Idea generation support device, idea generation support method, and idea generation support program

Abstract

To improve the novelty and inventiveness of ideas. [Solution] The idea generation support device has access to a database that stores a set of known sentences, a domain independence vector for each known sentence having a component indicating its independence from the domain to which the known sentence belongs, and a set of subject, object, and predicate combinations that constitute the known sentences. The device accepts an input sentence and, based on a domain correlation parameter that indicates the correlation between the known sentence and multiple domains, converts the vector representation of the input sentence into a domain independence vector having a component indicating its independence from the domain to which the input sentence belongs. Based on the domain independence vector of the input sentence and the domain independence vector of the known sentences, it identifies known similar sentences from the set of known sentences that are similar to the input sentence, extracts the subject and object that constitute the known similar sentences from the database, and outputs the subject and object that constitute the known similar sentences in a displayable format.

Description

[Technical Field]

[0001] This invention relates to an idea generation support device, an idea generation support method, and an idea generation support program. [Background technology]

[0002] When novel and progressive ideas that differ from known concepts are required in research and development or project planning, search systems are used to gather information via the internet or retrieve information from literature databases. These search systems, when given keywords, present existing examples related to those keywords.

[0003] However, selecting appropriate keywords is not easy in search systems. Therefore, a technique called a large-scale language model is used, which involves training a massive machine learning model with a large amount of natural language sentences, such as Japanese, to predict the continuation of the input sentence. This allows for the prediction of responses to sentences entered by humans, and chatbot systems that enable people to converse with artificial intelligence have been put into practical use.

[0004] This chatbot system allows users to perform literature searches through dialogue by incorporating keyword search results from databases and other sources into its responses. This has the advantage of making keyword selection easier.

[0005] Thus, it is possible to use publicly available technologies to search for information that a person desires. However, all of these publicly available technologies only have the function of obtaining information that is close to the person's own thoughts. Taking this into consideration, Patent Document 1 below has been proposed to enable people to take their own thoughts on a leap to something new and broaden their ideas.

[0006] Patent Document 1 describes an invention generation device that can automatically construct an invention, comprising: a storage unit that stores conceptual information which is a pair of lower-level information which is a string of characters describing the invention and higher-level information which is a higher-level concept of the lower-level information; a receiving unit that receives original invention information which is a string of characters describing the invention; a partial information acquisition unit that acquires partial information which is a part that constitutes the original invention information; a higher-level information acquisition unit that acquires higher-level information from the storage unit that is paired with the lower-level information corresponding to each partial information; an invention information construction unit that constructs invention information which is information of a higher-level conceptualized invention by concatenating the higher-level information acquired by the higher-level information acquisition unit, or any of the partial information among the partial information, with the higher-level information acquired by the higher-level information acquisition unit in the order of the corresponding partial information; and an output unit that outputs invention information. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2018-13920 [Overview of the project] [Problems that the invention aims to solve]

[0008] However, since the idea itself is no different from one devised by a human, Patent Document 1 does not take into consideration how to enhance the novelty or inventiveness of the idea.

[0009] The purpose of this invention is to improve the novelty and inventiveness of the idea. [Means for solving the problem]

[0010] An idea generation support device, which is one aspect of the invention disclosed in this application, is an idea generation support device having a processor that executes a program and a storage device that stores the program, and is able to access a database that stores a set of known sentences, a field independence vector for each known sentence having a component indicating independence from the field to which the known sentence belongs, and a set of subject, object, and predicate combinations that constitute the known sentence, and the processor is characterized by executing an input process that accepts an input sentence, a conversion process that converts the vector representation of the input sentence into a field independence vector having a component indicating independence from the field to which the input sentence belongs based on a field correlation parameter indicating the correlation between the known sentence and a plurality of fields, an extraction process that identifies known similar sentences similar to the input sentence from the set of known sentences based on the field independence vector of the input sentence and the field independence vector of the known sentence, and extracts the subject and object that constitute the known similar sentence from the database, and a first output process that outputs the subject and object that constitute the known similar sentence extracted by the extraction process in a displayable format. [Effects of the Invention]

[0011] According to a typical embodiment of the present invention, it is possible to improve the novelty and inventiveness of the idea. Problems, configurations, and effects other than those mentioned above will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0012] [Figure 1] Figure 1 is an explanatory diagram showing an example configuration of the idea generation support system according to Example 1. [Figure 2] Figure 2 is a block diagram showing an example of the hardware configuration of the idea generation support device according to Example 1. [Figure 3] Figure 3 is an explanatory diagram showing an example of a management database according to Example 1. [Figure 4] Figure 4 is an explanatory diagram showing an example of a literature knowledge database related to Example 1. [Figure 5]FIG. 5 is an explanatory diagram showing an example of data extraction from the knowledge graph according to Example 1. [Figure 6] FIG. 6 is an explanatory diagram showing an example of a set of literature noun phrase vectors according to Example 1. [Figure 7] FIG. 7 is an explanatory diagram showing an example of a literature triple DB according to Example 1. [Figure 8] FIG. 8 is a block diagram showing a functional configuration example of the idea support device according to Example 1. [Figure 9] FIG. 9 is a flowchart showing an example of an idea support processing procedure by the idea support device according to Example 1. [Figure 10] FIG. 10 is an explanatory diagram showing an example of an input screen according to Example 1. [Figure 11] FIG. 11 is a flowchart showing a detailed processing procedure example of the knowledge graph generation process (step S902) according to Example 1. [Figure 12] FIG. 12 is an explanatory diagram showing an example of creating a knowledge graph for an input sentence according to Example 1. [Figure 13] FIG. 13 is a flowchart showing a detailed processing procedure example of the vector conversion process (step S903) according to Example 1. [Figure 14] FIG. 14 is an explanatory diagram showing an example of knowledge data according to Example 1. [Figure 15] FIG. 15 is a flowchart showing a detailed processing procedure example of the amendment generation process (step S904) according to Example 1. [Figure 16] FIG. 16 is an explanatory diagram showing Example 1 of a display screen according to Example 1. [Figure 17] FIG. 17 is an explanatory diagram showing an example of predicate acquisition according to Example 1. [Figure 18] FIG. 18 is an explanatory diagram showing Example 2 of a display screen according to Example 1. [Figure 19] FIG. 19 is a flowchart showing a detailed processing procedure example of the verbalization process (step S905) according to Example 1. [Figure 20]Figure 20 is an explanatory diagram showing an example of extracting a triple data set from a proposed modification of the input sentence knowledge graph according to Example 1. [Figure 21] Figure 21 is an explanatory diagram showing an example of the display screen according to Example 1. [Figure 22] Figure 22 is a flowchart showing a detailed example of the processing procedure for the revision proposal generation process (step S904) according to Example 2. [Figure 23] Figure 23 is an explanatory diagram showing an example of identifying a similar structure in step S2003. [Figure 24] Figure 24 is an explanatory diagram showing an example of extracting a difference graph in step S2004. [Figure 25] Figure 25 is an explanatory diagram showing an example of a combination of difference graphs for the input sentence knowledge graph in step S2004. [Figure 26] Figure 26 is an explanatory diagram showing an example of a revised proposal screen for Example 2. [Figure 27] Figure 27 is an explanatory diagram showing the revised proposal presentation screen after the transition to the revised proposal according to Example 2. [Modes for carrying out the invention] [Examples]

[0013] <Figure 1: Example of an idea generation support system configuration> Figure 1 is an explanatory diagram showing an example configuration of an idea generation support system according to Embodiment 1. The idea generation support system 100 includes an idea generation support device 101, a management DB 102, a generation AI (Artificial Intelligence) 103, and a terminal 104. The management DB 102 and the generation AI 103 may be implemented within the idea generation support device 101, or they may be implemented on a computer accessible from the idea generation support device 101 via a network 110 such as the Internet, LAN (Local Area Network), or WAN (Wide Area Network).

[0014] The idea generation support device 101 is a computer that performs idea generation support processing. The management DB 102 is a database that stores data necessary for idea generation support processing. The generation AI 103 uses a dataset and includes a language model learned through natural language processing, and generates sentences using this language model.

[0015] A language model is a type of probabilistic model used in natural language processing, designed to probabilistically predict how likely a given word or sentence is to occur in natural language. Specifically, in the field of natural language processing, a language model is a mathematical model that learns language patterns and grammatical rules to generate and understand natural language. For example, the Generative AI 103 uses a language model to calculate the probability of occurrence of a given word sequence or sentence, and to compare the probability of occurrence of multiple word sequences or sentences, thereby automatically generating the most likely word or sentence based on the context when predicting the next word or sentence.

