Information processing apparatus, information processing method, storage medium, and program product

Abstract

The invention provides an information processing apparatus, an information processing method, a storage medium, and a program product. The purpose of the present invention is to analyze data of each category with high accuracy. The information processing apparatus includes: one or more memories storing instructions; and one or more processors in communication with the one or more memories. Upon execution of the instructions, the one or more processors cooperate with the one or more memories to operate including: identifying a category of the input data using a category identification model, based on goodness-of-fit values of data belonging to the Nth category with respect to analysis models for the first category to the Mth category used for analysis of data of the first category to the Mth category, the Nth category is selected from among the analysis models for the first category to the Mth category, and the Mth category is selected from among the analysis models for the first category to the Mth category. An analysis model to be applied to the input data identified as being in the Nth category by the identification is selected.

Description

Technical Field

[0001] The present disclosure relates to an information processing device, an information processing method, and a storage medium. Background Art

[0002] In recent years, technologies for extracting useful information from various data such as images, text, and time series data and analyzing the information have been discussed. In particular, by using machine learning technology, it has become possible to analyze information with high precision. Japanese Patent Application No. 2020-512631 describes a system that uses a machine learning model to identify which category unknown data belongs to, and performs an analysis task for data suitable for the identified category (hereinafter referred to as a specific category).

[0003] As described in Japanese Patent Application No. 2020-512631, in the case of performing an analysis task suitable for each specific category, it is necessary to prepare an analysis model for each specific category to perform the analysis task. At this time, in order to increase the type of the specific category, a new analysis model is required. However, there are cases where it is difficult to generate a new analysis model due to, for example, hardware resource limitations. Summary of the invention

[0004] Embodiments of the present disclosure are directed to analyzing data of various categories with high accuracy.

[0005] According to one aspect of the present disclosure, an information processing device includes: one or more memories storing instructions; and one or more processors communicating with the one or more memories. When executing the instructions, the one or more processors cooperate with the one or more memories to perform operations, the operations including: identifying the category of input data; and selecting an analysis model to be applied to the input data identified as being in the first category by the identification from the analysis models for each category based on the goodness of fit value of the data belonging to the first category relative to each of the analysis models for each category used to analyze the data of each category.

[0006] Further features of various embodiments will become apparent from the following description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a diagram showing a hardware configuration example of an information processing device.

[0008] Figure 2 is a diagram showing a functional configuration example of an information processing device.

[0009] Figure 3 is a flow chart describing the process of calculating the goodness-of-fit value for the Nth category.

[0010] Figure 4 is a flowchart describing the process of generating supervisory data for the Nth category.

[0011] Figure 5 is a flow chart describing the training process of the class recognition model.

[0012] Figure 6 is a flow chart describing the process of applying a detection model.

[0013] Figure 7 is a diagram showing a functional configuration example of an information processing device.

[0014] Figure 8 is a flowchart describing the process of generating supervisory data for the Nth category. DETAILED DESCRIPTION

[0015] Exemplary embodiments will be described below with reference to the accompanying drawings.

[0016] In the first exemplary embodiment, a system will be described which recognizes the category of an object in an input image, applies a detection model for detecting a specific part (e.g., a face) of the recognized category, and detects the specific part from the image. The process of detecting a specific part of a category from an image is an example of analysis processing. The input data may be not only an image but also text or time series data, etc., and there is no specific limitation.

[0017] Figure 1 An example of a hardware configuration of an information processing device according to the present exemplary embodiment is shown. The information processing device 1 is configured to include a central processing unit (CPU) 11, a read-only memory (ROM) 12, a random access memory (RAM) 13, an auxiliary storage device 14, an input device 15, and a display device 16. These constituent units are connected to each other via a connection bus 17. The CPU 11 controls the entire information processing device 1. The CPU 11 executes a control program stored in the ROM 12, etc. to perform the processing in each flowchart to be described below. A graphics processing unit (GPU) may be used instead of the CPU, or a GPU may be used together with the CPU.

[0018] The ROM 12 is a nonvolatile memory, and stores a control program and various parameter data. The RAM 13 is a volatile memory, and temporarily stores images, control programs, and execution results of the control programs. The auxiliary storage device 14 is a rewritable auxiliary storage device such as a hard disk or a flash memory, and stores various data used for each flowchart to be described below. For example, the auxiliary storage device 14 stores input data, training data sets, and processing results. These information are output to the RAM 13 and used by the CPU 11 to execute the control program. The input device 15 is a keyboard, a mouse, a touch panel device, etc., and inputs various instructions from the user.

