Information processing method and computer program

Abstract

To provide a technique capable of effectively reducing leakage of confidential information that is caused when a text generation model outputs text including the confidential information.SOLUTION: An information processing method includes: a) a step of acquiring input text 21; b) a step of acquiring an abstracted similar document 25 based on a document registered in a document database 41, the abstracted similar document being similar to the input text 21 and including confidential information abstracted by abstraction processing; and c) a step of acquiring output text 27 by using a text generation model M, the output text being answer text for the input text 21 in inputting the input text 21 and the abstracted similar document 25, and the text generation model being trained to generate answer text based on text and external information associated with the text.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The subject matter disclosed herein relates to information processing methods and computer programs. [Background technology]

[0002] Conventionally, a document creation support device that outputs text using a machine learning model is known (for example, Patent Document 1). In the document creation support device, documents similar to a text (description of an invention) input by a user are extracted from documents (patent documents) stored in a database, and an invention description document that paraphrases the text input by the user is created using the extracted similar documents. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2021 / 152712 Summary of the Invention [Problem to be solved by the invention]

[0004] If confidential information is included in documents stored in a database, the confidential information may be included in the output text of the sentence generator. Therefore, there is a risk that the confidential information may be leaked if a user who should not disclose the confidential information views the output text containing the confidential information.

[0005] An object of the present invention is to provide a technique that can effectively reduce the leakage of confidential information caused by a sentence generation model outputting sentences that include confidential information. [Means for solving the problem]

[0006] In order to solve the above problem, a first aspect is an information processing method executed by a computer, comprising: a) a step of acquiring an input text; b) a step of acquiring, based on a document registered in a document database, a similar-abstraction document that is similar to the input text and from which confidential information has been abstracted by an abstraction process; and c) a step of acquiring, using a sentence generation model trained to generate a response sentence based on text and external information related to the text, an output text that is a response sentence to the input text when the input text and the similar-abstraction document are used as input.

[0007] A second aspect is the information processing method of the first aspect, wherein the step b) includes: b11) a step of obtaining similar documents by searching the document database for documents similar to the input text; and b12) a step of generating the abstracted similar documents by abstracting the confidential information contained in the similar documents.

[0008] A third aspect is the information processing method of the second aspect, wherein the step b11) includes a step of generating an abstract input text by abstracting words contained in the input text, and a step of obtaining the similar documents by searching the document database for documents similar to the abstract input text.

[0009] A fourth aspect is an information processing method of the first or second aspect, wherein the step b) includes: b21) a step of abstracting confidential information of each of a plurality of documents held in the document database; and b22) a step of obtaining the abstract similar documents by searching for documents similar to the input text from among the abstract documents abstracted by the step b21).

[0010] A fifth aspect is an information processing method of the fourth aspect, wherein the step c) includes a step of inputting the document searched from the abstracted documents in the step b22) into the sentence generation model as the abstracted similar document.

[0011] A sixth aspect is an information processing method of the first or second aspect, further comprising: d) a step of generating an abstracted output text by abstracting confidential information contained in the output text obtained by step c).

[0012] A seventh aspect is an information processing method of the first or second aspect, wherein in step b), the abstraction process includes a process of abstracting the confidential information into concepts corresponding to a predetermined concept hierarchy using concept hierarchy information that defines the hierarchical relationships between multiple concepts.

[0013] An eighth aspect is a computer-readable computer program that causes the computer to execute the information processing method of the first or second aspect. [Effects of the Invention]

[0014] According to the first to seventh aspects, even if a document in a document database contains confidential information, the probability of outputting output text containing confidential information can be reduced by inputting the abstracted document into a text generation model.

[0015] According to the information processing method of the third aspect, by searching for similar documents using abstract input text that is an abstraction of the input text, it is possible to broadly search for documents similar to the input text without being limited by specific words.

[0016] According to the information processing method of the fourth aspect, by searching for documents similar to the input text from among abstract documents, it is possible to widely search for documents similar to the input text without being limited by specific words.

