Information processing device, information processing system, and information processing method
The system securely processes confidential document data by converting it into non-confidential feature amounts, enabling secure collaboration between devices and preventing data leaks.
Patent Information
- Application Number
- JP2023078058
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-10
- Publication Date
- 2024-11-21
- Estimated Expiration
- 2043-05-10
AI Technical Summary
Existing methods for processing document data using machine learning do not adequately address the secure handling of confidential information when multiple devices are involved, leading to potential leaks.
An information processing system and method that includes a first device for acquiring and analyzing document data, determining feature amounts using a hash function, and exporting non-confidential feature data to a second device for processing, while maintaining confidentiality by avoiding direct transmission of sensitive information.
Enables secure cooperative processing between devices by converting confidential document data into non-confidential feature amounts, preventing data leaks and allowing integrated processing across multiple systems.
Smart Images

Figure 2024162503000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to an information processing device, an information processing system, an information processing method, and the like. [Background technology]
[0002] Conventionally, there are known methods for processing document data using machine learning. For example, Patent Literature 1 discloses a document information extraction system that updates parameters based on the action type or weight of a feature to be evaluated when determining a feature of a model. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2022-148430 A Summary of the Invention [Problem to be solved by the invention]
[0004] The technique of Patent Document 1 evaluates feature amounts by taking into consideration similarity relationships based on a similar dictionary or the like, but does not take into consideration the case where processing related to document data, which is confidential information, is performed in an external device.
[0005] According to some aspects of the present disclosure, it is possible to provide an information processing device, an information processing system, an information processing method, and the like that are capable of securely executing cooperative processing between a plurality of information processing devices. [Means for solving the problem]
[0006] One aspect of the present disclosure relates to an information processing device that includes an acquisition unit that acquires document data, an analysis processing unit that performs morphological analysis of the document data, a feature determination unit that determines features corresponding to the document data based on the results of the morphological analysis and a hash function, an export unit that outputs export data, which is data based on the features, to a second information processing device, an import unit that acquires processing results based on the export data in the second information processing device, and a processing unit that performs processing based on the processing results acquired by the import unit.
[0007] Another aspect of the present disclosure relates to an information processing system including a first information processing device and a second information processing device, wherein the first information processing device includes an acquisition unit that acquires document data, an analysis processing unit that performs morphological analysis of the document data, a feature determination unit that determines features representing the document data based on the results of the morphological analysis and a hash function, an export unit that outputs export data, which is data based on the features, to the second information processing device, an import unit that acquires processing results based on the export data in the second information processing device, and a processing unit that performs processing based on the processing results acquired by the import unit.
[0008] Yet another aspect of the present disclosure relates to an information processing method in which an information processing device acquires document data, performs a morphological analysis of the document data, determines features representing the document data based on the results of the morphological analysis and a hash function, exports export data that is data based on the features to a second information processing device, imports processing results based on the export data in the second information processing device, and performs processing based on the imported processing results. [Brief description of the drawings]
[0009] [Figure 1] 1 is a configuration example of an information processing system including an information processing device. [Diagram 2] 2 illustrates an example of the configuration of a first information processing device. [Diagram 3]13 is a configuration example of a second information processing device. [Figure 4] FIG. 4 is a sequence diagram illustrating the processing of the present embodiment. [Diagram 5] 2 is an example of document data. [Figure 6] 1 is a diagram showing an example of a feature (vector) obtained by hashing document data. [Figure 7] 1 is a diagram illustrating an example of a relationship between document data and feature amounts. [Figure 8] 2 is a specific configuration example of an information processing system. [Figure 9] 13 illustrates another configuration example of the second information processing device. [Figure 10] FIG. 11 is another sequence diagram illustrating the process of the present embodiment. [Figure 11] 13 illustrates another configuration example of the second information processing device. [Figure 12] FIG. 11 is another sequence diagram illustrating the process of the present embodiment. [Figure 13] 13 illustrates another configuration example of the first information processing device. [Figure 14] FIG. 11 is another sequence diagram illustrating the process of the present embodiment. [Figure 15] 2 is a specific configuration example of an information processing system. [Figure 16] 11 is a flowchart illustrating an automatic update process of a control parameter. [Figure 17A] 1 is an example of a metadata feature. [Figure 17B] FIG. 11 is an explanatory diagram of a correction process of a metadata feature amount based on a first norm. [Figure 17C] FIG. 11 is an explanatory diagram of a correction process of a metadata feature amount based on a second norm. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0010] Hereinafter, the present embodiment will be described with reference to the drawings. In the drawings, the same or equivalent elements are given the same reference numerals, and duplicated explanations are omitted. Note that the present embodiment described below does not unduly limit the contents described in the claims. In addition, all of the configurations described in the present embodiment are not necessarily essential configurations of the present disclosure.
[0011] 1. System Configuration Fig. 1 is a diagram showing an example of the configuration of an information processing system 1 including a first information processing device 100 which is an information processing device according to this embodiment. The information processing system 1 includes a first information processing device 100 and a second information processing device 200. However, the configuration of the information processing system 1 is not limited to that shown in Fig. 1, and various modifications are possible, such as omitting some components or adding other components.
[0012] The first information processing device 100 is a device that acquires document data. The document data here broadly includes information related to text. For example, the document data can include various data such as e-mail, data written using a chat tool, data posted to a social networking service (SNS), and data created using a document creation application. The document data may also include data acquired as a result of a voice recognition process on voice data, data resulting from an OCR process performed on the results of reading handwritten documents and forms, and the like.
[0013] The second information processing device 200 is a device different from the first information processing device 100, and is a device that executes processing based on document data. For example, the second information processing device 200 generates a trained model used for processing document data by performing machine learning based on document data for learning. Alternatively, the machine learning for generating the trained model may be executed in the first information processing device 100, and the second information processing device 200 may perform analysis processing related to the machine learning. Note that, as described later, the processing based on document data here may be processing using data (feature amounts described later) obtained by hashing document data.
[0014] In this embodiment, the processing of document data executed using the trained model may be, for example, a classification process. The classification process may be, for example, a process for determining the degree of relationship between a predetermined case and target document data. Various predetermined cases are possible here. For example, the information processing system 1 of this embodiment may be included in various systems such as a discovery support system described below, and the predetermined case may be any of the various cases described below. However, the predetermined cases are not limited to those listed below.
[0015] - The case requires discovery procedures in a discovery-assisted system Crimes that are the subject of investigation in the forensic system - Misconduct in email monitoring systems (e.g., information leaks, collusion, etc.) ·Case studies and cases related to medicine in medical application systems ·Case studies on risks and opportunities in the Economic Security Support System ·Case studies · Cases and incidents related to the Internet in Internet application systems, etc. -Past projects in the project evaluation system · Products and services that are the subject of marketing in the marketing support system Intellectual property subject to evaluation in the intellectual property evaluation system -Fraudulent financial transactions in the fraudulent transaction monitoring system Past cases of call center escalation systems · Subjects of credit investigation in the credit investigation system - Matters related to vehicle operation using driving assistance systems ·Sales performance in sales support systems
[0016] As shown in the above medical application system, the information processing system 1 according to this embodiment may process medical / nursing and care information and personal information. For example, the information processing system 1 may determine the degree of association between the patient and disease, nursing, medicine, care, etc., by processing document data including the patient's electronic medical record as input data. The determination here may be to estimate a disease that the patient may suffer from, or to estimate the medical / nursing treatment and care treatment required for the patient. Alternatively, the determination may be to estimate the medical / nursing treatment, care treatment, patient attributes, patient environment, etc. that caused the problem when a problem occurs to the patient. In addition, various modifications are possible for specific examples of processing related to medical / nursing and care information.
[0017] Furthermore, the processing of document data in this embodiment is not limited to classification processing. For example, in this embodiment, the trained model may be a model that performs clustering processing of document data. Alternatively, the trained model may perform keyword extraction processing to extract keywords from document data based on given extraction conditions (search parameters). Alternatively, the trained model may be a model that performs summarization processing to summarize document data into relatively short content. In addition, various modifications of the processing of document data in this embodiment are possible.
