Knowledge delivery system, knowledge delivery method, and program

Abstract

To provide a knowledge sharing system that does not transmit confidential information to external parties, even when using AI generation services provided by external sources. [Solution] The knowledge provision system comprises: an inquiry acquisition unit that acquires the text of an inquiry; a search unit that searches for parts related to the text of the inquiry from knowledge information including confidential information; a conversion table storage unit that stores a conversion table containing confidential words related to confidential information and corresponding words associated with said confidential words; an inquiry conversion unit that uses the conversion table to convert the confidential words contained in the text of the inquiry and parts related to the text of the inquiry into corresponding words; an inquiry unit that inputs the text of the inquiry and parts related to the text of the inquiry converted by the inquiry conversion unit into a large-scale language model; and an answer conversion unit that acquires the answer to the input from the inquiry unit to the large-scale language model and uses the conversion table to convert the corresponding words contained in the answer into confidential words.

Description

Technical Field

[0001] The present invention relates to a knowledge providing system, a knowledge providing method, and a program.

Background Art

[0002] In recent years, Internet-based generative AI services using large language models (LLMs) such as ChatGPT (registered trademark) of OpenAI (registered trademark) and GEMINI (registered trademark) of Google have been provided. In these generative AI services, when a user sends a query text called a prompt, a response to the query is sent back.

[0003] Furthermore, there is also a technique called RAG (Retrieval-Augmented Generation) that improves the answer accuracy by combining external information retrieval with these large language models. In this RAG, information related to a query input by a user is retrieved from a database or documents, and a prompt combining the retrieval result and the query is input into the generative AI service.

[0004] Patent Document 1 discloses a search system that stores, as knowledge information, the results of machine learning of information useful to users such as information related to insurance and support information and how to use this search system, and receives questions from users, and presents answers created using generative AI technology from the knowledge information.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, when RAG is combined with a generative AI service provided externally, such as via the internet, there is a problem in that if user inquiries or search results from databases or documents contain confidential information such as company secrets or personal information, that confidential information may be transmitted externally.

[0007] This invention has been made in view of these circumstances, and provides a knowledge provision system, a knowledge provision method, and a program that do not transmit confidential information to external parties even when using an externally provided AI generation service. [Means for solving the problem]

[0008] This invention was made to solve the above-mentioned problems, and one aspect of the present invention is a knowledge provision system comprising: an inquiry acquisition unit that acquires an inquiry text; a search unit that searches for a portion related to the inquiry text from knowledge information including confidential information; a conversion table storage unit that stores a conversion table including confidential words related to the confidential information and corresponding words associated with the confidential words; an inquiry conversion unit that uses the conversion table to convert the confidential words contained in the inquiry text and the portion related to the inquiry text to the corresponding words; an inquiry unit that inputs the inquiry text and the portion related to the inquiry text converted by the inquiry conversion unit into a large-scale language model; and an answer conversion unit that acquires an answer to the input from the inquiry unit to the large-scale language model and uses the conversion table to convert the corresponding words contained in the answer into the confidential words.

[0009] Another aspect of the present invention is the knowledge provision system described above, comprising a conversion table creation unit that extracts the confidential words from the portion of the inquiry text and associates the corresponding words with the extracted confidential words to create the conversion table.

[0010] Another aspect of the present invention is the knowledge provision system described above, comprising: a knowledge information storage unit that stores the knowledge information; and a knowledge database generation unit that selects at least a portion of proper nouns, numerals, and verbs from the knowledge information as confidential words and stores them as confidential words in the knowledge information storage unit, wherein the conversion table creation unit creates the conversion table by referring to the knowledge information storage unit.

[0011] Another aspect of the present invention is the knowledge provision system described above, wherein the conversion table creation unit, if the extracted confidential word is a proper noun, uses the common noun of the proper noun, or another proper noun of the same type as the proper noun, as the corresponding word.