[0017] According to the information processing method of the fifth aspect, sentences searched from the abstracted documents are input to the sentence generation model as similar abstracted documents, thereby making it possible to quickly obtain output text.

[0018] According to the information processing method of the sixth aspect, the leakage of confidential information can be further reduced by abstracting the output text. [Brief explanation of the drawings]

[0019] [Figure 1] 1 is a diagram illustrating a configuration of an information processing device according to a first embodiment. [Figure 2] 2 is a block diagram schematically showing the flow of information processing according to the first embodiment. FIG. [Figure 3] FIG. 10 is a diagram showing a concept information tree. [Figure 4] FIG. 10 is a block diagram schematically illustrating the flow of information processing according to the second embodiment. [Figure 5] FIG. 11 is a block diagram showing the flow of information processing according to the third embodiment. [Figure 6] FIG. 10 is a diagram showing a GUI window for abstracting input text. [Figure 7] FIG. 11 is a diagram conceptually showing data used for a search by a similar document search unit in information processing according to the third embodiment. [Figure 8] FIG. 10 is a diagram showing a document database in which abstract documents are registered in advance. DETAILED DESCRIPTION OF THE INVENTION

[0020] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. Note that the components described in the embodiment are merely examples and are not intended to limit the scope of the present invention. In the drawings, the dimensions and numbers of each part may be exaggerated or simplified as necessary to facilitate understanding.

[0021] 1. First embodiment FIG. 1 is a diagram showing the configuration of an information processing device 1 according to the first embodiment. The information processing device 1 is a computer and includes a processor 11 and a memory 13. The processor 11 includes, for example, a CPU (Central Processing Unit). The memory 13 includes, for example, a ROM (Read-Only Memory) or a RAM (Random-Access Memory). The memory 13 may also include an auxiliary storage device such as a hard disk drive (HDD) or a solid-state drive (SSD). The memory 13 is connected to the processor 11 via a system bus.

[0022] The memory 13 stores a computer program P. The computer program P is executable by the processor 11 of the information processing device 1. When the processor 11 executes the computer program P, information processing, described below, is performed in the information processing device 1. The computer program P may be recorded on a non-transitory recording medium. The recording medium is, for example, optical media or a semiconductor memory such as a USB memory. The computer program P recorded on the recording medium can be read by a reading device (not shown). The computer program P may be stored in the memory 13 via a network line (not shown).

[0023] The information processing device 1 further includes a display 15 and an input device 17. The display 15 and the input device 17 are connected to the processor via a system bus. The display 15 is a device that visually displays the output of the information processing device 1, and is specifically a liquid crystal display. The input device 17 is a device that allows a user to input data and commands to the information processing device 1, and is specifically a keyboard or a mouse. Note that the display 15 may have the function of an input device by being equipped with, for example, a touch panel.

[0024] The information processing device 1 further includes a communication unit 19. The communication unit 19 is configured with a network adapter or the like. The communication unit 19 transmits and receives various data to and from external devices such as servers via a network (not shown) using wired or wireless communication protocols. The communication unit 19 is connected to the processor 11 via a system bus.

[0025] 2 is a block diagram showing an outline of the flow of information processing according to the first embodiment. The similar document search unit 31, concept abstraction unit 33, and sentence generation unit 35 shown in FIG. 2 are functional blocks realized by the processor 11 executing a computer program P. The information processing device 1 is configured to generate an output text 27 from an input text 21 in an interactive manner using a technique called retrieval expansion generation (RAG), as will be described below.

[0026] First, the similar document search unit 31 accepts input of text (question sentence) from the user. The user inputs the text via the input device 17. Then, the similar document search unit 31 searches for documents similar to the input text 21 in a document database 41 in which a plurality of documents are registered. This search process results in one or more similar documents 23. The document database 41 may be provided in the information processing device 1, or may be realized by an external device such as a server that can communicate with the information processing device 1.