[0018] The first information processing device 100 may be realized by, for example, a server system. The server system here may be one server, or may be configured to include multiple servers. For example, the server system may be configured to include a database server and an application server. The database server stores various data including document data and feature quantities described later. The application server executes the process described later with reference to FIG. 4 and the like. The multiple servers here may be physical servers or virtual servers. In addition, when a virtual server is used, the virtual server may be provided in one physical server, or may be distributed and arranged in multiple physical servers. In this way, the specific configuration of the server system in this embodiment can be variously modified. In other words, the first information processing device 100 according to this embodiment may be realized by a single device, or may be realized by distributed processing of multiple devices. In addition, the first information processing device 100 is not limited to a server system, and may be realized by other devices such as a PC (Personal Computer).
[0019] The second information processing device 200 may be realized by a server system or by other devices such as a PC. When the second information processing device 200 is realized by a server system, the specific configuration can be modified in various ways, similar to the first information processing device 100.
[0020] The first information processing device 100 and the second information processing device 200 are connected via a network, for example. The network here is a public communication network such as the Internet. However, the network may be a LAN (Local Area Network) or the like, and the specific configuration is not limited.
[0021] Fig. 2 is a diagram showing an example of the configuration of the first information processing device 100. The first information processing device 100 includes an acquisition unit 110, an analysis processing unit 120, a feature amount determination unit 130, a processing unit 140, an export unit 150, and an import unit 160. However, the configuration of the first information processing device 100 is not limited to that shown in Fig. 2, and various modifications such as adding some components or adding other components are possible.
[0022] The acquisition unit 110 acquires document data. As described above, the document data can include various data, and the acquisition unit 110 can acquire document data from various devices and applications. The acquisition unit 110 may be realized as a communication interface that communicates with a device that is a source of the document data, or as a processor that controls the communication. The communication interface here may be an interface that performs communication in accordance with IEEE802.11 or IEEE802.3, or an interface that performs communication according to another method. The communication interface may include, for example, an antenna, an RF (Radio Frequency) circuit, a baseband circuit, etc.
[0023] The analysis processing unit 120 acquires document data from the acquisition unit 110 and performs morphological analysis on the document data. Since morphological analysis is a method widely used in the field of natural language processing, a detailed description is omitted. By the morphological analysis, multiple morphemes contained in one document data are extracted from the document data.
[0024] The feature amount determining unit 130 determines feature amounts based on the document data, based on the results of the morphological analysis. The feature amount determining unit 130 may obtain feature amounts by using a hash function. For example, the feature amount may be data in which each morpheme is hashed while leaving the word order information intact, or may be data that does not include word order information, as will be described later with reference to Figures 5 to 7. The feature amount will be described in detail later.
[0025] The export unit 150 outputs export data, which is data based on the feature amount after the hash function is applied, to an external device. The export data may be the feature amount itself, or may be information obtained based on the feature amount (for example, a trained model or a search parameter described later). For example, the first information processing device 100 may include an export data determination unit (not shown) that determines what data to output as the export data. The export unit 150 outputs the data determined by the export data determination unit to the external device as the export data. The external device here is a device different from the first information processing device 100, for example, the second information processing device 200. The second information processing device 200 executes processing based on the export data and acquires the processing result. The import unit 160 acquires the processing result in the second information processing device 200.
[0026] The export unit 150 and the import unit 160 may be, for example, a communication interface with the second information processing device 200, or a processor that controls the communication interface. The communication method used by the communication interface here may be a communication method according to IEEE802.11, a communication method according to IEEE802.3, or another communication method.
[0027] The processing unit 140 performs processing based on the processing result in the second information processing device 200 acquired by the import unit 160. The processing here may be output processing (display processing, etc.) of the acquired processing result, or may be other processing. Details of the processing in the processing unit 140 will be described later.
[0028] Each unit of the first information processing device 100 according to this embodiment is configured by the following hardware. The hardware can include at least one of a circuit for processing digital signals and a circuit for processing analog signals. For example, the hardware can be configured by a single or multiple circuit devices mounted on a circuit board, or a single or multiple circuit elements. The single or multiple circuit devices are, for example, an integrated circuit (IC) or a field-programmable gate array (FPGA). The single or multiple circuit elements are, for example, a resistor, a capacitor, etc. Note that the units of the first information processing device 100 are, for example, an acquisition unit 110, an analysis processing unit 120, a feature determination unit 130, and a processing unit 140, but may also include an export unit 150 and an import unit 160.
[0029] Each unit of the first information processing device 100 may be realized by the following processor. The first information processing device 100 of this embodiment includes a memory that stores information and a processor that operates based on the information stored in the memory. The information is, for example, a program and various data. The program may include a program that causes the first information processing device 100 to execute the processing described in this specification. The processor includes hardware. The processor can use various processors such as a central processing unit (CPU), a graphics processing unit (GPU), and a digital signal processor (DSP). The memory may be a semiconductor memory such as a static random access memory (SRAM), a dynamic random access memory (DRAM), or a flash memory, a register, a magnetic storage device such as a hard disk drive (HDD), or an optical storage device such as an optical disk device. For example, the memory stores computer-readable instructions, and the functions of each unit of the first information processing device 100 are realized as processing by the processor executing the instructions. The instruction here may be an instruction from an instruction set that constitutes a program, or an instruction that instructs the hardware circuitry of the processor to operate.
[0030] Fig. 3 is a diagram showing a configuration example of the second information processing device 200. The second information processing device 200 includes a processing unit 240, an export unit 250, and an import unit 260. However, the configuration of the second information processing device 200 is not limited to that shown in Fig. 3, and various modifications such as adding some components or adding other components are possible.
[0031] The import unit 260 performs a process of acquiring the export data output by the first information processing device 100 .
[0032] The processing unit 240 performs processing based on the export data acquired by the import unit 260. For example, the processing unit 240 may acquire feature amounts as export data and perform machine learning (learning processing) using the acquired feature amounts as learning data. However, the export data is not limited to feature amounts, and the processing of the processing unit 240 based on the export data is not limited to learning processing. Details of the processing in the processing unit 240 will be described later.
[0033] The export unit 250 outputs the processing result of the processing unit 240 to the first information processing device 100. The export unit 250 and the import unit 260 may be, for example, a communication interface with the first information processing device 100, or a processor that controls the communication interface. The communication method used by the communication interface here may be a communication method according to IEEE802.11, a communication method according to IEEE802.3, or another communication method.
[0034] Each unit of the second information processing device 200 can be configured with hardware including at least one of a circuit for processing digital signals and a circuit for processing analog signals. For example, the hardware can be configured with a single or multiple circuit devices or single or multiple circuit elements mounted on a circuit board.
[0035] The second information processing device 200 may also include a processor and a memory. Each unit of the second information processing device 200 may be realized by a processor including hardware, and the processor may be various processors such as a CPU, a GPU, a DSP, etc. The memory may be a semiconductor memory, a register, a magnetic storage device, or an optical storage device. For example, the memory stores computer-readable instructions, and the processor executes the instructions to realize the functions of each unit of the second information processing device 200 as processing.
[0036] As described above, by using the export data based on the document data acquired by the first information processing device 100 for processing in the second information processing device 200, cooperative processing between the first information processing device 100 and the second information processing device 200 becomes possible. For example, the information processing system 1 of the present embodiment can be applied to various systems such as the above-mentioned discovery support system, and the system can be realized by combining the first information processing device 100 and the second information processing device 200.
[0037] However, the document data of this embodiment may contain information that is undesirable to be leaked to the outside. For example, when the first information processing device 100 acquires emails or business-related data created in the course of the activities of a given company as document data, the document data becomes confidential information that is undesirable to be leaked to the outside of the company. Alternatively, when targeting document data that is a transcription of conversation data between an employee and a customer, the document data may contain the contents of speech of a person (customer) outside the company. In this case, too, it is desirable to manage the document data as confidential information. For example, when the second information processing device 200 is a device of an external company, there is a possibility that confidential information may be leaked by transmitting the document data itself to the second information processing device 200.
[0038] Therefore, the first information processing device 100, which is an information processing device according to this embodiment, includes the acquisition unit 110, the analysis processing unit 120, the feature amount determination unit 130, the export unit 150, the import unit 160, and the processing unit 140 as described above. The feature amount determination unit 130 determines the feature amount based on a hash function. The export unit 160 outputs export data, which is data based on the feature amount, to the second information processing device 200.
