Computer-implemented method, computer system, and computer program

A method converts sensitive words in text to non-sensitive words, enabling the use of external processing services for language models like GPT without exposing sensitive information, thus securing data privacy and compliance with organizational policies.

JP2024160621APending Publication Date: 2024-11-14横山 諒一

Patent Information

Application Number
JP2023075834
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-01
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

Language models like GPT are often operated and managed by third parties, making it difficult to use their processing services for texts containing corporate trade secrets, national secrets, or privacy-sensitive information without sending such data externally.

Method used

A computer-implemented method that converts sensitive words in the text into non-sensitive words, stores the correspondence data, and uses an external system to generate processed data based on these converted words, allowing the use of external processing services without transmitting the original sensitive data.

Benefits of technology

Enables the use of external processing services while keeping sensitive information secure by masking and converting sensitive words, ensuring that only non-sensitive data is transmitted and processed externally.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024160621000001_ABST
    Figure 2024160621000001_ABST
Patent Text Reader

Abstract

To make it possible to use processing services by external systems for even documents that may include corporate trade secrets, national secrets, privacy information, or the like.SOLUTION: A disclosed method includes generating second data by converting a first word included in first data into a second word, storing correspondence data between the first word and the second word in a storage device, transmitting the second data to an external system, receiving third data from the external system, and generating fourth data from the third data. The external system is configured to generate the third data including the second word based on the received second data, and to transmit the generated third data. Generating the fourth data includes generating the fourth data by converting the second word included in the third data into the first word based on the correspondence data.SELECTED DRAWING: Figure 4
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates to a computer-implemented method, a computer system and a computer program product. [Background technology]

[0002] Patent Document 1 discloses a document summarization method using a language model such as a Generative Pre-trained Transformer (GPT). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2023-053867 A Summary of the Invention

[0004] Language models such as GPT are expected to be used in various language processing applications such as summarization, document generation, and translation.

[0005] However, language models such as GPT are often operated and managed by third parties. Therefore, in order to use language models such as GPT, it is necessary to send text to an external system operated and managed by a third party. For example, when generating a summary of a text using an external system owned by a third party, it is necessary to send the text to be summarized to the external system.

[0006] However, it is not desirable to transmit text containing corporate trade secrets, national secrets, or information related to privacy to an external party, as it is difficult to use processing services such as language processing by an external system for such text.

[0007] Therefore, it is desirable to be able to use processing services provided by external systems even for documents that may contain the above-mentioned information.

[0008] One aspect of the present disclosure is a computer-implemented method executed by a computer, the method includes generating second data by converting a first word included in first data into a second word, storing correspondence data between the first word and the second word in a storage device, transmitting the second data to an external system, receiving third data from the external system, and generating fourth data from the third data, the external system being configured to generate the third data including the second word based on the received second data and transmit the generated third data, and generating the fourth data includes generating the fourth data by converting the second word included in the third data into the first word based on the correspondence data.

[0009] Another aspect of the present disclosure is a computer system or a computer program, the details of which are described in the following embodiments. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a configuration diagram of a computer system and an external system according to an embodiment. [Diagram 2] FIG. 2 is a configuration diagram of a computer system (document creation support system) according to the embodiment. [Diagram 3] FIG. 3 is a functional block diagram and a data flow diagram of an external system. [Figure 4] FIG. 4 is a flow chart showing the procedure for creating a document by a computer system. [Diagram 5] FIG. 5 is a diagram showing examples of the first data to the fourth data. [Figure 6] FIG. 6 is a diagram illustrating an example of the correspondence data. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] <1. Overview of computer-implemented method, computer system, and computer program>

[0012] (1) A method according to an embodiment may be a computer-implemented method executed by a computer. The computer-implemented method may include generating second data by converting a first word included in first data into a second word, storing correspondence data between the first word and the second word in a storage device, transmitting the second data to an external system, receiving third data from the external system, and generating fourth data from the third data.

[0013] The external system may be configured to generate the third data including the second word based on the received second data, and transmit the generated third data. Generating the fourth data may include generating the fourth data by converting the second word included in the third data into the first word based on the corresponding data.

