Information processing device, information processing method, program, and recording medium

Abstract

Accurately extract data for building knowledge graphs. [Solution] The information processing device (100A, 100B) includes a generation unit (11) that generates a prompt for extracting data of candidate elements to be included in the knowledge graph from target document data using a language model, using data relating to the nodes or edges of the candidate elements, and an acquisition unit (12) that acquires output information output by the language model that has received the prompt.

Description

Technical Field

[0004] ,

[0006] , , , , , ,

[0005] , , , , ,

[0001] The present invention relates to a technique for extracting data for constructing a knowledge graph.

Background Art

[0002] Techniques for utilizing data constructed in the form of a knowledge graph including nodes and edges are known. For example, Patent Document 1 discloses an information processing method in which a query engine takes in a query using a machine learning model and evaluates the query via a knowledge graph.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In techniques that utilize a knowledge graph, it is necessary to construct a highly accurate knowledge graph in advance, and for this purpose, it is necessary to accurately extract data for constructing the knowledge graph from the original information. However, the above patent document had problems in this regard.

[0005] One aspect of the present disclosure has been made in view of the above problems, and one object thereof is to provide a technique capable of accurately extracting data for constructing a knowledge graph.

Means for Solving the Problems

[0006] To solve the above problems, an information processing device according to one aspect of the present invention includes a generation unit that generates a prompt for extracting data of candidate elements to be included in a knowledge graph from target document data using a language model, using data relating to the nodes or edges of the candidate elements, and an acquisition unit that acquires output information output by the language model that has received the prompt.

[0007] To solve the above problems, an information processing method according to one aspect of the present invention is an information processing method executed by an information processing device, comprising: a generation step of generating a prompt for extracting data of candidate elements to be included in a knowledge graph from target document data using a language model, using data relating to the nodes or edges of the candidate elements; and an acquisition step of acquiring output information output by a language model that has received the prompt.

[0008] Each aspect of the information processing device relating to this disclosure may be implemented by a computer, in which case a program that enables the computer to implement the information processing device by operating the computer as each part (software element) of the information processing device, and a computer-readable recording medium on which the program is recorded, also fall within the scope of this disclosure. [Effects of the Invention]

[0009] According to one aspect of this disclosure, data for constructing a knowledge graph can be extracted with high accuracy. [Brief explanation of the drawing]

[0010] [Figure 1] This is a block diagram showing the configuration of an information processing system according to Embodiment 1 of this disclosure. [Figure 2] This is a flowchart showing the processing flow by the information processing device according to Embodiment 1 of this disclosure. [Figure 3] This is a diagram illustrating the processing performed by the information processing device according to Embodiment 1 of this disclosure. [Figure 4]This is a diagram illustrating the processing performed by the information processing device according to Embodiment 1 of this disclosure. [Figure 5] This figure shows an example of display information displayed by the information processing device according to Embodiment 1 of this disclosure. [Figure 6] This is a diagram illustrating the processing performed by the information processing device according to Embodiment 1 of this disclosure. [Figure 7] This is a block diagram showing the configuration of an information processing system according to Embodiment 2 of this disclosure. [Figure 8] This is a flowchart showing the processing flow by the information processing device according to Embodiment 2 of this disclosure. [Figure 9] This is a diagram illustrating the processing performed by the information processing device according to Embodiment 2 of this disclosure. [Figure 10] This figure schematically shows the data formats that are processed by the information processing systems according to each embodiment of this disclosure. [Figure 11] This figure shows the hardware configuration of the information processing system according to each embodiment of the present disclosure. [Modes for carrying out the invention]

[0011] [Embodiment 1] <Overview of Information Processing System 1A> As will be described in detail later, an information processing system 1A according to one embodiment of this disclosure includes an information processing device 100A and a server 200. Here, the server 200 is equipped with a language model LM that generates output information OUT from a prompt PMT, which is input information provided by the information processing device 100.

[0012] On the other hand, the information processing device 100A can be broadly described as follows: • Using a language model (LM), prompts are generated to extract data for candidate elements to be included in the knowledge graph from the target document data, using data related to the nodes or edges of those candidate elements. • Obtain the output information OUT that the language model LM outputs when the generated prompt is input. It has the following configuration.

[0013] As described above, the information processing apparatus 100A generates a prompt for extracting data of candidate elements to be included in the knowledge graph from the target document data using the language model LM, using data related to the nodes or edges of the candidate elements, and acquires output information OUT output by the language model LM to which the generated prompt is input. Therefore, data for constructing a knowledge graph can be accurately extracted from the target document data.

[0014] <Configuration of Information Processing System 1A> Hereinafter, an information processing system 1A according to an embodiment of the present disclosure will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the information processing system 1A. As shown in FIG. 1, the information processing system 1A includes an information processing apparatus 100A and a server 200.

[0015] (Configuration of Server 200) First, the configuration of the server 200 included in the information processing system 1A will be described. As shown in FIG. 1, the server 200 includes a control unit 240, a storage unit 220, and a communication unit 230.

[0016] (Communication Unit 230) The communication unit 230 communicates with devices external to the server 200. As an example, the communication unit 230 communicates with the information processing apparatus 100A included in the information processing system 1A. The communication unit 230 transmits data supplied from the control unit 240 to the information processing apparatus 100A, or supplies data received from the information processing apparatus 100A to the control unit 240. Note that the data received by the communication unit 230 from the information processing apparatus 100A includes a prompt PMT that is input information generated by the information processing apparatus 100A. Further, the data provided by the communication unit 230 to the information processing apparatus 100 may include output information OUT generated by the language model LM referring to the prompt PMT.

[0017] (Storage Unit 220) The memory unit 220 stores the language model LM. More specifically, the memory unit 220 stores several parameters that define the language model LM. These parameters are, for example, parameters that have been pre-learned (updated) by machine learning, but this is not limited to this embodiment. Furthermore, the language model LM is a model that has been machine-learned using documents from various fields as training data, but the details of the language model LM are not limited to this embodiment.

[0018] (Control unit 240) The control unit 240 obtains the output result of the language model LM by executing the language model LM. For example, the control unit 240 inputs the prompt PMT received from the information processing device 100A to the language model LM and obtains the output information OUT generated by the language model LM. It also provides the output information OUT to the information processing device 100A via the communication unit 230.