[0019] The display device 16 is a monitor, and displays processing results, images, and the like.

[0020] In the present exemplary embodiment, the processing to be described below is implemented by software using the CPU 11, but part or all of the processing to be described below may be implemented by hardware. As hardware, a dedicated circuit (application-specific integrated circuit (ASIC)) or a processor (reconfigurable processor or digital signal processor (DSP)) or the like may be used. Additionally, the information processing device 1 includes a communication unit for communicating with an external device, and can acquire input data, a control program, or a training data set, etc. from the external device via the communication unit, and output the processing results, etc. to the external device via the communication unit.

[0021] Figure 2 1 is a diagram showing a functional configuration example of the information processing device 1 according to the present exemplary embodiment. The information processing device 1 is configured to implement the following functions by executing a control program stored in the ROM 12 or the like by the CPU 11. Figure 2 The information processing device 1 includes an input data acquisition unit 202 , a category identification unit 203 , and an analysis unit 206 .

[0022] The input data acquisition unit 202 acquires the image 201 as input data from an external device or the auxiliary storage device 14 .

[0023] The category recognition unit 203 recognizes the category of the image. In the present exemplary embodiment, serial numbers from the first to the Nth are assigned to the categories, and N categories (the first category, the second category, ..., the Mth category, and the Nth category (N>M)) are recognized. The category recognition unit 203 includes a category recognition model for recognizing the category of the object in the image.

[0024] The analysis unit 206 includes a detection model for detecting a region of a specific part of a category (hereinafter referred to as a specific region) from an image. In the present exemplary embodiment, the analysis unit 206 includes M detection models (a first category detection model 207, a second category detection model 208, ..., an Mth category detection model 209) for each of the M categories. On the other hand, the analysis unit 206 does not include a detection model for the Nth category (N>M).

[0025] Each detection model is a machine learning model that has been trained using a supervised data set belonging to each category. The detection model 207 for the first category is a machine learning model that has been trained using a supervised data set belonging to the first category. Similarly, the detection models for the second category to the Mth category are machine learning models that have been trained using supervised data sets belonging to each category. Each machine learning model is, for example, a model that has been trained using a multi-layer neural network based on ground truth information related to a specific area of ​​the category. As a training method, known methods such as "RetinaFace: Single-stage Dense Face Localisation in the Wild" written by JiankangDeng, Jia Guo, Evangelos Ververas, Irene Kotsia and Stefanos Zafeiriou at CVPR 2020 are adopted.

[0026] <Processing of Calculating Goodness-of-Fit Value for Nth Category>

[0027] The information processing device 1 includes a goodness-of-fit value calculation unit 214. The following describes the processing performed by the goodness-of-fit value calculation unit 214. In the present exemplary embodiment, the goodness-of-fit value calculation unit 214 uses the first category detection model to the M-th category detection model included in the analysis unit 206 and the data set 217 for calculating the goodness-of-fit value of the N-th category to generate the goodness-of-fit value calculation result 205. The generated goodness-of-fit value calculation result 205 is stored in the auxiliary storage device 14 or the like.

[0028] Figure 3 1 is a flowchart describing a process of calculating a goodness-of-fit value for the Nth category. The goodness-of-fit value of the Nth category data relative to the detection model for the mth category is an indicator for indicating whether the detection model for the mth category is suitable for detection of a specific area of ​​the Nth category. The goodness-of-fit value of the Nth category data relative to the detection model for the mth category is represented by a goodness-of-fit value A(N:m).

[0029] In the following description, each step is represented by prefixing S to the step number, and the mark of "step" is omitted.

[0030] In S301, the goodness-of-fit value calculation unit 214 obtains a data set 217 for calculating the goodness-of-fit value of the Nth category. The data set 217 for calculating the goodness-of-fit value of the Nth category includes one or more pairs of data consisting of an image of an object of the Nth category and information indicating a specific area in the image (ground truth information). The data set 217 for calculating the goodness-of-fit value of the Nth category can be stored in the auxiliary storage device 14 or the like. The goodness-of-fit value calculation unit 214 can obtain the data set 217 for calculating the goodness-of-fit value of the Nth category from the auxiliary storage device 14 or an external device.

[0031] In S302 , the goodness-of-fit value calculation unit 214 stores 1 in m indicating the class number.