[0014] According to the method of the embodiment, it is possible to utilize the processing services provided by the external system without transmitting data that is not desirable to be transmitted to the external system.

[0015] (2) The generating of the second data may include masking the first word included in the first data, and providing the first data in which the first word is masked to a language model to predict a masked word. The second word included in the second data may be a word predicted by the language model.

[0016] (3) Masking the first word included in the first data may include selecting a word to be masked from among the words included in the first data based on dictionary data indicating the words to be masked, and masking the selected word.

[0017] (4) Predicting the masked word may include transmitting the first data, in which the first word is masked, to a prediction system that predicts the masked word using the language model, and receiving the second data from the prediction system, the second data including the word predicted using the language model as the second word.

[0018] (5) The third data generated by the external system may include a summary of the second data.

[0019] (6) A system according to an embodiment may be a computer system configured to execute a process including: generating second data by converting a first word included in first data into a second word, storing correspondence data between the first word and the second word in a storage device, transmitting the second data to an external system, receiving third data from the external system, and generating fourth data from the third data.

[0020] The external system may be configured to generate the third data including the second word based on the received second data, and transmit the generated third data. Generating the fourth data may include generating the fourth data by converting the second word included in the third data into the first word based on the corresponding data.

[0021] A program according to the embodiment may be a computer program for causing a computer to execute a process, which may include generating second data by converting a first word included in first data into a second word, storing correspondence data between the first word and the second word in a storage device, transmitting the second data to an external system, receiving third data from the external system, and generating fourth data from the third data.

[0022] The external system may be configured to generate the third data including the second word based on the received second data, and transmit the generated third data. Generating the fourth data may include generating the fourth data by converting the second word included in the third data into the first word based on the corresponding data.

[0023] The computer program may be stored in a computer-readable, non-transitory storage medium.

[0024] 2. Examples of computer-implemented methods, computer systems, and computer programs

[0025] 1 shows computer systems 10A and 10B according to an embodiment. The computer systems 10A and 10B can execute processing related to text. The computer systems 10A and 10B according to the embodiment can use an external system 100.

[0026] The computer systems 10A and 10B may be installed in facilities 50A and 50B of users, such as individuals or organizations. The organizations may be, for example, companies, countries, local governments, professional offices, medical institutions, etc. The facilities 50A and 50B may be, for example, homes, buildings, offices, stores, hospitals, etc.

[0027] The computer systems 10A and 10B may be configured, for example, by a server computer. The computer systems 10A and 10B may be on-premise computers installed in facilities 50A and 50B of an individual or organization or in a location managed by an individual or organization, or may be cloud computers used by an individual or organization.

[0028] 1, as an example, computer system 10A is installed in a facility 50A of a first organization, and computer system 10B is installed in a facility 50B of a second organization separate from the first organization. Computer system 10A can be used only by those who belong to or are related to the first organization. Computer system 10B can be used only by those who belong to or are related to the second organization.

[0029] The computer systems 10A and 10B may be connected to the terminal 20 via a network 30. The network 30 may include, for example, a local area network (LAN). The LAN 30 is, for example, a computer network in facilities 50A and 50B. The terminal 20 may be used by a user. The terminal 20 is, for example, a personal computer, a smartphone, or a tablet. The user may use the computer systems 10A and 10B by operating the terminal 20. The user may directly operate the computer systems 10A and 10B. In other words, the computer systems 10A and 10B may also serve as the terminal 20.

[0030] In the following description, when there is no particular distinction between the multiple computer systems 10A, 10B, each of the computer systems 10A, 10B will be referred to as a "computer system 10."

[0031] The computer system 10 may be configured by a computer including a processor 11 and a storage device 12 connected to the processor 11. The computer system 10 may be configured by one or more computers. The storage device 12 includes, for example, a primary storage device and a secondary storage device (the same applies to the storage device 102 described below). The primary storage device is, for example, a RAM. The secondary storage device is, for example, a hard disk drive (HDD) or a solid state drive (SSD). The storage device 12 may include a computer program 12H (see FIG. 2) executed by the processor 11. The processor 11 reads and executes the computer program 12H stored in the storage device 12. The computer program 12H in the storage device 12 has program code indicating instructions for operating a computer as the computer system 10 according to the embodiment.