[0016] Furthermore, the language model also has an embedding function that converts a given text into a vector, which is a sequence of numerical values.

[0017] In this way, when the Generative AI 103 receives a query called a prompt, it uses a language model trained on a vast amount of data to output an answer to that query.

[0018] Terminal 104 transmits various data to the idea generation support device 101 and receives and displays various data from the idea generation support device 101.

[0019] <Figure 2: Example of hardware configuration of the idea generation support device 101> Figure 2 is a block diagram showing an example of the hardware configuration of the idea generation support device 101 according to Embodiment 1. The idea generation support device 101 includes a processor 201, a storage device 202, an input device 203, an output device 204, and a communication interface (communication IF) 205. The processor 201, storage device 202, input device 203, output device 204, and communication IF 205 are connected by a bus 206. The processor 201 controls the idea generation support device 101. The storage device 202 serves as the work area for the processor 201. The storage device 202 is also a non-temporary or temporary recording medium that stores various programs and data. Examples of storage devices 202 include ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), and flash memory. The input device 203 inputs data. Examples of input devices 203 include a keyboard, mouse, touch panel, numeric keypad, scanner, microphone, and sensor. Output device 204 outputs data. Examples of output devices 204 include a display, printer, and speaker. Communication IF 205 connects to network 110 and sends and receives data.

[0020] Furthermore, the computer that executes the generation AI 103 in terminal 104 also has the same hardware configuration as shown in Figure 2.

[0021] <Figure 3 Management DB102> Figure 3 is an explanatory diagram showing an example of the management DB 102 according to Example 1. The management DB 102 stores the literature knowledge DB 301, the literature noun phrase vector set 302, and the literature triple DB 303. The literature knowledge DB 301 is a database that stores knowledge data (hereinafter referred to as literature knowledge data) related to known literature. The literature noun phrase vector set 302 is a set of noun phrase vectors for nouns that appear in known literature. A noun phrase vector is a vector of a word or phrase whose part of speech is a noun. The literature triple DB 303 is a database that stores triple data contained in known literature. Triple data is a combination of subject, object, and predicate in a sentence.

[0022] <Figure 4: Literature Knowledge DB301> Figure 4 is an explanatory diagram showing an example of the bibliographic knowledge database 301 according to Example 1. The bibliographic knowledge database 301 has the following fields: bibliographic ID 401, bibliographic knowledge graph 402, bibliographic graph element vector list 403, and bibliographic knowledge text vector 404. The combination of each field in the same row constitutes an entry that defines the bibliographic knowledge data for that bibliographic document. A bibliographic document is text data that describes a document consisting of one or more sentences.

[0023] Document ID 401 is an identifier L1 to LK (where n is an integer greater than or equal to 1) that uniquely identifies a known document. Documents with document ID 401 being Lk (where i is an integer satisfying 1 ≤ i ≤ n) are denoted as document Lk. Document knowledge data with document ID 401 being Lk may also be denoted as document knowledge data Lk.

[0024] The bibliographic knowledge graph 402 is a directed graph relating to bibliographic data Lk. Hereafter, it will be referred to as bibliographic knowledge graph KG(Lk). Specifically, for example, bibliographic knowledge graph 402 is a directed graph in which the subject and object are nodes and the predicate is an edge, with the starting end of an edge without an arrowhead connected to the subject node and the ending end with an arrowhead connected to the object. If the same noun phrase (subject and object) exists across multiple sentences, it is represented by a single common node.

[0025] The bibliographic graph element vector list 403 is a list of graph element vectors. A graph element vector is a domain-independent vector of a graph element. A graph element is a component of the bibliographic knowledge graph 402, specifically, for example, a group of subject nodes, a group of object nodes, and a group of predicate edges. The bibliographic knowledge text vector 404 is, for example, a domain-independent vector of that bibliographic Lk.

[0026] A domain-independent vector is a vector that treats the string from which it is generated as equivalent to strings from other domains. Specifically, for example, a domain-independent vector is obtained by converting the vector representation of a string such as a sentence or phrase that is the target of the domain-independent vector generation into a domain representation vector that shows which domain the string depends on using a domain correlation parameter that shows the correlation between multiple domains, and then attempting to remove the component that indicates the dependency on a domain from that domain representation vector. This residual component is a component that shows independence from the domain in question and is called the independent component. By calculating the similarity between domain-independent vectors, it is possible to identify the cross-domain similarity relationships between words and phrases indicated by the domain-independent vectors.

[0027] The bibliographic graph element vector list 403 is a set of domain-independent vectors obtained by attempting to remove domain-dependent components from the subject vector representation, object vector representation, and predicate vector representation, which are embedded with generation AI 103, respectively.

[0028] In the bibliographic graph element vector list 403, the domain independence vectors for each subject within bibliographic Lk are denoted as bibliographic subject domain independence vectors VS(Lk)1, VS(Lk)2, ... (if these are not distinguished, they are denoted as bibliographic subject domain independence vector VS(Lk)).

[0029] In the bibliographic graph element vector list 403, the domain independence vectors of each object within bibliographic Lk are denoted as bibliographic object domain independence vectors VO(Lk)1, VO(Lk)2, ... (if these are not distinguished, they are denoted as bibliographic object domain independence vector VO(Lk)).

[0030] In the bibliographic graph element vector list 403, the domain independence vectors of each predicate within bibliographic Lk are denoted as bibliographic predicate domain independence vectors VP(Lk)1, VP(Lk)2, ... (if these are not distinguished, they are denoted as bibliographic predicate domain independence vector VP(Lk)).

[0031] The bibliographic knowledge text vector 404 is a domain-independent vector obtained by attempting to remove components indicating domain dependency from a vector representation embedded with the full text of bibliographic Lk using the generation AI 103. In the bibliographic knowledge text vector 404, the bibliographic domain-independent vectors of bibliographic L1 to LK are denoted as VT(L1) to VT(LK) (if they are not distinguished, they are denoted as bibliographic domain-independent vector VT(Lk)).

[0032] <Figure 5: Example of data extraction from the Literature Knowledge Graph KG(Lk)> Figure 5 is an explanatory diagram showing an example of data extraction from a literature knowledge graph KG(Lk) according to Example 1. The idea generation support device 101 extracts a literature noun phrase set 501(Lk) and a literature triple data set 502(Lk) from the literature knowledge graph KG(Lk). The literature knowledge graph KG(Lk) is a directed graph composed of nodes N1 to N6 and edges E12, E14, E32, E45, E46, and E53, for example. Nodes are denoted as node N when not distinguished, and edges are denoted as edge E when not distinguished. The start of an edge E without an arrowhead is connected to node N indicating the subject, and the end of an edge E with an arrowhead is connected to node N indicating the object.

[0033] The noun phrase set 501(Lk) is the set of noun phrases that function as subjects and objects within the document Lk, as identified by nodes N1 to N6. For example, N1 = passenger, N2 = train, and N4 = ticket.

[0034] The document triple data set 502(Lk) is a set of triple data where each sentence in document Lk is combined with a subject represented by node N, an object represented by node N, and a predicate represented by edge E. For example, in the sentence "The passenger gets on the train," the triple data becomes {N1,N2,E12}={passenger,train,get}, and in the sentence "The passenger buys a ticket," the triple data becomes {N1,N4,E14}={passenger,ticket,buy}.

[0035] <Figure 6: Set of 302 Noun Phrase Vectors from Literature> Figure 6 is an explanatory diagram showing an example of a set of noun phrase vectors 302 in the literature noun phrase set according to Example 1. The set of noun phrase vectors 302 is a set of vector representations of noun phrases in literature Lk. The set of noun phrase vectors 302 has, as fields, a literature noun phrase 601 and a literature noun phrase vector 602.

[0036] The nominal phrase 601 in the literature is the set of nominal phrases 501(L1) to 501(LK) extracted from the literature knowledge graph KG(L1) to KG(LK) by removing duplicates from literature L1 to LK.

[0037] The document noun phrase vector 602 is, for example, a vector representation of the document noun phrase 601 embedded in the generating AI103, the document subject domain-independent vector VS(Lk), the document object domain-independent vector VO(Lk), and the document predicate domain-independent vector VP(Lk), and is obtained from the document graph element vector list 403.

[0038] <Figure 7: Reference Triple DB303> Figure 7 is an explanatory diagram showing an example of the literature triple DB303 according to Example 1. The literature triple DB303 is a collection of literature triple data sets 502(L1) to 502(LK).

