Abstraction system, abstraction method, and abstraction program

Abstract

To provide an abstraction system that can conceal confidential information while securing information obtained from text. [Solution] The abstraction system 100 includes an extraction unit 10 that extracts a proper name P to be kept secret from an electronic document To, and a conversion unit 20 that converts the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than a higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. When the electronic document To has a first proper name P1 and a second proper name P2 that are classified into a common attribute, the conversion unit 20 converts the first proper name P1 and the second proper name P2 into a first abstract name Q1 and a second abstract name Q2 that are different from each other.

Description

[Technical Field]

[0001] The present invention relates to an abstraction system, an abstraction method, and an abstraction program. [Background technology]

[0002] Conventionally, there has been security masking technology for concealing confidential information such as personal information contained in text (for example, Patent Document 1).

[0003] Conventional security masking technologies simply abstract the confidential information contained in the original text and output it as anonymized text. As a result, the information obtained from the anonymized text is less than that obtained from the original text, making it less useful. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Special Publication No. 2017-503278 Summary of the Invention [Problem to be solved by the invention]

[0005] An object of the present invention is to provide an abstraction system, an abstraction method, and an abstraction program that can conceal confidential information while securing information obtained from text. [Means for solving the problem]

[0006] (1) An abstraction system according to one aspect of the present invention includes an extraction unit that extracts a proper name to be kept secret from an electronic document, and a conversion unit that converts the proper name into an abstract name abstracted at an appropriate abstraction level that is lower than a higher reference abstraction level in a group of abstraction levels hierarchically organized for each corresponding attribute. (2) In (1) above, when the electronic document has a first proper name and a second proper name that are classified into a common attribute, the conversion unit may convert the first proper name and the second proper name into a first abstract name and a second abstract name that are different from each other. (3) In the above (1) or (2), the conversion unit may convert the proper name into the abstract name using a language model. (4) In the above (3), the conversion unit may input a conversion prompt according to the attribute to the language model and obtain an output from the language model. (5) In (4) above, the system may include a determination unit that determines whether or not there is knowledge of the proper name in the language model, and a supplementation unit that, if it determines that there is no knowledge, obtains supplementary information about the proper name by searching and inputs the information into the language model. (6) In the above (5), the determination unit may perform the determination using the language model. (7) In the above (5), the supplementing unit may summarize the supplemental information using the language model. (8) An abstraction method according to one aspect of the present invention includes an extraction step of extracting a proper name to be kept secret from an electronic document, and a conversion step of converting the proper name into an abstract name abstracted at an appropriate abstraction level lower than a higher reference abstraction level in a group of abstraction levels hierarchically organized for each corresponding attribute. (9) An abstraction program according to one aspect of the present invention enables a computer to perform the following functions: extracting a proper name to be kept secret from an electronic document; and converting the proper name into an abstract name abstracted at an appropriate abstraction level lower than a higher reference abstraction level in a group of abstraction levels hierarchically organized for each corresponding attribute. [Effects of the Invention]

[0007] According to the present invention, it is possible to provide an abstraction system, an abstraction method, and an abstraction program that can conceal confidential information while ensuring the amount of information obtainable from text. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a schematic diagram of an abstraction system. [Figure 2] FIG. 1 is a diagram illustrating abstraction levels. [Figure 3] FIG. 10 is an explanatory diagram illustrating an example of an abstraction method. [Figure 4] FIG. 1 is an explanatory diagram illustrating abstraction of a proper name P for which a language model M has knowledge. [Figure 5] FIG. 10 is an explanatory diagram illustrating abstraction of a proper name P for which a language model M does not have knowledge. [Figure 6] FIG. 1 is a diagram illustrating a flow of an abstraction method. DETAILED DESCRIPTION OF THE INVENTION

[0009] (Embodiment) Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a schematic diagram of an abstraction system 100. FIG. 2 is a diagram explaining an abstraction level group L. FIG. 3 is an explanatory diagram showing an example of an abstraction method. Note that, hereinafter, parts having common functions may be given the same reference numerals or symbols. Note that, hereinafter, the operations or functions performed by each part can be realized by a computer through an abstraction program. Hereinafter, the term "system" may be read as "device."