[0039] According to the method of this embodiment, when transmitting data from the first information processing device 100 to the second information processing device 200, it is possible to transmit hashed features or data derived from the features as export data, rather than the original document data itself. Since the hash function performs irreversible (undecodable) conversion, transmitting the export data to an external device does not lead to leakage of confidential information. Therefore, it is possible to perform various processes based on the document data in cooperation with multiple devices while suppressing leakage of confidential information.
[0040] The method of the present embodiment can be applied to the information processing system 1 including the above-mentioned first information processing device 100 and the second information processing device 200. According to the method of the present embodiment, even if the second information processing device 200 is not in a secure environment for the first information processing device 100, it is possible to appropriately link a plurality of devices while suppressing leakage of confidential information.
[0041] The technique of this embodiment can also be applied to an information processing method in which an information processing device (first information processing device 100) executes the following steps. The information processing method includes steps in which the information processing device acquires document data, performs a morphological analysis of the document data, determines features representing the document data based on the results of the morphological analysis and a hash function, exports export data that is data based on the features to the second information processing device 200, imports a processing result for the export data in the second information processing device 200, and performs processing based on the imported processing result.
[0042] 2. Processing Details The details of the processing in the information processing system 1 will be described.
[0043] 2.1 Processing flow 4 is a sequence diagram illustrating the flow of processing in this embodiment. When this processing starts, first, in step S101, the acquisition unit 110 of the first information processing device 100 acquires document data. As described above, the processing in step S101 may be reception of an email, acquisition of posted data on an SNS, or other processing.
[0044] In step S102, the analysis processing unit 120 performs a morpheme analysis process on the acquired document data. Here, a morpheme refers to the smallest unit that has meaning in a language that constitutes a sentence. Morphological analysis includes a process of breaking down document data into a plurality of morphemes. The analysis processing unit 120 acquires a set of morphemes contained in the document data as a result of the morpheme analysis. The analysis processing unit 120 may determine the part of speech of the morpheme, and the determination results may be included in the result of the morpheme analysis. Since morpheme analysis is a method that is widely used in the field of natural language processing, a detailed description thereof will be omitted.
[0045] In step S103, the feature determining unit 130 determines a feature corresponding to the document data. For example, the feature determining unit 130 may perform a process of determining a value corresponding to a given morpheme based on the appearance state of the given morpheme in the target document data. The feature determining unit 130 may then determine a tensor (or, in a narrower sense, a vector) in which values determined for each morpheme are arranged as a feature representing the target document data.
[0046] 5 to 7 are diagrams for explaining the results of morphological analysis and the process of obtaining feature quantities based on a hash function. FIG. 5 shows simple examples of document data and the results of morphological analysis. For example, assume that there are four documents, document data A to document data D, as document data. For example, document data A is document data consisting of one sentence, "Impossible is nothing." In this case, three morphemes (Impossible, is, nothing) are obtained as a result of morphological analysis. The same applies to document data B to document data D.
[0047] In this case, a feature may be calculated by using the number of morphemes as elements and assigning the number of occurrences to the value of each element. For example, the feature of document data A is a vector consisting of three elements corresponding to (Impossible, is, nothing), and the value of each element is 1, which is the number of occurrences of the respective morpheme. However, in this case, since the types and numbers of morphemes contained in document data A to D are different, the number of elements of the vector that is the feature and the morphemes that correspond to each element will differ between document data.
[0048] Therefore, a method is used in which a tensor of size N (N is an integer of 2 or more) is set and a hash function is applied to each morpheme to associate each morpheme with one of the N elements. FIG. 6 is an example of a tensor in this case. Below, an example will be described in which a vector that is a first-order tensor is used as the tensor, but the tensor in this embodiment may be a matrix or a third-order or higher tensor. In the example of FIG. 6, the tensor size is 13203. For example, the feature determination unit 130 associates each morpheme with one of the elements of the tensor using a hash function whose output size (output range) is 1 or more and 13203 or less. In the example of FIG. 6, the result of applying the hash function to Impossible is 1, so the morpheme Impossible corresponds to the first element of the tensor. Similarly, consider an example in which the morpheme is corresponds to the second element of the tensor and the morpheme nothing corresponds to the third element of the tensor.
[0049] FIG. 7 is a diagram showing the feature amount obtained by converting the document data shown in FIG. 5 according to the rule shown in FIG. 6. For example, since the document data A includes one each of (Impossible, is, nothing) as a morpheme, the feature amount determining unit 130 determines a tensor in which the first to third elements are 1 and all other elements are 0 as the feature amount of the document data A. The same applies to the document data B to D. When only the feature amount obtained here is seen, it is impossible to obtain a relationship in which the first element corresponds to Impossible and the second element corresponds to is. Therefore, since the original document data cannot be restored from the feature amount obtained by the feature amount determining unit 130, the feature amount can be treated as non-confidential information. Furthermore, since the tensor size of all the document data is the same, it is possible to realize parallelization and commonization of processing, and as a result, it is also possible to execute the processing according to this embodiment at high speed. The parallelization of processing will be described later.
[0050] There is also a possibility that the results of applying a hash function to two different morphemes may coincide (collide). If a collision occurs, the two morphemes cannot be distinguished, which may reduce the accuracy of the processing. Therefore, the output size N of the hash function may be a value that is sufficiently larger than the expected number of morphemes (for example, tens to hundreds of times).
[0051] 5 to 7, an example was described in which a feature amount not including information on word order is obtained by performing processing focusing on the number of times a morpheme appears. For example, in document data A, (Impossible, is, nothing) appear in this order (FIG. 5), but even if the order is other than this, if the document data has (Impossible, is, nothing) each appearing once, the feature amount will be equal to that of document data A (FIG. 7). However, the feature amount of this embodiment is not limited to this, and may be data including information on word order. For example, the feature amount determining unit 130 may obtain, as a feature amount, information obtained by arranging the result of applying a hash function to a morpheme (FIG. 6) in the order of appearance of the morpheme. In addition, the feature amount of this embodiment can be extended to various information using the result of morpheme analysis and a hash function.
[0052] Returning to FIG. 4, the description will be continued. After the feature amount is obtained, in step S104, the export data determination unit determines data to be output as export data. For example, a storage unit (not shown) of the first information processing device 100 may store a plurality of candidates for the export data. The export data determination unit may determine one of the plurality of candidates as the export data based on, for example, a user input.
[0053] In step S105, the export unit 150 exports the export data based on the feature amount to the second information processing device 200. The import unit 260 of the second information processing device 200 imports the export data from the first information processing device 100.
[0054] In step S106, the processing unit 240 of the second information processing device 200 performs processing based on the export data. Details of the processing will be described later with reference to FIGS.
[0055] In step S107, the export unit 250 of the second information processing device 200 exports the processing result to the first information processing device 100. The import unit 160 of the first information processing device 100 imports the processing result.
[0056] In step S108, the processing section 140 of the first information processing device 100 performs processing based on the processing result.
[0057] Fig. 8 is a diagram showing a specific example of the configuration of the information processing system 1. An example in which the above-mentioned process flow is applied to the information processing system 1 in Fig. 8 will be described.
[0058] As shown in FIG. 8, the information processing system 1 may include, for example, a Japan server 10 located in Japan, a US server 20 located in the US, a European server 30 located in Europe, and a Korean server 40 located in Korea.
[0059] The Japanese server 10 accumulates document data acquired in Japan (corresponding to step S101 in FIG. 4). This document data is treated as confidential information as described above. Similarly, the US server 20 accumulates document data that is confidential information acquired in the US. The European server 30 accumulates document data that is confidential information acquired in Europe. The Korean server 40 accumulates document data that is confidential information acquired in Korea. In the following description, document data before hashing is also referred to as raw data (raw document data). Raw data is confidential information, and in the example of FIG. 8, it represents document data acquired within each country or region.
[0060] At this time, each server can expand the processing target to other countries by acquiring document data from other countries. For example, if a given company has its headquarters in Korea and branches in Japan, the United States, and Europe, the Korean server 40 acquires data from the Japanese server 10, the American server 20, and the European server 30, as shown in Fig. 8. This enables the Korean server 40 to process document data from each country in an aggregated manner.
