model and indicates that the model is highly effective for the posting service side. In other words, when multiple models are targeted, the model with the highest index value calculated by the information processing device 100 among the multiple models is the model with the highest deletion rate of violating posts and the model most desired by the posting service side. Therefore, by optimizing the model, such as selecting a model as described above, using the index value of the model calculated by the information processing device 100, it becomes possible to appropriately generate the model most desired by the posting service side.

[0034] It is also possible that the number of people checking violating posts on the posting service side will change. In this way, if the number of people checking violating posts on the posting service side changes, the threshold value (e.g., the value of X in the top X%) used when calculating the model evaluation will change in the above-mentioned process. In other words, in the above-mentioned process, the model selected will change depending on the threshold value that changes depending on the number of people checking violating posts on the posting service side.

[0035] In this way, in the process performed by the information processing device 100, even when the same data group is targeted, the optimal model to be selected (for example, the model with the highest index value) can be changed according to the number of people checking the violating posts on the posting service side. Therefore, even if the number of people checking the violating posts on the posting service side changes, the information processing device 100 can appropriately select the optimal model according to the changed number of people.

[0036] On the other hand, with conventional indices, even if conventional indices are combined, it is difficult to obtain the model most desired by the service side, such as the evaluation (index value, etc.) of the model calculated by the information processing device 100 by the above-mentioned processing. If Recall (recall rate) is used for a data group, it is possible to obtain the number of records whose predicted value (e.g., the output value of the model) is True (e.g., corresponding to a violating post) from a threshold value set between 0.0 and 1.0, but conversely, a search is required to calculate an appropriate threshold value from an arbitrary number of records. Therefore, with conventional indices, it is structurally difficult to perform processing (such as optimizing the model) similar to that in the case of using the evaluation of the model calculated by the information processing device 100 by the above-mentioned processing.

[0037] In response to this, the information processing device 100 can flexibly evaluate models by calculating the evaluation of the model through the above-mentioned process. Furthermore, the information processing device 100 optimizes the model, such as by selecting a model, using the calculated evaluation of the model. Even if the number of people who check violating posts on the posting service side changes, the information processing device 100 can select an optimal model according to the changed number of people, and can appropriately generate the model that the posting service side most wants.

[0038] [1-2. Configuration of information processing system] As shown in Fig. 2, the information processing system 1 includes a terminal device 10, an information providing device 50, and an information processing device 100. The terminal device 10, the information providing device 50, and the information processing device 100 are connected to each other via a predetermined network N so as to be able to communicate with each other by wire or wirelessly. Fig. 2 is a diagram showing an example of the configuration of the information processing system according to the embodiment. Note that the information processing system 1 shown in Fig. 2 may include a plurality of terminal devices 10, a plurality of information providing devices 50, and a plurality of information processing devices 100.

[0039] The terminal device 10 is an information processing device used by a user. The terminal device 10 accepts various operations by the user. In the following, the terminal device 10 may be referred to as a user. In other words, in the following, the user can also be read as the terminal device 10. The above-mentioned terminal device 10 is realized, for example, by a smartphone, a tablet terminal, a notebook PC (Personal Computer), a desktop PC, a mobile phone, a PDA (Personal Digital Assistant), or the like.

[0040] The information providing device 50 is an information processing device in which information for providing the information processing device 100 with various information used by the information processing device 100 for processing is stored. For example, the information providing device 50 is a computer that generates a model that has learned the characteristics of learning data, and is realized by, for example, a server device or a cloud system. For example, when the information providing device 50 receives a configuration file that indicates the type and behavior of the model to be generated and how to learn the characteristics of the learning data as a model generation index, the information providing device 50 automatically generates a model according to the received configuration file. Note that the information providing device 50 may learn the model using any model learning method. Also, for example, the information providing device 50 may be any of various existing services such as AutoML (Automated Machine Learning).

[0041] The information processing device 100 is a computer that executes information processing. The information processing device 100 executes a calculation process that calculates an index value that indicates an evaluation of a model. In addition, for example, the information processing device 100 is a computer that executes an index generation process that generates a generation index that is an index in the generation of a model (i.e., a recipe for a model) and a model generation process that generates a model according to the generation index, and provides the generated generation index and model, and may be realized by, for example, a server device, a cloud system, or the like.