[0012] Another aspect of the present invention is a knowledge provision method comprising: a first step of inquiry to obtain the text of an inquiry; a second step of searching for a portion related to the text of the inquiry from knowledge information including confidential information; a third step of converting the confidential words contained in the text of the inquiry and the portion related to the text of the inquiry to the corresponding words using a conversion table that includes confidential words related to the confidential information and corresponding words associated with the confidential words; a fourth step of inputting the text of the inquiry and the portion related to the text of the inquiry, converted in the third step, into a large-scale language model; and a fifth step of obtaining a response to the input to the large-scale language model in the fourth step, and converting the corresponding words contained in the response into confidential words using the conversion table.

[0013] Another aspect of the present invention is a program for causing a computer to function as: a query acquisition unit that acquires a query text; a search unit that searches for a portion related to the query text from knowledge information including confidential information; a conversion table storage unit that stores a conversion table including confidential words related to the confidential information and corresponding words associated with the confidential words; a query conversion unit that uses the conversion table to convert the confidential words contained in the query text and the portion related to the query text into the corresponding words; a query unit that inputs the query text and the portion related to the query text converted by the query conversion unit into a large-scale language model; and a response conversion unit that acquires a response to the input from the query unit to the large-scale language model and uses the conversion table to convert the corresponding words contained in the response into the confidential words. [Effects of the Invention]

[0014] According to this invention, the knowledge provision system, knowledge provision method, and program do not transmit confidential information to external parties even when using an externally provided AI generation service. [Brief explanation of the drawing]

[0015] [Figure 1] This is a schematic block diagram showing the configuration of a knowledge provision system 100 according to one embodiment of the present invention. [Figure 2] This table shows an example of the contents of the conversion table in the same embodiment. [Figure 3] This is a flowchart illustrating the operation of the knowledge provision system 100 in the said embodiment. [Figure 4] This figure shows an example of data illustrating an embodiment of the knowledge provision system 100 in the same embodiment. [Modes for carrying out the invention]

[0016] Embodiments of the present invention will be described below with reference to the drawings. Figure 1 is a schematic block diagram showing the configuration of a knowledge provision system 100 according to one embodiment of the present invention. The knowledge provision system 100 returns answers to inquiries from users, such as members of an organization, within an organization such as a company. The knowledge provision system 100 is connected to a network such as the Internet and uses a generation AI service provided by a generation AI device 200 installed outside the organization. However, even if the inquiry from the user, or the answer to the inquiry, contains confidential information that is confidential to those outside the organization, the knowledge provision system 100 does not transmit the confidential information to the outside, i.e., the generation AI device 200. The generation AI device 200 is a large language model (LLM) such as ChatGPT® or GEMINI®.

[0017] The knowledge provision system 100 comprises a query acquisition unit 101, a search unit 102, a knowledge database unit 103, a knowledge database generation unit 104, a conversion table creation unit 105, a conversion table storage unit 106, a query conversion unit 107, a query unit 108, a response conversion unit 109, and a response presentation unit 110. The knowledge provision system 100 may be implemented by one or more computers reading and executing a program. Furthermore, some of the query acquisition unit 101, search unit 102, knowledge database unit 103, knowledge database generation unit 104, conversion table creation unit 105, conversion table storage unit 106, query conversion unit 107, a query unit 108, a response conversion unit 109, and a response presentation unit 110 may be implemented on each of the multiple computers. For example, the query acquisition unit 101, search unit 102, conversion table creation unit 105, conversion table storage unit 106, query conversion unit 107, query unit 108, answer conversion unit 109, and answer presentation unit 110 may be implemented on multiple computers, and the search unit 102 implemented on each of those multiple computers may use a single knowledge DB unit 103.

[0018] The inquiry acquisition unit 101 acquires the text of the inquiry input by the user. The inquiry acquisition unit 101 may acquire the text of the inquiry by the user's input using input means such as a keyboard, mouse, touch panel, microphone, etc., or may acquire the text of the inquiry by receiving it from another device.