[0019] In this embodiment, the various components of the server 200 are illustrated as separate components from the information processing device 100A, but this does not limit this embodiment. For example, the control unit 240 of the server 200, or the function of the language model execution unit in the control unit 240, may be configured to be provided by the control unit of the information processing device 100A. Similarly, the language model LM stored in the storage unit 220 of the server 200 may be stored in the storage unit of the information processing device 100A, and the information processing device 100A itself may be able to execute the language model LM.

[0020] (Configuration of information processing device 100A) Next, the configuration of the information processing device 100A included in the information processing system 1A will be described. As shown in Figure 1, the information processing device 100A includes a control unit 110, a storage unit 120, a communication unit 130, and an input / output unit 140.

[0021] (Communications Section 130) The communication unit 130 communicates with devices outside the information processing device 100A. For example, the communication unit 130 communicates with the server 200. The communication unit 130 transmits data supplied from the control unit 110 to the server 200 and supplies data received from the server 200 to the control unit 110. The data transmitted by the communication unit 130 to the server 200 may include a prompt PMT generated by the control unit 110. The data received by the communication unit 130 from the server 200 may include output information OUT generated by the language model LM of the server 200 by referring to the prompt PMT.

[0022] (I / O section 140) The input / output unit 140 is configured to include at least one of the following input / output devices: a keyboard, mouse, display panel, touchpad, or touchscreen. Alternatively, the input / output unit 140 may be configured as an input / output interface to which input / output devices such as a keyboard, mouse, display panel, touchpad, or touchscreen are connected. In this configuration, the input / output unit 140 receives various types of information from the connected input device to the information processing device 100A. The input / output unit 140 also outputs various types of information to the connected output device.

[0023] (Storage unit 120) The memory unit 120 stores various data referenced by the control unit 110, as well as various data generated by the control unit 110. For example, the memory unit 120 stores: Knowledge Graph KG Document data group DG Candidate element CE • Prompt PMT • Output information OUT • Registration element RE The following are stored. Here, the knowledge graph KG (also called the knowledge graph) is the knowledge graph that is constructed by the information processing device 100A, and as an example, it includes multiple nodes and one or more edges (also called links) that connect the nodes.

[0024] The document data group DG is, for example, composed of multiple documents acquired in advance. Here, the type, language, source, size, etc. of the documents are not limited to this embodiment, and documents of any type, language, source, size, etc. can be used. Examples of such document group DG include publicly available free papers, paid papers, patent specifications, technical reports, and internal company documents.

[0025] Candidate element CE is an element that is a candidate for registration in the knowledge graph KG, and for example, it includes at least one of the following: data relating to one or more nodes, and data relating to one or more edges. Specific examples of candidate element CE will be described later. In this embodiment, "element" may also be expressed as "constituent element." Furthermore, the presentation of "elements" is not limited to the examples above, and in this embodiment, one or more nodes or one or more edges may also be referred to as "elements."

[0026] The prompt PMT is information generated by the generation unit 11 (described later) and is input information provided to the language model LM. Specific examples of the prompt PMT will be described later. The output information OUT is information output by the language model LM that has received the prompt PMT as input. Specific examples of the output information OUT will be described later.

[0027] The registration element RE is information generated by referring to the output information OUT output by the language model LM, and is registered in the knowledge graph KG by the registration unit 13, which will be described later. Specific examples of the registration element RE will be described later.

[0028] (Control Unit 110) As shown in Figure 1, the control unit 110 includes a first acquisition unit 111, a prompt generation unit 112, a second acquisition unit 12, and a registration unit 13. Here, the first acquisition unit 111 and the prompt generation unit 112 constitute the generation unit 11.

[0029] (Generation unit 11) The generation unit 11 generates a prompt PMT for extracting data of candidate elements CE to be included in the knowledge graph KG from the target document data TD using the language model LM, using data related to the nodes or edges of the candidate elements CE.

[0030] More specifically, the first acquisition unit 111 of the generation unit 11 acquires the target document data TD from the document data group DG. A specific example of how the first acquisition unit 111 acquires the target document data TD will be described later.

[0031] Then, the prompt generation unit 112 provided in the generation unit 11, for example, • Instruction IN containing data relating to the node or edge of the candidate element CE, • At least a portion of the aforementioned target document data TD and The prompt PMT is generated, which includes the above.

[0032] Here, the data relating to the node or edge of the candidate element CE may include at least one of the node label NL, edge label EL, and property name PN of the candidate element CE. Furthermore, the data relating to the node or edge of the candidate element CE may also include descriptive information relating to at least one of the node label NL, edge label EL, and property name PN of the candidate element CE. (Second acquisition section 12) The second acquisition unit 12 acquires the output information OUT that is output by the language model LM when the prompt PMT is input. The second acquisition unit 12 supplies the acquired output information OUT to the registration unit 13.

[0033] (Registration Section 13) The registration unit 13 generates registration elements RE for registration in the knowledge graph KG by referring to the output information OUT, and registers the generated registration elements RE in the knowledge graph KG. As an example, as shown in Figure 1, the registration unit 13 includes a processing unit 131, an element registration unit 132, and a presentation unit 133.

[0034] The processing unit 131 generates a registration element RE by processing the output information OUT. For example, the processing unit 131 extracts information about at least one of nodes and edges from the output information OUT, and generates a registration element RE using the extracted information. Then, the element registration unit 132 registers the registration element RE generated (extracted) by the processing unit 131 in the knowledge graph KG.

[0035] For example, the presentation unit 133 presents the output information OUT to the user via a display provided by the input / output unit 140. The input / output unit 140 then receives instructions from the user. The processing unit 131 may be configured to generate the registration element RE by referring to the user's instructions regarding the output information OUT.

[0036] According to the information processing device 100A configured as described above, • A prompt PMT is generated using data related to the nodes or edges of the candidate element CE to be included in the KnowledgeGraph KG, using the language model LM to extract the data of the candidate element CE from the target document data TD. • Obtain the output information OUT that the language model LM outputs when the generated prompt PMT is input. This configuration is employed. According to this configuration, data for constructing a knowledge graph can be accurately extracted from the target document data TD. In other words, the components to be included in the knowledge graph KG can be automatically and accurately extracted.