[0061] However, as described above, the Japanese server 10 does not send the raw document data as is, but sends export data to the Korean server 40 (corresponding to steps S102-S105 in FIG. 4). Here, an example will be described in which the export data is the feature amount itself obtained by hashing the document data. Because the feature amount is hashed data, it is non-confidential data, unlike raw data. The same is true for the American server 20 and the European server 30. This prevents confidential information obtained in each country from leaking out of the country.
[0062] When transmitting export data to the Korean server 40, the servers of each country may assign to each feature an ID that uniquely represents the original document data. In this way, the Korean server 40 can also identify the correspondence between the feature and the document data. The information assigned to the feature is not limited to the ID. For example, the Korean server 40 may manage metadata such as a document data ID, the country where the document data was obtained, and the date and time of acquisition in association with each feature. The Korean server 40 acquires such metadata for Korean document data, and the servers of each country transmit to the Korean server 40 for document data for other countries.
[0063] The Korean server 40 also performs processing based on the collected information (corresponding to step S106 in FIG. 4) and transmits the processing results to the servers in each country (corresponding to step S107). This makes it possible for the Japanese server 10, the American server 20, and the European server 30 to utilize the processing results.
[0064] For example, suppose that a lawsuit occurs at the US branch of the company, and documents need to be submitted for discovery procedures. In this case, for example, the Korean server 40 may create a trained model for discovery support by performing machine learning based on document data from each country in advance. Then, based on the classification results by the trained model, the Korean server 40 transmits information about document data determined to be related to the lawsuit to the US server 20. This makes it possible to appropriately support the discovery procedures at the target company.
[0065] In this case, the method of the present embodiment makes it possible to limit the data transmitted as a result of the processing to only the necessary data. In the example of the discovery procedure described above, the Korean server 40 only needs to transmit a portion of the multiple document data that is particularly related to the lawsuit.
[0066] For example, if the target data is Korean document data, the Korean server 40 performs a process to extract document data related to the lawsuit from the raw document data acquired in Korea, and transmits only the extracted document data to the US server 20. Alternatively, the Korean server 40 may transmit to the US server 20 image data of the document data or metadata attached to the document data, rather than the raw document data itself. This makes it possible to prevent confidential document data from leaking out of the country more than necessary. The country in which the document data related to discovery was originally acquired is identified from, for example, an ID or metadata.
[0067] Furthermore, if the target data is US document data, the Korean server 40 does not hold the raw data of the document data. Therefore, the Korean server 40 may transmit an ID for identifying the document data to the US server 20. Since the raw document data is stored in the US server 20, the US server 20 can identify the document data to be used in the discovery procedure based on the ID.
[0068] Furthermore, if the target document data is Japanese document data, the Korean server 40 may transmit to the Japanese server 10 an ID for identifying the document data and an instruction to transmit the document data corresponding to the ID to the US server 20. The Japanese server 10 identifies the target document data based on the ID and transmits the identified document data to the US server 20. Here, the information transmitted from the Japanese server 10 to the US server 20 may be raw data, information in which a part of the raw data has been imaged, or metadata added to the raw data, as in the above example. Similarly, if the target document data is European document data, the Korean server 40 transmits the necessary data to the US server 20 via the European server 30.
[0069] 8, the US server 20 that exports export data (feature amounts) and imports a processing result based on the export data corresponds to the information processing device of this embodiment (the first information processing device 100). Also, the Korean server 40 that performs processing based on the export data acquired from the US server 20 and transmits the processing result to the US server 20 corresponds to the second information processing device 200.
[0070] Although the Japanese server 10 and the European server 30 are not destinations for export data by the US server 20, they may be included in an external device (second information processing device 200) for the US server 20 since there is a possibility that the processing results of the Korean server 40 may be transmitted to the US server 20. In other words, the second information processing device 200 is not limited to one device, and may be realized by distributed processing of multiple devices.
[0071] 2.2 Specific examples of export data Next, some specific examples of the export data output from the first information processing device 100, the process in the second information processing device 200 using the export data, and the processing results of the second information processing device 200 will be described.
[0072] <Example 1> The second information processing device 200 may have already acquired the trained model at the time when the export data is output from the first information processing device 100. The trained model may be, for example, a model created based on document data acquired by the second information processing device 200. In this case, the first information processing device 100 may export the feature amount determined by the feature amount determining unit 130 as data to be subjected to the inference process. The second information processing device 200 inputs the feature amount imported from the first information processing device 100 into an existing trained model and acquires the output of the model. The output of the model is, for example, a score representing the degree of association between the document data corresponding to the feature amount and a given case. The import unit 160 of the first information processing device 100 may acquire the score as a processing result.
[0073] 9 is a diagram showing a configuration example of the information processing system 1 in this case, and in particular a diagram showing another configuration example of the second information processing device 200. As shown in Fig. 9, the second information processing device 200 may include an acquisition unit 210, an analysis processing unit 220, a feature amount determination unit 230, a processing unit 240, an export unit 250, and an import unit 260. The processing unit 240 may include a learning processing unit 241 and an inference processing unit 242.
[0074] The acquisition unit 210 acquires document data. The analysis processing unit 220 acquires the document data from the acquisition unit 210 and performs a morphological analysis of the document data. The feature determination unit 230 determines a feature representing the document data based on the result of the morphological analysis and a hash function. The acquisition unit 210, the analysis processing unit 220, and the feature determination unit 230 are similar to the acquisition unit 110, the analysis processing unit 120, and the feature determination unit 130 of the first information processing device 100.
[0075] The learning processing unit 241 performs machine learning to determine weights in a model based on the feature amount. Various models such as a linear model, a generalized linear model, a decision tree method, and a random forest can be applied as the model here. Alternatively, the model may be one using a neural network, for example, an RNN (Recurrent Neural Network) or a method developed from these. The machine learning here may be, for example, supervised learning, and correct answer data may be assigned to the feature amount. The correct answer data is, for example, a tag indicating whether or not the document data is related to a given case. For example, the learning processing unit 241 performs a process of estimating a likely weight with the feature amount as an explanatory variable and the correct answer data as an objective variable. The learning processing unit 241 outputs the model with the determined weights as a trained model. However, the trained model may be one that performs clustering processing, and the machine learning in this case may be unsupervised learning.
[0076] The inference processing unit 242 performs inference processing using the trained model output by the learning processing unit 241. Specifically, the inference processing unit 242 may obtain a score for the document data to be subjected to the inference processing by inputting a feature amount of the document data to the trained model.
[0077] For example, the import unit 260 may acquire, as export data, feature amounts of document data to be subjected to inference processing. The inference processing unit 242 inputs the feature amounts into the trained model. In this manner, the inference processing unit 242 can execute processing using the trained model on the document data acquired by the first information processing device 100. Specifically, the inference processing unit 242 executes classification processing to obtain information such as a score indicating the degree of association between the document data to be processed and a given case.
[0078] The import unit 160 of the first information processing device 100 imports the scores and the like obtained by the inference processing unit 242. The processing unit 140 of the first information processing device 100 performs processing to display the scores, for example.
[0079] FIG. 10 is a sequence diagram for explaining the flow of processing in this case. In step S201, the acquisition unit 110 of the first information processing device 100 acquires document data to be subjected to inference processing (for example, classification processing). For example, the acquisition unit 110 may determine document data to be subjected to classification processing based on an input from a user who uses the first information processing device 100. Alternatively, the acquisition unit 110 may automatically select one or more document data as document data to be subjected to classification processing based on the type, creator, creation date and time of the document data, etc. For example, when the method of this embodiment is used for auditing emails, the acquisition unit 110 of the first information processing device 100 may set all emails sent and received by the auditee within a predetermined period as document data to be subjected to classification processing.
[0080] In step S202, the analysis processing unit 120 performs a morphological analysis of the document data. In step S203, the feature amount determining unit 130 determines a feature amount based on the result of the morphological analysis and a hash function.
[0081] In step S204, the export unit 150 exports the feature amount to the second information processing device 200. Note that before step S204, a process in which the export data determination unit determines the feature amount as export data may be performed.
[0082] In step S205, the inference processing unit 242 of the second information processing device 200 obtains an output of the model by inputting the feature amount transmitted from the first information processing device 100 to an existing trained model. The export unit 250 exports the output of the model to the first information processing device 100.