[0010] The abstraction system 100 according to the embodiment is suitable for, for example, performing security masking to conceal confidential customer information when employees of a financial institution share the customer information obtained through sales activities. The abstraction system 100 is also suitable for, for example, performing security masking to conceal contact center logs in order to share the logs internally to improve services.

[0011] As shown in FIG. 1 or 3 , an abstraction system 100 according to an embodiment includes an extraction unit 10 that extracts (detects) a proper name P to be concealed from an electronic document To, and a conversion unit 20 that converts the proper name P into an abstract name Q. The abstraction system 100 also includes an output unit 30 that outputs a converted document Ta in which the proper name P is replaced with the abstract name Q based on the electronic document To. The abstraction system 100 also includes a search unit 40, a determination unit 50, and a supplementation unit 60, as appropriate. The abstraction system 100 may include a local language model M that can be read (called) by the abstraction system 100. The abstraction system 100 may include a language model M on a network N that can be read (called) by the abstraction system 100. The abstraction system 100 may also include a local database LD that stores information such as a local dictionary and a set of abstraction levels L that can be read (called) by the abstraction system 100. The abstraction system 100 may include a network database ND that stores information such as a dictionary and a set of abstraction levels L on a network N that can be read (called) by the abstraction system 100 .

[0012] The abstraction system 100 may be a computer system equipped with a CPU, memory, a network interface for controlling data communication with an external network, and a bus for interconnecting each functional unit. The memory stores programs for controlling each functional unit or unit in cooperation with the CPU. The memory stores a master that stores data necessary for controlling each functional unit or unit.

[0013] The electronic document To contains digitized text (description, document, sentence, phrase, or word). The text may be part of table data or part of text data. The electronic document To is in any file format, such as text or CSV, and is available for use on a computer.

[0014] A proper noun P is an expression containing a proper noun, which is confidential information to be masked. For example, a proper noun P can be classified into attributes (types, categories) such as a company name, a person's name, nationality, religion, political attribute, address or location (country, city, body of water, mountain), PHI (drug name, physical injury, treatment, test name), product, or service.

[0015] An abstract name Q is an expression defined as a single concept by removing details or specificity from a proper name P and extracting only essential elements or aspects of interest. An abstract name Q may also be constructed as a general concept by finding properties or elements common to multiple different abstract names Q and combining the commonalities. For example, an abstract name Q is "Company 1" for a proper name P such as "ACES Co., Ltd."

[0016] As shown in FIG. 2, the abstraction level group L is a hierarchical (staged) collection of multiple abstraction levels, from lower abstraction levels (subordinate concepts) to higher abstraction levels (superordinate concepts). The number of abstraction levels (number of hierarchies) constituting the abstraction level group L may be three or more. The lower abstraction levels represent attributes with a relatively low degree of abstraction (high specificity). The higher abstraction levels represent attributes with a relatively high degree of abstraction (low specificity). When the number of abstraction levels is three or more, the intermediate abstraction levels (intermediate concepts) are located between the lower and higher abstraction levels. The abstraction level group L may be stored in advance in the memory of the abstraction system 100, in the local database LD, or in the network database ND as a dictionary including a concept dictionary such as WordNet (registered trademark), or may be generated from information on the web using a language model M. Note that the data generated from the language model M does not necessarily need to be registered (stored on the network N) in advance.

[0017] The language model M may be a machine learning model including a deep neural network such as a large-scale language model (LLM) or a trained model. Natural language processing techniques such as syntactic analysis and TextRank may be used for the language model M. By using the language model M, it is possible to abstract a proper name P without having to prepare a complete dictionary in advance, and to generate a converted document Ta from the original electronic document To in which the proper name P is replaced with an abstract name Q.