[0019] The search unit 102 searches for a part related to the text of the inquiry from the knowledge information including confidential information. The confidential information is, for example, information designated as outside-company secret, customer information, personal information, etc. The knowledge information may include manuals within the organization, information on products and services provided by the organization, and information on customers.

[0020] The knowledge DB unit 103 (knowledge information storage unit) stores the knowledge information. The knowledge DB unit 103 may store information indicating words related to confidential information (confidential words) among the words included in the knowledge information. Confidential words are, for example, proper nouns such as organization names, product names, part names constituting products, service names, personal names, etc., numerals such as dates, amounts, quantities, performance values, etc., and specific verbs such as "acquire", "recall", "stall", etc., and may be set by the organization. The knowledge information may be vectorized and stored in the knowledge DB unit 103. The search unit 102 may search for a part related to the text of the inquiry from the vectorized knowledge information.

[0021] The knowledge DB generation unit 104 vectorizes the knowledge information and stores it in the knowledge DB unit 103. Also, the knowledge DB generation unit 104 may select at least a part of proper nouns, numerals, and predetermined verbs from the knowledge information as confidential words, and store information indicating the selected confidential words in the knowledge DB unit 103. The information indicating the confidential words may be, for example, a list of confidential words.

[0022] The conversion table creation unit 105 creates a conversion table by associating corresponding words with confidential words, and stores the created conversion table in the conversion table storage unit 106. The confidential words included in the conversion table may be confidential words extracted by the conversion table creation unit 105 from the portions related to the inquiry sentence retrieved by the retrieval unit 102. That is, the conversion table creation unit 105 may create a conversion table each time an inquiry sentence is input, and store it in the conversion table storage unit 106. The confidential words included in the conversion table may be confidential words extracted from the inquiry sentence in addition to the portions related to the inquiry sentence. For the extraction of these confidential words, the information indicating the confidential words stored in the knowledge DB unit 103 may be used.

[0023] The conversion table creation unit 105 may determine the corresponding words associated with the confidential words using language models such as a small language model and a large language model provided in the knowledge providing system 100. For example, if the confidential word is a proper noun, the conversion table creation unit 105 may query the language model for the general name of the confidential word and use the answer as the corresponding word, or query the language model for other proper nouns of the same category as the confidential word and use the answer as the corresponding word. Also, if the confidential word is a numeral, the conversion table creation unit 105 may query the language model for numerals of the same category as the confidential word and use the answer as the corresponding word. The conversion table creation unit 105 may use a preset corresponding word as the corresponding word associated with the confidential word.

[0024] The conversion table storage unit 106 stores a conversion table including confidential words related to confidential information and corresponding words associated with the confidential words. The inquiry conversion unit 107 uses the conversion table stored in the conversion table storage unit 106 to convert the confidential words included in the inquiry sentence acquired by the inquiry acquisition unit 101 and the portions related to the inquiry sentence retrieved by the retrieval unit 102 into the corresponding words associated with the confidential words.

[0025] The inquiry unit 108 inputs the inquiry text and related parts of the inquiry text, which have been converted by the inquiry conversion unit 107, into the generation AI device 200. The response conversion unit 109 obtains the response to the input from the inquiry unit 108 to the generation AI device 200, and uses the conversion table stored in the conversion table storage unit 106 to convert (reverse convert) the corresponding words included in the response to confidential words. The answer presentation unit 110 presents the answer converted (reverse converted) by the answer conversion unit 109 to the user. This presentation may be displayed on a screen, output as audio, or transmitted to another device that received the inquiry text.

[0026] Figure 2 is a table showing an example of the contents of the conversion table in this embodiment. In the example of the conversion table shown in Figure 2, the confidential word "Company Name A1" is associated with the corresponding word "Company B". Similarly, the confidential word "Product Name A2" is associated with the corresponding word "PC". The confidential word "Part Name A3" is associated with the corresponding word "CPU". The confidential word "Personal Name A4" is associated with the corresponding word "he". The confidential word "acquire" is associated with the corresponding word "deal with". The confidential word "recall" is associated with the corresponding word "contact". The confidential word "stall" is associated with the corresponding word "stop".