[0042] 1-3. Configuration of information processing device Next, the configuration of the information processing device 100 according to the embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the configuration of the information processing device 100 according to the embodiment. As shown in Fig. 3, the information processing device 100 has a communication unit 110, a storage unit 120, and a control unit 130. Note that the information processing device 100 may have an input unit (e.g., a keyboard, a mouse, etc.) that accepts various operations from an administrator of the information processing device 100, and a display unit (e.g., a liquid crystal display, etc.) that displays various information.

[0043] (Communication unit 110) The communication unit 110 is realized by, for example, a network interface card (NIC) etc. The communication unit 110 is connected to a network (for example, network N in FIG. 2) by wire or wirelessly, and transmits and receives information between the terminal device 10 and the information providing device 50.

[0044] (Storage unit 120) The storage unit 120 is realized by, for example, a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. As shown in FIG. 3, the storage unit 120 according to the embodiment has an index calculation information storage unit 121 and a model information storage unit 122.

[0045] (Indicator calculation information storage unit 121) The index calculation information storage unit 121 according to the embodiment stores various information used for index calculation. For example, the index calculation information storage unit 121 stores a function or the like used for index calculation. For example, the index calculation information storage unit 121 stores reference information indicating a reference for evaluating a model. For example, the index calculation information storage unit 121 stores a reference or the like used for determining a threshold value.

[0046] The index calculation information storage unit 121 may store a plurality of output values ​​output by the model to be evaluated. The index calculation information storage unit 121 may store information in which each output value output by the model to be evaluated is associated with a label (correct answer information) indicating whether or not the input data corresponding to the output value is a violation.

[0047] The index calculation information storage unit 121 is not limited to the above, and may store various kinds of information depending on the purpose.

[0048] (Model information storage unit 122) The model information storage unit 122 according to the embodiment stores information about a (machine learning) model. For example, the model information storage unit 122 stores information (model data) about a trained model (model) trained (generated) by a learning process. FIG. 4 is a diagram illustrating an example of a model information storage unit according to the embodiment. In the example shown in FIG. 4, the model information storage unit 122 includes items such as "model ID," "purpose," and "model data."

[0049] "Model ID" indicates identification information for identifying a model. "Use" indicates the use of the corresponding model. "Model Data" indicates the data of the model. Figure 4 shows an example in which conceptual information such as "MDT1" is stored in "Model Data", but in reality, various information constituting the model is included, such as information on the model configuration (network configuration) and information on parameters. For example, "Model Data" includes information including the nodes in each layer of the network, the functions employed by each node, the connection relationships between the nodes, and the connection coefficients set for the connections between the nodes.

[0050] In FIG. 4, the model (model M1) identified by the model ID "M1" has an application of "violation judgment", and for example, model M1 indicates that it is a model used to judge (estimate) whether the contents of input data such as input text (character information) correspond to a violation. For example, model M1 is a model that outputs, when text (character information) is input, a score (value) indicating the degree of possibility that the contents of the text correspond to a violation as output data. Also, it indicates that the model data of model M1 is model data MDT1.

[0051] The model information storage unit 122 may store various information according to the purpose, not limited to the above. The model such as the model M1 may be expected to be used as a program module that is a part of artificial intelligence software. The model such as the model M1 may be a program.

[0052] (Control unit 130) Returning to the explanation of Fig. 3, the control unit 130 is a controller, and is realized, for example, by a central processing unit (CPU), a micro processing unit (MPU), a graphics processing unit (GPU), or the like, by executing various programs (corresponding to an example of an information processing program) stored in a storage device inside the information processing device 100 using a RAM as a working area. The control unit 130 is also a controller, and is realized, for example, by an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0053] 3, control unit 130 has an acquisition unit 131, a determination unit 132, a processing unit 133, and a provision unit 134, and realizes or executes functions and actions of information processing described below. Note that the internal configuration of control unit 130 is not limited to the configuration shown in FIG. 3, and may be other configurations as long as they perform information processing described below.

[0054] (Acquisition part 131) The acquisition unit 131 acquires various information. For example, the acquisition unit 131 acquires various information from the storage unit 120. For example, the acquisition unit 131 acquires various information from the index calculation information storage unit 121, the model information storage unit 122, etc. In addition, the acquisition unit 131 acquires various information from an external information processing device. The acquisition unit 131 acquires various information from the terminal device 10 and the information providing device 50.