[0018] For example, if the number of abstraction levels that make up the abstraction level group L is 3, the lower abstraction level corresponding to the proper name P "ACES Inc." is "University of Tokyo AI startup," the middle abstraction level is "startup," and the higher abstraction level is "Company 1." In this way, by abstracting the proper name P hierarchically, it is possible to maintain confidentiality while preserving the amount of information.

[0019] The extraction unit 10 has a function of extracting a proper name P to be concealed from an electronic document To. The extraction unit 10 has a named entity extraction algorithm that mechanically extracts named entities, such as proper nouns (such as organizations, place names, and product names) plus date, quantity, and the like, from input natural language text. The extraction unit 10 may detect the position and attributes of the proper name P, which is a word to be concealed, using a technology called named entity extraction / NER. The extraction unit 10 may, for example, utilize open source software (OSS) having a named entity extraction algorithm. For example, when the natural language text "ACES is acquired by ABC Communications" is input, the extraction unit 10 removes unnecessary particles and extracts "ACES" and "ABC Communications" as an attribute called "organization." Note that this extraction function may be performed by a language model M. That is, for example, the extraction unit 10 may input the original natural language text together with a prompt PT into a large-scale language model, language model M, and obtain an output from language model M.

[0020] The output unit 30 has a function of outputting a converted document Ta in which the proper name P is replaced with an abstract name Q based on the electronic document To. For example, as shown in FIG. 3, based on the electronic document To, which is a natural sentence saying "ACES is acquired by ABC Communications," the output unit 30 outputs a converted document Ta such as "The University of Tokyo AI startup is acquired by a communications company" by replacing "ACES" with "The University of Tokyo AI startup is acquired by a communications company." In this way, the converted document Ta in which the proper name P is abstracted is output, and therefore the proper name P can be kept confidential while retaining a larger amount of information than when the proper name P is abstracted to a higher level, such as "Company 1 is acquired by Company 2."

[0021] The conversion unit 20 has a function of converting a proper name P into an abstract name Q obtained by abstracting the proper name P.

[0022] Here, the conversion unit 20 converts the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than the higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. In other words, the conversion unit 20 converts the proper name P at an abstraction level of a middle concept (middle concept). As a result, when the proper name P is abstracted at the reference abstraction level Lr (higher concept), less information can be obtained from the converted document Ta, whereas when the proper name P is abstracted at the appropriate abstraction level Ld (middle concept), more information can be obtained from the converted document Ta. Therefore, the abstraction system 100 can conceal confidential information while ensuring the amount of information obtainable from text.

[0023] Specifically, for example, the conversion unit 20 converts the proper name P (here, "ACES") into an abstract name Q abstracted at an appropriate abstraction level Ld (here, "University of Tokyo AI startup") that is lower than the higher reference abstraction level Lr (here, "company") in a group of abstraction levels L hierarchically organized by corresponding attribute (here, "organization") as shown in Figure 2.

[0024] The conversion unit 20 may perform the conversion using a dictionary that lists the relationships between proper names P and abstract names Q, which is prepared in advance.

[0025] The conversion unit 20 may convert the proper name P into the abstract name Q by using a language model M. This allows the proper name P to be appropriately abstracted into the abstract name Q. The conversion unit 20 may convert the proper name P into the abstract name Q by using a dictionary in combination with the language model M. The conversion unit 20 may convert a fixed proper name P such as a country name using a dictionary. The conversion unit 20 may change the conversion model (algorithm) depending on the attribute (type) of the proper name P, for example, by converting a fixed proper name P such as a country name using a dictionary and converting a fluid proper name P such as an organization (company name) using the language model M.