[0027] The concept abstraction unit 33 uses the concept information tree T to abstract confidential information (words) to be concealed that is included in the similar documents 23. The concept information tree T is concept system information that includes multiple concepts and defines the hierarchical relationships between the concepts.

[0028] Fig. 3 is a diagram showing a concept information tree T. The concept information tree T is made up of a plurality of concepts, and each concept is assigned a word that is a specific notation, such as "medicinal solution," "medicinal solution A," "medicinal solution B," "medicinal solution B1," or "medicinal solution B2." Note that notations such as "medicinal solution A" in Fig. 3 are notations for convenience, and in reality, a specific name (such as the name of a medical solution or a name including chemical properties (acidicity or basicity)) is assigned.

[0029] In the concept information tree T, the superordinate and subordinate relationships are described in a tree structure. For example, from the perspective of "medicinal solution B," "medicinal solution" is the concept one level higher, connected by a link. Also, from the perspective of "medicinal solution B," "medicinal solution B1" and "medicinal solution B2" are each concepts one level lower, connected by a link.

[0030] As shown in Figure 3, the concept information tree T defines a concept hierarchy for each concept. The concept hierarchy is information that indicates the depth of a concept from a base concept (here, the top-level root concept). In Figure 3, the hierarchical level of the root concept is "Level 1," and the number indicating the level increases by one as the concept depth increases, such as "Level 2," "Level 3," and so on.

[0031] 3 shows a tree structure for one root concept in the concept information tree T. However, the concept information tree T may have multiple types of root concepts, and each root concept may have its own tree structure.

[0032] The conceptual information tree T is prepared in advance by a user or the like and stored in the memory 13 together with the computer program P. Note that the conceptual information tree T may also be stored in an external device such as a server that can communicate with the information processing device 1.

[0033] Returning to FIG. 2, the concept abstraction unit 33 acquires the reader attribute R and abstracts confidential information contained in the similar document 23 according to the acquired reader attribute R. The abstraction process by the concept abstraction unit 33 generates an abstracted similar document 25. The reader attribute R is information indicating the attributes of the user who is the reader, and is information indicating the conceptual hierarchy that the user is permitted to view. The reader attribute R corresponds to the hierarchical level defined in the conceptual information tree T. That is, for example, if the reader attribute R is "Level-2," the reader is permitted to view the concepts of "Level-2" and higher levels (i.e., concepts of "Level-1") contained in the conceptual information tree T, but is prohibited from viewing concepts lower than "Level-2" (i.e., concepts of "Level-3").

[0034] The reader attribute R of each user is managed, for example, in a user database (not shown). The concept abstraction unit 33 may perform predetermined user authentication and acquire the reader attribute R of the user who has been successfully authenticated from the user database.

[0035] In the abstraction process, the concept abstraction unit 33 analyzes the similar document 23 to be processed and performs word segmentation. Then, the concept abstraction unit 33 identifies words contained in the analyzed similar document 23 that are not accessible to the user (reader) by referencing the words in the analyzed similar document 23 with the concept information tree T. More specifically, the concept abstraction unit 33 determines, for each word, whether the word corresponds to a concept registered in the concept information tree T. If the word is registered in the concept information tree T, the concept abstraction unit 33 acquires the hierarchical level of the word. Furthermore, if the acquired hierarchical level is lower than the reader attribute R, the concept abstraction unit 33 identifies the target word as an unaccessible word. After identifying the unaccessible word, the concept abstraction unit 33 abstracts the word (super-conceptualizes it) into a word in a conceptual hierarchy that is permitted to be accessed.

[0036] For example, if the reader attribute R is Level-2 and the similar document 23 contains a Level-3 word "liquid drug B1," the "liquid drug B1" is identified as a word that is prohibited from being viewed. The concept abstraction unit 33 then replaces the word with "liquid drug B," a Level-2 word that is permitted to be viewed. In this way, an abstracted similar document 25 is generated in which confidential information is abstracted according to the user's reader attribute R. Note that if there are multiple similar documents 23, an abstracted similar document 25 is generated for each similar document 23.