[0032] The external system 100 is composed of one or more computers. The computer constituting the external system includes a processor 101 and a storage device 102 connected to the processor 101. The storage device 102 may include a computer program executed by the processor 101. The processor 101 reads and executes the computer program stored in the storage device 102. The computer program in the storage device 102 has program code indicating instructions for causing a computer to operate as the external system 100 according to the embodiment.

[0033] The external system 100 may also function as a prediction system that predicts masked words using a language model. The external system 100 will be described in detail below.

[0034] As shown in FIG. 2, the computer system 10 of the embodiment may be, for example, a document creation support system. As an example, the computer system 10 of the embodiment may generate target text data 12G from initial text data 12C. The computer system 10 may support, for example, the creation of a summary sentence. That is, the computer system 10 may generate a summary of the initial text data 12C as the target text data 12G. Note that the computer system 10 may support the creation of other texts other than summaries. The generation of other texts may be, for example, a new text generated from the initial text data, a conversion of the style or appearance of the initial text data, a translation of the initial text data, etc.

[0035] As shown in Fig. 2, the storage device 12 may include dictionary data 12A. The dictionary data 12A may be used to select words to be masked by a masking process 11C described below. The dictionary data 12A may include words indicating corporate trade secrets, national secrets, or information related to privacy (hereinafter referred to as "sensitive information"). Of the words included in a sentence, the words included in the dictionary data 12A are masked by the masking process.

[0036] Words having sensitive information may include technical terms or industry terms in a specialized field to which a user such as an individual or an organization belongs. The dictionary data 12A may differ depending on the user such as an individual or an organization. Words included in the dictionary data 12A may be registered and edited by the user. The computer system 10 has multiple dictionary data 12A that may be used by the same user, and the user may select and use the dictionary data 12A according to the purpose, etc.

[0037] The dictionary data 12A may not only have a list of specific words to be masked, but also a list of "word types" to be masked. The "word types" here are, for example, "proper names" such as the names of people or organizations and place names, "titles," "words indicating date and time," "words indicating quantity," and "words indicating actions." The word types can be recognized by proper name recognition processing, or natural language processing such as recognition processing of words indicating date and time, recognition processing of words indicating quantity, and semantic analysis. For example, when the word "Yokoyama" indicating a person's name is present in a sentence, "Yokoyama" is recognized as a "person's name" by proper name recognition processing. In this case, if "person's name" is registered as a "word type" in the dictionary data 12A, "Yokoyama" is masked. In other words, even if the word "Yokoyama" itself is not registered in the dictionary data 12A, "Yokoyama" is masked.

[0038] The storage device 12 may include correspondence data 12B. The correspondence data indicates the correspondence between a portion masked by a masking process (mask word) and a complement of the masked portion (complement word). For example, when the word "Yokoyama" is masked and the masked portion is complemented by the word "Sato", "Yokoyama" and "Sato" are recorded in correspondence with each other in the correspondence data 12B.

[0039] The storage device 12 may store original data 12C (first data) which is initial data, and target data 12G (fourth data) generated by the computer system 10. The target data 12G is, for example, a summary of the original data 12C, as shown in Fig. 2. The original data 12C may include text data. The original data 12C may also include graphic data, image data, and the like other than text data.

[0040] The original data 12C may include a word indicating sensitive information (hereinafter, referred to as a "sensitive word"). Hereinafter, data including a sensitive word is referred to as "sensitive data". The original data 12C (first data) may be sensitive data. Sensitive data may be prohibited from being transmitted outside the computer system 10 due to organizational policies or protection of personal information, or the user may not wish to transmit the sensitive data outside the computer system 10.

[0041] The computer system 10 according to the embodiment can generate target data 12G based on original data 12C, which is sensitive data, by using an external system 100, without transmitting sensitive data to the outside. Note that the target data 12G (fourth data) may also be sensitive data.