[0055] The acquisition unit 131 inputs each of a plurality of input data to be subjected to inference processing to the model, and acquires a plurality of output values ​​indicating inference results corresponding to each of the plurality of input data output by the model, and reference information indicating a reference for evaluating the model. The acquisition unit 131 inputs each of a plurality of input data, which is character information to be judged as to whether or not it corresponds to a violation, to the model, and acquires a plurality of output values ​​indicating judgment results corresponding to each of the plurality of input data output by the model.

[0056] The acquisition unit 131 inputs each of a plurality of input data, which is information posted on the Internet, to the model, and acquires a plurality of output values ​​indicating a determination result corresponding to each of the plurality of input data output by the model. The acquisition unit 131 acquires a plurality of output values ​​output by the model that output a higher value the more likely it is that the input character information corresponds to a violation. The acquisition unit 131 acquires reference information indicating the number of pieces of input data that can be confirmed when the plurality of input data is confirmed by a predetermined person.

[0057] (Decision unit 132) The determination unit 132 executes a determination process for determining various pieces of information. The determination unit 132 stores the determined information in the storage unit 120. The determination unit 132 executes the determination process based on the various pieces of information stored in the storage unit 120. For example, the determination unit 132 executes the determination process based on the various pieces of information received from an external information processing device. For example, the determination unit 132 executes the determination process based on the various pieces of information acquired by the acquisition unit 131.

[0058] The determination unit 132 determines the threshold value based on the criteria indicated by the criteria information. The determination unit 132 determines the threshold value using the criteria information stored in the index calculation information storage unit 121.

[0059] (Processing section 133) The processing unit 133 executes a generation process for generating various information. The processing unit 133 stores the generated information in the storage unit 120. For example, the processing unit 133 executes the generation process based on various information acquired by the acquisition unit 131. The processing unit 133 executes the generation process based on various information stored in the storage unit 120. For example, the processing unit 133 executes the generation process based on various information received from an external information processing device. The processing unit 133 executes the generation process based on various information determined by the determination unit 132.

[0060] The processing unit 133 executes a calculation process to calculate various pieces of information. The processing unit 133 stores the calculated information in the storage unit 120. For example, the processing unit 133 executes the calculation process based on various pieces of information acquired by the acquisition unit 131. The processing unit 133 executes the calculation process based on various pieces of information stored in the storage unit 120. For example, the processing unit 133 executes the calculation process based on various pieces of information received from an external information processing device. The processing unit 133 executes the calculation process based on various pieces of information determined by the determination unit 132. The processing unit 133 executes the calculation process using a threshold determined by the determination unit 132.

[0061] The processing unit 133 uses a threshold value determined based on the criteria indicated by the reference information acquired by the acquisition unit 131 to select, from the multiple output values, a group of evaluation data to be used for evaluating the model, and calculates an index value indicating the evaluation of the model using the selected group of evaluation data. The processing unit 133 selects, from the multiple output values, output values ​​that are equal to or greater than a threshold value as an evaluation data group, and calculates a value indicating the proportion of data that falls under the violation included in the evaluation data group as an index value indicating the evaluation of the model.

[0062] The processing unit 133 uses a threshold value to select a numerical evaluation data group from the multiple output values, and calculates a value indicating the proportion of data corresponding to a violation included in the numerical evaluation data group as an index value indicating the evaluation of the model. The processing unit 133 sorts the multiple output values ​​in descending order of value. The processing unit 133 sets all output values ​​that are equal to or greater than the threshold value among the multiple output values ​​after sorting to a first value that is greater than the threshold value. The processing unit 133 sets all output values ​​that are less than the threshold value among the multiple output values ​​after sorting to a second value that is less than the threshold value. The processing unit 133 calculates an index value indicating the evaluation of the model using the evaluation data group set to the first value.

[0063] The processing unit 133 executes a model optimization process using the index value. The processing unit 133 executes a model optimization process using a plurality of index values ​​calculated for each of a plurality of models. The processing unit 133 executes a model optimization process by selecting a model having the highest index value from among the plurality of models. For example, the processing unit 133 may perform a process related to model optimization using the calculated model evaluation. For example, the processing unit 133 may repeat an update process for updating the model using the calculated model evaluation.

[0064] For example, the processing unit 133 performs model optimization processing by calculating the evaluation of each model with respect to a plurality of models as evaluation targets, selecting a model with a high evaluation from among the plurality of models, and updating the selected model. For example, the processing unit 133 performs an update processing to select a model with the highest calculated evaluation from among the plurality of models, and updates the selected model to generate a plurality of models, and selects a model with the highest index value from among the plurality of generated models. For example, the processing unit 133 performs model optimization processing by selecting a model with the highest calculated evaluation from among the plurality of learned models.