[0083] In step S206, the import unit 160 of the first information processing device 100 imports the output of the trained model and outputs it to the processing unit 140. In this way, it is possible to prevent leakage of confidential information when processing related to document data acquired by the first information processing device 100 is executed in an external device.
[0084] The process executed by the second information processing device 200 using the trained model is not limited to classification process, and may be clustering process, keyword extraction process, summarization process, or the like. For example, the learning processing unit 241 of the second information processing device 200 creates a trained model for clustering. The inference processing unit 242 performs clustering of document data by inputting the feature amount acquired by the import unit 260 into the trained model. Alternatively, the learning processing unit 241 creates a trained model for keyword extraction, and the inference processing unit 242 extracts keywords from document data according to a given extraction condition by inputting the feature amount acquired by the import unit 260 into the trained model. The extraction condition here may be, for example, a word set in advance and a word similar thereto, or may be a different condition. Alternatively, the learning processing unit 241 creates a trained model for summarization process, and the inference processing unit 242 outputs summarized document data that summarizes document data by inputting the feature amount acquired by the import unit 260 into the trained model.
[0085] <Example 2> The second information processing device 200 may use the feature amount exported from the first information processing device 100 in the learning process. In this case, the first information processing device 100 may export the feature amount as data for learning. The second information processing device 200 performs machine learning based on the feature amount imported from the first information processing device 100 to create a trained model. The import unit 160 of the first information processing device 100 may acquire the created trained model itself as a processing result, or may acquire the output of the created trained model as a processing result.
[0086] 11 is a diagram showing a configuration example of the information processing system 1 in this case, and in particular, a diagram showing another configuration example of the second information processing device 200. The second information processing device 200 includes an acquisition unit 210, an analysis processing unit 220, a feature amount determination unit 230, a learning processing unit 241, an inference processing unit 242, an export unit 250, and an import unit 260, as in FIG. 9. The difference from FIG. 9 is that the feature amount acquired by the import unit 260 from the first information processing device 100 is input to the learning processing unit 241. In addition, when the second information processing device 200 does not acquire raw document data (when acquiring learning data only from the outside), the acquisition unit 210, the analysis processing unit 220, and the feature amount determination unit 230 in FIG. 11 may be omitted.
[0087] FIG. 12 is a sequence diagram for explaining the flow of processing in this case. In step S301, the acquisition unit 110 of the first information processing device 100 acquires document data, which is data for learning. For example, the acquisition unit 110 may execute annotation processing for associating correct answer data with the document data. The annotation processing is performed, for example, by a user checking the details of the document data and then inputting whether or not the document data is related to a given case. Note that the annotation processing may be performed by a device other than the first information processing device 100. Furthermore, when unsupervised learning is performed, annotation is omitted.
[0088] In step S302, the analysis processing unit 120 performs a morphological analysis of the document data. In step S303, the feature determining unit 130 determines a feature based on the result of the morphological analysis and a hash function. In step S304, the export unit 150 exports the feature to the second information processing device 200 as export data.
[0089] In step S305, the learning processing unit 241 of the second information processing device 200 creates a learned model by performing machine learning using the feature amount transmitted from the first information processing device 100 as at least a part of the learning data. In this embodiment, since a wide range of known methods can be applied to the machine learning, detailed description will be omitted.
[0090] In this way, when the second information processing device 200 generates a trained model by machine learning based on features, the export unit 150 of the first information processing device 100 may output the features as export data, and the import unit 160 may acquire the trained model created using the features or information based on the trained model as a processing result. In this way, the first information processing device 100 can also execute processing using the results of machine learning.
[0091] For example, the import unit 160 may acquire the trained model itself as a processing result. In this way, by acquiring the trained model itself, it becomes possible to execute an inference process using the trained model in the first information processing device 100. For example, the processing unit 140 processes the document data by inputting the feature amount of the document data to be processed into the trained model. The process executed by the first information processing device 100 using the trained model can include various processes such as classification process, clustering process, keyword extraction process, and summarization process, as in the example of the second information processing device 200 described above.
[0092] 12 corresponds to a processing example in the case of importing a trained model. For example, in step S306, the import unit 160 of the first information processing device 100 imports the trained model created by the learning processing unit 241. The imported trained model is stored in, for example, a storage unit (not shown) of the first information processing device 100.
[0093] In step S307, the acquisition unit 110 acquires document data for inference, which is document data to be subjected to inference processing. In step S308, the analysis processing unit 120 performs a morphological analysis of the document data for inference. In step S309, the feature amount determination unit 130 determines a feature amount based on the result of the morphological analysis and a hash function.
[0094] In step S310, the processing unit 140 reads out the trained model from a storage unit (not shown), and inputs the feature amount calculated in step S309 to the trained model. In step S311, the processing unit 140 outputs the result of processing using the trained model. For example, as shown in FIG. 12, the processing unit 140 may perform classification processing of document data and output the result. Specifically, when the trained model outputs a score representing the degree of relevance between the document data and a given case, the processing unit 140 performs processing to display the score, etc.
[0095] However, the import unit 160 is not limited to importing the trained model itself. For example, the import unit 160 may import the result of processing performed by the inference processing unit 242 of the second information processing device 200 based on a trained model created based on the feature amount exported by the first information processing device 100, after which the trained model is created based on the feature amount exported by the first information processing device 100. For example, as described above with reference to Figs. 9 and 10, the export unit 150 of the first information processing device 100 may export the feature amount of document data to be newly subjected to inference processing, and the import unit 160 may import the processing result based on the feature amount (such as a score representing the classification processing result).
[0096] Alternatively, the export unit 150 may output the search parameters generated based on the feature amount to the second information processing device 200 as export data. According to the method of the present embodiment, the first information processing device 100 can not only simply provide learning data, but also specify the contents of the search process based on the learning data. Therefore, it is possible to cause the second information processing device 200 to execute the search process according to the intention of the user of the first information processing device 100.
[0097] The search parameters here are obtained, for example, by statistical processing of a plurality of document data acquired by the first information processing device 100. For example, the search parameters may be information representing attributes common to document data including a predetermined morpheme. The predetermined morpheme here may be a morpheme designated by a user, a morpheme having a large tf-idf (Term Frequency-Inverse Document Frequency) value, or a morpheme determined from other conditions. The attributes of the document data may be determined from metadata attached to the document data. For example, the type of document data, the creator, the creation date, etc. may be used as the attributes. For example, the inference processing unit 242 of the second information processing device 200 may perform classification processing by using a trained model, and then output data determined to be related to a given event and having attributes designated by the search parameters as a processing result.
[0098] Furthermore, when the trained model outputs a score, the search parameter may be a threshold value used in a comparison process with the score. For example, the inference processing unit 242 of the second information processing device 200 may obtain a score of the document data to be inferred by using the trained model, and output data in which the score exceeds a threshold value specified by the search parameter as a processing result. For example, the storage unit of the first information processing device 100 may store statistics such as a distribution range of the score for each metadata of the document data. The statistics are obtained, for example, based on past processing results. The threshold value may be obtained based on the metadata assigned to the document data to be processed and the statistics stored in the storage unit.
[0099] The search parameters may also be information that instructs a change in the characteristics of the trained model (for example, weights determined by a learning process). For example, the inference processing unit 242 of the second information processing device 200 may update the trained model based on the search parameters, and output a processing result using the trained model after the update. The weights to be changed and the amount of change in the weights may be calculated based on metadata added to the document data to be processed and past history (statistics, etc.) stored in the storage unit, as in the above example of the threshold value. In addition, various modifications of the search parameters and the processing using them are possible.
[0100] Steps S312-S315 in Fig. 12 correspond to a processing example in the case of outputting search parameters. For example, in step S312, the export unit 150 of the first information processing device 100 outputs the search parameters and the feature amounts of the inference document as export data to the second information processing device 200. For example, although not shown in Fig. 12, the export data determination unit may perform a process of determining search parameters based on the feature amounts.
[0101] In step S313, the processing unit 240 (the inference processing unit 242 in a narrow sense) of the second information processing device 200 performs processing based on the trained model and the search parameters. The export unit 250 of the second information processing device 200 exports the processing result to the first information processing device 100.
[0102] In step S314, the import unit 160 of the first information processing device 100 imports the processing result based on the trained model and the search parameters. In step S315, the processing unit 140 outputs the imported processing result.