[0042] The storage device 12 may store intermediate data 12D, 12E, and 12F that are generated from the initial data 12C to the target data 12G. The intermediate data 12D, 12E, and 12F may include, for example, mask data 12D, second data 12E, and third data 12F.

[0043] The masked data 12D is obtained by masking specific words contained in the original data 12C. In other words, the masked data 12D is a missing sentence in which a part of the original data 12C is masked. As described above, the masking is performed based on the dictionary data 12A. The masked words may be sensitive words. Since the sensitive words are masked, the masked data 12D becomes non-sensitive data.

[0044] The second data 12E is data generated from the first data 12C, which is the original data. In FIG. 2, the second data 12E is shown as complementary data 12E. The second data 12E is generated by converting a first word included in the first data 12C into a second word. The first word may be a sensitive word. The first word may be selected based on the dictionary data 12A. The second word may be a non-sensitive word. The second word may be a word that was not selected based on the dictionary data 12A.

[0045] In the embodiment, as an example, the second data 12E is generated based on the mask data 12D generated from the first data 12C. In this case, the conversion from the first word to the second word can be performed by masking the first word included in the first data 12C and complementing (predicting) the masked word. The complemented (predicted) word becomes the second word.

[0046] In this case, it is preferable that the completion or prediction is inaccurate, from the viewpoint of concealing sensitive information. In other words, the completion or prediction process may be sloppy. In addition, by increasing the number of masked words, the meaning of the sentence in the original data becomes difficult to understand, so that the accuracy of the completion or prediction is greatly reduced, and sensitive information can be concealed.

[0047] In addition, since the complemented or predicted word (second word) may coincide with the masked first word by chance, the first word and the second word may only partially coincide if they do not completely coincide throughout the entire sentence. In other words, the second word corresponding to some words (first words) among the masked words (first words) may be the same as the first word. Even if the first word and the second word coincidentally share some words, if it is merely a coincidence, the second word does not indicate sensitive information, and therefore sensitive information is concealed in the sentence data as a whole.

[0048] In the second data 12E, a first word that is a sensitive word included in the first data 12C is converted into a second word that is a non-sensitive word. Therefore, the second data 12E becomes non-sensitive data. The second data 12E does not have sensitive words like the masked data 12D, but the masked words are complemented. Therefore, the meaning of the sentence is easier to understand in the second data 12E than in the masked data 12D. Therefore, natural language processing based on the second data 12E is more easily performed appropriately than natural language processing on the masked data 12D.

[0049] The third data 12F is generated from the second data 12E. In Fig. 2, the third data 12F is shown as a summary of the complementary data, which is the second data 12E. The third data 12F, which is a summary of the second data 12E, which is non-sensitive data, is also non-sensitive data.

[0050] The processor 11 is capable of executing various processes 11A, 11B, 11C, 11D, 11E, and 11F according to instructions of a computer program 12H.

[0051] The processing executed by the processor 11 may include request transmission 11A. Request transmission 11A is to request some processing to the external system 100. The request is transmitted from the computer system 10 to the external system 100. The request may include a request for completion processing, which will be described later, or a request for text processing. The processor 11 may transmit the request to the external system 100 together with data to be used for processing in the external system 100.

[0052] The processing performed by the processor 11 may include response reception 11B. Response reception 11B is to receive a response from the external system 100 to the above-mentioned request. The response may include the result of execution performed by the external system 100 based on the request. The response may include a notification of the processing result of the completion processing described below, or a notification of the processing result of the text processing processing. The response may be transmitted from the external system 100 together with data generated by the processing in the external system 100. The processor 11 may receive a response including the generated data from the external system 100.

[0053] The processing executed by the processor 11 may include a masking process 11C. The masking process 11C includes masking words included in the first data 12C. The processor 11 refers to the dictionary data 12A and selects words to be masked from among the words included in the first data 12C. The words to be masked are, for example, words included in the dictionary data 12A or words of a type matching the "word type" included in the dictionary data 12A. The masking process 11C generates masked data 12D in which some of the words included in the first data 12C are masked.