[0032] In S303, the goodness-of-fit value calculation unit 214 determines whether the detection model for the m-th category exists in the analysis unit 206. In the case where it is determined that the detection model for the m-th category exists ("Yes" in S303), the goodness-of-fit value calculation unit 214 obtains the detection model for the m-th category from the analysis unit 206, and the process proceeds to S304. In the case where the goodness-of-fit value calculation unit 214 determines that the detection model for the m-th category does not exist ("No" in S303), the process proceeds to S307.

[0033] In S304, the goodness-of-fit value calculation unit 214 applies the detection model for the mth category to the data set 217 for calculating the goodness-of-fit value of the Nth category acquired in S301. Specifically, the goodness-of-fit value calculation unit 214 inputs the image in the data set 217 for calculating the goodness-of-fit value of the Nth category into the detection model for the mth category, and detects the area from the image.

[0034] In S305, the goodness-of-fit value calculation unit 214 calculates the goodness-of-fit value A(N:m) based on the comparison result between the detection area detected by the detection model for the mth category in S304 and the specific area indicated by the ground truth information in the data set 217 used to calculate the goodness-of-fit value of the Nth category. In this process, the goodness-of-fit value calculation unit 214 obtains the difference between the areas as the comparison result. It is assumed that the difference between the areas is quantitatively given by, for example, the position difference, size or area difference, or a combination thereof between the areas. When the difference between the areas is large, the specific area of ​​the Nth category cannot be detected using the detection model for the mth category. Therefore, it can be judged that the goodness-of-fit value A(N:m) is low. On the other hand, when the difference between the areas is small, the specific area of ​​the Nth category can be detected using the detection model for the mth category. Therefore, it can be judged that the goodness-of-fit value A(N:m) is high. The goodness of fit value A(N:m) can be represented by the inverse of the difference between the detection area detected by the detection model of the mth category and the specific area of ​​the Nth category, and can be represented by the size of the intersection over union (IoU) between the detection area and the specific area, etc.

[0035] In S306 , the goodness-of-fit value calculation unit 214 increments m indicating the class number by 1, and the process proceeds to S303 .

[0036] The goodness-of-fit value calculation unit 214 repeats the processing in S303 to S306, thereby calculating the goodness-of-fit value A (N: m) of each of the M detection models from the first class detection model 207 to the Mth class detection model 209.

[0037] Finally, in S307 , the goodness-of-fit value calculation unit 214 stores M goodness-of-fit values ​​A(N:m) (m=1, 2, . . . , M) as the goodness-of-fit value calculation result 205 in the auxiliary storage device 14 or the like. Thereafter, a series of processes in the flowchart ends.

[0038] <Training Process of Category Recognition Model>

[0039] The information processing device 1 includes a supervisory data set generation unit 215 and a training unit 213. The following describes the processing performed by the supervisory data set generation unit 215 and the training unit 213. The training unit 213 trains the category recognition model to be used in the category recognition unit 203. The supervisory data set used for training the category recognition model consists of a supervisory data set 211 of the first category to the Mth category and a supervisory data set 212 of the Nth category. The supervisory data set 211 is stored in the auxiliary storage device 14 or the like. The supervisory data set 212 is generated by the supervisory data set generation unit 215. The supervisory data sets 211 and 212 include a plurality of data sets each consisting of a pair of an image of a captured object and label information indicating to which category the object in the image belongs.

[0040] In the present exemplary embodiment, a case is described in which only the goodness-of-fit value A(N:1) is high among the M goodness-of-fit values ​​A(N:m). In this case, the supervised data set 212 of the Nth category can be labeled with the subcategories of the first category, and the category recognition model can be trained using the supervised data set 212 of the Nth category. In addition, the supervised data set 212 of the Nth category can be labeled with the Nth category as the M+1th category, and the category recognition model can be trained using the supervised data set 212 of the Nth category. In the former case, the training unit 213 performs training to recognize M categories from the first category to the mth category. In the latter case, the training unit 213 performs training to recognize the (M+1) categories obtained by adding the Nth category.

[0041] The following describes the processing performed by the supervisory dataset generation unit 215 in the former case described above. In the present exemplary embodiment, the supervisory dataset generation unit 215 generates the supervisory dataset 212 of the Nth category from the dataset 216 of the Nth category based on the goodness-of-fit value calculation result 205.