[0037] (Processing flow by information processing device 100A) Next, we will explain an example of the processing flow by the information processing device 100A. Figure 2 is a flowchart showing the processing flow (information processing method S100A) by the information processing device 100A.

[0038] (Step S101) First, in step S101, the first acquisition unit 111 acquires the target document data TD from the document data group DG that is to be extracted.

[0039] (Step S102) Next, in step S102, the first acquisition unit 111 determines whether or not there are any unprocessed candidate elements CE among the one or more candidate elements CE related to the target document data TD. If there are unprocessed candidate elements CE (YES in step S102), the process proceeds to step S103; otherwise (NO in step S102), the process proceeds to step S110.

[0040] If there are multiple types of candidate elements CE, in this step, the first acquisition unit 111 may be configured to determine whether or not there are any unprocessed types of candidate elements CE among the one or more candidate elements CE related to the target document data TD. Here, the type of candidate element CE refers, for example, to the type of node or edge of the candidate element CE. More specifically, it could refer to whether the candidate element CE relates to organic materials, or whether the document or node relates to inorganic materials, but this is not limited to this example. By performing such processing in this step, the information processing device 100A can suitably perform processing for each type of candidate element CE.

[0041] (Step S103) In step S103, the first acquisition unit 111 selects one unprocessed candidate element CE from the one or more candidate elements CE. It also acquires data related to the selected candidate element CE.

[0042] If there are multiple types of candidate elements CE, in this step, the first acquisition unit 111 selects one type of unprocessed candidate element CE from among the one or more candidate elements CE.

[0043] (Step S104) Then, in step S104, the prompt generation unit 112 generates a prompt PMT by referring to the target document data TD acquired in step S101 and the data relating to the candidate element CE selected in step S103.

[0044] Figure 3 shows a specific processing example A1 related to steps S101 to S104 described above. In the example shown in Figure 3, in step S101, the target document data TD is obtained, which reads, "A company's flexible display is realized by a display substrate with excellent flexibility, bending strength, and transparency."

[0045] Furthermore, in the example shown in Figure 3, the following data related to node 1 is obtained as data related to the candidate element CE selected in step S103. ·Node label NL1 of node 1 (in the example in Figure 3, "Function"), and • Node 1 property PR1.

[0046] Here, the property PR1 is: • The first property name for node 1 is PN11 (in the example in Figure 3, it is "Target"). • The second property name for Node 1 is PN12 (in the example in Figure 3, it is "Features"). • The third property name for Node 1 is PN13 (in the example in Figure 3, it is "Developer"). It includes [this]. On the other hand, the values ​​of the properties corresponding to the above property names are all blank (no data).

[0047] Similarly, the following data regarding node 2 is obtained as data related to the candidate element CE selected in step S103. • Node label NL2 of node 2 (in the example in Figure 3, "Function"), and • Node 2 property PR2.

[0048] Here, the property PR1 is: • The first property name for node 2 is PN21 (in the example in Figure 3, it is "Target"). • The second property name for node 2 is PN22 (in the example in Figure 3, it is "Features"). • The third property name for Node 2 is PN23 (in the example in Figure 3, it is "Developer"). It includes [this]. On the other hand, the values ​​of the properties corresponding to the above property names are all blank (no data).

[0049] Furthermore, as data related to the candidate element CE selected in step S103, the edge label EL of the edge connecting node 1 and node 2 is obtained as "realize".

[0050] In the example shown in Figure 3, in step S104, the prompt generation unit 112 generates a prompt PMT using the node label NL, edge label EL, and property name PN of the candidate element CE as data relating to the node or edge of the candidate element CE obtained in step S103. More specifically, the prompt generation unit 112, An instruction statement IN containing data relating to the node or edge of the candidate element CE, The aforementioned target document data TD and The prompt PMT is generated which includes the above.

[0051] Here, the instruction IN is an instruction to extract the data of the candidate element CE from the target document data TD using the language model LM, as shown in Figure 3, • Node label NL2 of node 2 (in the example in Figure 3, it is "Function") • Node label NL1 of node 1 (in the example in Figure 3, it is "Function") • Property names PN11 and PN21 (in the example in Figure 3, "Target") • Property names PN12 and PN22 (in the example in Figure 3, "Features") • Property names PN13 and PN23 (in the example in Figure 3, "Developer") It includes.

[0052] Thus, in this example, the generation unit 112 is configured to generate a prompt PMT using at least one of the node label NL, edge label EL, and property name PN of the candidate element CE as data relating to the node or edge of the candidate element CE. With this configuration, for various candidate elements (components) CE, a prompt PMT can be suitably generated for extracting the data of the candidate element CE to be included in the knowledge graph KG from the target document data using the language model LM.

[0053] (Step S105) Returning to Figure 2, in step S105, the second acquisition unit 12 supplies the prompt PMT generated as described above in step S104 to the language model LM of the server 200 via the communication unit 130, and acquires the output information OUT of the language model LM that received the prompt PMT.

[0054] (Step S106) Then, in step S106, the registration unit 13 generates a registration element RE for registration in the knowledge graph KG from the output information OUT acquired in step S105.

[0055] (Step S110) In this way, processing is performed on all of the one or more candidate elements CE related to the target document data TD, and in step S110, the registration unit 13 registers all the generated registration elements RE in the knowledge graph KG.

[0056] Figure 4 shows a specific processing example A2 relating to steps S105 to S110 described above. In the example shown in Figure 4, the output information OUT obtained from the language model LM in step S105 includes output information OUT1 and output information OUT2, and output information OUT1 is • Target: Display Features: Flexible • Developer: Company A Therefore, the output information OUT2 is, • Target: Display board Features: Flexibility / Bending strength / Transparency Here, output information OUT1 is the data corresponding to node 1 in candidate element CE, and output information OUT2 is the data corresponding to node 2 in candidate element CE.