[0054] The processing executed by the processor 11 may include a processing 11D for acquiring the first data 12C, which is the original data. The acquisition processing 11D is to acquire the first data 12C from the outside. For example, the processor 11 receives the first data 12C including a sentence input to the terminal 20 via the network 30 as the acquisition processing 11D. The processor 11 also receives the first data 12C transmitted from another computer such as the terminal 20 as the acquisition processing 11D. The first data 12C may be stored in the storage device 12 in advance.

[0055] The processing executed by the processor 11 may include a data output processing 11E. The output processing 11E is, for example, outputting the data 12A, 12B, 12C, 12D, 12E, 12F, and 12G stored in the storage device 12 to the display of the computer system 10 or the display of the terminal 20. The output processing 11E allows the user to refer to the data 12A, 12B, 12C, 12D, 12E, 12F, and 12G. The output processing 11E may also be data transmission to another computer.

[0056] The processing executed by the processor 11 may include a data editing processing 11F. The editing processing 11F is, for example, to present to the user an environment in which the user edits the data 12A, 12B, 12C, 12D, 12E, 12F, and 12G stored in the storage device 12. For example, the processor 11 can accept editing operations on the data 12A, 12B, 12C, 12D, 12E, 12F, and 12G displayed by the output processing 11E, and store the edited data 12A, 12B, 12C, 12D, 12E, 12F, and 12G in the storage device 12.

[0057] The data editing may include, for example, editing of the original data 12C, editing of the generated mask data 12D, editing of the generated second data 12E, editing of the generated third data 12F, and editing of the generated fourth data 12G.

[0058] FIG. 3 shows the configuration of the external system 100. The external system 100 may be managed by a third party other than the user of the computer system 10 or an organization to which the user belongs or is related. Here, the "outside" in the external system 100 refers to a range to which the first data 12C of the computer system 10 should not be transmitted. The "outside" typically refers to a range outside the individual or organization that has the computer system 10, but is not limited thereto. For example, in a certain organization, when there are a first group and a second group for which information blocking measures have been taken, the second group may be "outside" from the perspective of the first group. In the following, the external system 100 may be simply referred to as the system 100.

[0059] The system 100 may include, for example, a language model 120. For example, the language model 120 is an AI language model that has been trained by machine learning. The language model is used for recognizing, generating, summarizing, and the like of a sentence. When a sentence is given to a language model implemented in a computer, it can generate a summary of the given sentence, generate a response to the given sentence, or generate a new sentence.

[0060] The language model 120 is, for example, a Generative Pretrained Transformer (GPT) 121. The GPT 121 is a sentence generation model. The language model 120 may be a Masked Language Model (MLM) 122. The masked language model is used for completion processing of masked sentences (filling in missing sentences). Note that since the GPT 121 is also capable of completion processing, there is no need to prepare a masked language model separate from the GPT 121.

[0061] GPT 121 is a language model trained by unsupervised learning using a large-scale data set. The data set for machine learning is stored in database 112.

[0062] The system 100 receives text data via the interface 111. The received text data is, for example, mask data 12D or complement data 12E. The system 100 inputs the received text data to the language model 120. The language model 120 generates and outputs an appropriate text based on the input text data.

[0063] The system 100 transmits the generated text data via the interface 111. The text data to be transmitted is, for example, the supplementary data 12E or a summary 12F of the supplementary data 12E.

[0064] The interface 111 is a function for the system 100 to communicate with the outside, and is implemented as, for example, an interface for chat. The interface 111 for chat accepts input of a sentence from outside the system 100, and outputs to the outside a sentence generated by the language model 120 based on the input sentence.

[0065] In the embodiment, as an example, the computer system 10 executes sending a request 11A for creating a sentence to an interface 111 for chat, and receives a sentence output from the interface 111 as receiving a response 11B to the request.

[0066] The system 100 can provide the received data to the language model 120 and can also store the data in the database 112. The system 100 can improve the functionality and performance of the language model 120 by re-running the machine learning of the language model 120 based on the stored data.

[0067] Fig. 4 shows a procedure for generating fourth data 12G, which is target data, from first data 12C, which is initial data. Fig. 3 also shows the data flow according to the procedure shown in Fig. 4. Fig. 5 and Fig. 6 show an example of data generated by the procedure shown in Fig. 4.