[0037] The sentence generation unit 35 uses a sentence generation model M to obtain an output text 27 based on the input text 21 and the abstract similar document 25. The sentence generation model M is a trained model that is trained to generate a response sentence to the input text based on the input text and external information related to the input text. Specifically, the sentence generation model M is a large language model (LLM). The LLM is, for example, a deep neural network based on a self-attention mechanism called a Transformer. The Transfer is capable of capturing the relationship of the entire input sequence through the self-attention mechanism.

[0038] The abstract similar document 25 is a document similar to the input text 21. In other words, the abstract similar document 25 corresponds to external information related to the input text 21. The sentence generation unit 35 inputs the input text 21 and the abstract similar document 25 into the sentence generation model M, thereby obtaining an output text 27 that is a response sentence to the input text 21. The information processing device 1 displays the obtained output text 27 on the display 15. This allows the user to view the output text 27.

[0039] The generation of the output text 27 using the sentence generation model M may be realized in an external device such as a server that can communicate with the information processing device 1. In this case, the information processing device 1 transmits the input text 21 and the abstract similar document 25 to the external device. Then, the information processing device 1 may acquire the output text 27 by receiving the output text 27 generated by the external device.

[0040] As described above, the information processing device 1 searches the document database 41 for similar documents 23 that are similar to the input text 21, and obtains the output text 27, which is an answer sentence, from the input text 21 and the abstract similar document 25 based on the searched similar document 23 using the sentence generation model M.

[0041] According to the information processing device 1, even if the similar document 23 contains confidential information, the abstract similar document 25 obtained by abstracting the confidential information is input to the sentence generation model M. This significantly reduces the probability that text containing confidential information is output from the sentence generation model M. Therefore, it is possible to effectively reduce the leakage of confidential information.

[0042] 2. Second Embodiment Next, a second embodiment will be described. In the following description, elements having the same functions as elements already described will be given the same reference numerals or reference numerals with an additional alphabetical character, and detailed description thereof may be omitted.

[0043] FIG. 4 is a block diagram illustrating a schematic flow of information processing according to the second embodiment. In the information processing according to the second embodiment, the output text 27 is obtained by the same process as the information processing according to the first embodiment described with reference to FIG. 2. Then, the concept abstraction unit 33 executes a process of abstracting confidential information contained in the output text 27 based on the reader attribute R and the conceptual information tree T. This abstraction process is the same as the process in which the concept abstraction unit 33 abstracts the input text 21, and therefore details thereof will be omitted. If the output text 27 contains words of confidential information that are prohibited from being viewed for the reader attribute R, the abstraction process of the concept abstraction unit 33 generates an abstracted output text 29 in which the confidential information is abstracted. The information processing device 1 displays the generated abstracted output text 29 on the display 15. This allows the user to view the abstracted output text 29.

[0044] According to the information processing of the second embodiment, even if the text generation model M outputs the output text 27 including confidential information that is prohibited for the user to view, the confidential information is abstracted by the abstraction processing, which further reduces the leakage of confidential information.

[0045] 3. Third Embodiment 5 is a block diagram showing the flow of information processing according to the third embodiment. In the information processing according to the third embodiment, before searching for documents similar to the input text 21 in the document database 41, the concept abstraction unit 33 abstracts the input text 21 and the documents registered in the document database 41. Then, the similar document search unit 31 searches for similar documents 23 similar to the input text 21 by using the abstracted input text 21 (abstracted input text 21a) and the abstracted document (abstracted document 43). The data matching unit 37 shown in FIG. 5 is a function realized by the processor 11 executing the computer program P. The function of the data matching unit 37 will be described later.