[0103] <Example 3> The first information processing device 100 may also perform a learning process to create a trained model. In this case, the first information processing device 100 may export the feature amount as analysis data for the trained model. The second information processing device 200 executes an analysis process based on the feature amount acquired from the first information processing device 100. The analysis process here may include processes such as data assessment, label assessment, cleansing, teacher data adjustment, and AI model type selection. Note that these processes are widely used methods in the field of machine learning, so detailed explanations will be omitted. The import unit 160 of the first information processing device 100 may acquire the analysis result in the second information processing device 200 as a processing result.
[0104] 13 is a diagram showing a configuration example of the information processing system 1 in this case, and in particular a diagram showing another configuration example of the first information processing device 100. As shown in Fig. 13, the first information processing device 100 includes an acquisition unit 110, an analysis processing unit 120, a feature amount determination unit 130, a processing unit 140, an export unit 150, and an import unit 160. The processing unit 140 may include a learning processing unit 141 and an inference processing unit 142.
[0105] The learning processing unit 141 performs machine learning to determine weights in a model based on the feature amount. The learning processing unit 141 outputs the model with the determined weights as a trained model. The inference processing unit 142 performs inference processing using the trained model output by the learning processing unit 141.
[0106] 14 is a sequence diagram for explaining the flow of processing in this case. In step S401, the acquisition unit 110 of the first information processing device 100 acquires document data that is data for learning. For example, the acquisition unit 110 may execute annotation processing for associating the document data with correct answer data.
[0107] In step S402, the analysis processing unit 120 performs a morphological analysis of the document data. In step S403, the feature determining unit 130 determines a feature based on the result of the morphological analysis and a hash function. In step S404, the processing unit 140 (the learning processing unit 141) creates a learned model by performing machine learning based on the feature.
[0108] In step S405, the export unit 150 outputs the feature amounts as export data to the second information processing device 200. More specifically, the export unit 150 outputs the feature amounts determined by the feature amount determining unit 130 as data for analysis processing (adjustment processing, analysis processing). Note that the export data is not limited to the feature amounts themselves, and may include statistical information and metadata about the feature amounts.
[0109] In step S406, the processing unit 240 of the second information processing device 200 executes an analysis process based on the feature amount. Since the feature amount here is training data for machine learning, the second information processing device 200 can determine whether the machine learning is being performed based on appropriate data, and analyze the necessary adjustment contents when the data is inappropriate. The export unit 250 of the second information processing device 200 exports the result of the analysis process to the first information processing device 100.
[0110] In step S407, the import unit 160 imports the results of the analysis process. In step S408, the learning processing unit 141 updates the trained model based on the results of the analysis process. For example, the learning processing unit 141 may update the learning data or the model type based on the results of the analysis process, and then execute the learning process again.
[0111] In step S409, the inference processing unit 142 performs processing based on the updated trained model and outputs the results. Note that, although an example in which the learning processing is executed (step S404) before the export of the feature amount (before step S405) is shown in FIG. 14, the present invention is not limited to this. It is sufficient that the trained model is acquired before the update processing based on the processing result (step S408), and the processing order of the learning processing and the output processing of the export data may be reversed, or may be executed in parallel.
[0112] As described above, the processing unit 140 (learning processing unit 141) of the first information processing device 100 may perform a learning process (step S404) for generating a trained model by machine learning based on the feature amount determined by the feature amount determining unit 130, and a process (step S408) for updating the trained model based on the processing result. Then, the import unit 160 acquires the result of the analysis process based on the export data as the processing result (step S407).
[0113] According to the method of the present embodiment, when the first information processing device 100 performs machine learning, an analysis related to the machine learning can be executed in an external device (the second information processing device 200). At this time, the information provided to the second information processing device 200 is a feature amount, so that leakage of confidential information can be suppressed. For example, the user of the first information processing device 100 can use an external analysis application without leaking his / her own confidential information.
[0114] Furthermore, when the second information processing device 200 performs an analysis related to machine learning, the export data output by the first information processing device 100 is not limited to feature quantities. For example, the export unit 150 may output a trained model generated by the processing unit 140 (learning processing unit 141) as the export data. In this case, the second information processing device 200 can acquire an analysis process based on information such as the model type and weights set by the learning process by acquiring the trained model itself. The import unit 160 acquires the result of the analysis process related to the trained model as a processing result. In this case, when the first information processing device 100 performs machine learning, it becomes possible to execute the analysis related to the machine learning in an external device without leaking confidential information.
[0115] Furthermore, the export unit 150 may output, as export data, both the feature amounts determined by the feature amount determining unit 130 and the trained model generated by the processing unit 140 (the training processing unit 141).
[0116] 3. Variations Some modified examples will be described below.
[0117] 3.1 Other system configuration examples 8, the US server 20 functions as the first information processing device 100, and the Korean server 40 (and the Japanese server 10 and the European server 30) functions as the second information processing device 200 that is an external device of the first information processing device 100. However, the method of the present embodiment is not limited to this.
[0118] Fig. 15 is a diagram showing another example of the configuration of the information processing system 1. As in Fig. 8, the information processing system 1 includes a Japanese server 10, an American server 20, a European server 30, and a Korean server 40. Also, as in Fig. 8, the server of each country stores raw document data acquired in that country as confidential information.
[0119] 15, the Japanese server 10 transmits features obtained after hashing document data acquired in Japan to the U.S. server 20, the European server 30, and the Korean server 40. Similarly, the U.S. server 20, the European server 30, and the Korean server 40 also transmit features obtained after hashing document data acquired in their own countries to the other three servers.
[0120] The Japanese server 10 acquires features based on document data acquired in the United States, Europe, and Korea, and is therefore capable of executing processing targeting all document data in Japan, the United States, Europe, and Korea. The same is true for the US server 20, the European server 30, and the Korean server 40. For example, the Japanese server 10, the US server 20, the European server 30, and the Korean server 40 can each execute learning processing targeting document data acquired in multiple countries or regions, and inference processing using a trained model created by the learning processing.
[0121] The Japanese server 10 may also transmit the results of the executed processing to the US server 20, the European server 30, and the Korean server 40. The information transmitted here may be image data or metadata of document data acquired in Japan, or the results of analysis processing (data assessment, label assessment, cleansing, teacher data adjustment, AI model type selection) of document data acquired outside Japan, or other information. Similarly, each of the US server 20, the European server 30, and the Korean server 40 transmits the processing results to the other servers.
[0122] As described above, in the example of Fig. 15, the Japanese server 10, the American server 20, the European server 30, and the Korean server 40 can each function as the first information processing device 100 and can also function as the second information processing device 200. In the information processing system 1 of Fig. 15, each device can convert confidential information into non-confidential information and transmit and receive it between each other, and can transmit and receive processing results based on the non-confidential information between each other, making it possible to flexibly change the system configuration and processing procedures.
[0123] For example, by having the servers in each country execute different processes and share the results of the processes, it is possible to execute various processes at high speed. Specifically, the servers in each country may set cases based on the legal system of the country and execute a process to determine the relevance of the cases to document data. In this case, no matter which country a problem such as a lawsuit occurs in, the information processing system 1 can support an appropriate response. Alternatively, it is possible to increase the reliability of the process by executing the same process on servers in multiple countries.
[0124] 8 and 15, an example in which the first information processing device 100 and the second information processing device 200 are located in different countries or regions has been described, but the information processing system 1 of the present embodiment is not limited to this. For example, the first information processing device 100 and the second information processing device 200 may be devices managed by different companies. For example, in the example of FIG. 13, a given first company performs machine learning using the first information processing device 100, but does not obtain the expected accuracy. In this case, the first company considers requesting analysis from a second company that specializes in machine learning analysis, but since the document data for learning is confidential information, it is undesirable to leak it outside the company. In this respect, with the method of the present embodiment, the features after hashing and the trained model based on the features are transmitted to the second information processing device 200 used by the second company. Therefore, it becomes possible to appropriately use an application of an external company (in this case, a machine learning analysis application) without leaking the document data itself. In addition, the method of the present embodiment can be widely applied when the second information processing device 200 is in an unsecure environment for the first information processing device 100.