[0068] The procedure shown in Fig. 3 is executed by the processor 11, for example, using the external system 100. Here, for example, a summary 12G shown in Fig. 5 is generated from original data (original text) 12C shown in Fig. 5.

[0069] As shown in Fig. 4, the processor 11 first executes a masking process (step S41). The processor 11 generates masked data 12D by masking words selected based on the dictionary data 12A from among the words included in the first data 12C stored in the storage device 12. The masked data 12D is obtained by masking some of the words included in the first data 12C. In the masked data 12D shown in Fig. 5, the masked parts are indicated by "*". The unmasked parts are common to the masked data 12D and the first data 12C.

[0070] As shown in FIG. 5, the masked data 12D has the same sentence structure as the first data 12C, but sensitive words are masked, so that sensitive information is generally concealed.

[0071] The user may perform an editing operation on the generated mask data 12D to additionally mask words that are not masked in the generated mask data 12D, or to cancel masking and return masked words to their original state. By additional masking through editing, words that were not masked by the dictionary data 12A can be masked at the user's discretion. Also, by canceling masking through editing, masking of non-sensitive words can be removed. The edited mask data 12D is saved. When edited, the edited mask data 12D is used for the complementation process described below.

[0072] The processor 11 executes a complementation process on the mask data 12D (step S42). The complementation process is a process of predicting a masked portion in the mask data 12D (a process of filling in a worm-eaten sentence).

[0073] 3, as a complementation process, the processor 11 requests the external system 100, which serves as a prediction system for predicting a masked portion, to complement the mask data 12D, and receives the complemented data 12E generated by the external system 100. Since the mask data 12D is non-sensitive data that does not contain sensitive words, there is no problem in transmitting the mask data 12D to the external system 100.

[0074] When the interface of the external system 100 is a chat interface 111, the processor 11 of the computer system 10 generates a request message as a request for completion by adding a sentence requesting mask completion, such as "Please fill in the following sentence," before the sentence of the mask data 12D shown in FIG. 5.

[0075] The processor 11 transmits the generated request message to the interface 111 of the external system 100. The external system 100 provides the request message received via the interface 111 to the language model 120. The language model 120 performs a semantic analysis of the request message, and outputs a sentence (complementary data 12E in FIG. 5 ) that is mask-complemented (filling in the blank; predicting the word in the masked portion) of the sentence in the mask data 12D following the sentence "Please fill in the blank in the following sentence."

[0076] The external system 100 transmits the outputted complement data (second data) 12E to the computer system 10. The processor 11 of the computer system 10 receives the transmitted complement data 12E and stores it in the storage device 12. As shown in Fig. 5, the complement data 12E is obtained by complementing the masked portion of the mask data 12D with an appropriately predicted word, so that the structure of the sentence is the same as that of the original data 12C, but sensitive information is concealed. Furthermore, the complement data 12E is a more natural sentence than the mask data 12D, and is therefore suitable for semantic analysis by the external system 100.

[0077] The processor 11 generates the corresponding data 12B based on the original data 12C and the supplementary data 12E. The mask data 12D may be used to generate the corresponding data 12B if necessary. FIG. 6 shows an example of the corresponding data 12B. The corresponding data 12B shown in FIG. 6 is generated based on the supplementary data 12E and the mask data 12D shown in FIG. 5. In FIG. 6, the left side shows a list of words (first words) included in the first data 12C, and the right side shows a list of words (second words) included in the second data 12E. The corresponding data 12B shows a correspondence between the first word and the second word by which the first word is converted. For example, "Prime Minister Kishida" as the first word in the original data 12C is converted to "Mr. Tanaka" as the second word in the supplementary data 12E. In this case, the corresponding data 12B records "Prime Minister Kishida" as the first word and "Mr. Tanaka" as the second word in association with each other.

[0078] To generate the corresponding data 12B, the processor 11 compares the original data 12C with the complementary data 12E and identifies which word (second word) in the complementary data 12E has replaced a word (first word) in the original data 12C. Note that the mask data 12D may be used to identify the portion in the original data 12C and / or the complementary data 12E where the first word has been replaced with the second word.