[0039] <Figure 8: Example of Functional Configuration of Idea Generation Support Device 101> Figure 8 is a block diagram showing an example of the functional configuration of the idea generation support device 101 according to Embodiment 1. The idea generation support device 101 includes an input unit 801, a graph generation unit 802, a conversion unit 803, a revised proposal generation unit 804, a language processing unit 805, and an output unit 806. Specifically, these are realized, for example, by having the processor 201 execute a program stored in the memory device 202 shown in Figure 2.

[0040] The input unit 801 accepts the input of a string or the selection of a predetermined instruction by the user operating the input device 203.

[0041] The graph generation unit 802 generates a knowledge graph related to the input sentences input by the input unit 801. Specifically, for example, the graph generation unit 802 instructs the generation AI 103 to generate a knowledge graph related to the input sentences and obtains the generated knowledge graph related to the input sentences.

[0042] The conversion unit 803 converts input strings, such as input sentences and graph elements, into domain-independent vectors. Specifically, for example, the conversion unit 803 instructs the generation AI 103 to convert input sentences and graph elements into vector representations, and obtains the converted vector representations. The conversion unit 803 then converts the obtained vector representations into domain-independent vectors.

[0043] The conversion unit 803 converts the vector representation of the input string into a domain representation vector, which has a component indicating the dependency of the input string to the domain, based on a domain correlation parameter that shows the correlation between a known sentence and multiple domains. The domain representation vector is a vector that has a component indicating the domain dependency of which domain the input string belongs to. The conversion unit 803 converts the vector representation of the input string into a domain-independent vector of the input string by attempting to remove the component of the domain representation vector from the vector representation of the input string.

[0044] In other words, when the component indicating dependency on a domain is removed from the domain representation vector, the component indicating independence from that domain remains. The vector containing this remaining component is the domain independence vector of the input string. The domain correlation parameter is a pre-defined parameter. By using the domain correlation parameter, it is possible to identify which domains the input string depends on.

[0045] The revision proposal generation unit 804 generates revision proposals for the knowledge graph related to the input sentence. Specifically, for example, the revision proposal generation unit 804 generates noun phrases that can be used as revision candidates, and generates revision proposals for the knowledge graph related to the input sentence using the noun phrases that can be used as revision candidates.

[0046] The language processing unit 805 instructs the generation AI 103 to verbalize the knowledge graph related to the input sentence, and retrieves the verbalized sentence from the knowledge graph related to the input sentence from the generation AI 103.

[0047] The output unit 806 outputs various data in a format that can be displayed on the output device 204 or terminal 104. Specifically, for example, the output unit 806 outputs the input sentence input by the input unit 801, the data generated by the revised draft generation unit 804, and the sentence processed into language by the language processing unit 805 in a format that can be displayed.

[0048] <Figure 9 Idea generation support processing by the idea generation support device 101> Figure 9 is a flowchart showing an example of the idea generation support processing procedure by the idea generation support device 101 according to Embodiment 1.

[0049] (Step S901) The idea generation support device 101 performs input sentence reception processing using the input unit 801. Specifically, for example, the input unit 801 receives input sentences from the input screen by operating the input device 203.

[0050] [Figure 10 Example of input screen] Figure 10 is an explanatory diagram showing an example of an input screen according to Embodiment 1. The input screen 1000 is displayed on a display which is an example of an output device 204 of the idea generation support device 101 or terminal 104. The input screen 1000 has an input field 1001 and an input button 1002. The input field 1001 is a text box that accepts the input of a string and displays the entered string. In Figure 10, the input sentence 1010 is entered into and displayed in the input field 1001. The input button 1002 is a user interface for outputting the string in the input field 1001 to the graph generation unit 802 when pressed. In Figure 10, the input sentence 1010 is output to the graph generation unit 802 when the input button 1002 is pressed.

[0051] (Step S902) Returning to Figure 9, the idea generation support device 101 executes the input sentence knowledge graph generation process using the graph generation unit 802. The input sentence knowledge graph generation process is the process of generating a knowledge graph KG(IN) related to the input sentence 1010. The knowledge graph KG(IN) related to the input sentence 1010 is denoted as input sentence knowledge graph KG(IN).

[0052] [Figure 11 Input Text Knowledge Graph Generation Process (Step S902)] Figure 11 is a flowchart showing a detailed example of the input sentence knowledge graph generation process (step S902) according to Example 1.

[0053] (Step S1101) The idea generation support device 101 creates a triple data generation instruction statement for the input sentence 1010. The triple data generation instruction statement for the input sentence 1010 is a statement that instructs the generation of triple data for the input sentence 1010. Triple data for the input sentence 1010 (hereinafter referred to as input sentence triple data) is a combination of subject, object, and predicate contained in the input sentence 1010.

[0054] The triple data generation instruction for this input sentence 1010 is an instruction in a format that the generating AI 103 can accept, and the generated instruction becomes a sentence in which the input sentence 1010 is inserted into a predetermined standard phrase. For example, if a publicly known large-scale Japanese language model for chatbots is used as the generating AI 103, the following standard phrase will be prepared.

[0055] "System: You are an agent that helps the User understand documents. From the text entered by the User, extract the subject, object, and predicate pairs of sentences and return them in CSV format as "subject, object, predicate". / User:"

[0056] The idea generation support device 101 creates a triple data generation instruction statement related to the input statement 1010 by inserting the input statement 1010 after the above-mentioned standard phrase as a prompt to be given to the generation AI 103.

[0057] (Step S1102) The idea generation support device 101 sends a triple data generation instruction statement for the input sentence 1010 created in step S1101 to the generation AI 103. Upon receiving the triple data generation instruction statement for the input sentence 1010, the generation AI 103 generates the input sentence triple data. In the case of input sentence 1010, for example, four input sentence triple data items are generated, such as "Agent:Camera,Worker,Shoot\nWorker,Video,Confirm\nWorker,Server,Grant portrait rights". One or more input sentence triple data items are referred to as an input sentence triple data set.

[0058] (Step S1103) The idea generation support device 101 receives the input sentence triple data set from the generation AI 103.

[0059] (Step S1104) The idea generation support device 101 creates an input sentence knowledge graph KG(IN) based on the input sentence triple data set received in step S1103. This completes the input sentence knowledge graph generation process (step S1101), and the process proceeds to step S903.

[0060] [Figure 12: Example of creating an input text knowledge graph KG(IN)] Figure 12 is an explanatory diagram showing an example of creating an input sentence knowledge graph KG(IN) according to Embodiment 1. In Figure 12, the rectangle represents node N of the input sentence knowledge graph KG(IN). In Figure 12, the input sentence triple data set 1200 is created in step S1103, and the input sentence knowledge graph KG(IN) is created in step S1104. In the input sentence knowledge graph KG(IN), the idea generation support device 101 connects the subject node N and the object node N with a predicate edge E. The start of edge E is connected to the subject node N, and the end of edge E is connected to the object node N.

[0061] (Step S903) Returning to Figure 9, the idea generation support device 101 performs vector transformation processing using the transformation unit 803. Vector transformation processing is the process of generating domain-independent vectors from the vector representation of the input sentence 1010 and the vector representations of the graph elements in the knowledge graph KG(IN).

[0062] [Figure 13 Vector transformation process (step S903)] Figure 13 is a flowchart showing a detailed example of the vector transformation process (step S903) according to Example 1.

[0063] (Step S1301) The idea generation support device 101 creates vector representation conversion instructions for the input sentence 1010 and the graph elements of the input sentence knowledge graph KG(IN). Graph elements are the constituent elements (subject, object, predicate) of the input sentence knowledge graph KG(IN). The vector representation conversion instructions are created by inserting the input sentence 1010 and the graph elements of the input sentence knowledge graph KG(IN) as vector representation generation targets into a standard phrase that instructs the generation AI 103 to convert from the vector representation generation targets to vector representations.

[0064] (Step S1302) The idea generation support device 101 sends the vector representation conversion instruction statement created in step S1301 to the generation AI 103 as a prompt. Upon receiving the vector representation conversion instruction statement, the generation AI 103 converts the input sentence 1010 and the graph elements of the input sentence knowledge graph KG(IN) into vector representations. Specifically, for example, the generation AI 103 generates four vector representations by embedding: a vector representation of the input sentence 1010, a vector representation of the subject group, a vector representation of the object group, and a vector representation of the predicate group.

[0065] (Step S1303) The idea generation support device 101 receives the vector representation generated in step S1302 from the generation AI 103.