[0065] (Provider 134) The providing unit 134 executes a providing process for providing various kinds of information. The providing unit 134 executes the providing process based on various kinds of information stored in the storage unit 120. For example, the providing unit 134 executes the providing process based on various kinds of information received from an external information processing device. For example, the providing unit 134 transmits various kinds of information to the terminal device 10 and the information providing device 50.

[0066] For example, the providing unit 134 executes the providing process based on various information acquired by the acquiring unit 131. The providing unit 134 executes the providing process based on various information determined by the determining unit 132. The providing unit 134 executes the providing process based on various information generated by the processing unit 133. The providing unit 134 executes the providing process based on various information calculated by the processing unit 133. For example, the providing unit 134 transmits information indicating the index calculated by the processing unit 133 to an external device such as the terminal device 10. For example, the providing unit 134 transmits information indicating the processing result by the processing unit 133 to the information providing device 50.

[0067] [1-4. Information processing flow] Next, a procedure of information processing by the information processing system 1 according to the embodiment will be described with reference to Fig. 5. Fig. 5 is a flowchart showing an example of information processing according to the embodiment.

[0068] 5, the information processing device 100 inputs each of a plurality of input data to be subjected to inference processing to a model, and obtains a plurality of output values ​​indicating inference results corresponding to each of the plurality of input data output by the model (step S101). The information processing device 100 obtains reference information indicating a reference for evaluating the model (step S102).

[0069] The information processing device 100 selects, from among the multiple output values, a group of evaluation data to be used for evaluating the model, using a threshold determined based on the criteria indicated by the reference information (step S103).The information processing device 100 calculates an index value indicating the evaluation of the model, using the selected group of evaluation data (step S104).

[0070] [1-5. Example of Experimental Results] From here, an example of experimental results when a model trained using the evaluation calculated by the above-mentioned process is used will be described with reference to Figs. 6 and 7. Figs. 6 and 7 are diagrams showing an example of experimental results. For example, the experimental results shown in Figs. 6 and 7 show an example of experimental results when a model trained (generated) by performing the above-mentioned process for optimizing the model is used. As a specific example, Fig. 6 shows an example of experimental results showing the relationship between the new evaluation index and the standard. Also, Fig. 7 shows an example of experimental results showing the relationship between the deletion rate of violating posts and the number of confirmations.

[0071] First, an example of an experimental result regarding the relationship between the new evaluation index and the standard will be described with reference to Fig. 6. In graph RS1 in Fig. 6, the horizontal axis indicates the standard value (standard value), and the vertical axis indicates the new evaluation index. The standard in Fig. 6 indicates the proportion of data used to calculate the new evaluation index among multiple output values ​​output by the model. For example, a standard of "0.06" corresponds to a case where the top 6% of data from the largest values ​​among multiple output values ​​output by the model is used to calculate the new evaluation index.

[0072] Line LN11 in FIG. 6 shows the experimental results for a model (question model) when the question is the target. Line LN12 in FIG. 6 shows the experimental results for a model (answer model) when the answer is the target. For the question model, when the standard is set to "0.06", the new evaluation index is "0.893", and when the standard is set to "0.08", the new evaluation index is "0.929". For the answer model, when the standard is set to "0.02", the new evaluation index is "0.602".

[0073] Thus, in graph RS1 of Figure 6, as shown by line LN11, 92.9% of all violating posts are included in the top 8% of question model scores. In this way, it was shown that violating posts can be appropriately extracted when using a model trained based on the evaluation calculated by the above-mentioned process.

[0074] Next, an example of an experimental result regarding the relationship between the deletion rate of violating posts and the number of confirmations will be described with reference to Fig. 7. Graph RS2 in Fig. 7 shows the number of posts corresponding to questions on the horizontal axis and the deletion rate of violating posts on the vertical axis. Fig. 7 shows a case where the number of confirmations is the visual review number of "9000", which is the number of posts that can be visually reviewed by a person in one day, and Fig. 7 shows a case where the number of questions posted per day is 50,000, and the number of answers posted per day is 100,000.

[0075] The lines LN21 and LN22 in Fig. 7 show the deletion ratios of violating posts for each of the questions and answers when the confirmation number "9000" is allocated to each of the questions and answers. The line LN23 in Fig. 7 shows the total deletion ratios of violating posts for each of the questions and answers.