[0057] In the example shown in Figure 4, the registration information RE generated by the registration unit 13 from the output information OUT in step S106 includes: In addition to the data on the candidate element CE mentioned above (node ​​label NL, edge label EL, property name PN), • Values ​​for each property PR1 of Node 1 (Display, Flexible, Company A) • Values ​​for each property PR2 of node 2 (display substrate, flexibility / bending strength / transparency) This includes the following. Here, each value of property PR1 of node 1 and each value of property PR2 of node 2 are information extracted by the processing unit 131 of the registration unit 13 from the output information OUT1 and output information OUT2 described above, respectively. As shown in Figure 4, the element registration unit 132 of the registration unit 13 registers the registration information RE generated in this way into the knowledge graph KG.

[0058] Figure 5 shows an example in which the presentation unit 133 of the registration unit 13 presents output information OUT to the user via the display of the input / output unit 140 and receives instructions from the user. In the example shown in Figure 5, the presentation unit 133, along with the output information OUT described above, Whether or not to generate the registration element RE based on the output information OUT. • Correction details if "No" is selected A user interface is displayed that can receive user instructions regarding the above. The processing unit 131 may be configured to refer to the user instructions received by the presentation unit 133 via the above user interface and generate a registration element RE from the output information OUT. By using such a configuration, a more suitable knowledge graph KG can be constructed while reflecting the user's instructions.

[0059] Note that the processing performed by the information processing device 100A is not limited to the example above. Figure 6 shows a specific processing example A3 relating to steps S101 to S104 described above. In the example shown in Figure 6, similar to processing example A1 shown in Figure 3, in step S101, the document "A company's flexible display is realized by a display substrate with excellent flexibility, bending strength, and transparency." is acquired as the target document data TD.

[0060] On the other hand, in processing example A3 shown in Figure 6, in addition to the data explained in processing example A1 shown in Figure 3, descriptive information EI relating to at least one of the node label NL, edge label EL, and property name NL of the candidate element CE is obtained as data related to the candidate element CE selected in step S103. More specifically, as shown in Figure 6, • Descriptive information EI1 regarding property names PN11 and PN21 (shown in Figure 6 as "Target: Target implementing the function") • Description information for property names PN12 and PN22 EI2 (shown in Figure 6 as "Features: Functional Characteristics") • Description information for property names PN13 and PN23 EI3 (shown in Figure 6 as "Developer: Name of the company that developed the function") It includes.

[0061] In the example shown in Figure 6, the prompt PMT generated by the prompt generation unit 112 in step S104 includes the information described in processing example A1 shown in Figure 3, as well as the descriptive information EI1, EI2, and EI3 mentioned above. In other words, in this example, the prompt generation unit 112 is configured to generate the prompt PMT using descriptive information EI relating to at least one of the node label NL, edge label EL, and property name PN of the candidate element CE as data relating to the node or edge of the candidate element CE.

[0062] By including explanatory information in the prompt PMT in this way, the language model LM can be used to extract data for building the knowledge graph with greater accuracy.

[0063] [Embodiment 2] Other embodiments of this disclosure are described below. For convenience of explanation, components having the same function as those described in the above embodiments may be denoted by the same reference numerals, and their descriptions may be omitted.

[0064] <Configuration of Information Processing System 1B> Figure 7 is a block diagram showing the configuration of the information processing system 1B according to this embodiment. As shown in Figure 7, the information processing system 1B comprises an information processing device 100B and a server 200. The server 200 has the same configuration as in Embodiment 1, so a redundant explanation will be omitted.

[0065] (Configuration of information processing device 100B) As shown in Figure 7, the information processing device 100B according to this embodiment includes a search unit 113 in the control unit 110, in addition to the configurations of the information processing device 100A according to Embodiment 1. Regarding the configurations of the information processing device 100B, redundant explanations of configurations similar to those of the information processing device 100A according to Embodiment 1 will be omitted, and the following explanation will focus on the differences from the information processing device 100A.

[0066] (Search section 113) The search unit 113 searches for multiple elements already included in the knowledge graph KG and retrieves data related to the searched elements. For example, the search unit 113 searches for multiple elements already included in the knowledge graph KG and retrieves data related to the searched elements. • Search for multiple elements already included in the Knowledge Graph KG. Identify one or more elements that have the same node label, edge label, or property name as at least one of the node label NL, edge label EL, and property name PN of candidate element CE, and that are already included in the KnowledgeGraph KG. • Obtain actual data on the one or more of the identified elements mentioned above. This process is performed.

[0067] Here, the actual data mentioned above may include, as an example, the document data from which the one or more elements already included in the KnowledgeGraph KG are extracted. However, the term "actual data" is not limited to this embodiment.

[0068] Furthermore, the prompt generation unit 112 according to this embodiment uses the actual data, in addition to the data referenced by the prompt generation unit 112 according to Embodiment 1, as data relating to the node or edge of the candidate element CE, to generate a prompt PMT.

[0069] The information processing device 100B according to this embodiment, with the configuration described above, can extract data for constructing a knowledge graph with greater accuracy using the language model LM.

[0070] (Processing flow by information processing device 100B) Figure 8 is a flowchart showing the processing flow (information processing method S100B) by the information processing device 100B. As shown in Figure 8, the processing flow by the information processing device 100B includes step S107 in addition to the steps performed by the information processing device 100A shown in Figure 2. Regarding each step performed by the information processing device 100B, redundant explanations of steps that are the same as those of the information processing device 100A in Embodiment 1 will be omitted, and the following explanation will focus on the differences from the information processing device 100A.

[0071] (Step S107) In step S103, the first acquisition unit 111 selects one unprocessed candidate element CE from among one or more candidate elements CE and acquires data related to the selected candidate element CE. Then, in step S107, the search unit 113 selects further data related to the candidate element CE. This process involves, as described above, the search unit 113, • Search for multiple elements already included in the Knowledge Graph KG. Identify one or more elements that have the same node label, edge label, or property name as at least one of the node label NL, edge label EL, and property name PN of candidate element CE, and that are already included in the KnowledgeGraph KG. • Obtain actual data on the one or more of the identified elements mentioned above. This process is included.

[0072] Figure 9 shows a specific processing example B1 relating to steps S101, S103, S107, and S104 according to this embodiment. In the example shown in Figure 9, similar to processing example A1 shown in Figure 3, in step S101, the document "A company's flexible display is realized by a display substrate with excellent flexibility, bending strength, and transparency." is acquired as the target document data TD.