[0042] Figure 4 is a flowchart describing the process of generating supervisory data for the Nth category.

[0043] In S401, the supervisory dataset generation unit 215 obtains the dataset 216 of the Nth category. The dataset 216 of the Nth category may be stored in the secondary storage device 14 or the like. The supervisory dataset generation unit 215 may obtain the dataset 216 of the Nth category from the secondary storage device 14 or an external device.

[0044] In S402 , the supervisory data set generating unit 215 obtains the goodness-of-fit value A(N:m) of the data of the Nth category with respect to the detection model for the mth category from the goodness-of-fit value calculation result 205 .

[0045] As described above, the case where only the goodness-of-fit value A(N:1) is acquired in S402 is described assuming that only the goodness-of-fit value A(N:1) is high among the M goodness-of-fit values ​​A(N:m).

[0046] In S403, the supervisory data set generation unit 215 determines whether the goodness of fit value A(N:1) is above a predetermined threshold value. When the supervisory data set generation unit 215 determines that the goodness of fit value A(N:1) is above the threshold value ("Yes" in S403), the process proceeds to S404. When the supervisory data set generation unit 215 determines that the goodness of fit value A(N:1) is less than the threshold value ("No" in S403), the process proceeds to S405.

[0047] In S404, the supervisory data set generation unit 215 labels the data set 216 of the Nth category with the subcategories of the first category, and generates a data set as the supervisory data set 212 of the Nth category. Through this process, it becomes possible to train the category recognition model so that the Nth category approximates the first category. The generated supervisory data set 212 is stored in the auxiliary storage device 14 or the like. Thereafter, the process in this flowchart ends.

[0048] In S405, the supervisory data set generation unit 215 labels the data set 216 of the Nth category with a category different from the first category, and generates a data set as the supervisory data set 212 of the Nth category. Through this process, it becomes possible to train the category recognition model so that the Nth category is away from the first category. The generated supervisory data set 212 is stored in the auxiliary storage device 14 or the like. Thereafter, the process in this flowchart ends.

[0049] Figure 5 The training unit 213 uses the supervisory data set 211 of the first to M-th categories prepared in advance and the supervisory data set 212 of the N-th category generated by the supervisory data set generating unit 215 to train the class identification model.

[0050] In S501 , the training unit 213 obtains the supervision data set 211 of the first category to the Mth category.

[0051] In S502 , the training unit 213 obtains the supervision data set 212 of the Nth category.

[0052] In S503, the training unit 213 obtains a class recognition model to be used in the class recognition unit 203. In the present exemplary embodiment, the class recognition model is a machine learning model for recognizing the class of an object in an image. It is assumed that the machine learning model is a multi-layer neural network. However, the machine learning model is not limited to the multi-layer neural network, and known machine learning models such as random forest and AdaBoost can be used.

[0053] In S504 , the training unit 213 acquires mini batch images as input images from the supervised dataset acquired in S501 and S502 .

[0054] In S505 , the training unit 213 performs inference processing on the image acquired in S504 using the class identification model acquired in S503 .

[0055] In S506 , the training unit 213 acquires label information corresponding to the image acquired in S504 .

[0056] In S507 , the training unit 213 calculates a loss value (loss function).

[0057] In S508 , the training unit 213 calculates a gradient by applying a back propagation method to the loss value calculated in S507 , and obtains an update amount of the weight of the class recognition model.

[0058] In S509, the training unit 213 updates the weight of the class recognition model. Specifically, only a known learning method for a multi-layer neural network needs to be applied, and a detailed description is omitted.

[0059] In S510 , the training unit 213 outputs the class identification model whose weights have been updated to the class identification unit 203 .

[0060] The training unit 213 repeatedly performs the processing in S501 to S510 until the loss value or the recognition accuracy converges, thereby determining the weight of the class recognition model. Thereafter, a series of processing in the flowchart ends.

[0061] <Processing of Application Detection Model>

[0062] Figure 6 2 is a flowchart describing a process of applying a detection model. The process of applying a detection model is a process of identifying a category of input data of an unknown category, selecting a detection model corresponding to the identified category, and applying the detection model to the input data. The information processing device 1 includes an analysis application unit 204. Figure 6The flowchart in describes the processing performed by the analysis application unit 204. It is assumed that the category identification unit 203 includes a category identification model capable of identifying the first category to the Nth category. Additionally, it is assumed that the analysis unit 206 includes a detection model for the first category to the Mth category. In addition, it is assumed that the goodness-of-fit value calculation result 205 includes M goodness-of-fit values ​​A(N:m) (m=1, 2, ..., M) of the data of the Nth category relative to each detection model for the first category to the Mth category.