[0079] When completing the masked data 12D, the language model 120 may also change the sentences other than the masked portions in order to make the sentences in the generated completion data 12E more natural. In this case, the processor 11 may record the changes to the sentences other than the masked portions in the correspondence data 12B.

[0080] The user may perform an editing operation on the generated complement data 12E. If the complemented word (second word) happens to be the same as the original first word, the user may change the complemented second word to another second word by the user editing operation. Also, if the sentence in the complement data 12E is unnatural, the user may edit the sentence by the user editing operation. The processor 11 also records the correspondence between the word (second word) in the complement data 12E after the user editing and the word (first word) in the original data 12C in the correspondence data 12B.

[0081] The processor 11 stores the generated corresponding data 12B in the storage device 12 (step S43).

[0082] 4 (step S42), the second data 12E may be generated directly from the first data 12C by an appropriate algorithm or a user editing operation. That is, the first word included in the first data 12C may be converted into the second word by an appropriate algorithm or a user editing operation.

[0083] Next, as shown in Fig. 4, the processor 11 executes a document processing process (step S44). As shown in Fig. 3, the processor 11 requests the external system 100 to process the second data 12E as the document processing process, and receives the third data 12F generated by the external system 100. Since the second data 12E is non-sensitive data that does not contain sensitive words, there is no problem in transmitting it to the external system 100.

[0084] When the text processing is the generation of a summary, the processor of the computer system 10 generates a request message as a request for generating a summary by adding a sentence requesting text processing (text summarization), such as "Please summarize the following text," before the text in the supplementary data 12E shown in FIG. 5.

[0085] The processor 11 transmits the generated request message to the interface 111 of the external system 100 to generate a summary. The external system 100 provides the request message received via the interface 111 to the language model 120. The language model 120 performs a semantic analysis of the request message, and outputs a sentence (summary 12F of the complementary data in FIG. 5) that summarizes the sentence in the second data 12E that follows the sentence "Please summarize the following sentence." Unlike the mask data 12D, the second data 12E is a natural sentence, so the language model 120 can appropriately perform a semantic analysis of the second data 12E, and as a result, can generate an appropriate summary.

[0086] The external system 100 transmits the output summary (third data) 12F to the computer system 10. The processor 11 of the computer system 10 receives the transmitted summary 12F and stores it in the storage device 12. The third data 12F, which is a summary of the second data 12E, which is non-sensitive data, is non-sensitive data, and therefore there is no problem in exchanging it between the external system 100 and the computer system 10.

[0087] The user may perform editing operations on the generated summary 12 F. The edited summary 12 F is stored in the storage device 12 .

[0088] The processor executes a restoration process on the third data 12F to generate the fourth data 12G (step S45). The restoration process is a process of converting a second word included in the third data 12F into a first word. The processor 11 performs the process of converting the second word into the first word by referring to the corresponding data 12B. The processor 11 identifies a word included in the third data 12F as a second word in the corresponding data 12B, and converts the word identified as the second word in the third data 12F into the first word associated with the second word in the corresponding data 12B. For example, the processor 11 identifies "Mr. Tanaka" as one of the second words among the words included in the third data 12F by referring to the corresponding data 12B. "Mr. Tanaka" is associated with "Prime Minister Kishida" in the corresponding data 12B. Therefore, the processor 11 converts "Mr. Tanaka" included in the third data 12F into "Prime Minister Kishida". Similarly, the processor 11 generates the fourth data 12G by converting each second word included in the third data 12F into a corresponding first word.

[0089] The processor 11 stores the fourth data 12G generated by the restoration process in the storage device 12. The user may perform an editing operation on the generated fourth data 12G. The edited fourth data 12G is stored in the storage device 12.

[0090] When the sentence processing in step S44 is translation, it is preferable that the corresponding data 12B includes translations of the first word and the second word. The processor 11 can generate the fourth data by converting the translation of the second word in the third data 12F, which is a translated sentence, into the translation of the first word by referring to the translations of the first and second words included in the corresponding data 12B.