[0073] On the other hand, in processing example B1 shown in Figure 9, the data related to the candidate element CE selected in step S103 includes, in addition to the data explained in processing example A1 shown in Figure 3, • The property values ​​of each property PR1 in Node 1 ("Display", "Bendable", "Company B" in Figure 9) • Property values ​​for each property of Node 2's PR2 ("Display Conductive Layer", "Flexibility / Bending Strength", "Company C" in Figure 9) This includes the following. In this example, each of these property values ​​is information already included in KnowledgeGraph KG.

[0074] Furthermore, in processing example B1 shown in Figure 9, the data relating to the above candidate element CE is as follows: • The source document data from which one or more elements (for example, each of the property values ​​mentioned above) are extracted from the aforementioned KnowledgeGraph KG (in Figure 9, "Source text: The bendable display developed by Company B is realized using a display conductive layer from Company C that possesses flexibility and bending strength.") This includes the source document data, which is an example of the actual data described above. This type of processing can be achieved, for example, by the registration unit 13 storing the source document data for the registration element RE in association with the registration element RE when registering the registration element RE in the knowledge graph KG.

[0075] Furthermore, in the processing example B1 shown in Figure 9, the prompt PMT generated by the prompt generation unit 112 in step S104 includes: The instruction statement IN_REF, used to extract one or more elements (the property values ​​of properties PR1 and PR2 mentioned above) from the actual data (the source document data mentioned above) using the language model LM, is included as an example question. The response OUT_REF of the language model LM to the instruction IN_REF is included as an example response.

[0076] Alternatively, instead of using IN_REF as the example question, you may create a document by adding the source text containing the extracted property values ​​of properties PR1 and PR2 to the instruction IN, and then use that document.

[0077] Thus, in this example, by including the instructions and answers from a previously executed data extraction process as example questions and answers in the prompt PMT, the language model LM can be used to extract data for knowledge graph construction with greater accuracy.

[0078] In this embodiment, various methods can be used for selecting the actual data to be referenced when generating the prompt PMT. For example, these are as follows.

[0079] (Example 1) The prompt generation unit 112 may use as the actual data data data for one or more elements among the multiple elements already included in the KnowledgeGraph KG that have a higher percentage of their property values ​​filled in.

[0080] For example, if each element (node) A, B, and C has three properties, and three properties are filled in element A, one property is filled in element B, and one property is filled in element C, the prompt generation unit 112 may be configured to use the data relating to element A as the actual data.

[0081] (Example 2) The prompt generation unit 112 may use as the actual data data data one or more elements that have a higher similarity to the target document data TD from among the multiple elements already included in the knowledge graph KG.

[0082] (Example 2-1) In the above example 2, the prompt generation unit 112 is: • Extract one or more keywords contained in the target document data TD, Using the extracted keywords, search for multiple nodes already included in KnowledgeGraph KG. • Identify nodes with a relatively high score indicating similarity to the keywords extracted above. The document data from which the data of the identified node was extracted may be used as the actual data mentioned above.

[0083] (Example 2-2) Alternatively, in the above example 2, the prompt generation unit 112 is • Vectorize one or more texts contained in the target document data TD into a feature space (text vectorization), Each of the multiple nodes already included in the KnowledgeGraph KG is also vectorized (node ​​vectorized) within the feature space. In the feature space described above, identify node vectors that are close in distance (highly similar as vectors) to one or more text vectors contained in the target document data TD. The document data from which the data of the node indicated by the identified node vector was extracted may be used as the actual data mentioned above.

[0084] (Example 3) Alternatively, the prompt generation unit 112, The source document and multiple nodes already included in KnowledgeGraph KG are classified into one of several pre-defined classes. • The document data from which the data of nodes belonging to the same class as the target document data (TD) was extracted may be used as the actual data mentioned above.

[0085] As an example, the prompt generation unit 112 may classify the source document and multiple nodes already included in the KnowledgeGraph KG into one of several classes according to the type of node. Here, the type could be, for example, whether the document or node relates to organic materials or inorganic materials, but this example is not limited to this embodiment.

[0086] (Example 4) Furthermore, the prompt generation unit 112 may be configured to include multiple sets (K items, where K is a natural number greater than or equal to 2) of the above-mentioned example questions and answer sets (multiple example questions and answer sets) in the prompt PMT. As an example, the prompt generation unit 112 may be configured to include multiple sets (multiple example questions and answer sets) of the above-mentioned example questions and answer sets. • Execute each of the processes shown in Examples 1 to 3 above, or a combination thereof. Regarding the multiple actual data obtained through this process, each actual data is ranked according to its degree of similarity to the target document data TD. • Include example questions and answers corresponding to each of the top K real data points in the prompt PMT. You may perform the following process.

[0087] Alternatively, the prompt generation unit 112, • Execute each of the processes shown in Examples 1 to 3 above, or a combination thereof. From the multiple real data obtained through this process, randomly select K items. • Include example questions and answers corresponding to each of the K selected real data points in the prompt PMT. You may perform the following process.

[0088] (Example 5) Alternatively, the prompt generation unit 112, in a configuration in which K sets of the above-mentioned example questions and answer examples are included in the prompt PMT, • Execute each of the processes shown in Examples 1 to 3 above, or a combination thereof. The multiple nodes obtained through this process are clustered into K clusters. By selecting one node from each created cluster, a total of K nodes are selected. The document data from which the data of the selected K nodes was extracted may be used as the actual data.

[0089] Here, the clustering process described above can be performed, for example, by vectorizing the document data from which the data for each node was extracted, and then referencing that vector.

[0090] By performing clustering in this way, a diverse range of question and answer examples can be included in the prompt PMT, allowing for more optimal data extraction for knowledge graph construction.

[0091] (Example 6) Alternatively, the prompt generation unit 112 may use data relating to an element randomly selected from a plurality of elements already included in the KnowledgeGraph KG as the actual data.

[0092] (Example 7) Alternatively, the prompt generation unit 112 may be configured to identify the type of candidate element CE and use predetermined actual data for each identified type.