[0066] (Step S1304) The idea support device 101 converts each of the vector representation of the input sentence 1010 and the vector representation of the graph element into a field-dependent reduced vector using a projection matrix which is one of the field correlation parameters. The field correlation parameter is a parameter indicating the correlation between known documents L1 to LK and a plurality of fields. The field-dependent reduced vector consists of components that depend on the field to which the input sentence 1010 belongs, and is a vector having a smaller number of dimensions than its vector representation.

[0067] The projection matrix is a non-square matrix generated based on known canonical correlation analysis using various sentences in advance and vectors (hereinafter referred to as field representation vectors) representing the corresponding fields for each of them.

[0068] Here, as the field representation vector, for example, when classification codes such as electricity, information, and machinery are assigned to a sentence, it is a vector obtained by associating each dimension with those classification codes. For example, if the number of classifications is three, namely electricity, information, and machinery, the field representation vector will be a three-dimensional vector. That is, when the first dimension is associated with electricity, the second dimension with information, and the third dimension with machinery, the field representation vector for a sentence with codes for electricity and information will be (1, 1, 0) in such a way that only the dimensions corresponding to the coded fields are set to "1".

[0069] Also, for example, any vector capable of specifying a field may be used, such as a vector obtained by textifying the name of a field such as "electricity, information" and embedding it with the generation AI 103.

[0070] Here, when the number of dimensions of the vector representation is n and the number of dimensions of the field-dependent reduced vector is m (<n), the projection matrix A is an n×m matrix, and by multiplying this projection matrix A by the vector representation, the field-dependent reduced vector is calculated.

[0071] [Projection matrix A] The analysis for determining the projection matrix is ​​known as canonical correlation analysis and is performed using the following procedure. Assume that there are various sentences as described above and corresponding domain representation vectors. Let the vector representations be x1, x2, x3, ... and the domain representation vectors be y1, y2, y3, ... The set of vector representations x1, x2, x3, ... is denoted as the vector representation group x, and if the vector representations x1, x2, x3, ... are not distinguished, it is denoted as the vector representation xi. The set of domain representation vectors y1, y2, y3, ... is denoted as the domain representation vector group y, and if the domain representation vectors y1, y2, y3, ... are not distinguished, it is denoted as the domain representation vector yj (j may also be i).

[0072] At this time, we assume the existence of an n-dimensional vector a, which has the same number of dimensions as the vector representation xi, and a vector b, which has the same number of dimensions as the domain representation vector yj. The idea generation support device 101 calculates the vectors a and b that maximize the correlation coefficient ρ(a,b) between the inner product {a·xi} of vector a and the vector representation xi, and the inner product {b·yj} of vector b and the domain representation vector yj, using the following equation (1).

[0073]

number

[0074] Thus, vector a can be considered a vector for extracting the component with the highest correlation between the vector representation group x and the domain representation vector group y. The calculation in equation (1) above is a calculation to obtain one vector a and one vector b, but by expanding it to m vectors at a time using a known procedure for canonical correlation analysis, an n × m-dimensional transformation matrix can be obtained by arranging m types of n-dimensional vector a.

[0075] Specifically, for example, in the iterative calculation method, the component in the direction of vector a is subtracted from the vector representation xi using equation (2) below, and the component in the direction of vector b is subtracted from the field representation vector yj using equation (3) below.

[0076] [Number]

[0077] The xi on the left side of the above formula (2) (new) is the vector representation xi after subtraction, and the yj on the left side of the above formula (3) (new) is the field representation vector yj after subtraction.

[0078] Furthermore, the idea support device 101 calculates the above formula (1) using the new vectors a and b for each of the vector representation xi after subtraction (new) and the field representation vector yj after subtraction. (new) When the idea support device 101 repeatedly executes this calculation m times, m vectors a are obtained. By arranging the m vectors a thus obtained, an n×m-dimensional projection matrix A is obtained. Note that this procedure is an example, and any method can be used as long as it can extract components dependent on the field.

[0079] In this way, the idea support device 101 calculates the field-dependent reduction vector A·xi by multiplying the n×m-dimensional projection matrix A and the vector representation xi.

[0080] (Step S1305) Returning to FIG. 13, the idea support device 101 generates a field-dependent vector by inverse-transforming each of the field-dependent reduction vectors A·xi. Specifically, for example, the idea support device 101 multiplies the field-dependent reduction vector A·xi by the inverse projection matrix A -1 , which is one of the field correlation parameters, to calculate the field-dependent vector A -1 ·A·xi. The inverse projection matrix A -1 is the generalized inverse matrix of the projection matrix A obtained by calculating a known generalized inverse matrix for the projection matrix A. By multiplying the field-dependent vector A -1 , the field-dependent vector A -1 ·A·xi becomes the same dimension as the vector representation xi.

[0081] (Step S1306) The idea generation support device 101 uses a field-dependent vector A -1 For each of A and xi, calculate the domain-independent vector zi. The domain-dependent vector A obtained in step S1305 -1 ·A·xi can be considered a vector that has only the domain-dependent components of the vector representation xi. Therefore, the idea generation support device 101 generates the domain-dependent vector A from the original vector representation xi. -1 By subtracting A·xi, the domain-independent vector zi, from which the domain-dependent components have been attempted to be removed, is calculated (see equation (4) below). This completes the vector transformation process (step S903), and the process moves to step S904.

[0082] zi = xi - A -1 ·A·xi···(4)

[0083] [Figure 14 Input Sentence Knowledge Data] Figure 14 is an explanatory diagram showing an example of input sentence knowledge data according to Example 1. The input sentence knowledge data 1400 has an input sentence graph element vector 1401 and an input sentence knowledge text vector 1402 for the input sentence knowledge graph KG(IN).

[0084] The input sentence graph element vector 1401 is a domain-independent vector zi converted from each graph element of the input sentence knowledge graph KG(IN). The input sentence graph element vector 1401 converted from the subject group vector representation xi is denoted as input sentence graph element vector VS(IN), the input sentence graph element vector 1401 converted from the object group vector representation xi is denoted as input sentence graph element vector VO(IN), and the input sentence graph element vector 1401 converted from the predicate group vector representation xi is denoted as input sentence graph element vector VP(IN).

[0085] Furthermore, the input text knowledge text vector 1402 is the domain-independent vector zi converted from input text 1010, and is denoted as input text knowledge text vector VT(IN).

[0086] (Step S904) Returning to Figure 9, the idea generation support device 101 executes the revision proposal generation process using the revision proposal generation unit 804. The revision proposal generation process is the process of generating revision proposals for the input sentence knowledge graph KG(IN) using the input sentence knowledge data 1400.

[0087] [Figure 15 Revision Proposal Generation Process (Step S904)] Figure 15 is a flowchart showing a detailed example of the processing procedure for the revision proposal generation process (step S904) according to Example 1.

[0088] (Step S1501) The idea generation support device 101 acquires the input sentence 1010, the input sentence knowledge graph KG(IN) of the input sentence 1010, and the input sentence knowledge text vector 1402 in the input sentence knowledge data 1400 of the input sentence 1010. For example, the idea generation support device 101 acquires the input sentence knowledge text vector VT(IN) of the input sentence 1010.

[0089] (Step S1502) The idea generation support device 101 extracts from the literature knowledge DB 301 literature knowledge data Lk that has a literature field independent vector VT(Lk) similar to the input text knowledge vector VT(IN) acquired in step S1501 (hereinafter referred to as similar literature knowledge data).

[0090] (Step S1503) The idea generation support device 101 displays the input sentence 1010, the input sentence knowledge graph KG(IN) of the input sentence 1010, and the set of noun phrases in the similar literature knowledge data Lk, namely 501(Lk).

[0091] [Figure 16 Example of display screen 1] Figure 16 is an explanatory diagram showing an example of a display screen according to Embodiment 1. The display screen 1600 is displayed on a display which is an example of an output device 204 of the idea generation support device 101 or terminal 104. In Figure 16, rectangles are nodes of the input sentence knowledge graph KG(IN). The display screen 1600 is displayed in step S1503. The display screen 1600 includes an input sentence display unit 1601, a graph display unit 1602, a word display unit 1603, a predicate display unit 1604, and a confirmation button 1605.

[0092] The input text display section 1601 is an area for displaying the input text 1010. The graph display section 1602 is an area for displaying the input text knowledge graph KG(IN). The word display section 1603 is an area for displaying words and accepting the selection of those words. The words displayed in the word display section 1603 are the subject and object in the graph elements of the document noun phrase set 501(Lk), that is, the similar document knowledge data Lk.

