Processor system, knowledge graph generation method, and program

Abstract

By generating a knowledge graph using modified reference examples, it is possible to generate a highly accurate knowledge graph. [Solution] A processor system having one or more processors and one or more memory resources, wherein the memory resource stores a defect knowledge database containing case knowledge information including corrected knowledge graphs related to defects, a language processing model, a target defect report, and a program for generating a knowledge graph related to the target defect report. The processor extracts similar case knowledge information from the defect knowledge database, and inputs the target defect report and the extracted case knowledge information into the language processing model to identify named entities in the target defect report and the relationships between named entities, and generates a knowledge graph.

Description

[Technical Field] 【0001】 The present invention relates to a processor system, a knowledge graph generation method, and a program that extract the causes of product defects from defect reports describing product defects and accurately generate a knowledge graph about defects using reference cases of similar defects. [Background technology] 【0002】 In the field of industrial products, the current situation is that information about defects is not being fully utilized for defect prevention or next-generation design because the wording of defect reports varies from person to person. 【0003】 Therefore, it is considered an effective method to generate a knowledge graph from defect-related information by utilizing language processing models such as LLMs (Large Language Models), which have high processing capabilities for differences in expression. On the other hand, language processing models do not learn domain knowledge such as industrial product or part names, and their response accuracy and reliability regarding these are not high. One way to improve the response accuracy of language processing models is to provide them with a high-quality knowledge graph, which is data containing the necessary domain knowledge. 【0004】 Furthermore, Patent Document 1 discloses technology relating to a case search device that enables highly flexible searching. Specifically, Patent Document 1 states that "The case search device according to an embodiment has a first acquisition unit, a second acquisition unit, a calculation unit, a search unit, and a presentation unit. The first acquisition unit acquires search conditions, which are data of cases to be searched. The second acquisition unit acquires meta-search conditions, which are descriptions of viewpoints to be considered when searching for cases similar to the search conditions. The calculation unit calculates the similarity between the search conditions and each of a plurality of reference cases, which are data of cases to be searched, based on the meta-search conditions. The search unit searches for similar reference cases that are similar to the search conditions from the viewpoint of the meta-search conditions, based on the similarity. The presentation unit presents the search results from the search unit." [Prior art documents] [Patent Documents] 【0005】 [Patent Document 1] Japanese Patent Publication No. 2023-39656 [Overview of the Initiative] [Problems that the invention aims to solve] 【0006】 Generating highly accurate, high-quality knowledge graphs sometimes requires manual correction, but such corrections are time-consuming and labor-intensive. To minimize the manual effort required to correct knowledge graphs, it is considered effective to generate subsequent knowledge graphs using reference examples that include domain knowledge and have undergone appropriate corrections, thereby generating knowledge graphs that reflect the correction patterns in those reference examples. 【0007】 Patent Document 1 discloses a technique for converting search target case data into meta-search conditions, calculating similarity in the feature space, and searching for similar cases. However, the technique in Patent Document 1 does not consider using the retrieved case data to reflect past modification patterns in the generation of knowledge graphs for subsequent searches. 【0008】 This invention has been made in view of the above problems, and aims to generate a knowledge graph with high accuracy by generating a knowledge graph using modified reference examples. [Means for solving the problem] 【0009】 The present invention includes several means for solving at least part of the above problems, but an example thereof is as follows. A processor system according to one aspect of the present invention for solving the above problems is a processor system having one or more processors and one or more memory resources, wherein the memory resources store a defect knowledge database which stores case knowledge information including a corrected knowledge graph relating to defects of a product or part, a language processing model, a target defect report, and a program which generates a knowledge graph relating to the target defect report, and the processor, by executing the program, extracts the case knowledge information which has registered case knowledge similar to the defect indicated by the target defect report from the defect knowledge database, and by inputting the target defect report and the extracted case knowledge information into the language processing model, identifies named entities of the target defect report and the relationships between named entities, and generates a knowledge graph relating to the target defect report based on the identified named entities and the relationships. [Effects of the Invention] 【0010】 According to the present invention, a knowledge graph with high accuracy can be generated by generating a knowledge graph using modified reference examples. [Brief explanation of the drawing] 【0011】 [Figure 1] This diagram shows an example of a general configuration of a processor system. [Figure 2] This is a schematic diagram illustrating the defect knowledge database. [Figure 3] This is a diagram illustrating an example of a knowledge graph. [Figure 4] This figure shows an example of a case defect report. [Figure 5] This is a flowchart illustrating an example of a knowledge database update process. [Figure 6] This is an overall diagram illustrating the knowledge database update process. [Figure 7] This is an explanatory diagram of the reference case search process. [Figure 8]It is a flowchart showing the generation process of a knowledge graph. [Figure 9] It is an explanatory diagram of the knowledge graph generation process. [Figure 10] FIG. 10A is a diagram showing an example of a prompt. FIG. 10B is a diagram showing an example of named entity data. [Figure 11] It is a diagram showing an example of a named entity list, a relationship list, and an example of a case defect report sentence. [Figure 12] FIG. 12A is a diagram showing an example of a prompt. FIG. 12B is a diagram showing an example of relationship data. [Figure 13] It is a diagram showing an example of a prompt. [Figure 14] It is a diagram showing a modified example of a knowledge graph. [Figure 15] FIG. 15A is a diagram showing an example of knowledge graph features. FIG. 15B is a diagram showing an example of each element of an arithmetic expression used for generating modified features. [Figure 16] It is a diagram showing an example of a user interface screen. [Figure 17] It is a flowchart showing an example of a modification process. [Figure 18] It is an overall explanatory diagram of the knowledge DB update process. [Figure 19] It is a diagram showing an example of a user interface screen. 