[0091] The processor 11 may obtain the translation of the first word from the dictionary data 12A. In this case, the dictionary data 12A preferably also includes a translation of the word to be masked. The processor 11 may obtain the translation of the second word from the second data 12E before translation and the third data 12F after translation. The processor 11 may compare the second data 12E before translation with the third data 12F after translation, and extract a translation corresponding to the second word included in the second data 12E from the third data 12F. The processor 11 may store the obtained first word and second translation as the corresponding data 12B.

[0092] In this manner, converting the second word included in the third data into the first word to generate the fourth data 12G may include converting a translation of the second word into a translation of the first word.

[0093] According to the embodiment, it is possible to obtain the fourth data 12G obtained by processing the first data 12C, which is sensitive data, by using the external system 100 without transmitting the sensitive data to the external system 100.

[0094] The present invention is not limited to the above embodiment, and various modifications are possible. For example, the external system 100 that predicts the masked portion in step S41 and the external system 100 that processes the text such as summarizing in step S44 may be separate systems. [Explanation of symbols]

[0095] 10: Computer Systems 10A: Computer Systems 10B: Computer Systems 11: Processor 11A: Send request 11B: Response received 11C: Mask processing 11D: Acquisition process 11E: Data output process 11F: Data editing process 12:Storage device 12A: Dictionary data 12B: Corresponding data 12C: First Data 12D: Mask data 12E: Second data 12F: 3rd Data 12G: 4th data 12H: Computer Program 20: Terminal 30: Network 50A: Facilities 50B: Facilities 100: External system 101: Processor 102: Storage device 111: Interface 112: Database 120: Language Model 121: Generative Models 121: Masked Language Model

Claims

1. 1. A computer-implemented method for executing a computer, comprising: generating second data by converting a first word included in the first data into a second word; storing the correspondence data between the first word and the second word in a storage device; Transmitting the second data to an external system; receiving third data from the external system; generating fourth data from the third data; Including, The external system includes: generating the third data including the second word based on the received second data; Transmitting the generated third data It is configured as follows: generating the fourth data includes generating the fourth data by converting the second word included in the third data into the first word based on the corresponding data; Computer-implemented method.

2. Generating the second data includes: masking the first word included in the first data; providing the first data, with the first words masked, to a language model to predict masked words; Including, The second word included in the second data is a word predicted by the language model. The computer-implemented method of claim 1 .

3. Masking the first word included in the first data includes: selecting a word to be masked from among the words included in the first data based on dictionary data indicating the word to be masked; Mask selected words, Including, The computer-implemented method of claim 2 .

4. Predicting the masked words sending the first data, in which the first words are masked, to a prediction system that predicts the masked words using the language model; receiving the second data from the prediction system, the second data including the second words predicted using the language model; Including, The computer-implemented method of claim 2 .

5. The third data generated by the external system includes a summary of the second data. A computer-implemented method according to any one of claims 1 to 4.

6. 1. A computer system comprising: generating second data by converting a first word included in the first data into a second word; storing the correspondence data between the first word and the second word in a storage device; Transmitting the second data to an external system; receiving third data from the external system; generating fourth data from the third data; The method is configured to perform a process including: The external system includes: generating the third data including the second word based on the received second data; Transmitting the generated third data It is configured as follows: generating the fourth data includes generating the fourth data by converting the second word included in the third data into the first word based on the corresponding data; Computer system.

7. A computer program for causing a computer to execute a process, The process comprises: generating second data by converting a first word included in the first data into a second word; storing the correspondence data between the first word and the second word in a storage device; Transmitting the second data to an external system; receiving third data from the external system; generating fourth data from the third data; Including, The external system includes: generating the third data including the second word based on the received second data; Transmitting the generated third data It is configured as follows: generating the fourth data includes generating the fourth data by converting the second word included in the third data into the first word based on the corresponding data; Computer program.

Citation Information

Patent Citations

  • Method and system for summarizing document using hyperscale language model

    JP2023053867A

Cited By

  • Bearing assembly

    US12416333B2

  • Drive module assembly

    US20250075780A1