[0093] The information processing device 100B according to this embodiment can extract data for constructing a knowledge graph with greater accuracy using the language model LM by performing the processing described in each of the above examples.

[0094] (Additional Note 1 concerning Embodiments 1 and 2) In embodiments 1 and 2, the knowledge graph KG to be processed is primarily, • A graph format (also called a property graph) in which each node has one or more attributes (properties). Although this is given as an example, this does not limit the embodiments. For example, embodiments 1 and 2 are, • A graph format (also called RDF (Resource Description Framework) format) that represents all graph structures using sets of subject-predicate-object triples. That's fine. When using the RDF format as the Knowledge Graph KG, the term "property" in each of the embodiments described above should be read as "the triple corresponding to that property."

[0095] Figure 10 shows an example of the correspondence between the property graph and the RDF format. As shown in Figure 10, in the RDF format, the property graph • Node labels NL(NL1, NL2) • Edge label EL • Property name PN (PN11~PN13, PN21~PN23) In RDF format, triple • Node, or Edge This corresponds to the example shown in Figure 10. • Node labels NL1 and NL2 correspond to nodes in RDF format (the "Function" node in Figure 10), • Property names PN11-PN13 and PN21-PN23 correspond to each edge in the RDF format (the "Target," "Feature," and "Developer" edges in Figure 10). • The edge label EL corresponds to the edge in the RDF format (the "Implement" edge in Figure 10).

[0096] Furthermore, the values ​​of each property in the property graph (OBJ1, FTR1, DEV1, OBJ2, FTR2, DEV2 in Figure 10) can, for example, correspond to nodes in RDF format.

[0097] Therefore, in the matters described in Embodiments 1 and 2, as an example, • Candidate element CE: node label NL, edge label EL, and property name PN The matter is as follows: • Nodes corresponding to the node label NL of candidate element CE, edges corresponding to the edge label EL, and edges corresponding to the property name PN It can also be expressed as follows.

[0098] Furthermore, in Embodiments 1 and 2, the Knowledge Graph KG to be processed is primarily... • A directed graph where each edge has an orientation. The example given is that the above case does not limit the disclosures described herein. For example, the knowledge graph KG may include one or more undirected edges.

[0099] (Additional Note 2 concerning Embodiments 1 and 2) The generation unit 11 may pre-divide the document data group DG into multiple chunks, or pre-extract multiple chunks from the document data group DG, and use one or more chunks from the multiple chunks obtained by such division or extraction as the target document data TD. More specifically, the first acquisition unit 111 of the generation unit 11 may acquire one or more chunks extracted from the document data group DG as the target document data TD.

[0100] Here, "chunk" refers to the document itself included in the document data group DG, or one or more components included in the document. Examples of chunks include the document itself, the text, paragraphs, sentences, phrases, words, characters, numbers, symbols, etc., included in the document, but these examples are not limited to those described herein.

[0101] (Additional Note 3 concerning Embodiments 1 and 2) The information processing device 100A or 100B may be configured to extract one or more partial texts from the document data group DG in advance and retain those partial texts. This reduces the amount of data in the source data from which the data was extracted.

[0102] As an example, the generation unit 11 may be configured to extract one or more partial texts from the document data group DG in advance and use those partial texts as the target document data TD. For example, the generation unit 11 may, - From the target document data TD, extract partial text containing descriptions corresponding to at least one of the node label NL, edge label EL, and property name PN of the candidate element CE. • Include the aforementioned partial text as at least part of the target document data TD in the prompt PMT. The following processing may be performed. More specifically, the first acquisition unit 111 provided in the generation unit 11, • Generate prompts to extract document data set DG from one or more subtexts that describe components that may be embedded in the knowledge graph KG. • The generated prompt is supplied to the language model LM via the communication unit 130. • The system retains multiple subtexts included in the output of the language model LM that received the prompt. The above partial text may be used as the target document data (TD).

[0103] Thus, the generation unit 11 may be configured to extract the subtext using a language model LM. By adopting such a configuration, data for constructing the knowledge graph can be extracted more efficiently.

[0104] (Additional Note 4 concerning Embodiments 1 and 2) If there are multiple types of candidate element CE, In step S102 of Embodiments 1 and 2, the first acquisition unit 111 determines whether there are any unprocessed candidate element types among the one or more candidate elements CE related to the target document data TD. In step S103, one unprocessed candidate element CE is selected from the one or more candidate elements CE. The above process was given as an example, but this does not limit the matters described in this specification. As an example, the first acquisition unit 111 is • Information indicating one or more types specified by the user is obtained via the input / output unit 140. • Determine whether there are any unprocessed candidate element CE types among the one or more types indicated by the information. This configuration is also acceptable. Alternatively, the configuration may determine whether or not there are any unprocessed candidate element CE types among a predetermined set of one or more types from among multiple types of candidate elements CE.

[0105] (Additional Note 5 concerning Embodiments 1 and 2) In Embodiments 1 and 2, the output information OUT output by the language model LM was given as an example in the format of "property name:property value" enumeration, but it is not limited to this. For example, the output information OUT may be in tabular format or JSON format. In the case of tabular format, for example, the output format may be such that the columns are properties and one set of data is on one line. Alternatively, the columns may be separated by a specific string. Furthermore, the text specifying the data format of the output information OUT may be included in the prompt PMT.

[0106] (Additional Note 6 concerning Embodiments 1 and 2) The formats of the prompt PMTs illustrated in Embodiments 1 and 2 are not limited to those described herein, and other formats may be used. For example, the format of the prompt PMT may be other than natural language. Furthermore, the prompt PMT may include multimodal information, such as figures, photographs, and graphs.

[0107] (Additional Note 7 concerning Embodiments 1 and 2) The information processing devices 100A and 100B may provide the knowledge graph KG generated by the device to other devices via the network N, and the other devices may perform search processing, information presentation processing, or information generation processing by referring to the knowledge graph KG. Alternatively, the first acquisition unit 111 of the information processing devices 100A and 100B may be configured to acquire a query from the user, perform search processing, information presentation processing, or information generation processing by referring to the query and the knowledge graph KG, and present the execution results to the user via the input / output unit 140. With such a configuration, suitable search processing, information presentation processing, or information generation processing can be performed using the knowledge graph KG that has been accurately constructed by the information processing devices 100A and 100B.