[0046] 6A and 6B are diagrams showing a GUI window 5 for abstracting input text 21. FIG. 6A shows the initial state of the GUI window 5. FIG. 6B shows the GUI window 5 that has transitioned from the initial state to a pop-up state. First, as shown in FIG. 6A, the GUI window 5 in the initial state has an input area section 51, a layer specification section 53, and a search button 55.

[0047] The input area section 51 defines an area for the user to input the target input text 21. When the user inputs text with the input area section 51 selected, the input text is displayed in the input area section 51.

[0048] When the input text 21 is entered, the user operates the hierarchy specification unit 53 to specify the hierarchy level at which abstraction is to be performed. The hierarchy levels are displayed so that they can be selected, for example, in a pull-down format. The hierarchy level selected by the hierarchy specification unit 53 corresponds to the hierarchy levels (hierarchy-1, hierarchy-2, etc.) defined in the conceptual information tree T.

[0049] When the user presses the layer specification section 53 in the GUI window 5 in the initial state, the GUI window 5 transitions to the pop-up state shown in Fig. 6(b). In the pop-up state, the GUI window 5 displays a pull-down of the layer specification section 53 and a preview section 57. When the user selects a specific layer level on the pull-down of the layer specification section 53, an abstract input text 21a obtained by abstracting the input text 21 up to the selected layer level is displayed in the preview section 57.

[0050] When abstracting the input text 21, the concept abstraction unit 33 first identifies words at a hierarchical level lower than the specified hierarchical level by looking up words included in the input text 21 in the concept information tree T. Then, the concept abstraction unit 33 replaces the identified words with words at the specified hierarchical level. As a result, an abstract input text 21a is generated in which the input text 21 is abstracted to the conceptual level of the specified hierarchical level, and the abstract input text 21a is displayed in the preview unit 57. Note that if the hierarchy specification unit 53 is pressed before the input text 21 is input, a sample sentence prepared in advance may be displayed in the preview unit 57.

[0051] In the pop-up GUI window 5, a decision button 59 is displayed. When the user presses the decision button 59, the GUI window 5 transitions to the initial state shown in Fig. 6(a). When the search button 55 is pressed after the input text 21 has been entered and a hierarchical level has been selected, the input text 21 and the hierarchical level are confirmed. Then, a search for similar documents is performed using the document database 41.

[0052] FIG. 7 is a conceptual diagram showing data used for a search by the similar document search unit 31 in information processing according to the third embodiment. When a hierarchical level is determined by an operation on the GUI window 5, the concept abstraction unit 33 performs a process of abstracting each document in the document database 41 up to the determined hierarchical level, thereby generating a plurality of abstract documents 43. This abstraction process is the same as the process of abstracting the input text 21. The concept abstraction unit 33 also assigns ID information to each generated abstract document 43 to identify the original document. This ID information is the document ID assigned to each document in the document database 41. Once generation of the abstract document 43 is complete, the similar document search unit 31 searches for a document similar to the abstract input text 21a from among the plurality of abstract documents 43.

[0053] 5, when an abstract document 43 similar to the abstract input text 21a is found by the search, the similar document search unit 31 acquires a similar document ID 231, which is ID information of the abstract document 43. The similar document ID 231 is passed to the data matching unit 37. Note that when multiple similar abstract documents 43 are found, the similarity may be calculated, and only the similar document ID 231 of the abstract document 43 with the highest similarity may be passed to the data matching unit 37.

[0054] By comparing the similar document ID 231 in the highly similar document database 41, the document corresponding to the similar document ID 231 is acquired as the similar document 23. Note that the process up to acquiring the output text 27 from the similar document 23 is the same as the process described in the first embodiment, and therefore the description will be omitted.

[0055] As described above, according to the information processing of the third embodiment, by performing a similar document search using the abstracted abstract input text 21a and the abstract document 43, it is possible to obtain a wide range of similar documents without being limited by specific words.