[0093] The predicate display unit 1604 displays predicates that can be selected when the word pair selected in the graph display unit 1602 and the word display unit 1603 is used as the subject and object.

[0094] The confirmation button 1605 is a user interface that allows the user to confirm the display content of the graph display unit 1602 by pressing the input device 203.

[0095] The graph regeneration button 1606 is a user interface that, upon pressing the input device 203, generates a proposed revision of the input sentence knowledge graph KG(IN) using the word pairs (subject and object) selected in the graph display unit 1602 and the word display unit 1603, and the predicate selected from the predicate display unit 1604.

[0096] (Step S1504) Returning to Figure 15, the idea generation support device 101 accepts the selection of one node N (hereinafter referred to as the first selected term) from the input sentence knowledge graph KG(IN) and accepts the selection of one word (hereinafter referred to as the second selected term) from the set of document noun terms 501(Lk) of the similar document knowledge data Lk. In Figure 16, as an example, the first selected term is "worker" and the second selected term is "fingerprint" in the word display unit 1603.

[0097] (Step S1505) The idea generation support device 101 obtains the domain-independent vector of the first selected term from the input sentence graph element vector 1401 of the input sentence knowledge data 1400, and obtains the domain-independent vector of the second selected term from the literature graph element vector list 403 of the literature knowledge DB 301.

[0098] (Step S1506) The idea generation support device 101 searches the literature noun phrase vector set 302 for the first synonym of the first selected term, and searches the literature noun phrase vector set 302 for the second synonym of the second selected term. Specifically, for example, the idea generation support device 101 searches the literature noun phrase vector set 302 for literature noun phrases 601 of literature noun phrase vectors 602 whose Euclidean distance from the domain-independent vector of the first selected term is less than or equal to a predetermined threshold, and uses this as the first synonym. Similarly, the idea generation support device 101 searches the literature noun phrase vector set 302 for literature noun phrases 601 of literature noun phrase vectors 602 whose Euclidean distance from the domain-independent vector of the second selected term is less than or equal to a predetermined threshold, and uses this as the second synonym.

[0099] The first and second synonyms are retrieved from the literature noun phrase vector set 302, and therefore may differ from the fields of the first and second selected terms. Consequently, the first and second synonyms become cross-disciplinary synonyms that span across different fields.

[0100] (Step S1507) The idea generation support device 101 obtains first triple data from the document triple DB 303, where the first synonym is the subject and the second synonym is the object, and obtains second triple data from the document triple DB 303, where the second synonym is the subject and the first synonym is the object.

[0101] (Step S1508) The idea generation support device 101 obtains predicates from the first triple data and the second triple data.

[0102] [Figure 17 Example of predicate acquisition] Figure 17 is an explanatory diagram showing an example of predicate acquisition according to Example 1. In Figure 17, for the sake of explanation, a two-dimensional space is used as an example. Domain-independent vectors 1702 and 1703 are arranged in a two-dimensional spatial coordinate system 1701. Domain-independent vector 1702 represents the subject, and domain-independent vector 1703 represents the object. The arrow 1704 connecting the subject and object, which represent domain-independent vectors 1702 and 1703, represents the predicate. Hereafter, these will be referred to as subject 1702, object 1703, and predicate 1704.

[0103] Step S1506, the synonym search, represents the process of selecting other noun phrases within predetermined ranges 1720 and 1730 surrounding the subject 1702 and object 1703. Then, obtaining the first and second synonyms in step S1507 is equivalent to seeking arrows connecting the other noun phrases within the predetermined range 1720 and the other noun phrases within the predetermined range 1730.

[0104] In other words, steps S1506 and S1507 involve searching for triple data indicated by the domain-independent vectors 1705 and 1706 and the arrow 1707 connecting them. Thus, the processing in steps S1506 and S1507 seeks examples of predicates that are similar in terms of domain-independent vectors to the subject and object specified by the user. For example, when considering connecting the noun phrases "pedestrian" and "face," the synonym selection would involve selecting "worker" as the first synonym for "pedestrian" and "fingerprint" as the second synonym for "face," and then selecting the predicate "authenticate" from the triple data {worker, fingerprint, authenticate} in the document triple DB303 that includes the selected first and second synonyms.

[0105] (Step S1509) The idea generation support device 101 aggregates the number of occurrences for each predicate acquired in step S1508 and determines recommended predicates based on the aggregation results. Specifically, for example, the idea generation support device 101 determines recommended predicates if the aggregate count is equal to or greater than a predetermined number. Alternatively, the idea generation support device 101 may determine recommended predicates to be those with the highest aggregate count up to the Nth position (where N is a predetermined integer of 1 or more). In this way, recommended predicates are determined based on the number of occurrences of predicates.

[0106] (Step S1510) The idea generation support device 101 displays the recommended predicate determined in step S1509 on the predicate display unit 1604 of the display screen 1600. The recommended predicate is displayed as an icon that can be selected by pressing a button.

[0107] [Figure 18 Example of display screen 2] Figure 18 is an explanatory diagram showing an example of the display screen 2 according to Embodiment 1. In Figure 18, the rectangle represents the node of the proposed modification KG(IN)R of the input sentence knowledge graph KG(IN). Figure 18 shows the state in which the predicate display unit 1604 displays "imaging," "authentication," and "discrimination" as recommended predicates, compared to the state in Figure 16.

[0108] (Step S1511) The idea generation support device 101 accepts the selection of a recommended predicate upon pressing a button. Figure 18 shows the state in which "Authenticate" in the predicate display unit 1604 has been selected as the recommended predicate.

[0109] (Step S1512) The idea generation support device 101 accepts the generation of a proposed revision for the input text knowledge graph KG(IN) when the graph regeneration button 1606 is pressed. As a result, the proposed revision KG(IN)R of the input text knowledge graph KG(IN) is generated and displayed on the graph display unit 1602. In this example, the first selected word is "worker", the second selected word is "fingerprint" on the word display unit 1603, and the selected recommended predicate is "authenticate".

[0110] Therefore, a "fingerprint" node that does not exist in the input text knowledge graph KG(IN) is generated. Also, an "authenticate" edge that does not exist in the input text knowledge graph KG(IN) is generated. The start of this edge is connected to "worker" and the end is connected to "fingerprint". In this way, a revised version of the input text knowledge graph KG(IN) KG(IN)R is generated and displayed in the graph display unit 1602.

[0111] (Step S1513) The idea generation support device 101 confirms the display content of the graph display unit 1602 when the confirmation button 1513 is pressed. That is, if the graph regeneration button 1606 is not pressed, the input sentence knowledge graph KG(IN) is output to the language processing unit 805, and if the graph regeneration button 1606 is pressed, the revised version KG(IN)R of the input sentence knowledge graph KG(IN) is output to the language processing unit 805. This completes the revised version generation process (step S904), and the process moves to step S905.

[0112] Figure 18 illustrates a modified version of the input sentence knowledge graph KG(IN) called KG(IN)R, in which a node for "worker" indicating the subject, a node for "fingerprint" indicating the object, and an edge for the predicate indicating "authenticate" are added to the input sentence knowledge graph KG(IN). However, modifications may involve not only additions but also deletions of nodes and edges. Furthermore, by combining the addition and deletion of nodes and edges, it is possible to specify a change from triple data composed of deleted nodes and edges to triple data composed of added nodes and edges.

[0113] (Step S905) Returning to Figure 9, the idea generation support device 101 performs the verbalization process. The verbalization process (step S905) is the process of presenting a draft text based on the revised version KG(IN)R of the input text knowledge graph KG(IN).

[0114] [Figure 19 Language Processing (Step S905)] Figure 19 is a flowchart showing a detailed example of the processing procedure for the language processing (step S905) according to Example 1.

[0115] (Step S1901) The idea generation support device 101 creates a language expression instruction sentence that instructs the device to verbalize the proposed revision KG(IN)R of the input text knowledge graph KG(IN). Specifically, the language expression instruction sentence is in a format that the generation AI 103 can receive. For example, if a publicly known large-scale Japanese language model for chatbots is used as the generation AI 103, a standard sentence like the following would be prepared.

[0116] "System: You are an agent assisting User in creating a document. Respond with a Japanese sentence corresponding to the list of characters "subject, object, predicate" entered by User. / User:"

[0117] [Figure 20: Example of extracting a triple data set from the proposed revision KG(IN)R of the input text knowledge graph KG(IN)] Figure 20 is an explanatory diagram showing an example of extracting a triple data set from the proposed revision KG(IN)R of the input sentence knowledge graph KG(IN) according to Example 1. The idea generation support device 101 extracts the triple data set 2000 from the proposed revision KG(IN)R of the input sentence knowledge graph KG(IN). The triple data {worker, fingerprint, authenticate} of the triple data set 2000 corresponds to the nodes and edges added in the proposed revision generation process (step S904).