[0108] When the information processing devices 100A and 100B are applied to the field of materials informatics, the properties of the knowledge graph may include, as an example, at least one of the following: substance name, composition name, physical properties, process, process conditions, synthesis equipment, compounding equipment, and processing equipment.

[0109] [Examples of implementation using software] Each functional block of the information processing devices 100A and 100B (especially each block of the control unit 110), and each functional block of the server 200 (especially the control unit 240) may be implemented by logic circuits (hardware) formed on an integrated circuit (IC chip) or the like, or by software. In the latter case, each of the information processing devices 100A, 100B, and the server 200 can be configured using, for example, a computer (electronic computer). Figure 11 is a block diagram illustrating the physical configuration of a computer used as the information processing devices 100A, 100B, and the server 200.

[0110] (Physical configuration of the information processing device 100) As shown in Figure 11, the information processing devices 100A and 100B can be configured by a computer comprising a bus 110, a processor 101, a main memory 102, an auxiliary memory 103, a communication interface 104, and an input / output interface 105. The processor 101, main memory 102, auxiliary memory 103, communication interface 104, and input / output interface 105 are connected to each other via the bus 110. An input device 40 and an output device 50 are connected to the input / output interface 105.

[0111] For the processor 101, for example, a microprocessor, a digital signal processor, a microcontroller, or a combination thereof may be used. For the main memory 102, for example, semiconductor RAM (random access memory) may be used.

[0112] For example, the auxiliary memory 103 may be flash memory, an HDD (Hard Disk Drive), an SSD (Solid State Drive), or a combination thereof. The auxiliary memory 103 stores a program that causes the processor 101 to execute the operations of the information processing devices 100A and 100B described above. The processor 101 expands the program stored in the auxiliary memory 103 onto the main memory 102 and executes each instruction contained in the expanded program. The auxiliary memory 103 also stores various data that the processor 101 refers to in order to operate the computer as the information processing devices 100A and 100B.

[0113] The communication interface 104 is an interface that connects to the network 91. As the input / output interface 105, for example, a USB (Universal Serial Bus) interface, a short-range communication interface such as infrared or Bluetooth (registered trademark), or a combination thereof can be used.

[0114] The input device 40 may be, for example, a keyboard, mouse, touchpad, microphone, or a combination thereof. The output device 50 may be, for example, a display, printer, speaker, or a combination thereof.

[0115] (Physical configuration of Server 200) As shown in Figure 11, the server 200 can be configured by a computer comprising a bus 210, a processor 201, main memory 202, auxiliary memory 203, communication interface 204, and communication interface 205. The processor 201, main memory 202, auxiliary memory 203, communication interface 204, and communication interface 205 are connected to each other via the bus 210.

[0116] For the processor 201, for example, a microprocessor, a digital signal processor, a microcontroller, or a combination thereof may be used. For the main memory 202, for example, semiconductor RAM may be used.

[0117] Auxiliary memory 203 can be, for example, flash memory, an HDD, an SSD, or a combination thereof. The auxiliary memory 203 stores a program for operating the computer as a server 200. The processor 201 loads the program stored in the auxiliary memory 203 into the main memory 202 and executes each instruction contained in the loaded program. The auxiliary memory 203 also stores various data that the processor 201 references in order to operate the computer as a server 200.

[0118] Communication interface 204 is an interface that connects to network 91. Communication interface 205 is an interface that connects to network 92.

[0119] Each of the above-described programs may, instead of being stored in auxiliary memory 103 or auxiliary memory 203, be recorded on an external recording medium and supplied to the computer by being read from the external recording medium. As the external recording medium, a computer-readable "non-temporary tangible medium" such as tape, disk, card, semiconductor memory, or programmable logic circuit can be used. Furthermore, each of the above-described programs may be supplied to the computer via any transmission medium (such as a communication network or broadcast wave). In addition, one aspect of this disclosure may be realized in the form of data signals embedded in a carrier wave, in which each program is embodied by electronic transmission.

[0120] 〔summary〕 The matters described herein include at least the following components:

[0121] (Composition 1) A generation unit generates prompts for extracting data of candidate elements to be included in the KnowledgeGraph from the target document data using a language model, using data related to the nodes or edges of the candidate elements. The acquisition unit acquires output information output by the language model that received the prompt. An information processing device equipped with the following features.

[0122] The above configuration allows for accurate extraction of data for constructing a knowledge graph.

[0123] (Configuration 2) The generating unit is The prompt is generated using at least one of the node label, edge label, and property name of the candidate element as data relating to the node or edge of the candidate element. The information processing device described in Configuration 1.

[0124] According to the above configuration, with respect to a variety of candidate elements (components), prompts can be suitably generated for extracting data for those candidate elements to be included in the knowledge graph from the target document data using a language model.

[0125] (Composition 3) The generation unit generates the prompt using descriptive information relating to at least one of the node label, edge label, and property name of the candidate element as data relating to the node or edge of the candidate element. The information processing device described in Configuration 2.

[0126] With the above configuration, it is possible to extract data for building a knowledge graph with greater accuracy using a language model.

[0127] (Composition 4) The generating unit is An instruction statement including data relating to the node or edge of the candidate element, At least a portion of the aforementioned target document data and Generate the prompt including the above An information processing device as described in any one of items 1 to 3 of the configuration.

[0128] The above configuration allows for more accurate extraction of data for constructing a knowledge graph.

[0129] (Composition 5) The generating unit is From the aforementioned target document data, extract partial text containing descriptions corresponding to at least one of the node label, edge label, and property name of the candidate element. The partial text is included in the prompt as at least part of the target document data. The information processing device described in Configuration 4.

[0130] According to the above configuration, data for constructing a knowledge graph can be extracted efficiently.

[0131] (Composition 6) The generation unit extracts the subtext using a language model. The information processing device described in configuration 5.

[0132] The above configuration allows for more efficient extraction of data for constructing knowledge graphs.