【Embodiments for Carrying Out the Invention】 【0012】 The following embodiments are examples for explaining the present invention, and for the sake of clarity of explanation, omissions and simplifications are made as appropriate. The present invention can be implemented in various other forms. Also, unless otherwise particularly limited, each component may be singular or plural. 【0013】 Furthermore, the position, size, shape, and range of each component shown in the drawings may not represent the actual position, size, shape, and range in order to facilitate understanding of the invention. For this reason, the present invention is not necessarily limited to the position, size, shape, and range disclosed in the drawings. 【0014】 Furthermore, while various types of information may be described using terms such as "table," "list," and "queue," these types of information may also be represented by other data structures. For example, various types of information such as "XX table," "XX list," and "XX queue" may be referred to as "XX information." When describing identification information, terms such as "identification information," "identifier," "name," "ID," and "number" are used, and these terms are interchangeable. 【0015】 Furthermore, if there are multiple components with the same or similar function, they may be described using the same symbol but with different subscripts. Also, if there is no need to distinguish between these multiple components, the subscripts may be omitted in the description. 【0016】 Furthermore, in the embodiments, the processes performed by executing the program may be described. Here, the computer executes the program using a processor (e.g., CPU, GPU) and performs the processing defined in the program using memory resources (e.g., memory) and interface devices (e.g., communication ports). Therefore, the processor may be the main entity performing the processing by executing the program. 【0017】 Similarly, the entity performing the processing by executing the program may be a controller, device, system, computer, or node having a processor. The entity performing the processing by executing the program may be an arithmetic unit, and may include dedicated circuits that perform specific processing. Here, dedicated circuits include, for example, FPGAs (Field Programmable Gate Arrays), ASICs (Application Specific Integrated Circuits), CPLDs (Complex Programmable Logic Devices), etc. 【0018】 The program may be installed on the computer from the program source. The program source may be, for example, a program distribution server or a storage medium readable by the computer. If the program source is a program distribution server, the program distribution server includes a processor and storage resources for storing the program to be distributed, and the processor of the program distribution server may distribute the program to other computers. In addition, in the embodiment, two or more programs may be implemented as one program, or one program may be implemented as two or more programs. 【0019】 Embodiments of the present invention will be described below with reference to the drawings. 【0020】 <Outline configuration of processor system 100> Figure 1 shows an example of the schematic configuration of the processor system 100. The processor system 100 is a system that updates the database by generating knowledge graphs and their features related to the defect report text to be processed, using case knowledge graphs etc. stored in the defect knowledge database (database), and storing them in the defect knowledge database. 【0021】 Specifically, the processor system 100 extracts similar reference cases related to defects from a defect knowledge database, for example, using defect report documents that describe defects in products (devices) or parts in fields such as industrial equipment. 【0022】 Furthermore, the processor system 100 uses a language processing model to identify named entities and the relationships between them from the target defect report text based on knowledge information from reference cases (such as a case knowledge graph), and generates a knowledge graph for them. 【0023】 Furthermore, the processor system 100 modifies the named entities and the relationships between named entities in the generated knowledge graph according to predetermined rules (constraints). 【0024】 Furthermore, the processor system 100 calculates the features of the modified knowledge graph and the modified features. 【0025】 Furthermore, the processor system 100 updates the database by storing the generated (calculated) knowledge graph and each feature quantity for the target defect report in the defect knowledge database. 【0026】 This type of processor system allows for the incorporation of past correction patterns into subsequent knowledge graph generation, enabling the creation of highly accurate knowledge graphs from bug reports. As a result, the manual effort required to correct knowledge graphs can be minimized. 【0027】 While the technical field of the product or component indicated in the defect report is not particularly limited, in this embodiment, the processing of a defect report for a diesel generator will be explained as an example. 【0028】 <Configuration of Processor System 100> The processor system 100 is a computer that generates knowledge information about product defects and stores it in a database. Specifically, the processor system 100 reads programs and information stored in the memory resource 20, generates a knowledge graph for the target defect report, calculates the feature quantities of the knowledge graph and their corrected feature quantities, and updates the database by registering this information as case knowledge information. Details of each process performed by the processor system 100 will be described later. 【0029】 The processor system 100 is, for example, a personal computer, a server computer, a cloud server, or a computer such as a tablet terminal or smartphone, and is a system that includes at least one of these computers. 【0030】 As shown in Figure 1, the processor system 100 includes a processor 10, memory resources 20, a Network Interface Device (NI) 30, and a User Interface Device (UI) 40. 【0031】 The processor 10 is an arithmetic unit that reads various programs stored in the memory resource 20 and executes the processing corresponding to each program. The processor 10 is a device capable of performing arithmetic processing, such as a microprocessor, CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), or semiconductor device. 【0032】 Memory resource 20 is a storage device that stores various types of information. Specifically, memory resource 20 is a non-volatile or volatile storage medium such as RAM (Random Access Memory) or ROM (Read Only Memory). Alternatively, memory resource 20 may be a rewritable storage medium such as flash memory, HDD (Hard Disk Drive), or SSD (Solid State Drive), or a USB (Universal Serial Bus) memory, memory card, or hard disk. 