[0056] In the third embodiment, the concept abstraction unit 33 generates the abstract document 43 at a specified hierarchical level from the document database 41, but this is not essential. Fig. 8 is a diagram showing the document database 41 in which the abstract documents 43 are registered in advance. As shown in Fig. 8, the abstract documents 43 for each hierarchical level may be stored in advance in the document database 41 along with the original document before non-abstraction. By generating the abstract document 43 once in this way, it is no longer necessary to generate the abstract document 43 again, and therefore the similar document search can be started quickly.

[0057] Furthermore, in the third embodiment, the original similar document 23 before abstraction is first acquired, and then the concept abstraction unit 33 performs abstraction according to the reader attribute R, thereby generating the abstract similar document 25. However, the abstract document 43 discovered by the similar document search unit 31 may also be used as the abstract similar document 25. In this case, it is possible to omit the acquisition of the similar document 23 before abstraction through collation by the data collation unit 37 and the abstraction of the similar document 23, thereby simplifying information processing.

[0058] <4. Variations> Although the embodiments have been described above, the present invention is not limited to the above and various modifications are possible.

[0059] For example, in the above embodiment, hierarchical levels are defined in the conceptual information tree T, but this is not essential. For example, in the conceptual information tree T, instead of hierarchical levels, a disclosure range indicating the range within which disclosure is permitted may be defined for each concept, and the reader attribute R may be used as information indicating whether or not the reader attribute R is included in the disclosure range. In this case, the concept abstraction unit 33 may identify words whose reader attribute R is not included in the disclosure range, and abstract (superordinately conceptualize) the words into words in the disclosure range that include the reader attribute R.

[0060] Although the present invention has been described in detail, the above description is merely illustrative in all respects and does not limit the present invention. It is understood that countless variations not illustrated can be envisioned without departing from the scope of the present invention. The configurations described in the above embodiments and variations can be combined or omitted as appropriate as long as they are not mutually inconsistent. [Explanation of symbols]

[0061] 1: Information processing device (computer) 21: Input text 21a: Abstract input text 23: Similar documents 25: Abstracted Similar Documents 27: Output text 29: Abstract output text 41: Document database 43: Abstract Document M: Sentence generation model P: Computer Program R: Reader attributes T: Concept information tree (concept system information)

Claims

1. 1. A computer-implemented information processing method, comprising: a) obtaining input text; b) acquiring, based on documents registered in a document database, abstracted similar documents that are similar to the input text and from which confidential information has been abstracted by an abstraction process; c) using a sentence generation model trained to generate a response sentence based on a text and external information related to the text, when the input text and the abstract similar document are input, to obtain an output text that is a response sentence to the input text; An information processing method, including:

2. 2. The information processing method according to claim 1, The step b) b11) obtaining similar documents by searching the document database for documents similar to the input text; b12) generating the abstracted similar document by abstracting the confidential information included in the similar document; An information processing method, including:

3. 3. The information processing method according to claim 2, The step b11) generating an abstract input text by abstracting words included in the input text; obtaining documents similar to the abstract input text by searching the document database for documents similar to the abstract input text; An information processing method, including:

4. 3. The information processing method according to claim 1, further comprising: The step b) b21) abstracting confidential information from each of a plurality of documents stored in the document database; b22) a step of searching for documents similar to the input text from among the abstract documents abstracted in step b21) to obtain the abstract similar documents; An information processing method, including:

5. 5. The information processing method according to claim 4, The information processing method, wherein the step c) includes a step of inputting the document searched from the abstract documents in the step b22) into the sentence generation model as the abstract similar document.

6. 3. The information processing method according to claim 1, further comprising: d) generating an abstracted output text by abstracting confidential information contained in the output text obtained by step c); The information processing method further includes:

7. 3. The information processing method according to claim 1, further comprising: In the step b), the abstraction process includes a process of abstracting the confidential information into concepts corresponding to a predetermined conceptual hierarchy using concept hierarchy information that defines the hierarchical relationships between multiple concepts.

8. A computer-readable computer program comprising: A computer program causing the computer to execute the information processing method according to claim 1 or 2.

Images