[0076] For example, as shown by line LN23, if the confirmation number "9000" is allocated to questions at "2000" and to answers at "7000", the deletion rate of violating posts is "0.835". Also, in graph RS3 of FIG. 7, if the confirmation number "9000" is allocated to questions at "3000" and to answers at "6000", the deletion rate of violating posts is "0.837". Also, in graph RS3 of FIG. 7, if the confirmation number "9000" is allocated to questions at "4000" and to answers at "5000", the deletion rate of violating posts is "0.827". Thus, graph RS3 of FIG. 7 shows that the deletion rate of violating posts was the largest when the confirmation number "9000" was allocated to questions at "3000" and to answers at "6000".

[0077] For example, in the previous model, the maximum deletion rate of violating posts when the number of confirmations was 9,000 was 0.531, whereas in the model trained (generated) by performing the model optimization process described above, the deletion rate of violating posts when the number of confirmations was 9,000 could be improved to a maximum of 0.837, achieving an improvement in accuracy of +0.306 (+30.6%).

[0078] 2. Modifications An example of the information processing has been described above. However, the embodiment is not limited to this. Below, a modified example of the providing processing will be described.

[0079] [2-1. Equipment configuration] In the above embodiment, an example has been described in which the information processing system 1 has the information processing device 100 that evaluates a model and the information providing device 50 that generates a model, but the embodiment is not limited to this. For example, the information processing device 100 may have the functions of the information providing device 50. Furthermore, the functions performed by the information processing device 100 may be included in the terminal device 10. In such a case, the terminal device 10 will evaluate the model and automatically generate the model using the information providing device 50.

[0080] [2-2.Other] In addition, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by a known method. In addition, the information including the processing procedures, specific names, various data and parameters shown in the above documents and drawings can be changed arbitrarily unless otherwise specified. For example, the various information shown in each drawing is not limited to the illustrated information.

[0081] In addition, each component of each device shown in the figure is a functional concept, and does not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads, usage conditions, etc.

[0082] Furthermore, the above-described embodiments can be appropriately combined as long as the processing contents are not contradictory.

[0083] [2-3. Program] The information processing device 100 according to the embodiment described above is realized by a computer 1000 having a configuration as shown in Fig. 8. Fig. 8 is a diagram showing an example of a hardware configuration. The computer 1000 is connected to an output device 1010 and an input device 1020, and has a configuration in which a calculation device 1030, a primary storage device 1040, a secondary storage device 1050, an output IF (Interface) 1060, an input IF 1070, and a network IF 1080 are connected by a bus 1090.

[0084] The arithmetic device 1030 operates based on programs stored in the primary storage device 1040 and the secondary storage device 1050 and programs read from the input device 1020, and executes various processes. The primary storage device 1040 is a memory device, such as a RAM, that temporarily stores data used by the arithmetic device 1030 for various calculations. The secondary storage device 1050 is a storage device in which data used by the arithmetic device 1030 for various calculations and various databases are registered, and is realized by a ROM (Read Only Memory), HDD, flash memory, or the like.

[0085] The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various types of information, such as a monitor or a printer, and is realized by a connector conforming to a standard such as USB (Universal Serial Bus), DVI (Digital Visual Interface), or HDMI (registered trademark) (High Definition Multimedia Interface). The input IF 1070 is an interface for receiving information from various input devices 1020, such as a mouse, keyboard, and scanner, and is realized by a USB, for example.

[0086] The input device 1020 may be a device that reads information from, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), or a PD (Phase change rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory. The input device 1020 may also be an external storage medium such as a USB memory.

[0087] The network IF 1080 receives data from other devices via the network N and sends it to the arithmetic device 1030, and also transmits data generated by the arithmetic device 1030 to other devices via the network N.

[0088] The arithmetic unit 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic unit 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040 and executes the loaded program.

[0089] For example, when the computer 1000 functions as the information processing device 100 , the arithmetic unit 1030 of the computer 1000 executes a program loaded onto the primary storage device 1040 to realize the functions of the control unit 130 .