【0033】 NI30 is a communication device that communicates information with an external device (for example, external device 400). NI30 communicates information with the external device via a predetermined communication network N, such as a LAN (Local Area Network) or the Internet. Unless otherwise specified below, information communication between the processor system 100 and the external device is assumed to be performed via NI30. 【0034】 UI40 is an interface between an input device that inputs user (operator) instructions to the processor system 100 and an output device that outputs information generated by the processor system 100 (hereinafter, the input device and output device may be collectively referred to as the input / output device 300). The input device may include, for example, a keyboard, touch panel, a mouse, or other pointing devices, or a microphone or other voice input device. The output device may include, for example, a display, a printer, or a speech synthesis device. Unless otherwise specified below, user operations on the processor system 100 (e.g., inputting and outputting information, and instructions for processing) are assumed to be received via UI40. 【0035】 Furthermore, some or all of the configurations, functions, and processing means of the processor system 100 may be implemented in hardware, for example, by designing them as integrated circuits. Also, some or all of the functions of the processor system 100 may be implemented in software, or through the collaboration of software and hardware. Additionally, the processor system 100 may use hardware with fixed circuits, or it may use hardware in which at least some of the circuits can be modified. 【0036】 Furthermore, the processor system 100 can also be realized by having a user (operator) perform some or all of the functions and processes realized by each program. 【0037】 Furthermore, the database and various information within the memory resource 20 described below may be files or other data structures besides databases, as long as they are areas capable of storing data. 【0038】 Furthermore, various types of information, including the database within the memory resource 20, do not need to be stored in the memory resource 20 in advance; they may be obtained from, for example, an external device (external device 400) each time a corresponding process is performed. 【0039】 <<Language Processing Model 110>> The language processing model 110 is an algorithm or learning model (information model) capable of predicting or generating appropriate words from text data according to context using natural language processing. In this embodiment, for example, an LLM (Large Language Model: deep learning model) is assumed. 【0040】 <<Trouble Knowledge Database 120>> The Defect Knowledge Database 120 is a database that stores knowledge information from reference cases. Specifically, the Defect Knowledge Database 120 stores case knowledge graphs, corresponding case defect reports, case knowledge graph features, and case correction features. 【0041】 Figure 2 is a schematic diagram of the defect knowledge database. The case knowledge graph stored in the defect knowledge database 120 is generated for each defect report and shows the relationships between nodes, representing information such as the location, components, and status of defects in the target product (device) or part. It hierarchically represents the inclusion relationship between the defect location and the units that constitute it, and also expresses the reasons and consequences of phenomena that may occur in other locations, based on the phenomena that may occur in each of those parts (failures and conditions that may cause them). 【0042】 Figure 3 shows an example of a knowledge graph (case knowledge graph). As shown in the figure, in the knowledge graph of this embodiment, the relationship between a phenomenon (Status) and the location where the phenomenon occurs (Component) is represented as an "is edge" with an arrow (start: location, end: phenomenon). In addition, the knowledge graph represents the causal relationship between the cause and effect of a phenomenon as a "cause edge" with an arrow (start: cause, end: effect). Furthermore, the knowledge graph represents the inclusion relationship between a device and its components (parts) as a "part of edge" with an arrow (start: component, end: device containing the component). Note that the causal relationships represented by the cause edges in the knowledge graph may also include coefficient information, and this coefficient information may be in the form of a Bayesian network containing probability information such as conditional probability. 【0043】 Figure 4 shows an example of a case defect report. The case defect reports stored in the defect knowledge DB120 are defect reports corresponding to each case knowledge graph, and are information that contains text sentences describing defects in products (devices) or parts. 【0044】 The case knowledge graph features stored in the defect knowledge DB120 are information that represents the case knowledge graph as vector-based features (vector format, hereinafter the same). The case correction features, on the other hand, are information that represents the corrections made to the case knowledge graph as vector-based features (correction burden). The corrections made to the case knowledge graph refer to, for example, corrections made by the processor system 100 or manually to named entities in the knowledge graph or the relationships between named entities. The case correction features are features generated based on a comparison between the knowledge graph before and after the correction. 【0045】 Furthermore, the case knowledge graph, the case defect report, the case knowledge graph features, and the case correction features each have a one-to-one correspondence. 【0046】 <<Bug Report Document DB130>> The Bug Report Database (DB130) is a database that stores bug reports that are targeted for processing, using case knowledge graphs and case bug reports stored in the Bug Knowledge Database (DB120) to generate knowledge graphs and their features. Such bug reports are created, for example, by extracting relevant documents related to bugs from bug reports created by maintenance personnel. Since the content of the target bug reports is the same as that of case bug reports, a detailed explanation is omitted. 【0047】 <<Knowledge DB Update 210>> Knowledge DB Update Program 210 is a program that generates a knowledge graph and calculates features related to the target bug report by executing the knowledge DB update process, and stores them in the bug knowledge DB 120. Details of the knowledge DB update process will be described later. 【0048】 The details of the processor system 100 have been explained above. <<External device 400>> External device 400 is a device that transmits input information to the processor system 100. External device 400 is also a device that acquires information generated by the processor system 100. Specifically, external device 400 may transmit knowledge information of reference cases to the processor system 100. In addition to such information, external device 400 may also provide (transmit) various types of information used in processing performed by the processor system 100. Furthermore, external device 400 may store information generated by the processor system 100 and display it on a display built into the external device 400. 