[0125] 3.2 Hash size As described above with reference to FIGS. 5-7, the feature determination unit 130 may determine control parameters including the output size of the hash function, and apply the hash function according to the control parameters to the result of the morphological analysis, thereby converting the document data into a tensor, which is a feature. Note that a tensor is one type of feature, and other types of features may be used. The feature determination unit here is the feature determination unit 130 of the first information processing device 100, but the same processing may also be performed in the feature determination unit 230 of the second information processing device 200. The first information processing device 100 and each unit thereof will be described below, but may be replaced with the second information processing device 200 and each unit thereof as appropriate.
[0126] Here, the output size of the hash function, i.e., the size of the tensor (the number of elements in FIG. 6), N, may be a value that does not depend on the content of the document data. For example, as described above, N is set to a value that is sufficiently larger than the expected number of morphemes. In this way, the format of the features will be constant regardless of the document data, making it possible to standardize the processing.
[0127] Furthermore, if the output size of the hash function is set (if a fixed tensor is used), the format of the features will not change even if a new morpheme is added. Therefore, there is no need to reload the processed document data or re-run the conversion to features. For example, when document data containing a new morpheme is added, it is possible to obtain the features of the entire document data by finding the features corresponding to the document data and merging the found features with the existing features.
[0128] In addition, by using stylized tensors, it becomes possible to execute the process of converting a large amount of document data into features in parallel. For example, if a large amount of document data is divided into n groups (n is an integer equal to or greater than 2) and features are calculated in parallel using n devices, it is possible to calculate the features of all document data by simply merging the outputs of the n devices. As a result, high-speed processing can be easily achieved.
[0129] However, while it is necessary to increase N in order to improve accuracy by suppressing morpheme collisions, it is necessary to reduce N in order to reduce the amount of calculations, and it is not easy for users to set an appropriate value for N that balances accuracy and amount of calculations.
[0130] Therefore, the feature determining unit 130 may perform a process of updating the control parameters including the output size of the hash function in the evaluation process of the trained model until the performance of the trained model is determined to be equal to or higher than a predetermined level. In this way, the setting of N, which is the output size of the hash function, can be automated, thereby reducing the burden on the user.
[0131] 16 is a flowchart for explaining the process of the first information processing device 100 in this case. First, in step S501, the acquisition unit 110 acquires document data. The document data here is document data for learning, and in the narrow sense, is document data to which correct answer data has been added. In step S502, the analysis processing unit 120 performs a morphological analysis of the document data.
[0132] In step S503, the feature determining unit 130 determines formal parameters as control parameters including the value of N, which is the output size of the hash function. For example, the feature determining unit 130 selects one of a plurality of formal parameters that are preset based on the target language, the type of document data, etc. Alternatively, the formal parameters may be arbitrarily set by the user.
[0133] In step S504, the feature determining unit 130 converts the document data into features using the formal parameters. Specifically, the feature determining unit 130 obtains features by applying a hash function, whose output size is N included in the formal parameters, to the result of the morphological analysis.
[0134] In step S505, the learning processing unit 141 creates a trained model by performing machine learning based on the feature amount. In step S506, the learning processing unit 141 performs an evaluation process of the trained model created in step S505. For example, the learning processing unit 141 may obtain an index value representing the learning accuracy by performing cross-validation, and evaluate the trained model based on the index value. Cross-validation is a method of dividing multiple training data into k (k is an integer of 2 or more) units, updating weights as training data using k-1 units, and performing a process of obtaining the above-mentioned index value using the remaining unit as test data (validation data). Since cross-validation is a known method, a detailed description will be omitted. In addition, various index values such as recall, accuracy, precision, and AUC (Area Under the Curve) can be used as the index value here.
[0135] In step S507, the learning processing unit 141 determines whether the formal parameters are appropriate. For example, when the index value calculated in step S506 is equal to or greater than a given threshold, the learning processing unit 141 determines that the formal parameters are appropriate.
[0136] If it is determined that the formal parameters are not appropriate (step S507: No), the first information processing device 100 returns to step S503 and continues the process. That is, the feature amount determination unit 130 performs a process of updating the formal parameters. Specifically, the feature amount determination unit 130 may perform a process of selecting any one of the multiple formal parameters that has not been selected. Alternatively, the feature amount determination unit 130 may update the formal parameters by performing an operation (e.g., addition or multiplication of a predetermined value) on N before the update according to a predetermined rule.
[0137] The process after updating the formal parameters is the same as the above example, and the feature amount is determined using the formal parameters (step S504), a learning process (step S505), and an evaluation process (step S506) are executed.
[0138] If it is determined that the formal parameters are appropriate (step S507: Yes), the first information processing device 100 determines the formal parameters at that time as the control parameters, and in subsequent processes, the determined control parameters are used to execute the processes.
[0139] 3.3 Special characters Morphological analysis may result in the detection of symbols or garbled characters, but in the past these were treated as noise and removed. Symbols may be periods (.) or commas (,) in Japanese, commas (,), periods (.), colons (:), etc. in English, or other symbols. Garbled characters refer to characters that are not properly represented due to differences in character codes, etc.
[0140] However, if symbols or garbled characters are output as morphemes, it is possible that they were determined to be morphemes according to some rule. Therefore, the feature determination unit 130 may convert document data into tensors by applying a hash function to special characters, including at least one of symbols and garbled characters, from the results of morphological analysis. In this way, it becomes possible to include information about special characters in the features. Since the types of input information are increased, it becomes possible to improve processing accuracy.
[0141] In this embodiment, whether or not to use special characters in the process may be treated as a control parameter. For example, in the process described above with reference to FIG. 16, the control parameter may include a parameter that determines whether or not to use special characters. In this way, it is possible to automatically determine whether or not to use special characters, thereby reducing the burden on the user and improving accuracy.
[0142] 3.4 Metadata In the above, an example has been described in which metadata of features is included in export data, and an example in which metadata of given document data is output when the document data is output as a processing result of the second information processing device 200. The metadata here is treated as data different from the features, for example. However, the method of the present embodiment is not limited to this, and the features may include metadata of the document data.
[0143] The metadata in this embodiment includes, for example, the number of characters and the number of lines of the document data, and their distribution and statistics (average, median, standard deviation, etc.). The document data in this embodiment may also be data obtained by transcribing a conversation between multiple people. For example, the acquisition unit 110 may acquire voice data of a recorded conversation and perform a voice recognition process on the voice data to acquire the document data. In this case, the metadata of the document data includes the number of spoken characters, the number of spoken lines, the speaking time, etc. of each person. For example, if the document data corresponds to a conversation between a customer and an employee, the number of spoken characters of the customer, the number of spoken characters of the employee, the time distribution, etc. are used as the metadata. In addition, the ratio of the number of spoken characters of the customer or the number of spoken characters of the employee to the total number of spoken characters may also be used as the metadata. For example, the file path name where the document data was stored and the date and time of the email exchange may be included.
[0144] The feature amount determination unit 130 determines metadata features, which are features related to the metadata, based on the metadata added to the document data. Then, both the feature amount corresponding to the result of the morphological analysis described above and the metadata feature amount may be defined as new feature amounts. For example, the export unit 150 outputs export data, which is data based on the feature amount corresponding to the result of the morphological analysis and the metadata feature amount. In this way, it becomes possible to use not only morpheme information but also metadata information in various processes using the feature amounts described above. Since it is possible to use a variety of information, it becomes possible to improve the processing accuracy, etc.
[0145] For example, the learning processing unit 141 (or the learning processing unit 241) performs machine learning based on features corresponding to morphemes and metadata features. In this way, metadata different from morphemes can be included in the features, which makes it possible to improve the learning accuracy.
[0146] However, metadata values can vary greatly for each data set. For example, the number of spoken characters tends to be larger than the number of lines of speech. Also, the value of the speaking time can change depending on whether seconds or minutes are used as the unit. Therefore, if metadata values are used directly as features, the learning model may be strongly influenced by features with large values, and it may not be possible to learn all features evenly.
[0147] Therefore, in this embodiment, instead of the metadata value itself, the post-correction information may be used as the metadata feature. For example, consider a case where first to P pre-correction feature values are acquired as pre-correction feature values corresponding to the metadata, and first to Q documents are acquired as document data. The number of types of feature values corresponding to the metadata is P, and the number of document data is Q. Here, P and Q are each integers equal to or greater than 1. However, since it is expected that the number of types of metadata and the number of document data are both plural in practice, P and Q may each be integers equal to or greater than 2.