[0026] The conversion unit 20 inputs a conversion prompt PT according to the attribute to the language model M and obtains an output from the language model M. This allows the language model M to appropriately convert a fluid proper name P that is difficult to include in a dictionary, such as an organization (company name), into an abstract name Q.

[0027] When the electronic document To has a first proper name P1 and a second proper name P2 classified as a common attribute, the conversion unit 20 may convert the first proper name P1 and the second proper name P2 into a first abstract name Q1 and a second abstract name Q2 that are different from each other. Specifically, when the electronic document To has a first proper name P1 and a second proper name P2 classified as a common attribute, the conversion unit 20 reduces the abstraction level in the abstraction level group L until the first proper name P1 and the second proper name P2 become the first abstract name Q1 and the second abstract name Q2 that are different from each other. As a result, even when the first proper name P1 and the second proper name P2 classified as a common attribute are abstracted, they are not converted into the same abstract name Q, but are converted into a first abstract name Q1 and a second abstract name Q2 that are distinguishable from each other. This allows for confidentiality while maintaining the amount of information.

[0028] Specifically, for example, if "ACES" (first proper name P1) and "ABC Bio" (second proper name P2) are extracted from the original text (electronic document To) "ACES is acquired by ABC Bio" and classified as a common attribute (here, organization) and abstracted at an appropriate abstraction level Ld lower than the higher reference abstraction level Lr, "ACES" could be converted to "University of Tokyo venture" (abstract name Q) and "ABC Bio" could be converted to "University of Tokyo venture" (abstract name Q), resulting in low usefulness even if the information is concealed. Therefore, "ACES" (first proper name P1) and "ABC Bio" (second proper name P2) are converted to a lower abstraction level. For example, the conversion unit 20 converts "ACES" to "University of Tokyo AI venture" (first abstract name Q1) and "ABC Bio" to "University of Tokyo bio venture" (second abstract name Q2). As a result, even if the first proper name P1 and the second proper name P2, which are classified as having a common attribute, are abstracted, they are not converted into the same abstract name Q, but are converted into the first abstract name Q1 and the second abstract name Q2, which are distinguishable from each other. Therefore, the amount of information can be secured while maintaining confidentiality. Note that the conversion unit 20 does not have to continue the process of converting at a lower abstraction level as described above until a different abstract name Q is obtained. For example, if there are 100 words (proper names P) to be abstracted in the same sentence (electronic document To), 10 of them may become the same abstract word (abstract name Q). Also, for example, as a result of converting the original text (electronic document To), multiple identical abstract words (abstract names Q) may be included in the same sentence (converted document Ta).

[0029] In addition, when the electronic document To has a first proper name P1 and a second proper name P2 classified into a common attribute, the conversion unit 20 may execute a series of functions, from the function of converting the first proper name P1 and the second proper name P2 into an abstract name Q abstracted at an appropriate abstraction level Ld lower than the higher reference abstraction level Lr in an abstraction level group L hierarchically structured for each corresponding attribute, to the function of converting the first proper name P1 and the second proper name P2 into a first abstract name Q1 and a second abstract name Q2 that are different from each other by lowering the abstraction level in the abstraction level group L until they become different from each other, by inputting the functions into the language model M using a prompt PT and acquiring the output from the language model M all at once.

[0030] The abstraction system 100 may include a determination unit 50 that determines whether or not there is knowledge about the proper name P in the language model M. The abstraction system 100 may also include a supplementation unit 60 that, if the determination unit 50 determines that there is no knowledge, acquires supplemental information a about the proper name P through a search (e.g., dictionary search, internet search, etc.) by the search unit 40 and inputs the information a into the language model M. As a result, even if the language model M does not have knowledge about the proper name P (the information has not been updated, learning has not been completed), such as when the proper name P is the name of a new company, the proper name P can be appropriately converted into an abstract name Q by the language model M to which the supplemental information a has been added.