【0049】 <Knowledge DB update process> Next, we will describe the knowledge database update process performed by the processor system 100. 【0050】 Figure 5 is a flowchart illustrating an example of the knowledge database update process. Figure 6 is an overall diagram illustrating the knowledge database update process, including the data flow. This process begins when the processor system 100 receives an instruction input from the user. At this time, the processor 10 reads the knowledge database update program 210 from the memory resource 20, and retrieves and uses the necessary information from the memory resource 20 according to the processing stage to execute the following processes. 【0051】 When processing begins, the processor 10 retrieves the bug report to be processed from the bug report DB 130 in the memory resource 20 (step S10). 【0052】 Next, the processor 10 searches for reference cases (step S20). Specifically, the processor 10 uses the acquired defect report as input data to search the defect knowledge DB 120 for knowledge information of reference cases that will be used to generate a knowledge graph of the defect report. 【0053】 Figure 7 is an explanatory diagram of the reference case search process, including the data flow. During the search, processor 10 generates feature quantities for the defect report text. Specifically, processor 10 inputs the input defect report text into a predetermined learning model such as a neural network and generates report text feature quantities in vector embedding format (vector embedding format) based on the words contained in the defect report text and their positional relationships. 【0054】 Furthermore, the processor 10 uses the generated report features to search the defect knowledge DB 120 for knowledge information on reference cases similar to the defect report. Specifically, the processor 10 performs a vector search (vector search, hereafter the same) on the defect knowledge DB 120 based on the report features and obtains a group of case knowledge graphs (including a case knowledge graph and its corresponding case defect report, case knowledge graph features, and case correction features) with a high cosine similarity between the case knowledge graph features and the report features. The processor 10 may obtain multiple groups of case knowledge graphs with high cosine similarity. 【0055】 Furthermore, processor 10 narrows down the reference cases. Specifically, processor 10 selects reference cases from the acquired reference cases that have a larger weighted sum of cosine similarity and case modification features (modification burden). The weight values ​​for similarity and case modification features are arbitrary. For example, the weight values ​​may be set so that one or more (e.g., two) sets of case knowledge graphs with larger case modification features are selected. By using reference cases with larger case modification features to generate the knowledge graph, it is possible to generate a knowledge graph about the target defect report based on reference cases that reflect a larger number of modification patterns. 【0056】 Next, the processor 10 generates a knowledge graph (step S30). Specifically, the processor 10 generates a knowledge graph using knowledge information of reference cases retrieved from the defect knowledge DB 120 with respect to the target defect report. 【0057】 Figure 8 is a flowchart illustrating the knowledge graph generation process. Figure 9 is an explanatory diagram of the knowledge graph generation process, including the data flow. First, the processor 10 extracts named entities from the target defect report (step S31). Specifically, the processor 10 generates a prompt 1 containing instructions for extracting named entities from the defect report and inputs it into the language processing model 110 (e.g., LLM). 【0058】 Figure 10A shows an example of prompt 1. The example prompt 1 includes instructions that tell the LLM to extract named entities related to the configuration and defects of a product (device), etc., from the target defect report, classify them into component and status, and output named entity data in list format with number, named entity, and classification as items. 【0059】 Furthermore, the example prompt 1 includes, as a reference example of named entity extraction, the case knowledge graph retrieved in step S20 and the corresponding case defect report. The processor 10 generates a named entity list and a relationship list from the case knowledge graph, converts the named entity list to a predetermined format (e.g., CSV format: Comma Spaced Value format), and writes it into prompt 1 along with the case defect report. 【0060】 Figure 11 shows an example of a named entity list, a relationship list, and a case defect report. As shown in the figure, the named entity list is generated in a list format with the items index, entity, and category associated. Similarly, the relationship list is generated in a list format with the items index, relationship, entity1, and entity2 associated. Processor 10 generates this list information from the case knowledge graph, converts it to CSV format, and writes it to prompt 1. 【0061】 Furthermore, the example prompt 1 includes the relevant defect report (incident record). In step S10, the processor 10 writes the defect report obtained from the defect report DB 130 into prompt 1. 【0062】 The processor 10 obtains named entity data in list format by inputting such prompt 1 to the language processing model 110. 【0063】 Figure 10B shows an example of named entity data. As shown in the figure, named entity data is output from LLM as a list format data with associated items: index (number), entity (named entity), and category (classification), based on the input of prompt 1. 【0064】 Next, the processor 10 extracts the relationships between named entities (step S32). Specifically, the processor 10 generates a prompt 2 containing instructions for extracting the relationships between named entities from the generated named entity data, and inputs it into the language processing model 110 (e.g., LLM). 【0065】 Figure 12A shows an example of prompt 2. The example prompt 2 includes instructions that tell the LLM to extract the relationships between named entities from the named entity data, classify the extracted relationships into "is," "cause," and "part of," and output relationship data in list format with items: number, relationship, component 1 (named entity 1), and component 2 (named entity 2). 【0066】 Furthermore, prompt 2 includes a case knowledge graph and a case defect report corresponding to the case knowledge graph as reference examples for relationship extraction. The processor 10 converts the named entity list and relationship list shown in Figure 11 into a predetermined format (e.g., CSV format) and writes them into prompt 2 along with the case defect report. 【0067】 Furthermore, Prompt 2, like Prompt 1, contains the relevant defect report (incident record). Prompt 2 also contains named entity data generated by Prompt 1. Processor 10 converts the named entity data into a predetermined format (e.g., CSV format) and writes it to Prompt 2. 【0068】 The processor 10 obtains relational data in list format by inputting a prompt 2 containing such information to the language processing model 110. 