[0118] Returning to Figure 19, the idea generation support device 101 creates a verbalization instruction sentence by inserting the triple data set 2000 extracted from the revised version KG(IN)R of the input sentence knowledge graph KG(IN) as a prompt to be given to the generating AI 103, following the above-mentioned standard sentence.

[0119] (Step S1902) The idea generation support device 101 sends a verbalization instruction to the generation AI 103. Upon receiving the verbalization instruction, the generation AI 103 generates a draft text that is a revised version of the input text knowledge graph KG(IN)KG(IN)KG(IN)KG(IN)KG(IN)KG(IN)KG(IN)KG(IN)KR. Specifically, for example, the generation AI 103 creates a sentence for each triple data in the triple data set 2000, according to the part of speech of each graph element. For example, in the case of the triple data {camera, worker, photograph}, it generates the sentence "The camera photographs the worker". The generation AI 103 concatenates the sentences it has generated for each triple data and sends them to the idea generation support device 101 as a draft text.

[0120] (Step S1903) The idea generation support device 101 receives the draft text from the generation AI 103.

[0121] (Step S1904) The idea generation support device 101 outputs the draft text received in step S1903 in a displayable format. This completes the input text knowledge graph generation process (step S1101), and the process proceeds to step S903.

[0122] [Figure 21 Example of a display screen] Figure 21 is an explanatory diagram showing an example of a presentation screen according to Embodiment 1. The presentation screen 2100 is displayed on a display which is an example of an output device 204 of the idea generation support device 101 or terminal 104. The presentation screen 2100 has a draft source information display area 2101 and a draft text display area 2102.

[0123] The draft source information display area 2101 is the area that displays the information (draft source information) that forms the basis of the draft text 2120. Draft source information is, for example, the revised draft KG(IN)R of the input text knowledge graph KG(IN) that is the source of the triple data set 2000. The draft text display area 2102 is the area that displays the draft text 2120.

[0124] Thus, in Example 1, users can visually confirm their own idea in the form of an input text knowledge graph KG(IN) and make revisions based on hints from existing literature. In particular, in Example 1, similarity searches based on domain-independent vectors make it possible to select similar structures in different fields. Similarity search results may generate structures different from existing ones, making it more efficient than a person generating idea ideas and their revisions on their own. [Examples]

[0125] Example 2 shows another example of the revision proposal generation process (step S904). Since everything except the revision proposal generation process (step S904) is the same as in Example 1, the explanation is omitted.

[0126] In Example 2, the process up to steps S1501 and S1502, which involve obtaining similar literature knowledge data Lk, is the same as in Example 1. However, in Example 2, the number of selections of similar literature knowledge data Lk is increased compared to Example 1 to suppress cases where no recommendation predicates are found.

[0127] <Figure 22 Revision Proposal Generation Process (Step S904)> Figure 22 is a flowchart showing a detailed example of the processing procedure for the revised proposal generation process (step S904) according to Example 2. Processes identical to those in Figure 15 are given the same step numbers, and their explanations are omitted.

[0128] (Step S2003) After step S1502, the idea generation support device 101 identifies the similar structure between the input sentence knowledge graph KG(IN) and the similar literature knowledge graph KG(Lk)S. The similar literature knowledge graph KG(Lk)S is a literature knowledge graph KG(Lk) composed of similar literature knowledge data Lk obtained from the literature knowledge DB301 in step S1502.

[0129] [Figure 23: Examples of identifying similar structures] Figure 23 is an explanatory diagram illustrating an example of identifying a similar structure in step S2003. In Figure 23, the circles represent nodes of the input sentence knowledge graph KG(IN).

[0130] (A) The idea generation support device 101 selects one node from the input sentence knowledge graph KG(IN). In (A), it is assumed that node Na is selected. The selected node is designated as the source node. The idea generation support device 101 searches for nodes similar to the source node Na in the similar literature knowledge graph KG(Lk)S. Specifically, for example, the idea generation support device 101 searches for a similar node in the similar literature knowledge graph KG(Lk)S where the distance of the domain independence vector from the source node Na is less than or equal to a predetermined threshold, and the node with the smallest distance is selected. In (A), it is assumed that node No is found as a similar node.

[0131] The idea generation support device 101 associates node Na and node No as similar node pairs. The source node Na and node No, which have been associated as node pairs, are designated as explored nodes.

[0132] (B) Once the explored nodes Na and node No are determined, the idea generation support device 101 sets a candidate source from the input sentence knowledge graph KG(IN). The candidate source is the object node connected to the predicate edge if the explored node in the input sentence knowledge graph KG(IN) indicates the subject, and the subject node connected to the predicate edge if the explored node in the input sentence knowledge graph KG(IN) indicates the object.

[0133] In this example, the explored node in the input sentence knowledge graph KG(IN) is node Na, which represents the subject. Therefore, the predicate edges are edges Eab and Eac, whose start ends are connected to node Na. The object nodes are nodes Nb and Nc, whose ends are connected to the predicate edges Eab and Eac. These nodes Nb and Nc are the search source candidates. The idea generation support device 101 selects one new search source node from the search source candidates. In (B), it is assumed that node Nb has been selected as the search source node.

[0134] The idea generation support device 101 sets the search target for the source node Nb from the similar document knowledge graph KG(Lk)S. In this example, the already searched node Na in the input sentence knowledge graph KG(IN) represents the subject. Therefore, the search target is the object node connected to the predicate edge when the already searched node No in the similar document knowledge graph KG(Lk)S that pairs with the source node Nb is taken as the subject. In this example, the predicate edges are edges Eop and Eoq, whose start is connected to node No. The object nodes are nodes Np and Nq, whose end is connected to the predicate edges Eop and Eoq. These nodes Np and Nq are the search targets.

[0135] Furthermore, if a searched node in the input sentence knowledge graph KG(IN) indicates an object, the search target will be the subject node connected to the predicate edge, when the object is the searched node No. in the similar document knowledge graph KG(Lk)S that is paired with the source node Nb.

[0136] The idea generation support device 101 searches for nodes similar to the source node Nb from the search destination. Specifically, for example, the idea generation support device 101 searches for a similar node among the search destination nodes Np and Nq such that the distance of the domain-independent vectors to the source node Nb is less than or equal to a predetermined threshold, and the node with the smallest distance. In (B), it is assumed that node Np was found to be a similar node.

[0137] The idea generation support device 101 associates node Nb and node Np as a similar node pair. The source node Nb and node Np, which have been associated as a node pair, are designated as explored nodes.

[0138] The idea generation support device 101 selects any unselected nodes from the list of search source candidates for already explored nodes in the input sentence knowledge graph KG(IN). In (B), the unselected node Nc is selected as the search source node. Similarly for this search source node Nc, the idea generation support device 101 sets a search destination for the search source node Nc from the similar document knowledge graph KG(Lk)S, excluding already explored nodes, and searches for nodes similar to the search source node Nc from the search destination. In the case of the search source node Nc, the search destination excluding already explored nodes Np is, for example, node Nq. If the search source node Nc and node Nq are similar, the idea generation support device 101 associates them as a node pair. The search source node Nc and node Nq, which are associated as a node pair, are considered explored nodes.

[0139] (C) The process in (B) is repeated until no more similar pairs of nodes appear. The idea generation support device 101 identifies the similar structure between the input sentence knowledge graph KG(IN) and the similar document knowledge graph KG(Lk)S. Specifically, for example, a subgraph 2301 within the input sentence knowledge graph KG(IN) and a subgraph 2302 within the similar document knowledge graph KG(Lk)S are identified as having a similar structure.

[0140] The identified similar structure subgraphs 2301 and 2302 are the subgraphs with the maximum number of nodes and edges among the similar structures in the input text knowledge graph KG(IN) and the similar literature knowledge graph KG(Lk)S. However, the idea generation support device 101 may terminate process (B) when the number of nodes and edges reaches a predetermined number and identify the similar structure. Furthermore, the idea generation support device 101 may exclude from the similar structure any subgraphs whose number of nodes and edges does not reach a predetermined number even after executing process (B) a predetermined number of times.