[0031] The determination unit 50 may make the determination using the language model M. In other words, the determination unit 50 causes the language model M itself to check whether or not it has knowledge about the proper name P, and acquires the result. This allows the supplementation unit 60 to automatically input (or not input) the supplemental information a to the language model M according to the result acquired by the determination unit 50 from the language model M. The determination unit 50 may also make the determination using a dictionary. That is, the determination unit 50 refers to the dictionary to check whether the proper name P is included in the dictionary, and if it is not included in the dictionary, can cause the search unit 40 to perform a search.

[0032] The supplementing unit 60 may summarize the supplemental information a using the language model M. In other words, the language model M may summarize the supplemental information a. That is, as shown in FIG. 5, the supplementing unit 60 may use the language model M to summarize the search information S related to the proper name P obtained by the search unit 40 through a search, and then add the summary to the prompt PT as supplemental information a, specifically as summary information SS, and input the summary to the language model M. This allows the supplemental information a to be input to the language model M, so that the proper name P, which was not in the knowledge of the language model M, can be abstracted by the language model M in a more appropriate expression than if the supplemental information a is not summarized. Note that the supplementing unit 60 may input the search information S (supplemental information a) directly to the language model M without summarizing it.

[0033] Next, an example of a case where an abstraction method using the abstraction system 100 is executed using a language model M will be described. FIG. 3 is an explanatory diagram showing an example of the abstraction method. FIG. 4 is an explanatory diagram showing abstraction of a proper name P that the language model M has as knowledge. Note that FIG. 4 shows a case where a search for information about the proper name P is not required. FIG. 5 is an explanatory diagram showing abstraction of a proper name P that the language model M does not have as knowledge. Note that FIG. 5 shows a case where a search for information about the proper name P is required. FIG. 6 is a diagram showing the flow of the abstraction method. Unless otherwise specified below, the abstraction system 100 is the subject of operations or functions.

[0034] The abstraction method includes a method of converting a proper name P included in the electronic document To into an abstract name Q. The abstraction method may be included in a method of converting the electronic document To into a transfer document Ta.

[0035] (1) As shown in FIGS. 3 and 4, the text "ACES acquired by ABC Communications" is input to the abstraction system 100 as an example of an electronic document To.

[0036] (2) Next, the proper name P to be kept secret is extracted from the electronic document To (extraction step S1). <aces>And <ABC Communication> is extracted as the proper name P.

[0037] (3) Next, the proper name P is converted into an abstract name Q abstracted at an appropriate abstraction level Ld lower than the upper reference abstraction level Lr in the group of abstraction levels L hierarchically organized for each corresponding attribute (conversion step S2).

[0038] (4) In the conversion step, when the extracted proper name P (here, <ABC Communication>) is possessed by the language model M (for example, a large language model LLM) as knowledge (the case of "known" in Fig. 6), as shown in Fig. 4, an instruction book C with content such as "Please abstract the company name ABC Communication referring to the following examples" and a prompt PT1 with examples E for reference of the appropriate abstraction level Ld when abstracting the proper name P at an appropriate abstraction level Ld lower than the upper reference abstraction level Lr to output the abstract name Q, such as "Toyota → large automobile company", "Airbnb → emerging IT company", "DeNA → megaventure", are input to the language model M. Note that Toyota, Airbnb, and DeNA are registered trademarks.

[0039] (5) When the prompt PT1 is input, as shown in Fig. 4, the language model M outputs text indicating an abstract name Q such as "large communication company". Note that, as shown in Fig. 4, text indicating a combination of the proper name P and the abstract name Q such as "ABC Communication → large communication company" may be output. That is, an abstract name Q abstracted at the appropriate abstraction level Ld corresponding to the proper name P is output.

[0040] (6) Whether the language model M has the extracted proper name P as knowledge may be determined by the language model M. For example, as shown in Fig. 5, a prompt PT2 with content such as "Please answer that you don't know if you don't know the company ACES" is input to the prompt PT together with the example E for reference of the appropriate abstraction level Ld when abstracting the proper name P to output the abstract name Q.