【0069】 Figure 12B shows an example of relational data. As shown in the figure, based on the input of prompt 2, relational data is output from LLM as a list-formatted data with the items index (number), relationship (relationship), entity1 (component 1 / named entity 1), and entity2 (component 2 / named entity 2) associated. 【0070】 Next, the processor 10 performs tuning (step S33). Specifically, the processor 10 generates a prompt 3 that describes instructions to review named entity data and relational data, and inputs it to the language processing model 110 (e.g., LLM). 【0071】 Figure 13 shows an example of prompt 3. The illustrated prompt 3 is an example of an instruction sentence that instructs the language processing model 110 to review relational data in the form of a question. In addition to reviewing relational data, prompt 3 may also contain instruction sentences that instruct the review of named entity data. For example, regarding named entity data, there may be instruction sentences that instruct the review of whether "Component" is registered in the category corresponding to named entity of a product or part, or whether "status" is registered in the category corresponding to named entity of a phenomenon. 【0072】 Furthermore, for relational data, for example, there are instructions that instruct you to review whether the appropriate named entity for category is registered for entity1 and entity2, depending on the relationship ("is", "cause", "part of") (for example, if the relationship is "is", then the named entity for Component is registered for entity1 and the named entity for status is registered for entity2). 【0073】 There are various types of such instructions, and the processor 10 can, for example, select the appropriate one from a database (not shown) in the memory resource 20 where these instructions are registered and write it to prompt 3. 【0074】 Next, processor 10 generates a knowledge graph (step S34). Specifically, processor 10 generates a knowledge graph (illustrated in Figure 3) with Component and status as nodes based on named entity data, and the relationships between nodes based on relationship data, and then terminates the processing of this flow. 【0075】 This knowledge graph generation process allows for the creation of a highly accurate knowledge graph that reflects the correction patterns of the target technical field indicated by the bug report, even if a large amount of training data on the relevant technical field has not been used to train the language processing model 110. This is possible because, if at least one correction has been made in the past, that example can be used as a reference. 【0076】 Let's return to Figure 5 for explanation. After generating the knowledge graph, the processor 10 determines whether or not the knowledge graph needs to be modified (step S40). For example, the processor 10 determines whether or not modifications are needed to the name of the component, the causal relationship between the cause and effect of a defect, or the inclusion relationship between products and parts. 【0077】 For example, processor 10 references basic data (such as product and component names, rules regarding component inclusion relationships within a product, and rules regarding causal relationships between the cause and effect of a defect; not shown) from memory resource 20. Furthermore, if the generated knowledge graph contains component names, causal relationships between the cause and effect of a defect, or component inclusion relationships that contradict the basic data, processor 10 determines that correction is necessary (Yes in step S40) and proceeds to step S50. On the other hand, if no inconsistencies with the basic data are found, processor 10 determines that correction is not necessary (No in step S40) and proceeds to step S60. 【0078】 In step S50, the processor 10 modifies the knowledge graph. Specifically, the processor 10 modifies the knowledge graph according to the rules indicated by the underlying data. For example, the processor 10 modifies the names of products and parts to match the names of products, etc., included in the underlying data. Alternatively, the processor 10 modifies the connection relationships of edges indicating causal relationships (cause edges) and edges indicating inclusion relationships (part of edges) according to the rules of the underlying data. 【0079】 Figure 14 shows an example of a corrected knowledge graph. As illustrated, in the pre-correction knowledge graph generated from the target defect report, "engine" and "diesel" are extracted separately as named entities for the Component. In addition, the attribute of the cooling pump is extracted as "status". Furthermore, the automatic shutdown of the diesel engine is caused by "damage" to the cooling pump, and the automatic shutdown is the result, but in the pre-correction knowledge graph, the cooling pump and the automatic shutdown are connected by a cause edge without going through the cause of "damage". 【0080】 In this case, processor 10 corrects the named entities for "engine" and "diesel" to "diesel engine". Also, processor 10 corrects the attribute of the cooling pump from status to component. Furthermore, processor 10 makes corrections by connecting an is edge to "broken" which indicates the status of the "cooling pump", and by connecting a cause edge from "broken" (cause) to "automatic stop" (result). 【0081】 Furthermore, the correction of the knowledge graph is not limited to this method; any method is acceptable as long as it can identify and correct the errors related to product and component names, the inclusion relationships between products and components, and the causal relationships between causes and effects of defects, and restore them to their correct form and relationships. 【0082】 Next, in step S60, the processor 10 generates (calculates) features. Specifically, the processor 10 generates knowledge graph features. If modifications have been made (via step S50), the processor 10 also generates (calculates) knowledge graph features of the modified knowledge graph and modified features. 【0083】 Specifically, the processor 10 inputs the knowledge graph into a predetermined learning model such as a neural network and generates knowledge graph features in vector embedding format (vector embedding format) based on named entities and the relationships between named entities contained in the knowledge graph. 【0084】 Figure 15A shows an example of knowledge graph features. The processor 10 obtains knowledge graph embedding vectors, as shown in the figure, from a predetermined learning model as knowledge graph features. 【0085】 Furthermore, the processor 10 generates modified features of the knowledge graph using a predetermined arithmetic formula. 【0086】 Figure 15B shows an example of each element of the calculation formula used to generate the corrected features. Specifically, the processor 10 calculates the correction burden as a corrected feature based on the following formula (1). The correction burden is a value that represents the degree of correction burden and is expressed as a floating-point number between 0 and 1. γ is a positive floating-point number and is a weighting value that determines whether to give more weight to the NER or RE F-value (F-Measure). 【0087】 【number】 【0088】 Here, F NER And, F RE F is an evaluation metric for binary classification tasks. NER And, F RE This can be calculated using the following equations (2) and (3). 