[0063] In S601 , the input data acquisition unit 202 acquires an image 201 of an unknown category.

[0064] In S602 , the analysis application unit 204 obtains the goodness-of-fit value calculation result 205 .

[0065] In S603 , the category identification unit 203 identifies the category of the image 201 using the category identification model and outputs the identification result.

[0066] In S604, the analysis application unit 204 determines whether the recognition result is in any of the first to M-th categories. In the case where the analysis application unit 204 determines that the recognition result is in any of the first to M-th categories ("Yes" in S604), the process proceeds to S605. In the case where the analysis application unit 204 determines that the recognition result is a category other than the first to M-th categories ("No" in S604), the process proceeds to S606.

[0067] In S605, the analysis application unit 204 selects a detection model for the category output as the recognition result. Thereafter, the process proceeds to S608. In this way, in the case where the recognition result indicates any category (specific category) among the first category to the Mth category, the analysis application unit 204 selects a detection model for the corresponding category.

[0068] In S606, the analysis application unit 204 determines whether the recognition result is in the Nth category. In the case where the analysis application unit 204 determines that the recognition result is not in the Nth category ("No" in S606), a series of processes in the flowchart ends. In the case where the analysis application unit 204 determines that the recognition result is in the Nth category ("Yes" in S606), the process proceeds to S607.

[0069] In S607, the analysis application unit 204 searches for the category n having the highest goodness of fit value A(N:n) (1≤n≤M) from the goodness of fit calculation result 205 acquired in S602, and selects the detection model for the nth category. In this way, in the case where the recognition result indicates the Nth category (a category other than a specific category), the analysis application unit 204 selects a detection model for the corresponding category from the detection model for the first category to the detection model for the Mth category based on the goodness of fit value relative to the Nth category.

[0070] In S608, the analysis unit 206 applies the detection model selected in S605 or S607 to the image 201 acquired in S601, and outputs the detection result 210. Thereafter, the processing in this flowchart ends.

[0071] As in the present exemplary embodiment, when the goodness of fit value A(N:1) is high and the category recognition model is trained by labeling the supervision data set 212 of the Nth category with subcategories of the first category, the training is performed to approximate the first category with the Nth category.

[0072] Therefore, in Figure 6 In S603 , the Nth category is identified as the first category, and a detection model for the first category is selected in S605 .

[0073] On the other hand, in the case where the class recognition model is trained by labeling the supervision data set 212 of the Nth class with the Nth class, the training is performed to recognize the Nth class as the Nth class. Figure 6 In S603 of the method, the Nth category is identified as the Nth category, and in S607, a detection model is selected from the detection model for the first category to the detection model for the Mth category based on the goodness of fit value. In the present exemplary embodiment, since the goodness of fit value A(N:1) is high, the detection model for the first category is selected.

[0074] In this exemplary embodiment, even when there are only detection models for a specific category and images of categories other than the specific category are acquired, a detection model for the specific category corresponding to the goodness-of-fit value is selected, thereby making it possible to perform high-quality detection tasks. Additionally, by training the category recognition model using images of categories other than the specific category as supervisory data for identifying a specific category with a high goodness-of-fit value relative to the category, it becomes easier to identify the appropriate category to which the detection task is applied. That is, by effectively utilizing known analysis models, it becomes possible to perform analysis tasks with high precision on data of categories other than the specific category.

[0075] In the first exemplary embodiment, a configuration is described in which, for an input image, the category of an object in an image is identified and a detection model for detecting a specific part of the identified category is applied. In the second exemplary embodiment, a configuration is described in which an input document is classified into categories such as news (e.g., business, science and technology, and entertainment) and a generalization model for summarizing documents classified into the category is applied. The process of summarizing documents of a category is an example of analytical processing. Note that the document summarization task is to perform abstract generalization. Since abstract generalization is a task involving document generation, high-quality generalization should be difficult compared to extractive generalization. As a method of improving quality, it is conceivable to use a machine learning model that has been trained specifically for a category. In the following description, matters common to the first exemplary embodiment are not described, and matters different from the above exemplary embodiments are mainly described.