[0041] (7) When prompt PT2 is input, as shown in FIG. 5, language model M outputs text indicating the content such as "ACES is not included in my knowledge...." In this way, language model M can be made to determine whether it has knowledge of proper noun P, which is the target of abstraction. Note that, to facilitate subsequent processing, prompt PT2 may instruct language model M to return a predetermined character string if it has no knowledge of the extracted proper noun P. For example, the predetermined character string may be output by language model M as "unknown," and information search may be performed using this output as described below.

[0042] (8) Returning to Figure 3, in the conversion step, the extracted proper name P (here <aces>) as knowledge that the language model M (e.g., the large language model LLM) does not have (in the case of “unknown” in FIG. 6), as shown in FIG. 5, the search unit 40 may perform a search for information (search information S) related to the proper name P (e.g., a search from the local database LD or the network database ND, or a web search).

[0043] (9) Here, as shown in Figure 5, the information about the retrieved proper name P (retrieval information S) may be summarized as appropriate. Then, the summarized information about the proper name P (summary information SS) is input to the language model M together with the instruction C and the example E. This allows the language model M to abstract the proper name P by referring to the information about the proper name P obtained by the search as new knowledge. Therefore, even if the language model M does not have knowledge of the proper name P, it can abstract the proper name P.

[0044] (10) When prompt PT3 is input, as shown in Figure 5, language model M outputs text indicating abstract name Q, such as "Tokyo University AI venture." That is, it outputs abstract name Q abstracted at appropriate abstraction level Ld corresponding to proper name P (output step).

[0045] (11) Finally, based on the electronic document To, a converted document Ta is output in which the proper name P is replaced with the abstract name Q output from the language model M.

[0046] As described above, the abstraction method includes an extraction step S1 for extracting a proper name P to be concealed from an electronic document To, and a conversion step S2 for converting the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than a higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. Therefore, by executing the abstraction method using the abstraction system 100, confidential information can be concealed while ensuring the amount of information obtainable from text.

[0047] (abstraction program) The operations or functions of each step in the above-described abstraction method can be implemented in a computer by an abstraction program.

[0048] The abstraction program enables a computer to perform the following functions: extract a proper name P to be concealed from an electronic document To; and convert the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than a higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. This makes it possible to conceal confidential information while maintaining the amount of information obtainable from the text.

[0049] The technical scope of the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.

[0050] In addition, the elements in the above-described embodiments may be replaced with common general technical knowledge or well-known matters without departing from the spirit of the present invention. Furthermore, the above-described elements may be appropriately combined without departing from the spirit of the present invention.

[0051] As described above, the abstraction system 100 according to the embodiment includes an extraction unit 10 that extracts a proper name P to be concealed from an electronic document To, and a conversion unit 20 that converts the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than a higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. This makes it possible to conceal confidential information while maintaining the amount of information obtainable from text.

[0052] The abstraction method according to the embodiment includes an extraction step S1 for extracting a proper name P to be concealed from an electronic document To, and a conversion step S2 for converting the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than a higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. This makes it possible to conceal confidential information while maintaining the amount of information obtainable from the text.

[0053] The abstraction program according to the embodiment causes a computer to perform the following functions: extracting a proper name P to be concealed from an electronic document To; and converting the proper name P into an abstract name Q abstracted at an appropriate abstraction level Ld that is lower than a higher reference abstraction level Lr in a group of abstraction levels L hierarchically organized for each corresponding attribute. This makes it possible to conceal confidential information while maintaining the amount of information obtainable from text. [Explanation of symbols]