[0027] In the example in Figure 2, the corresponding word "Company B" for the confidential word "Company Name A1" is an example where the corresponding word is a proper noun "Company B" which is the same type of proper noun as the confidential word "Company Name A1". The corresponding word "PC" for the confidential word "Product Name A2" is an example where the corresponding word is a common noun "PC" which is the proper noun of the confidential word "Product Name A2". The corresponding word "to trade" for the confidential word "to acquire" is an example where the corresponding word is a pre-set "to trade". The conversion table creation unit 105 creates a conversion table so that the conversion by the answer conversion unit 109 is the reverse conversion of the conversion by the inquiry conversion unit 107, and so that the same corresponding word is not associated with multiple confidential words.

[0028] Figure 3 is a flowchart illustrating the operation of the knowledge provision system 100 in this embodiment. First, the query acquisition unit 101 acquires the query text (step Sa1). Next, the search unit 102 searches the knowledge information stored in the knowledge DB unit 103 for the portion related to the query text acquired in step Sa1 (step Sa2). Next, the conversion table creation unit 105 extracts confidential words contained in the portion related to the query text, which is the search result from step Sa2, creates a conversion table consisting of the extracted confidential words and corresponding words, and stores it in the conversion table storage unit 106 (step Sa3).

[0029] Next, the query conversion unit 107 converts the query text obtained in step Sa1 and the part related to the query text, which is the search result from step Sa2, using the conversion table stored in step Sa3 (step Sa4). As a result, confidential words contained in the query text and the part related to the query text are converted to corresponding words, so that the converted query text and the part related to the query text do not contain confidential information.

[0030] Next, the inquiry unit 108 inputs the inquiry text and related parts of the inquiry text, which were converted in step Sa4, into the generation AI device 200 to perform the inquiry (step Sa5). Next, the answer conversion unit 109 obtains the answer to the inquiry from step Sa5 from the generation AI device 200 and performs a reverse conversion using the conversion table stored in step Sa3 (step Sa6). As a result, the corresponding words included in the answer are converted into confidential words, so the answer may contain confidential information. Next, the answer presentation unit 110 presents the answer that was reverse-converted in step Sa6 to the user (step Sa7).

[0031] Figure 4 is a diagram illustrating an example of data illustrating an embodiment of the knowledge provision system 100 in this embodiment. In this embodiment, the inquiry unit 108 obtains inquiry D1, "Please tell me the release date of product A5." The knowledge database unit 103 stores knowledge information D2, "Product A5 went on sale on X year Y month Z day." The search unit 102 searches the knowledge information stored in the knowledge database unit 103 for the portion related to inquiry D1 and obtains knowledge information D2 as a search result.

[0032] The conversion table creation unit 105 extracts "Product A5" and "X year Y month Z day" as confidential words contained in the search result knowledge information D2. Furthermore, the conversion table creation unit 105 creates a conversion table "Product A5 ⇒ PC, X year Y month Z day ⇒ A year B month C day" consisting of "Product A5" and "X year Y month Z day" and their corresponding words "PC" and "A year B month C day", respectively, and stores it in the conversion table storage unit 106.

[0033] The inquiry conversion unit 107 converts inquiry D1 using conversion table D3 to obtain the converted inquiry D4, "Please tell me the release date of the PC." The inquiry conversion unit 107 also converts the knowledge information D2, which is related to inquiry D1, using conversion table D3 to obtain the converted related part D5, "The PC went on sale on year A, month B, day C."