[0076] Figure 7 is a diagram showing a functional configuration example of the information processing apparatus according to the present exemplary embodiment. Figure 7 Function configuration and Figure 2 The functional configuration in is different in that the input data acquisition unit 202 acquires the document 701 as input data, and the analysis unit 706 has a summary model dedicated to each category. Additionally, Figure 7 Function configuration and Figure 2 The difference in the functional configuration in is that the analysis application unit 704 selects a summary model and outputs a summary result 710 as a result of applying the summary model. In the present exemplary embodiment, the analysis unit 706 includes M summary models (a summary model 707 for the first category, a summary model 708 for the second category, ..., a summary model 709 for the Mth category) for the M categories respectively. On the other hand, the analysis unit 706 does not include a summary model for the Nth category (N>M).

[0077] The category identification unit 203 identifies the category of the document. The category identification model included in the category identification unit 203 is, for example, a machine learning model fine-tuned using a known training method for category identification based on a model according to non-patent literature.

[0078] Each summarization model included in the analysis unit 706 is a machine learning model fine-tuned using a known training method for abstract summarization based on the model of "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" written by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova in arXiv:1810.04805.

[0079] As the goodness-of-fit value A(N:m), for example, a score of Recall-Oriented Understudy for Gisting Evaluation (ROUGE) which is a type of evaluation indicator for document summarization is used. There are some variations of ROUGE. For example, in the case of ROUGE-L, a document summarized using a summarization model and a summarization model indicated by ground truth information related to an input document are compared with each other, and a score calculated based on the maximum sequence (number of words) consistent between the two documents is used. It is assumed that the higher the score of ROUGE, the higher the goodness-of-fit value.

[0080] In the present exemplary embodiment, the following case will be described: among the M goodness-of-fit values ​​A(N:m), the goodness-of-fit value A(N:1) of the Nth category with respect to the summary model 707 for the first category and the goodness-of-fit value A(N:2) of the Nth category with respect to the summary model 708 for the second category are above the threshold. In this case, both the summary model 707 for the first category and the summary model 708 for the second category can be applied to the document of the Nth category. In contrast, it is assumed that the goodness-of-fit values ​​A((N:3) to (N:M)) of the Nth category data with respect to the summary models for the third category to the summary models for the Mth category are less than the threshold, and none of the summary models for the third category to the summary models for the Mth category are applicable to the document of the Nth category.

[0081] The following describes the processing performed by the supervisory data set generating unit 215 in the present exemplary embodiment.

[0082] Figure 8 is a flowchart describing the process of generating supervisory data for the Nth category.

[0083] In S801 , the supervisory data set generation unit 215 obtains the data set 216 of the Nth category.

[0084] In S802 , the supervisory data set generation unit 215 obtains the goodness-of-fit values ​​A(N:m) (m=1, 2, . . . , M) of the data of the Nth category with respect to each summary model from the summary model for the first category to the summary model for the Mth category from the goodness-of-fit value calculation result 205 .

[0085] In S803 , the supervisory data set generating unit 215 stores 1 in m indicating the class number.

[0086] In S804, the supervisory data set generation unit 215 determines whether the goodness of fit value A(N:m) is above a predetermined threshold value. If the supervisory data set generation unit 215 determines that the goodness of fit value A(N:m) is above the threshold value ("Yes" in S804), the process proceeds to S805. If the supervisory data set generation unit 215 determines that the goodness of fit value A(N:m) is less than the threshold value ("No" in S804), the process proceeds to S806.

[0087] In S805, the supervisory dataset generation unit 215 labels the dataset 216 of the Nth category with the subcategories of the mth category and generates a dataset as the supervisory dataset 212 of the Nth category. The generated supervisory dataset 212 is stored in the secondary storage device 14, etc. Thereafter, the process proceeds to S807.

[0088] In S806, the supervisory dataset generation unit 215 labels the dataset 216 of the Nth category with a category different from the mth category, and generates a dataset as the supervisory dataset 212 of the Nth category. The generated supervisory dataset 212 is stored in the secondary storage device 14, etc. Thereafter, the process proceeds to S807.

[0089] In S807 , the supervisory data set generation unit 215 increments m indicating the class number by 1, and the process proceeds to S808 .

[0090] In S808, the supervisory data set generation unit 215 determines whether the relationship of m>M holds. In the case where the relationship of m≤M holds ("No" in S808), the process proceeds to S804. In the case where the relationship of m>M holds ("Yes" in S808), a series of processes in the flowchart ends. In the present exemplary embodiment, the data of the Nth category is handled as supervisory data of a subcategory of the first category and a subcategory of the second category.