[0133] (Composition 7) The generating unit is Obtain actual data for one or more elements already included in the KnowledgeGraph that have the same node label, edge label, or property name as at least one of the node label, edge label, and property name of the candidate element, respectively. The actual data is further used as data relating to the nodes or edges of the candidate elements. An information processing device as described in any one of items 2 to 6 of the configuration.

[0134] With the above configuration, it is possible to extract data for building a knowledge graph with greater accuracy using a language model.

[0135] (Composition 8) The actual data acquired by the generation unit includes the source document data of the one or more elements already included in the KnowledgeGraph. The information processing device described in Configuration 7.

[0136] With the above configuration, it is possible to extract data for building a knowledge graph with greater accuracy using a language model.

[0137] (Composition 9) The generating unit is A language model is used to extract one or more elements from the actual data, The language model's response to the instruction statement and Generate the prompt including the above The information processing device described in configuration 8.

[0138] With the above configuration, it is possible to extract data for building a knowledge graph with greater accuracy using a language model.

[0139] (Composition 10) The generation unit uses as the actual data data data for one or more elements among the multiple elements already included in the knowledge graph that have a higher percentage of their property values ​​filled in. An information processing device as described in any one of items 7 to 9.

[0140] With the above configuration, it is possible to extract data for building a knowledge graph with greater accuracy using a language model.

[0141] (Composition 11) The generation unit uses data relating to one or more elements that have a higher similarity to the target document data from among the multiple elements already included in the knowledge graph as the actual data. An information processing device according to any one of items 7 to 10.

[0142] With the above configuration, it is possible to extract data for building a knowledge graph with greater accuracy using a language model.

[0143] (Composition 12) A processing unit that generates registration elements by processing the output information, A registration unit that registers the aforementioned registration elements in the knowledge graph. It is equipped with An information processing device as described in any one of items 7 to 11 of the configuration.

[0144] According to the above configuration, a knowledge graph can be constructed effectively.

[0145] (Composition 13) The registration unit further comprises a presentation unit that presents the output information to the user, The processing unit generates the registration elements by referring to user instructions regarding the output information. According to the above configuration, it is possible to construct a more suitable knowledge graph while reflecting user instructions.

[0146] (Composition 14) An information processing method performed by an information processing device, A generation step that generates prompts for extracting data of candidate elements to be included in the KnowledgeGraph from the target document data using a language model, using data related to the nodes or edges of those candidate elements, The acquisition step involves obtaining the output information output by the language model that received the prompt. An information processing method that includes this.

[0147] The above configuration allows for accurate extraction of data for constructing a knowledge graph.

[0148] (Composition 15) A program for causing a computer to function as an information processing device as described in Configuration 1, wherein the program causes the computer to function as the generation unit and the acquisition unit.

[0149] (Composition 16) A computer-readable recording medium on which the program described in Configuration 15 is recorded.

[0150] This disclosure is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of this disclosure. [Explanation of Symbols]

[0151] 100A, 100B... Information processing equipment 110 ··· Control Unit 11...Generation part 111 ··· First acquisition section 112 ··· Prompt generation unit 12 ··· Second acquisition section (acquisition section) 13 ··· Registration Department 131... Processing section 132 ··· Element Registration Section 133 ... Presentation part 200... Server

Claims

1. A generation unit generates prompts for extracting data of candidate elements to be included in the KnowledgeGraph from the target document data using a language model, using data related to the nodes or edges of the candidate elements. The acquisition unit acquires output information output by the language model that received the prompt. An information processing device equipped with the following features.

2. The generating unit is The prompt is generated using at least one of the node label, edge label, and property name of the candidate element as data relating to the node or edge of the candidate element. The information processing apparatus according to claim 1.

3. The generation unit generates the prompt using descriptive information relating to at least one of the node label, edge label, and property name of the candidate element as data relating to the node or edge of the candidate element. The information processing apparatus according to claim 2.

4. The generating unit is An instruction statement including data relating to the node or edge of the candidate element, At least a portion of the aforementioned target document data and Generate the prompt that includes the above The information processing apparatus according to claim 3.

5. The generating unit is From the aforementioned target document data, extract partial text containing descriptions corresponding to at least one of the node label, edge label, and property name of the candidate element. The partial text is included in the prompt as at least part of the target document data. The information processing apparatus according to claim 4.

6. The generation unit extracts the subtext using a language model. The information processing apparatus according to claim 5.

7. The generating unit is Obtain actual data for one or more elements already included in the KnowledgeGraph that have the same node label, edge label, or property name as at least one of the node label, edge label, and property name of the candidate element, respectively. The actual data is further used as data relating to the nodes or edges of the candidate elements. The information processing apparatus according to any one of claims 2 to 6.

8. The actual data acquired by the generation unit includes the source document data from which the one or more elements already included in the KnowledgeGraph are extracted. The information processing apparatus according to claim 7.

9. The generating unit is An instruction sentence used to extract the one or more elements from the actual data using a language model, The language model's response to the aforementioned instruction and Generate the prompt that includes the above The information processing apparatus according to claim 8.

10. The generation unit uses as the actual data data data for one or more elements among the multiple elements already included in the knowledge graph that have a higher percentage of their property values ​​filled in. The information processing apparatus according to claim 7.

11. The generation unit uses data relating to one or more elements that have a higher similarity to the target document data from among the multiple elements already included in the knowledge graph as the actual data. The information processing apparatus according to claim 7.

12. A processing unit that generates registration elements by processing the output information, A registration unit that registers the aforementioned registration elements in the knowledge graph. It is equipped with The information processing apparatus according to claim 7.

13. The registration unit further comprises a presentation unit that presents the output information to the user, The processing unit generates the registration elements by referring to user instructions regarding the output information. The information processing apparatus according to claim 12.

14. An information processing method performed by an information processing device, A generation step that generates prompts for extracting data of candidate elements to be included in the KnowledgeGraph from the target document data using a language model, using data related to the nodes or edges of the candidate elements, The acquisition step involves obtaining the output information output by the language model that received the prompt. An information processing method that includes this.

15. A program for causing a computer to function as an information processing device according to claim 1, wherein the program causes the computer to function as the generation unit and the acquisition unit.

16. A computer-readable recording medium having the program described in claim 15 recorded on it.

Images