[0141] (Step S2004) Returning to Figure 22, the idea generation support device 101 creates a revised version KG(IN)R of the input sentence knowledge graph KG(IN) based on the similar structure identified in step S2003. In step S2004, the idea generation support device 101 performs the extraction of difference graphs (Figure 24) and the combination of difference graphs for the input sentence knowledge graph KG(IN) (Figure 25). A difference graph is a subgraph obtained by deleting the similar structure subgraphs 2301 and 2302 from the input sentence knowledge graph KG(IN) and the similar document knowledge graph KG(Lk)S, respectively.

[0142] [Figure 24 Example of extracting a difference graph] Figure 24 is an explanatory diagram showing an example of difference graph extraction in step S2004. Difference graph 2401 is a subgraph obtained by deleting the similar structure subgraph 2301 from the input sentence knowledge graph KG(IN). Difference graphs 2421-2423 are subgraphs obtained by deleting the similar structure subgraph 2302 from the similar document knowledge graph KG(Lk)S.

[0143] [Figure 25: Example of combination of difference graphs for input text knowledge graph KG(IN)] Figure 25 is an explanatory diagram showing an example of a combination of difference graphs for the input sentence knowledge graph KG(IN) in step S2004. The idea generation support device 101 treats difference graph 2401 as an element to be deleted from the input sentence knowledge graph KG(IN), and difference graphs 2421 to 2423 as elements to be added to the input sentence knowledge graph KG(IN).

[0144] (A) shows the proposed modification KG(IN)R1 of the input sentence knowledge graph KG(IN) by removing the difference graph 2401 from the input sentence knowledge graph KG(IN).

[0145] (B) shows the proposed modification KG(IN)R2 of the input sentence knowledge graph KG(IN), with the difference graph 2422 added to the input sentence knowledge graph KG(IN).

[0146] (C) shows the proposed modification KG(IN)R3 of the input sentence knowledge graph KG(IN), in which difference graph 2401 is removed from the input sentence knowledge graph KG(IN) and difference graphs 2421 and 2423 are added.

[0147] Figure 25 shows three possible revisions KG(IN)R1 to KG(IN)R3 as examples, but a revised KG(IN)R is generated that covers all cases, including whether or not to delete deleted elements and whether or not to add added elements.

[0148] (Step S2005) Returning to Figure 22, the idea generation support device 101 displays the proposed revision KG(IN)R of the input sentence knowledge graph KG(IN) generated in step S2004.

[0149] [Figures 26 and 27: Examples of screens presenting revised proposals] Figure 26 is an explanatory diagram showing Example 1 of the suggested revision screen according to Embodiment 2. The suggested revision screen 2600 is displayed on a display which is an example of the output device 204 of the idea generation support device 101 or terminal 104. In Figure 26, the rectangle represents the node of suggested revision KG(IN)R of the input sentence knowledge graph KG(IN). The suggested revision screen 2600 is displayed in step S2005. The suggested revision screen 2600 includes an input sentence display unit 1601, a suggested revision transition button 2601, a suggested revision display unit 2602, and a selection button 2603.

[0150] The proposed revision transition button 2601 is a user interface that transitions from proposed revision KG(IN)Rx of the input text knowledge graph KG(IN) displayed in the proposed revision display unit 2602 to another proposed revision KG(IN)Ry. The proposed revision display unit 2602 is the area that displays proposed revision KG(IN)R of the input text knowledge graph KG(IN). In proposed revision KG(IN)Rx of the input text knowledge graph KG(IN), nodes and edges indicating additional elements are displayed with thick lines to distinguish them from the nodes and edges of the input text knowledge graph KG(IN). The selection button 2603 is a user interface for selecting a proposed revision KG(IN)Rx of the input text knowledge graph KG(IN) currently displayed in the proposed revision display unit 2602.

[0151] Figure 27 is an explanatory diagram showing the revised proposal presentation screen 2600 after the revised proposal transition according to Embodiment 2. In Figure 27, pressing the revised proposal transition button 2601 shows that the display content of the revised proposal display unit 2602 has transitioned from revised proposal KG(IN)Rx of the input text knowledge graph KG(IN) to another revised proposal KG(IN)Ry. In revised proposal KG(IN)Ry, the node indicating "server," which is a deleted element from revised proposal KG(IN)R, and the edge indicating "grant permission for portrait rights" are displayed with dotted lines, making it possible to visually confirm that they are deleted elements.

[0152] Thus, in Example 2, a revised version KG(IN)R can be presented, in which added and deleted elements are combined in the input sentence knowledge graph KG(IN) of input sentence 1010.

[0153] As explained above, the idea generation support device 101 described above can improve the novelty and inventiveness of ideas.

[0154] For example, by displaying the subject and object of the graph elements of the similar literature knowledge data Lk in the word display unit 1603, it is possible to provide users with noun phrases that enhance the novelty and inventiveness of their ideas.

[0155] Furthermore, the conversion unit 803 performs a first conversion that converts the vector representation of the input sentence 1010 into a domain representation vector having a component indicating its dependency on the domain to which the input sentence 1010 belongs, and a second conversion that converts the domain representation vector into a domain-independent vector of the input sentence 1010 by removing the domain representation vector from the vector representation of the input sentence 1010. This enables the automatic generation of a domain-independent vector of the input sentence 1010.

[0156] Furthermore, by displaying input sentence 1010, users can compare input sentence 1010 with noun phrases that enhance the novelty and inventiveness of the idea.

[0157] Furthermore, by generating an input sentence knowledge graph KG(IN) in the word display unit 1603 and displaying the subject and object in the graph elements of the similar document knowledge data Lk, the system can provide users with noun phrases to use in modifying the input sentence knowledge graph KG(IN).

[0158] Furthermore, by displaying predicates in the predicate display unit 1604 that use the word pairs selected in the graph display unit 1602 and the word display unit 1603 as the subject and object, it is possible to provide users with predicates that connect subjects and objects, thereby increasing the novelty and inventiveness of ideas.

[0159] Furthermore, by determining recommended predicates based on the frequency of occurrence of predicates, it is possible to provide users with predicates that are suitable for subjects and objects, thereby enhancing the novelty and inventiveness of their ideas.

[0160] Furthermore, by presenting a revised version of the input text knowledge graph KG(IN) KG(IN) R, it is possible to present users with revisions to their ideas that are easy to understand intuitively.

[0161] Furthermore, by accepting selections from the word display unit 1603 and the predicate display unit 1604, a revised version KG(IN)R of the input sentence knowledge graph KG(IN) can be automatically generated, thereby improving usability.

[0162] Furthermore, by verbalizing the proposed revisions KG(IN)R to the input text knowledge graph KG(IN), it is possible to facilitate understanding of the content of the proposed revisions KG(IN)R to the input text knowledge graph KG(IN).

[0163] Furthermore, by identifying the similar structures between the input sentence knowledge graph KG(IN) and the bibliographic knowledge graph KG(Lk) of known similar documents Lk, and generating a revised version KG(IN)R of the input sentence knowledge graph KG(IN), it is possible to clearly present to the user the common and different structures between the input sentence knowledge graph KG(IN) and the bibliographic knowledge graph KG(Lk) of known similar documents Lk.

[0164] By determining the similarity between nodes based on the distance between the domain-independent vector of a node in the input text knowledge graph KG(IN) and the domain-independent vector of a node in the known similar document knowledge graph KG(Lk) of Lk, it is possible to identify cross-domain similarity structures.

[0165] By removing the difference graph 2401, which has no similarity relationship with the known similar document Lk's document knowledge graph KG(Lk), from the input document knowledge graph KG(IN), it is possible to eliminate elements unnecessary for improving novelty and inventiveness, thereby supporting the generation of ideas that enhance novelty and inventiveness.

[0166] By adding a difference graph 2402, which has no similarity relationship with the known similar document Lk's document knowledge graph KG(Lk), to the input document knowledge graph KG(IN) and modifying the input document knowledge graph KG(IN), it is possible to support the generation of ideas that enhance novelty and inventiveness.

[0167] It should be noted that the present invention is not limited to the embodiments described above, but includes various modifications and equivalent configurations within the spirit of the attached claims. For example, the embodiments described above are described in detail to make the present invention easier to understand, and the present invention is not necessarily limited to having all of the described configurations. Furthermore, some of the configurations of one embodiment may be replaced with those of another embodiment. Furthermore, some of the configurations of one embodiment may be added to those of another embodiment. Furthermore, some of the configurations of each embodiment may be added, deleted, or replaced with other configurations.

[0168] Furthermore, each of the aforementioned configurations, functions, processing units, and processing means may be implemented in hardware, for example, by designing them as integrated circuits, or they may be implemented in software by having a processor interpret and execute programs that realize each function.