[0091] In the present exemplary embodiment, the Nth category is handled as a subcategory of the first category and a subcategory of the second category. However, in the case where there are two or more summary models having a goodness of fit value A(N:m) above a threshold value, a category of any summary model among the summary models may be selected as a subcategory. For example, a category of n summary models having goodness of fit values ​​A(N:m) ranked from the highest to the nth may be selected as a subcategory. In this case, when n is 1, the category of the summary model having the highest goodness of fit value is used as a subcategory.

[0092] Then, refer to Figure 5 The training process of the class identification model according to the present exemplary embodiment is described. The differences from the first exemplary embodiment are now described.

[0093] In S506, the training unit 213 treats the Nth category as a subcategory of the first category and a subcategory of the second category. Therefore, in S507, even if the training unit 213 identifies the category of the data of the Nth category as the first category or the second category in step S505, the training unit 213 calculates the loss value by treating it as the ground truth. On the other hand, in the case where the category of the data of the Nth category is identified as any category from the third category to the Mth category, errors in applying generalization are not allowed, and the training unit 213 calculates the loss value as a higher value to prevent misidentification. For example, the training unit 213 may multiply the calculated loss value by a correction coefficient greater than 1, or correct the loss value according to the goodness of fit value so that the loss value becomes higher as the goodness of fit value becomes lower.

[0094] In the present exemplary embodiment, the category identification unit 203 uses the category identification model obtained by training in this manner to identify the category of the input document 701. In the case where the category of the document 701 is identified as the first category or the first subcategory, the analysis application unit 704 controls to apply the summary model 707 for the first category to the input document 701. In the case where the category of the document 701 is identified as the second category or the second subcategory, the analysis application unit 704 controls to apply the summary model 708 for the second category to the input document 701.

[0095] In the present exemplary embodiment, in the case where there are only summary models of a specific category and documents of categories other than the specific category are obtained, the summary model of the specific category corresponding to the goodness-of-fit value is selected, thereby making it possible to perform a high-quality summary task. Additionally, by training the category recognition model using documents of categories other than the specific category as supervisory data for identifying documents of the specific category having a high goodness-of-fit value relative to the category, it becomes easier to identify the appropriate category to which the summary task is applied. That is, by effectively utilizing known analysis models, it becomes possible to perform analysis tasks with high accuracy for input data of categories other than the specific category.

[0096] As described above, when there are only analysis models for a specific category, it becomes possible to perform high-precision analysis for categories other than the specific category without newly training analysis models for categories other than the specific category. With this configuration, the types of categories applicable to the analysis model can be increased.

[0097] (Variation of the Second Embodiment)

[0098] The training unit 213 according to the present exemplary embodiment trains the category recognition model with the Nth category as a subcategory of the first category and a subcategory of the second category. On the other hand, in the category recognition model that has been trained in this manner, there is a case where the recognition performance of the first category or the second category is reduced. In this case, the supervisory data set generation unit 215 does not label the supervisory data of the Nth category with the subcategory, but adds the category and labels the supervisory data of the Nth category with the Nth category itself. The training unit 213 uses the supervisory data set 212 of the Nth category that has been generated in this manner for training. The analysis application unit 704 applies the summary model for the category with the highest goodness of fit value relative to the Nth category (the summary model for the first category or the summary model for the second category) to the document 701 that is identified as being in the Nth category using the category recognition model, and summarizes the document 701 of the Nth category.

[0099] (Other Exemplary Embodiments)

[0100] Although the present disclosure describes exemplary embodiments, the above exemplary embodiments are only specific examples for implementing the present disclosure, and the technical scope of the present disclosure should not be interpreted in a limited manner by the exemplary embodiments. That is, the present disclosure can be implemented in various modes without departing from the technical concept or main features of the present disclosure.

[0101] The embodiments of the present disclosure may be implemented in the following manner: a computer executable instruction (e.g., a program) that implements one or more functions of the above exemplary embodiments is supplied to a system or device via a network or storage medium, and one or more processors in a computer of the system or device load and execute the program. In addition, the embodiments of the present disclosure may be implemented by a circuit (e.g., an ASIC) that implements one or more functions.

[0102] The disclosure of the above-described respective exemplary embodiments includes the following configurations, methods, and programs.