【0089】 【number】 【0090】 【number】 【0091】 Here, β is a positive floating-point number and is a weighting value that determines whether to prioritize recall or precision. NER The named entity recall rate is the number of correctly extracted named entities divided by the number of corrected named entities, and can be calculated using the following formula (4). 【0092】 【number】 【0093】 Also, Precision NERThe named entity precision is the number of correctly extracted named entities divided by the number of uncorrected named entities, and can be calculated using the following formula (5). 【0094】 【number】 【0095】 Also, Recall RE The (relationship recall rate) is the number of relationships correctly extracted / the number of relationships after correction, and can be calculated using the following formula (6). 【0096】 【number】 【0097】 Also, Precision RE The (relationship precision) is the number of relationships correctly extracted / the number of relationships before correction, and can be calculated using the following formula (7). 【0098】 【number】 【0099】 The number of named entities before correction refers to the number of named entities in the knowledge graph before correction. The number of named entities after correction refers to the number of named entities in the knowledge graph after correction. The number of relationships before correction refers to the number of relationships in the knowledge graph before correction. The number of relationships after correction refers to the number of relationships in the knowledge graph after correction. The number of correctly extracted named entities refers to the number of named entities that remain unchanged before and after correction. The number of correctly extracted relationships refers to the number of relationships that remain unchanged before and after correction. 【0100】 For example, in the case of correcting the knowledge graph illustrated in Figure 14, the number of added named entities is 1 (addition of "damaged"), the number of corrected errors in named entities is 3 (for example, corrections to "diesel" and "engine", correction of "cooling pump" from status to component: 3 in total), and the recall rate indicates the named entity recall rate.NER is 3 / 4, which is the precision indicating the fitting rate of idiomatic expressions NER is 1 / 4. The number of additional relationships is 2 (adding "The cooling pump is damaged" and "The damage causes automatic stop": total 2), the number of relationship error corrections is 1 (correcting "The cooling pump causes automatic stop": total 1), and recall indicating the relationship reproduction rate RE is 1 / 3, which is the precision indicating the relationship fitting rate RE is 1 / 2, and the correction burden is 0.6125 (when β = 1 and γ = 1). These are obtained by the above formulas (1) to (7). 【0101】 Next, the processor 10 updates the defect knowledge DB120 (step S70). Specifically, the processor 10 stores the generated knowledge graph (if the knowledge graph is modified in step S50, the modified knowledge graph), the calculated knowledge graph feature amount, and the correction feature amount in the defect knowledge DB120 and updates the database. Also, when the processor 10 updates the defect knowledge DB120, it ends this flow. 【0102】 The above is the description of the knowledge DB update process. 【0103】 According to such a processor system, since the pattern of past corrections can be reflected in the generation of the knowledge graph in subsequent times, it is possible to generate a highly accurate knowledge graph from the defect report text. As a result, the man-hours for manually correcting the knowledge graph can be minimized. 【0104】 Also, according to the processor system, since the knowledge graph feature amount is generated and registered in the database in association with the knowledge graph, when generating the knowledge graph in subsequent times, an appropriate reference example can be retrieved from the database. 【0105】 Furthermore, according to the processor system, if the generated knowledge graph is modified, the modified features are registered in the database. This allows the system to select reference cases with more modifications from among reference cases similar to the problem being processed when generating the knowledge graph. High-accuracy knowledge graphs usually have more detailed and numerous modifications, often made manually. Therefore, by generating a knowledge graph using similar reference cases with large modified features, it is possible to generate a high-accuracy knowledge graph that reflects a wider range of modification patterns. 【0106】 Next, the user interface in the processor system 100 will be described. The user interface is screen information generated by the processor 10 using various information in the memory resources 20 and information generated by the processor system 100, and is displayed on an output device connected via the UI 40. The user interface may also be displayed on an output device of an external device 400 connected via the network N, for example. 【0107】 Figure 16 shows an example of a user interface screen. As shown in the figure, the user interface screen 500 has a bug report input area 501, a reference case display area 502, a knowledge graph display area 503, and a feature quantity display area 504. 【0108】 The bug report input area 501 is an area for receiving input of bug reports to be processed. Specifically, the bug report input area 501 displays multiple bug reports stored in the bug report DB 130, and the user selects the bug report to be processed. When a bug report is selected or the "Execute Processing" button is pressed, the processor 10 executes the knowledge DB update process. 【0109】 The reference case display area 502 is an area where knowledge information of reference cases extracted from the defect knowledge DB 120 based on the features of the defect report text (for example, one or more of the following: case knowledge graph, case knowledge graph features, and correction features) is displayed. 【0110】 Knowledge graph display area 503 is the area for displaying knowledge graphs generated based on reference examples. If the generated knowledge graph is modified, both the pre-modification and post-modification knowledge graphs may be displayed. In this case, the modified sections may be highlighted or bolded. 【0111】 The feature display area 504 is an area that displays at least one of the knowledge graph features and the modified features. Specifically, the feature display area 504 displays the knowledge graph features in vector embedding format and the modified feature called the modification burden. 【0112】 In the illustrated example, each area is included in one screen, but this is not the only option. The processor 10 may display each area on a separate screen, or it may display a predetermined combination of areas (for example, a combination of the bug report input area 501 and the knowledge graph display area 503) on one screen. 【0113】 <Second Embodiment> In the first embodiment described above, the processor 10 performed correction processing (step S50) on named entities and the relationships between named entities based on basic data. However, the processor system 100 according to the second embodiment performs correction of named entities and the relationships between named entities based on user operations when it receives a correction instruction from the user, asynchronously with the knowledge DB update process. 