[0148] The feature amount determination unit 130 may determine the metadata feature amount by correcting the first to P-th pre-correction feature amounts based on the number P of pre-correction feature amounts, the number Q of document data, a first norm obtained from the i-th pre-correction feature amount (i is an integer between 1 and P) appearing in the first to Q-th documents, and a second norm obtained from the first to P-th pre-correction feature amounts appearing in the j-th document (j is an integer between 1 and Q). In this way, it is possible to properly normalize the metadata feature amount. Specifically, the correction based on the first norm can suppress the difference in values between metadata. Furthermore, by performing the correction based on the second norm, it is possible to align information (for example, the sum of squares) corresponding to the sum of the feature amounts for each document. As a result, the format of the feature amount obtained is the same as that in the case of only linguistic information (morphemes), so that even when metadata is used, learning processing and the like can be performed by the same processing as for linguistic information.
[0149] 17A to 17C are diagrams specifically explaining the correction process (normalization process) of metadata features. FIG. 17A shows metadata features before correction. The metadata features before correction are, for example, the metadata values themselves. Here, an example will be explained in which four types of metadata features and three document data, documents 1 to 3, are the targets. That is, an example will be considered in which P=4 and Q=3.
[0150] As shown in FIG. 17A, the values of metadata feature 1 are 100, 300, and 500 in documents 1 to 3, respectively. The values of metadata feature 2 are 3, 2, and 1 in documents 1 to 3, respectively. The values of metadata feature 3 are 5000, 300, and 1 in documents 1 to 3, respectively. The values of metadata feature 4 are 0, 1, and 0 in documents 1 to 3, respectively. In this example, the influences of metadata features 1 and 3 are relatively strong. ||L2|| in FIG. 17A represents the L2 norm, which is the square root of the sum of squares. The L2 norm in the vertical direction corresponds to the first norm, since it is a norm calculated from the i-th pre-correction feature (i is an integer between 1 and P) that appears in the first to Q-th documents.
[0151] FIG. 17B is a diagram for explaining the correction process using P, Q, and the first norm. FIG. 17B is a diagram showing the result of performing the correction process represented by the following formula (1) for each element in FIG. 17A. For example, for metadata feature 1, this is the result of multiplying by (1 / 591)×√(3 / 4). By performing the correction process shown in FIG. 17B, the squared value of the L2 norm in the vertical direction is unified to 0.75 for all metadata features. This makes it possible to suppress the influence of the scale of each metadata feature.
[0152]
number
[0153] Also, the horizontal L2 norm in Fig. 17B corresponds to the second norm since it is a norm found from the first to Pth pre-correction features appearing in the jth document (j is an integer between 1 and Q). As is clear from Fig. 17B, there is a variation in which the metadata feature values of document 1 are generally large, while the metadata feature values of document 3 are generally small. In this embodiment, the variation may be suppressed by performing a correction process using the second norm.
[0154] FIG. 17C is a diagram for explaining the correction process using the second norm. FIG. 17C is a diagram showing the result of performing the correction process of multiplying each element in FIG. 17B by 1 / (second norm). For example, for document 1, this is the result of multiplying each element by (1 / √1.25). By performing the correction process shown in FIG. 17C, the squared value of the horizontal L2 norm is unified to 1 for all documents. This makes it possible to make the format of the feature equivalent to that when only language information is targeted. The correction process for metadata features can be summarized as the following formula (2).
[0155]
number
[0156] Although the present embodiment has been described in detail as above, it will be easily understood by those skilled in the art that many modifications are possible that do not substantially deviate from the novel matters and effects of the present embodiment. Therefore, all such modifications are intended to be included in the scope of the present disclosure. For example, a term described at least once in the specification or drawings together with a different term having a broader meaning or synonymy can be replaced with that different term anywhere in the specification or drawings. All combinations of the present embodiment and modifications are also included in the scope of the present disclosure. Furthermore, the configurations and operations of the information processing system, information processing device (first information processing device), second information processing device, etc. are not limited to those described in the present embodiment, and various modifications are possible. [Explanation of symbols]
[0157] 1...information processing system, 10...Japan server, 20...US server, 30...European server, 40...Korea server, 100...first information processing device, 110...acquisition unit, 120...analysis processing unit, 130...feature amount determination unit, 140...processing unit, 141...learning processing unit, 142...inference processing unit, 150...export unit, 160...import unit, 200...second information processing device, 210...acquisition unit, 220...analysis processing unit, 230...feature amount determination unit, 240...processing unit, 241...learning processing unit, 242...inference processing unit, 250...export unit, 260...import unit
Claims
1. an acquisition unit for acquiring document data; an analysis processing unit that performs a morphological analysis of the document data; a feature determining unit that determines a feature corresponding to the document data based on a result of the morphological analysis and a hash function; an export unit that outputs export data, which is data based on the feature amount, to a second information processing device; an import unit that acquires a processing result based on the export data in the second information processing device; a processing unit that performs processing based on the processing result acquired by the import unit; An information processing device comprising:
2. In claim 1, The export unit includes: outputting the feature amount as the export data; The second information processing device is A trained model is generated by machine learning based on the feature amount; The import unit is Obtaining the trained model or information based on the trained model as the processing result; The processing unit includes: An information processing device that processes document data to be processed, which is document data to be processed, by inputting the feature amounts of the document data to be processed into the trained model.
3. In claim 2, The export unit includes: outputting, as the export data, search parameters generated based on the feature amount; The import unit is An information processing device that acquires information identified based on the search parameters and the trained model as the processing result.
4. In claim 1, The processing unit includes: A learning process for generating a trained model by machine learning based on the feature determined by the feature determination unit, and a process for updating the trained model based on a result of the learning process are performed; The import unit is The information processing device acquires a result of an analysis process based on the export data as the processing result.
5. In claim 4, The export unit includes: Outputting the trained model generated by the processing unit as the export data; The import unit is An information processing device that acquires a result of the analysis processing regarding the trained model as the processing result.
6. In claim 4, The export unit includes: outputting the feature amount as the export data; The import unit is The information processing device acquires, as the processing result, a result of the analysis processing on the feature, which is training data for the machine learning.
7. In any one of claims 1 to 6, The feature amount determining unit is determining a metadata feature that is a feature corresponding to the metadata based on the metadata assigned to the document data; The export unit includes: an information processing device that outputs the export data, which is data based on the feature amount corresponding to the result of the morphological analysis and the metadata feature amount;
8. In any one of claims 2 to 6, The feature amount determining unit is An information processing device that determines control parameters including an output size of the hash function, and converts the document data into a tensor, which is one of the types of features, by applying the hash function in accordance with the control parameters to the results of the morphological analysis.
9. In claim 8, The feature amount determining unit is An information processing device that performs a process of updating the control parameters including the output size of the hash function in a process of evaluating the trained model until the performance of the trained model is determined to be equal to or greater than a predetermined level.
10. In claim 8, The feature amount determining unit is An information processing device that converts the document data into the tensor by applying the hash function to special characters, including at least one of symbols and garbled characters, from the results of the morphological analysis.
11. A first information processing device; A second information processing device; Including, The first information processing device, an acquisition unit for acquiring document data; an analysis processing unit that performs a morphological analysis of the document data; a feature determining unit that determines a feature representing the document data based on a result of the morphological analysis and a hash function; an export unit that outputs export data, which is data based on the feature amount, to the second information processing device; an import unit that acquires a processing result based on the export data in the second information processing device; a processing unit that performs processing based on the processing result acquired by the import unit; An information processing system comprising:
12. An information processing device, Acquire document data, performing a morphological analysis of the document data; determining a feature quantity representing the document data based on the result of the morphological analysis and a hash function; exporting export data based on the feature amount to a second information processing device; importing a processing result based on the export data in the second information processing device; performing processing based on the imported processing results; Information processing methods.
Citation Information
Patent Citations
Terminal device, transmission method, transmission program and information processing system
JP2023027697A
Method and device for identifying object
US20200193207A1
Secure search method and secure search device
WO2013080365A1
Natural language processing system, natural language processing method, and natural language processing program
WO2015166606A1
Document information extraction system and document information extraction method
JP2022148430A