[0034] The inquiry unit 108 inputs the converted inquiry D4, "Please tell me the release date of the PC," and the converted related part D5, "The PC went on sale on A year B month C day," into the generation AI device 200. The answer conversion unit 109 obtains the answer D6, "The PC was released on A year B month C day," from the generation AI device 200. The answer conversion unit 109 then performs a reverse conversion on the answer D6 using the conversion table D3 to obtain the reverse-converted answer D7, "The release date of product A5 is X year Y month Z day."

[0035] In this way, the converted query D4, the converted related part D5, and the answer D6 transmitted to and from the generating AI device 200 have their confidential words converted to corresponding words. Therefore, even if query D1 is related to confidential information, the confidential information will not be transmitted externally.

[0036] The present invention may also be in the following embodiments. (1) One embodiment of the present invention is a knowledge provision system comprising: an inquiry acquisition unit that acquires an inquiry text; a search unit that searches for a portion related to the inquiry text from knowledge information including confidential information; a conversion table storage unit that stores a conversion table including confidential words related to the confidential information and corresponding words associated with the confidential words; an inquiry conversion unit that uses the conversion table to convert the confidential words included in the inquiry text and the portion related to the inquiry text into the corresponding words; an inquiry unit that inputs the inquiry text and the portion related to the inquiry text converted by the inquiry conversion unit into a large-scale language model; and an answer conversion unit that acquires an answer to the input from the inquiry unit to the large-scale language model and uses the conversion table to convert the corresponding words included in the answer into the confidential words.

[0037] (2) Another embodiment of the present invention is the knowledge provision system described in (1), comprising a conversion table creation unit that extracts the confidential words from the part of the inquiry text and associates the corresponding words with the extracted confidential words to create the conversion table.

[0038] (3) Another embodiment of the present invention is a knowledge provision system as described in (2), comprising: a knowledge information storage unit that stores the knowledge information; and a knowledge DB generation unit that selects at least a portion of proper nouns, numerals, and verbs from the knowledge information as confidential words and stores them as confidential words in the knowledge information storage unit, wherein the conversion table creation unit creates the conversion table by referring to the knowledge information storage unit.

[0039] (4) Another embodiment of the present invention is the knowledge provision system described in (2) or (3), wherein the conversion table creation unit, if the extracted confidential word is a proper noun, uses the common noun of the proper noun or another proper noun of the same type as the proper noun as the corresponding word.

[0040] (5) Another embodiment of the present invention is a knowledge provision method comprising: a first step of querying to obtain a query text; a second step of searching for a portion related to the query text from knowledge information including confidential information; a third step of converting the confidential words contained in the query text and the portion related to the query text to the corresponding words using a conversion table that includes confidential words related to the confidential information and corresponding words associated with the confidential words; a fourth step of inputting the query text and the portion related to the query text converted in the third step into a large-scale language model; and a fifth step of obtaining a response to the input to the large-scale language model in the fourth step and converting the corresponding words contained in the response into the confidential words using the conversion table.

[0041] (6) Another embodiment of the present invention is a program for causing a computer to function as: a query acquisition unit that acquires a query text; a search unit that searches for a portion related to the query text from knowledge information including confidential information; a conversion table storage unit that stores a conversion table including confidential words related to the confidential information and corresponding words associated with the confidential words; a query conversion unit that uses the conversion table to convert the confidential words contained in the query text and the portion related to the query text into the corresponding words; a query unit that inputs the query text and the portion related to the query text converted by the query conversion unit into a large-scale language model; and a response conversion unit that acquires a response to the input from the query unit to the large-scale language model and uses the conversion table to convert the corresponding words contained in the response into the confidential words.

[0042] Alternatively, the knowledge provision system 100 may be realized by recording a program for realizing the functions of the knowledge provision system 100 in Figure 1 onto a computer-readable recording medium, loading the program recorded on this recording medium into a computer system, and executing it. The term "computer system" here includes hardware such as the operating system and peripheral devices.