[0090] 3. Effects As described above, the information processing device 100 has an acquisition unit (the acquisition unit 131 in the embodiment) that inputs each of a plurality of input data to be subjected to the inference process to a model, acquires a plurality of output values ​​indicating an inference result corresponding to each of the plurality of input data output by the model, and reference information indicating a reference for evaluating the model, and a processing unit (the processing unit 133 in the embodiment) that selects an evaluation data group to be used for evaluating the model from among the plurality of output values ​​using a threshold determined based on the reference indicated by the reference information acquired by the acquisition unit, and calculates an index value indicating the evaluation of the model using the selected evaluation data group. This allows the information processing device 100 to evaluate the model according to the evaluation reference, thereby enabling flexible evaluation of the model.

[0091] The acquisition unit inputs each of a plurality of input data, which is character information to be judged as to whether or not it corresponds to a violation, to the model, and acquires a plurality of output values ​​indicating judgment results corresponding to each of the plurality of input data output by the model. This enables the information processing device 100 to appropriately evaluate the model that judges whether or not character information corresponds to a violation, thereby enabling flexible evaluation of the model.

[0092] The acquisition unit inputs each of a plurality of input data, which is information posted on the Internet, to the model and acquires a plurality of output values ​​indicating a determination result corresponding to each of the plurality of input data output by the model. This enables the information processing device 100 to appropriately evaluate the model that determines whether or not information posted on the Internet corresponds to a violation, thereby enabling flexible evaluation of the model.

[0093] The acquisition unit also acquires a plurality of output values ​​output by a model that outputs a higher value as the likelihood that the input character information corresponds to a violation increases. This allows the information processing device 100 to appropriately evaluate a model that outputs a higher value as the likelihood that the input character information corresponds to a violation increases, thereby enabling flexible evaluation of the model.

[0094] Furthermore, the processing unit selects, from among the multiple output values, output values ​​equal to or greater than a threshold value as an evaluation data group, and calculates a value indicating the proportion of data corresponding to violations included in the evaluation data group as an index value indicating the evaluation of the model. This allows the information processing device 100 to calculate a value indicating the proportion of data corresponding to violations included in the evaluation data group as an index value indicating the evaluation of the model, thereby enabling flexible evaluation of the model.

[0095] The acquisition unit also acquires reference information indicating the number of pieces of input data that can be confirmed when the input data are confirmed by a predetermined number of people. This allows the information processing device 100 to appropriately evaluate the model based on the number of pieces of input data that can be confirmed when the input data are confirmed by a predetermined number of people, thereby enabling flexible evaluation of the model.

[0096] The processing unit also uses a threshold value to select a set of evaluation data of the number from the multiple output values, and calculates a value indicating the ratio of data corresponding to violations included in the set of evaluation data of the number as an index value indicating the evaluation of the model. This allows the information processing device 100 to appropriately evaluate the model based on the number that can be confirmed when multiple input data are confirmed by a predetermined person, thereby enabling flexible evaluation of the model.

[0097] The processing unit also sorts the multiple output values ​​in descending order of value, sets all output values ​​that are equal to or greater than a threshold among the multiple output values ​​after sorting to a first value that is greater than the threshold, sets all output values ​​that are less than the threshold among the multiple output values ​​after sorting to a second value that is less than the threshold, and calculates an index value indicating an evaluation of the model using the evaluation data group set to the first value. This allows the information processing device 100 to calculate an index value indicating an evaluation of the model using the evaluation data group set to the first value, thereby enabling flexible evaluation of the model.

[0098] The processing unit also uses the index value to perform optimization processing of the model. This allows the information processing device 100 to appropriately evaluate the model using a dynamically changed threshold value, thereby enabling flexible evaluation of the model according to a criterion for determining the threshold value, and allowing the optimization processing of the model to be appropriately performed according to the criterion.

[0099] The processing unit also executes a model optimization process using the multiple index values ​​calculated for each of the multiple models. This allows the information processing device 100 to appropriately evaluate each of the multiple models using a dynamically changed threshold, enabling flexible evaluation of the multiple models according to a criterion for determining the threshold, and enabling the optimization process to be executed appropriately for the multiple models according to the criterion.

[0100] In addition, the processing unit selects the model with the highest index value from among the multiple models, thereby executing the optimization process for the model. As a result, the information processing device 100 can appropriately execute the optimization process for the model by selecting the model with the highest index value and executing the optimization process for the model, so that the model with the highest evaluation can be left.

[0101] Although some of the embodiments of the present application have been described in detail above with reference to the drawings, these are merely examples, and the present invention can be embodied in other forms that incorporate various modifications and improvements based on the knowledge of those skilled in the art, including the forms described in the Disclosure of the Invention section.