[0054] 100 Abstract Systems 10 Extraction part 20 Conversion unit 30 Output section 40 Search Section 50 Judgment section 60 Supplementary part C Instructions E example L Abstraction Levels M language model N Network P proper name P1 1st proper name P2 Second proper name PT,PT1,PT2,PT3 prompt Q abstract name Q1 1st abstract name Q2 Second abstract name S Search Information SS Summary Information Ta conversion document To electronic documents a Supplementary information S1 Extraction step S2 Conversion Step< / aces> < / aces>

Claims

1. an extraction unit that extracts proper names from electronic documents; a conversion unit that converts the proper name into an abstract name abstracted at an appropriate abstraction level that is lower than a reference abstraction level that is a higher abstraction level in a group of abstraction levels hierarchically divided into a plurality of abstraction levels for each corresponding attribute; a determination unit that inputs a prompt to a language model itself to determine whether the language model has knowledge about the proper name, and obtains the output thereof, thereby determining whether the language model has knowledge about the proper name; a supplementation unit that, when it is determined that there is no knowledge, acquires supplemental information about the proper name by searching and inputs the supplemental information into the language model as reference knowledge for the proper name; an output unit that outputs a converted document in which the proper names are replaced with the abstract names based on the electronic document; An abstraction system comprising: The abstraction system inputs the electronic document to the extraction unit, inputs a prompt containing an instruction to abstract the proper name to the appropriate abstraction level to the language model by the conversion unit, and outputs the converted document by the output unit. Abstraction system.

2. When the electronic document has a first proper name and a second proper name classified into the common attribute, the conversion unit converts the first proper name and the second proper name into the first abstract name and the second abstract name that are different from each other by lowering the abstraction level in the group of abstraction levels until the first proper name and the second proper name become the first abstract name and the second abstract name that are different from each other. The abstraction system of claim 1 .

3. The conversion unit inputs a prompt describing a reference example of the appropriate abstraction level that changes depending on the attribute to the language model, and obtains the converted document output from the language model. The abstraction system according to claim 1 or claim 2.

4. The supplemental information is summarized using the language model, and the summary information is added to the prompt and input to the language model, and the converted document is output. The abstraction system of claim 1 .

5. an extraction step of extracting a proper name from the electronic document; a conversion step of converting the proper name into an abstract name abstracted at an appropriate abstraction level, which is an abstraction level lower than a reference abstraction level, which is a higher abstraction level, in a group of abstraction levels hierarchically divided into a plurality of abstraction levels for each corresponding attribute; a determining step of determining whether or not the language model has knowledge of the proper name by inputting a prompt to the language model itself to determine whether or not the language model has knowledge of the proper name and obtaining an output thereof; a supplementation step of acquiring supplemental information about the proper name by searching if it is determined that there is no knowledge, and inputting the supplemental information into the language model as reference knowledge for the proper name; an output step of outputting a converted document in which the proper names are replaced with the abstract names based on the electronic document; An abstraction method comprising: In the abstraction method, the electronic document is input in the extraction step, a prompt describing an instruction for abstracting the proper name to the appropriate abstraction level is input to the language model in the conversion step, and the converted document is output in the output step. Abstraction method.

6. On the computer, A function to extract proper names from electronic documents; a function of converting the proper name into an abstract name abstracted at an appropriate abstraction level, which is an abstraction level lower than a reference abstraction level, which is a higher abstraction level, in a group of abstraction levels hierarchically divided into a plurality of abstraction levels for each corresponding attribute; a function of inputting a prompt to a language model itself to determine whether the language model has knowledge of the proper name, and obtaining the output thereof, thereby determining whether the language model has knowledge of the proper name; a function of acquiring supplementary information about the proper name by searching if it is determined that there is no knowledge, and inputting the supplementary information into the language model as reference knowledge for the proper name; a function of outputting a converted document in which the proper names are replaced with the abstract names based on the electronic document; An abstraction program that realizes the above. The abstraction program causes a computer to input the electronic document, input a prompt containing an instruction to abstract the proper name to the appropriate abstraction level to the language model, and output the converted document. Abstraction program.

Images