【0114】 Figure 17 is a flowchart showing an example of the modification process according to the second embodiment. Figure 18 is an overall diagram illustrating the knowledge database update process, including the data flow. As shown in the figure, processor 10 executes the correction process when it receives a correction instruction from the user for the knowledge graph, and corrects the knowledge graph based on the user's operation regarding the correction. Processor 10 also generates the feature quantities and correction feature quantities of the corrected knowledge graph and registers them together with the corrected knowledge graph in the defect knowledge database 120. 【0115】 Specifically, in step S51, the processor 10 accepts user modifications to named entities in the knowledge graph and the relationships between named entities via the user interface. 【0116】 Figure 19 shows an example of a user interface screen 600. As shown in the figure, the user interface screen has a knowledge graph search area 601, a search results display area 602, a named entity modification area 603, and a relationship modification area 604. 【0117】 The knowledge graph search area 601 is an area that accepts input of information for searching the knowledge graph to be corrected (in this case, the case knowledge graph, as it is registered in the defect knowledge DB 120; hereafter, it may be referred to as the "knowledge graph to be corrected") from the defect knowledge DB 120. The user enters an identification number or name that identifies the knowledge graph to be corrected via an input device connected to the UI 40. At this time, the processor 10 searches the defect knowledge DB 120 based on the input information, extracts the relevant knowledge graph, and displays it in the search results display area 602. The processor 10 also extracts the corresponding defect report document along with the knowledge graph to be corrected. 【0118】 As mentioned above, the search results display area 602 is the area that displays the knowledge graph to be corrected. 【0119】 The named entity correction area 603 is an area that displays bug reports (case bug reports) corresponding to the searched knowledge graph to be corrected and accepts corrections of named entities. The processor 10 displays the bug reports corresponding to the knowledge graph to be corrected in area 603. At this time, the processor 10 generates a named entity list (exemplified in Figure 11) from the knowledge graph to be corrected, identifies the named entities contained in the bug reports using this list, and highlights them using bold or highlighting. The processor 10 may also attach attributes such as component or status to the displayed named entities. 【0120】 Furthermore, the user can modify displayed named entities by manipulating the named entity correction area 603. Specifically, the user selects a named entity that is highlighted or otherwise displayed and corrects it to the correct named entity. When the processor 10 receives a correction based on the user's operation, it reflects it in the knowledge graph of the entity being corrected. 【0121】 Furthermore, the relationship modification area 604 is an area that displays the knowledge graph to be modified and accepts modifications to the relationships between nodes. Specifically, the user selects an edge (arrow line) connecting nodes and performs operations such as changing the start or end point or adding a new edge between nodes. When the processor 10 accepts a modification based on the user's operation, it reflects it in the knowledge graph to be modified. 【0122】 After receiving such corrections, processor 10 generates knowledge graph features and corrected features for the corrected knowledge graph. Processor 10 also registers the corrected knowledge graph, knowledge graph features, and corrected features in the faulty knowledge DB 120 and terminates the correction process. Note that the processes in steps S61 and S71 correspond to steps S60 and S70 described above, so a detailed explanation is omitted. 【0123】 The processor system 100 of the second embodiment has been described above. 【0124】 Such a processor system allows for the incorporation of past correction patterns into the generation of subsequent knowledge graphs, making it possible to generate highly accurate knowledge graphs from bug reports. 【0125】 In particular, the processor system accepts user modifications and updates the knowledge graph accordingly, allowing it to accumulate a more accurate knowledge graph as case knowledge information. Furthermore, the knowledge graph being modified is highly accurate because it reflects past modification patterns for reference cases during its generation stage, resulting in users being able to make modifications with minimal burden. 【0126】 It should be noted that the present invention is not limited to the embodiments and modifications described above, and various modifications are included within the scope of the same technical idea. For example, the embodiments described above are described in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace parts of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add configurations from other embodiments to the configuration of one embodiment. In addition, it is possible to add, delete, or replace parts of the configuration of each embodiment with other configurations. 【0127】 Furthermore, the control lines and information lines shown above are those deemed necessary for the explanation, and do not necessarily represent all control lines and information lines present in the actual product. In reality, it is safe to assume that almost all components are interconnected. [Explanation of Symbols] 【0128】 100...Processor system, 10...Processors, 20...Memory resources, 30...NI (Network Interface Device), 40...UI (User Interface Device), 110...Language processing model, 120...Trouble knowledge database, 130...Trouble report database, 210...Knowledge database update program, 300...Input / output devices, 400...External devices, N...Network

Claims

[Claim 1] A processor system having one or more processors and one or more memory resources, The aforementioned memory resources are It stores a defect knowledge database containing case knowledge information including corrected knowledge graphs related to product or component defects, a language processing model, a target defect report, and a program that generates a knowledge graph related to the target defect report. The processor executes the program, The case knowledge information, which contains case knowledge similar to the defect indicated in the aforementioned defect report, is extracted from the defect knowledge database. By inputting the target defect report and the extracted case knowledge information into the language processing model, the named entities of the target defect report and the relationships between named entities are identified, and a knowledge graph related to the target defect report is generated based on the identified named entities and relationships. A processor system characterized by the following features. [Claim 2] A processor system according to claim 1, The aforementioned processor, In accordance with predetermined constraints, modify at least one of the named entity and the relationship of the knowledge graph relating to the target