[0043] Furthermore, "computer system" shall also include the homepage provisioning environment (or display environment) if a WWW system is being used. Furthermore, "computer-readable recording media" refers to portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard disks built into computer systems. Moreover, "computer-readable recording media" also includes those that dynamically hold programs for a short period of time, such as communication lines used when transmitting programs over networks such as the Internet or communication lines such as telephone lines, and those that hold programs for a certain period of time, such as volatile memory inside computer systems that act as servers or clients in such cases. In addition, the above-mentioned programs may be for the purpose of realizing some of the functions described above, and may also be able to realize the above-mentioned functions in combination with programs already recorded in the computer system.

[0044] While embodiments of this invention have been described in detail above with reference to the drawings, the specific configuration is not limited to these embodiments and may include design modifications and the like that do not depart from the spirit of this invention. [Explanation of Symbols]

[0045] 100 Knowledge Provision System 101 Inquiry Acquisition Department 102 Search Section 103 Knowledge Database Department 104 Knowledge DB Generation Unit 105 Conversion Table Creation Section 106 Conversion Table Storage Unit 107 Inquiry Conversion Section 108 Inquiry Department 109 Answer Conversion Section 110 Answer presentation part 200 Generation AI device

Claims

1. The inquiry retrieval unit retrieves the text of the inquiry, A search unit that searches for parts of the aforementioned inquiry text from knowledge information, including confidential information, A conversion table storage unit stores a conversion table that includes confidential words related to the confidential information and corresponding words associated with the confidential words. A query conversion unit that uses the conversion table to convert the confidential words contained in the query text and the parts related to the query text into corresponding words, A query unit inputs the query text and related parts of the query text, which have been converted by the query conversion unit, into a large-scale language model. The query unit obtains a response to the input to the large-scale language model, and the response conversion unit uses the conversion table to convert the corresponding words included in the response into confidential words. A knowledge-sharing system equipped with [features / equipment].

2. The knowledge provision system according to claim 1, further comprising a conversion table creation unit that extracts the confidential words from the part of the inquiry text related to the confidential words, and creates the conversion table by associating the extracted confidential words with the corresponding words.

3. A knowledge information storage unit that stores the aforementioned knowledge information, A knowledge database generation unit selects at least some of the proper nouns, numerals, and verbs from the aforementioned knowledge information as confidential words and stores them as confidential words in the knowledge information storage unit. Equipped with, The knowledge provision system according to claim 2, wherein the conversion table creation unit creates the conversion table by referring to the knowledge information storage unit.

4. The knowledge provision system according to claim 2, wherein the conversion table creation unit, if the extracted confidential word is a proper noun, uses the common noun equivalent of the proper noun, or another proper noun of the same type as the proper noun, as the corresponding word.

5. The first step of the query is to obtain the query text, A second step involves searching for the portion of the query text that is relevant to the knowledge information, which may contain confidential information. A third step involves using a conversion table that includes confidential words related to the confidential information and corresponding words associated with those confidential words to convert the confidential words contained in the query text and the parts related to the query text into the corresponding words. A fourth step involves inputting the query text and the parts related to the query text, which have been converted in the third step, into a large-scale language model. A fifth step involves obtaining a response to the input to the large-scale language model in the fourth step, and using the conversion table, converting the corresponding words included in the response into confidential words. A knowledge-providing method having [a certain characteristic].

6. Computers, Inquiry retrieval unit that retrieves the text of the inquiry, A search unit that searches for parts of the aforementioned inquiry text from knowledge information, including confidential information. A conversion table storage unit that stores a conversion table including confidential words related to the confidential information and corresponding words associated with the confidential words, A query conversion unit that uses the conversion table to convert the confidential words contained in the query text and the parts related to the query text into corresponding words. A query unit inputs the query text and related parts of the query text, which have been converted by the query conversion unit, into a large-scale language model. A response conversion unit obtains a response to the input to the large-scale language model by the query unit and uses the conversion table to convert the corresponding words included in the response into confidential words. A program designed to function as such.

Images