[0169] Information such as programs, tables, and files that implement each function can be stored in memory, hard disks, SSDs (Solid State Drives), or on recording media such as IC (Integrated Circuit) cards, SD cards, and DVDs (Digital Versatile Discs).

[0170] Furthermore, the control lines and information lines shown are those deemed necessary for explanation purposes and do not necessarily represent all control lines and information lines required for implementation. In reality, it can be assumed that almost all components are interconnected. [Explanation of Symbols]

[0171] 101 Idea Generation Support Device 201 Processor 202 Storage Devices 302. Set of Noun Phrase Vectors from Literature 402 Literature Knowledge Graph 403 List of graph element vectors in the literature 404 Bibliographic Knowledge Text Vector 501(Lk) Literature noun phrase set 502-reference triple data set 601 Literature noun phrase 602 Noun phrase vectors from literature 801 Input section 802 Graph Generation Unit 803 Conversion Unit 804 Amendment generator 805 Language Processing Section 806 Output Section

Claims

1. An idea generation support device having a processor for executing a program and a storage device for storing the program, It is possible to access a database that stores a set of known sentences, a domain-independent vector for each known sentence having a component indicating its independence from the domain to which the known sentence belongs, and a set of subject, object, and predicate combinations that constitute the known sentence. The aforementioned processor, An input process that accepts input text, A transformation process that converts the vector representation of the input sentence into a field-independent vector having a component indicating independence from the field to which the input sentence belongs, based on a field correlation parameter that shows the correlation between a known sentence and multiple fields, Based on the domain-independent vector of the input sentence and the domain-independent vector of the known sentence, an extraction process is performed to identify known similar sentences similar to the input sentence from the set of known sentences, and to extract the subject and object constituting the known similar sentence from the database. A first output process that outputs the subject and object constituting the known similar sentence extracted by the extraction process in a displayable format, An idea generation support device characterized by performing the following actions.

2. An idea support device according to claim 1, In the conversion process, the processor performs a first conversion, which converts the vector representation of the input statement into a domain representation vector having a component indicating the dependency on the domain to which the input statement belongs, based on the domain correlation parameter, and a second conversion, which converts the domain representation vector into a domain-independent vector of the input statement by removing the domain representation vector from the vector representation of the input statement. An idea generation support device characterized by the following features.

3. An idea support device according to claim 1, In the first output processing, the processor outputs the input statement in a displayable format. An idea generation support device characterized by the following features.

4. An idea support device according to claim 1, The aforementioned processor, A first generation process is executed to generate a knowledge graph of the input sentence, where each of the subject and object constituting the input sentence is a node, and the predicates constituting the input sentence are edges indicating the direction from the node representing the subject to the node representing the object. In the first output process, the processor outputs the knowledge graph of the input sentence generated by the first generation process in a displayable format. An idea generation support device characterized by the following features.

5. An idea support device according to claim 4, The database stores domain-independent vectors for each of the subject, object, and predicate that constitute the known sentence. In the conversion process, the processor converts the vector representation of the first selected word phrase selected from the knowledge graph of the input sentence into a field-independent vector of the first selected word phrase, based on the field correlation parameter. The aforementioned processor, A decision process that searches for a first synonym similar to the first selected word from the known sentence based on the domain independence vector of the first selected word and the domain independence vector of the known sentence, searches for a second synonym similar to the second selected word from the known sentence based on the domain independence vector of the second selected word selected from the known similar sentence and the domain independence vector of the known sentence, obtains a specific combination from the set of combinations that corresponds when the first synonym is a noun phrase that is either the subject or the object and the second synonym is a noun phrase that is the other, and determines the predicate of the specific combination as the recommended predicate. A second output process that outputs the recommendation predicate determined by the aforementioned decision process in a displayable format, An idea generation support device characterized by performing the following actions.

6. An idea support device according to claim 5, In the decision process, the processor determines the recommended predicate based on the number of occurrences of the specific combination of predicates. An idea generation support device characterized by the following features.

7. An idea support device according to claim 5, The aforementioned processor, A second generation process generates a proposed revision of the knowledge graph of the input sentence by setting the node representing the first selected phrase in the knowledge graph of the input sentence to a first node representing either of the noun phrases, generating a second node representing the second selected phrase, and generating an edge representing the recommendation predicate and connecting it to the first and second nodes. A third output process that outputs a suggested revision of the knowledge graph of the input sentence generated by the second generation process in a displayable format, An idea generation support device characterized by performing the following actions.

8. An idea support device according to claim 7, In the second generation process, the processor accepts the selection of a specific recommendation predicate from among the recommendation predicates, generates an edge indicating the specific recommendation predicate, and connects it to the first node and the second node, thereby generating a proposed modification of the knowledge graph of the input sentence. An idea generation support device characterized by the following features.

9. An idea support device according to claim 7, The aforementioned processor, A language processing process that verbalizes proposed revisions to the knowledge graph of the aforementioned input sentence, A fourth output process that outputs the sentence verbalized by the verbalization process in a displayable format, An idea generation support device characterized by performing the following actions.

10. An idea support device according to claim 5, The database stores a knowledge graph of each known sentence, where each of the subject and object constituting the known sentence is a node, and the predicate constituting the known sentence is an edge indicating the direction from the node representing the subject to the node representing the object. The aforementioned processor, The process involves extracting the knowledge graph of known similar sentences from the database, and identifying the similar structure between the knowledge graph of the input sentence and the knowledge graph of the known similar sentences. A third generation process generates a proposed revision of the knowledge graph of the input sentence based on a first subgraph in the knowledge graph of the input sentence identified as having a similar structure by the specified process, and a second subgraph in the knowledge graph of the known similar sentence identified as having a similar structure by the specified process. A fifth output process that outputs a suggested revision of the knowledge graph of the input sentence generated by the third generation process in a displayable format, An idea generation support device characterized by performing the following actions.

11. An idea support device according to claim 10, In the aforementioned specific processing, the processor identifies the similarity structure based on the domain-independent vectors of noun phrases indicated by the nodes constituting the knowledge graph of the input sentence and the domain-independent vectors of noun phrases indicated by the nodes constituting the knowledge graph of the known similar sentence. An idea generation support device characterized by the following features.

12. An idea support device according to claim 10, In the third generation process, the processor generates a proposed revision of the knowledge graph of the input sentence by deleting the first residual graph, which is obtained by removing the first subgraph from the knowledge graph of the input sentence, from the knowledge graph of the input sentence. An idea generation support device characterized by the following features.

13. An idea support device according to claim 10, In the third generation process, the processor generates a proposed revision of the knowledge graph of the input sentence by adding a second residual graph, obtained by removing the second subgraph from the knowledge graph of known similar sentences, to the knowledge graph of the input sentence. An idea generation support device characterized by the following features.

14. An idea generation support method performed by an idea generation support device having a processor for executing a program and a storage device for storing the program, The idea generation support device has access to a database that stores a set of known sentences, a domain-independent vector for each known sentence having a component indicating its independence from the domain to which the known sentence belongs, and a set of subject, object, and predicate combinations that constitute the known sentences. The aforementioned processor, An input process that accepts input text, A transformation process that converts the vector representation of the input sentence into a field-independent vector having a component indicating independence from the field to which the input sentence belongs, based on a field correlation parameter that shows the correlation between a known sentence and multiple fields, Based on the domain-independent vector of the input sentence and the domain-independent vector of the known sentence, an extraction process is performed to identify known similar sentences similar to the input sentence from the set of known sentences, and to extract the subject and object constituting the known similar sentence from the database. A first output process that outputs the subject and object constituting the known similar sentence extracted by the extraction process in a displayable format, An idea generation support method characterized by performing the following.

15. An idea generation support program that causes a processor to perform idea generation support, The processor has access to a database that stores a set of known sentences, a domain-independent vector for each known sentence having a component indicating its independence from the domain to which the known sentence belongs, and a set of subject, object, and predicate combinations that constitute the known sentences. The aforementioned processor, An input process that accepts input text, A transformation process that converts the vector representation of the input sentence into a field-independent vector having a component indicating independence from the field to which the input sentence belongs, based on a field correlation parameter that shows the correlation between a known sentence and multiple fields, Based on the domain-independent vector of the input sentence and the domain-independent vector of the known sentence, an extraction process is performed to identify known similar sentences similar to the input sentence from the set of known sentences, and to extract the subject and object constituting the known similar sentence from the database. A first output process that outputs the subject and object constituting the known similar sentence extracted by the extraction process in a displayable format, An idea generation support program characterized by its ability to execute certain actions.

Images