defect report. A processor system characterized by the following features. [Claim 3] A processor system according to claim 2, The aforementioned processor, Knowledge graph features, which are the features of the knowledge graph related to the aforementioned defect report document, and correction features, which are the features related to the correction, are generated. The generated knowledge graph, the knowledge graph features, and the modified features are stored in the defect knowledge database as case knowledge information. A processor system characterized by the following features. [Claim 4] A processor system according to claim 1, The aforementioned case knowledge information includes a knowledge graph, knowledge graph features, and modified knowledge graph features. The aforementioned processor, The case knowledge information that has a higher similarity to the features of the aforementioned defect report and has a larger modified feature is extracted from the defect knowledge database as case knowledge information to be used to generate a knowledge graph related to the aforementioned defect report. A processor system characterized by the following features. [Claim 5] A processor system according to claim 1, The aforementioned case knowledge information includes a knowledge graph and a case defect report corresponding to the knowledge graph. The aforementioned processor, Based on the knowledge graph of the extracted case knowledge information and the case defect report text, instructions are input to the language processing model to extract named entities related to defects from the target defect report text. A processor system characterized by the following features. [Claim 6] A processor system according to claim 5, The aforementioned processor, Based on the knowledge graph of the extracted case knowledge information and the case defect report text, instructions are input to the language processing model to extract the relationships between named entities using the target defect report text and the named entities extracted from the target defect report text. A processor system characterized by the following features. [Claim 7] A processor system according to claim 3, The aforementioned processor, Using a predetermined learning model, the knowledge graph features are generated, which represent named entities and the relationships between named entities included in the knowledge graph related to the target defect report in vector embedding format. A processor system characterized by the following features. [Claim 8] A processor system according to claim 3, The aforementioned processor, By comparing the knowledge graph before and after correction for the aforementioned bug report, a correction feature is generated that indicates the degree of correction burden, based on the number of named entities, the number of relationships, the number of added named entities and relationships, and the number of errors corrected. A processor system characterized by the following features. [Claim 9] A processor system according to claim 1, The aforementioned case knowledge information includes a knowledge graph and a case defect report corresponding to the knowledge graph. The aforementioned processor, Based on user instructions, a search is performed within the aforementioned defect knowledge database. The system accepts user requests for corrections regarding at least one of the named entities and relationships between named entities identified from the knowledge graph of the retrieved case knowledge information and the case defect report text. A processor system characterized by the following features. [Claim 10] A processor system according to claim 1, The aforementioned processor, The system generates screen information for displaying at least one of the following areas: an area for receiving input of the target defect report; an area for displaying the case knowledge information extracted from the defect knowledge database; an area for displaying the generated knowledge graph; and an area for displaying at least one of the feature quantities of the generated knowledge graph and the modified feature quantities of the knowledge graph. A processor system characterized by the following features. [Claim 11] A processor system according to claim 1, The aforementioned case knowledge information includes a knowledge graph and a case defect report corresponding to the knowledge graph. The aforementioned processor, The system generates screen information to display at least one of the following areas: an area for receiving input information that identifies the aforementioned case knowledge information to be corrected; an area for displaying the aforementioned case knowledge information extracted from the defect knowledge database based on the received input information; an area for displaying named entities in the aforementioned case defect report text included in the aforementioned case knowledge information and accepting correction operations; and an area for displaying the aforementioned knowledge graph included in the aforementioned case knowledge information and accepting correction operations for the relationships between named entities. A processor system characterized by the following features. [Claim 12] A method for generating a knowledge graph performed by a processor system having one or more processors and one or more memory resources, The aforementioned memory resources are It stores a defect knowledge database containing case knowledge information including corrected knowledge graphs related to product or component defects, a language processing model, a target defect report, and a program that generates a knowledge graph related to the target defect report. The processor executes the program, The steps include: extracting the case knowledge information from the defect knowledge database, which contains case knowledge similar to the defect indicated in the aforementioned defect report; The process involves inputting the target defect report and the extracted case knowledge information into the language processing model to identify named entities in the target defect report and the relationships between named entities, and generating a knowledge graph related to the target defect report based on the identified named entities and relationships. A method for generating a knowledge graph characterized by the following features. [Claim 13] A program that runs on a processor system having one or more processors and one or more memory resources, The aforementioned memory resources are It stores a defect knowledge database containing case knowledge information, including corrected knowledge graphs related to product or component defects, a language processing model, and the target defect report text. The aforementioned processor, The case knowledge information, which contains case knowledge similar to the defect indicated in the aforementioned defect report, is extracted from the defect knowledge database. By inputting the target defect report and the extracted case knowledge information into the language processing model, named entities in the target defect report and the relationships between named entities are identified, and a knowledge graph related to the target defect report is generated based on the identified named entities and the relationships. A program to perform the following process.

Images