[0102] Furthermore, the above-mentioned "section, module, unit" can be read as "means" or "circuit", etc. For example, an acquisition section can be read as an acquisition means or an acquisition circuit. [Explanation of symbols]

[0103] 1. Information Processing Systems 10 Terminal Equipment 50 Information provision device 100 Information processing device 120 Storage section 121 Information storage unit for index calculation 122 Model information storage unit 130 Control section 131 Acquisition Department 132 Decision Section 133 Processing section 134 Provision Department

Claims

1. 1. A computer-implemented information processing method, comprising: an acquisition step of inputting each of a plurality of input data to be subjected to an inference process into a model, and acquiring a plurality of output values ​​indicating inference results corresponding to each of the plurality of input data output by the model, and reference information indicating a standard for evaluating the model; a processing step of selecting, from the plurality of output values, a group of evaluation data to be used for evaluating the model, using a threshold value determined based on a criterion indicated by the reference information acquired in the acquisition step, and calculating an index value indicating an evaluation of the model, using the selected group of evaluation data; 13. An information processing method comprising:

2. The obtaining step includes: Each of the plurality of input data, which is character information to be judged as to whether or not it corresponds to a violation, is input to the model, and the plurality of output values ​​indicating judgment results corresponding to each of the plurality of input data output by the model are obtained.

2. The information processing method according to claim 1,

3. The obtaining step includes: Each of the plurality of pieces of input data, which are information posted on the Internet, is input to the model, and the plurality of output values, which indicate a determination result corresponding to each of the plurality of pieces of input data output by the model, are obtained.

3. The information processing method according to claim 2.

4. The obtaining step includes: The model outputs a higher value as the likelihood that the input character information corresponds to the violation increases. The model outputs the higher value.

3. The information processing method according to claim 2.

5. The processing step comprises: Among the plurality of output values, output values ​​equal to or greater than the threshold value are selected as the evaluation data group, and a value indicating a ratio of data corresponding to the violation included in the evaluation data group is calculated as the index value indicating an evaluation of the model.

5. The information processing method according to claim 4.

6. The obtaining step includes: The reference information indicating the number of pieces of input data that can be confirmed when the pieces of input data are confirmed by a predetermined person is acquired.

3. The information processing method according to claim 2.

7. The processing step selects the number of evaluation data groups from the plurality of output values ​​using the threshold value, and calculates a value indicating a ratio of data corresponding to the violation included in the number of evaluation data groups as the index value indicating an evaluation of the model.

7. The information processing method according to claim 6,

8. The processing step comprises: sorting the plurality of output values ​​in descending order of value, setting all output values ​​of the plurality of output values ​​after sorting that are equal to or greater than the threshold to a first value greater than the threshold, setting all output values ​​of the plurality of output values ​​after sorting that are less than the threshold to a second value less than the threshold, and calculating the index value indicating an evaluation of the model using the evaluation data group set to the first value; 2. The information processing method according to claim 1,

9. The processing step performs an optimization process for a model using the calculated index value.

2. The information processing method according to claim 1,

10. The processing step performs an optimization process for the model using a plurality of index values ​​calculated for each of a plurality of models.

10. The information processing method according to claim 9.

11. The processing step selects a model having the highest index value from among the plurality of models, thereby performing an optimization process for the model.

11. The information processing method according to claim 10.

12. an acquisition unit that inputs each of a plurality of input data to be subjected to an inference process into a model, and acquires a plurality of output values ​​indicating inference results corresponding to each of the plurality of input data output by the model, and reference information indicating a standard for evaluating the model; a processing unit that selects, from the plurality of output values, a group of evaluation data to be used for evaluating the model, using a threshold value determined based on a criterion indicated by the reference information acquired by the acquisition unit, and calculates an index value indicating an evaluation of the model, using the selected group of evaluation data; An information processing device comprising:

13. an acquisition step of inputting each of a plurality of input data to be subjected to an inference process into a model, and acquiring a plurality of output values ​​indicating inference results corresponding to each of the plurality of input data output by the model, and reference information indicating a reference for evaluating the model; a processing procedure of selecting, from the plurality of output values, a group of evaluation data to be used for evaluating the model, using a threshold value determined based on a criterion indicated by the reference information acquired by the acquisition procedure, and calculating an index value indicating an evaluation of the model, using the selected group of evaluation data; An information processing program characterized by causing a computer to execute the above.

Images