[0103] According to the present disclosure, data of various categories can be analyzed with high accuracy.

[0104] Other embodiments

[0105] The embodiments of the present invention may also be implemented by providing software (program) for performing the functions of the above-described embodiments to a system or device via a network or various storage media, and a computer or a central processing unit (CPU) or a microprocessing unit (MPU) of the system or device reads and executes the program.

[0106] Although the disclosure describes exemplary embodiments, it is to be understood that some embodiments are not limited to the disclosed exemplary embodiments.The scope of the following claims is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures and functions.

Claims

1. An information processing device, comprising: one or more memories storing instructions; as well as One or more processors communicate with the one or more memories and, when executing the instructions, cooperate with the one or more memories to perform operations including: Identify the categories of input data; as well as Based on the goodness of fit value of the data belonging to the first category relative to the analysis models used for each category for analyzing the data of each category, an analysis model to be applied to the input data identified as being in the first category by the identification is selected from the analysis models used for each category.

2. The information processing device according to claim 1, wherein: The operations also include: If an analysis model for the first category exists, applying the analysis model for the first category to input data identified as being in the first category; and In the absence of an analysis model for the first category, the selected analysis model is applied.

3. The information processing device according to claim 1, wherein: The operation also includes: selecting an analysis model with a highest goodness of fit value from the analysis models for each category.

4. The information processing device according to claim 1, wherein: The operation also includes calculating the goodness-of-fit value based on a comparison result between a result obtained by inputting the data belonging to the first category into an analysis model for each category and ground truth information included in the data belonging to the first category.

5. The information processing device according to claim 1, wherein: The operations also include: training a recognition model to be used in the recognition, wherein the training uses supervised data obtained by labeling data belonging to the first category with the first category to train the recognition model; and In the absence of an analysis model for the first category, the selected analysis model is applied to input data identified as being in the first category.

6. The information processing device according to claim 1, wherein: The operation also includes: identifying the category of the image, and The analysis model is a model configured to detect a specific part of the category from the image.

7. The information processing device according to claim 1, wherein: The operations also include: identifying the category of the document, and The analysis model is a model configured to summarize documents of the category.

8. The information processing device according to claim 1, wherein: The analytical model is a neural network.

9. An information processing device, comprising: one or more memories storing instructions; as well as One or more processors communicate with the one or more memories and, when executing the instructions, cooperate with the one or more memories to perform operations including: training a recognition model configured to recognize the class of the data, In which, the training of the recognition model uses supervised data, and the supervised data is obtained by labeling the data belonging to the first category using categories based on the goodness of fit value of the analysis model for each category used to analyze the data of each category.

10. The information processing device according to claim 9, wherein: The operations also include: Using the trained recognition model to recognize the category of input data; and An analysis model for the class identified using the recognition model is applied to the input data.

11. The information processing device according to claim 9, wherein: The supervised data is also obtained by labeling data belonging to the first category with a second category, the second category being a category of the analysis model for which the goodness of fit value is greater than or equal to a threshold.

12. The information processing device according to claim 9, wherein: The supervisory data is also obtained by labeling data belonging to the first category with a category different from a category of the analysis model having a goodness of fit value less than a threshold value.

13. The information processing device according to claim 11, wherein: The training of the recognition model uses supervised data obtained by labeling the data belonging to the first category with the first category to train the recognition model in a case where the recognition performance of the recognition model for the data belonging to the second category decreases as a result of the training.

14. The information processing device according to claim 9, wherein: When the class of the analysis model whose goodness of fit value is smaller than a threshold value is output as a result of identification for data belonging to the first class using the identification model, the training of the identification model is corrected so that the loss value becomes higher.

15. The information processing device according to claim 11, wherein: In the case where there are multiple analysis models each having the goodness of fit value greater than or equal to the threshold, the training of the recognition model is labeled using the category of the analysis model with the higher goodness of fit value.

16. The information processing device according to claim 9, wherein: The recognition model is a neural network.

17. An information processing method, comprising: Identify the categories of input data; as well as Control is performed to select an analysis model to be applied to input data identified as being in the first category from analysis models for each category based on the goodness of fit value of the data belonging to the first category relative to each analysis model in the analysis models for each category used to analyze the data of each category.

18. A non-transitory computer-readable storage medium storing computer-executable instructions for causing a computer to perform the method according to claim 17.

19. A computer program product comprising computer executable instructions for causing a computer to perform the method according to claim 17.

Images