Generation method, generation program, and information processing device.

Abstract

The present invention provides a generation method, a generation program, and an information processing device that improve the performance of answering questions. [Solution] The information processing device selects a first entity from a real failure knowledge graph 101 which includes a plurality of first entities that represent events linked by cause-and-effect relationships based on failure case document 103, generates a virtual failure knowledge graph 200 which includes second entities that represent events linked by cause-and-effect relationships based on non-failure case document 102 that the first entity is related to, and combines the real failure knowledge graph 101 and the virtual failure knowledge graph 200 to generate a combined knowledge graph 300.

Description

Technical Field

[0001] The present invention relates to a generation method, a generation program, and an information processing apparatus.

Background Art

[0002] In recent years, in IT systems (Information Technology), due to virtualization and multi-vendor support, the configuration has become complex, and the causes of failures have diversified. For example, in an IT system having multi-vendor RUs (Radio Units) and DUs (Distributed Units), a failure may occur when a new vendor's RU is added and connected to an existing DU. In the failure cause analysis in such a case, various classifications and analyses may be performed. For example, in the failure cause analysis, various failure classifications such as hardware failures, compatibility issues between RUs and DUs, compatibility issues between RU switches and DU switches, and ID (Identifier) setting mistakes are assumed, and each is examined.

[0003] However, relying on the experience and knowledge of analysts to perform such complex failure cause analysis consumes a lot of time, and it is unclear whether accurate analysis can be performed, and it may be difficult to achieve appropriate failure cause analysis. Therefore, for the stable operation of IT systems and networks, the development of failure cause analysis technology that can speed up failure recovery is important. For example, in large-scale networks, etc., both identifying the failure location and analyzing the failure cause are important, but here, the failure cause analysis is mainly considered.

[0004] For example, as a failure cause analysis technology, a technology has been proposed that uses an LLM (Large Language Model) to answer the failure cause from a failure case document such as a failure report or a non-failure case document such as a specification document. In this technology, the cause of the failure is estimated based on a given specific document for the input failure, and the estimation result is answered.

[0005] Furthermore, the following techniques exist for troubleshooting. For example, a technique has been proposed to generate an integrated graph of the first and second knowledge graphs when the semantic similarity between the knowledge items in the first knowledge graph, which contains troubleshooting-related graph information, and the knowledge items in the second knowledge graph, which is based on knowledge from a different domain, satisfies certain conditions. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2024-005871 [Non-patent literature]

[0007] [Non-Patent Document 1] Y. Chen et.al., “Automatic Root Cause Analysis via Large Language Models for Cloud Incidents,” arXiv:2305.15778, 2023. [Overview of the project] [Problems that the invention aims to solve]

[0008] However, conventional technologies that use LLM to generate answers based on specific documents may have low answer performance because they rely on a single document. For example, a failure case document often describes a single failure that actually occurred, making it difficult for LLM to answer about unoccurred failures based on a single failure case document. Furthermore, non-failure case documents such as specifications are not documents intended to describe failure cases, so the causal relationship of failures is often not clearly indicated. Therefore, the amount of information about the causal relationship of failures that can be obtained by referring to a non-failure case document alone is small, making it difficult for LLM to answer about unoccurred failures based on a single non-failure case document. Thus, conventional technologies that use LLM to generate answers based on specific documents have difficulty answering about the causal relationship of failures that can only be recognized by comprehensively referring to multiple documents, making it difficult to improve answer performance.

[0009] Furthermore, the technique for generating an integrated graph when semantic similarity meets certain conditions generates an integrated graph from knowledge graphs obtained from independent documents, and it is unclear whether both documents contain information about the same disorder. Moreover, attempting to extract words related to the disorder from documents other than those that mention the disorder may result in many omissions and may not yield appropriate results. Additionally, this technique does not generate an integrated graph to obtain new disorder causal relationships for a specific disorder. Therefore, even using this technique, it is difficult to obtain new disorder causal relationships for a specific disorder, and thus difficult to improve response performance.

[0010] The disclosed technology was developed in view of the above, and aims to provide a generation method, generation program, and information processing device that improve the performance of answering questions. [Means for solving the problem]

[0011] In one embodiment of the generation method, generation program, and information processing device disclosed herein, the information processing device selects a first entity from a first knowledge graph which includes a plurality of first entities representing events linked by cause-and-effect relationships based on a first input document, generates a second knowledge graph which includes second entities representing events linked by cause-and-effect relationships based on a second input document in which the first entity is related, based on the selected first entity and a second input document, and executes a process to generate a third knowledge graph by combining the first knowledge graph and the second knowledge graph. [Effects of the Invention]

[0012] In one respect, the present invention can improve the performance of answering questions. [Brief explanation of the drawing]

[0013] [Figure 1] Figure 1 is a block diagram of the information processing device according to Example 1. [Figure 2] Figure 2 shows an example of a disability case document and an actual disability knowledge graph. [Figure 3] Figure 3 shows an example of data representation in the actual failure knowledge graph. [Figure 4] Figure 4 shows an example of a non-fault case document and a virtual fault knowledge graph. [Figure 5] Figure 5 shows an overview of the processing performed by the information processing device. [Figure 6] Figure 6 is a diagram illustrating an example of prompt options. [Figure 7] Figure 7 shows an example of how to combine the actual failure knowledge graph and the virtual failure knowledge graph. [Figure 8] Figure 8 shows a concrete example of how the actual failure knowledge graph and the virtual failure knowledge graph are combined. [Figure 9] Figure 9 shows the coupling pattern when both the source and destination entities are faulty entities. [Figure 10] FIG. 10 is a diagram showing a coupling pattern when both the coupling source and the coupling destination are failure cause entities. [Figure 11] FIG. 11 is a diagram showing a coupling pattern when the entity of the coupling source is an intermediate event entity and the entity of the coupling destination is a failure cause entity or a failure entity. [Figure 12] FIG. 12 is a flowchart of the coupling knowledge graph generation process by the information processing apparatus according to Embodiment 1. [Figure 13] FIG. 13 is a diagram for explaining the effect by the generation of the coupling knowledge graph. [Figure 14] FIG. 14 is a block diagram of the information processing apparatus according to Embodiment 2. [Figure 15] FIG. 15 is a diagram showing an example of information held by the information processing apparatus according to Embodiment 2. [Figure 16] FIG. 16 is a diagram showing the generation of a virtual failure knowledge graph by the information processing apparatus according to Embodiment 2. [Figure 17] FIG. 17 is a flowchart of the coupling knowledge graph generation process by the information processing apparatus according to Embodiment 3. [Figure 18] FIG. 18 is a hardware configuration diagram of the information processing apparatus.

BEST MODE FOR CARRYING OUT THE INVENTION

[0014] Hereinafter, embodiments of the generation method, generation program, and information processing apparatus disclosed in the present application will be described in detail based on the drawings. Note that the generation method, generation program, and information processing apparatus disclosed in the present application are not limited by the following embodiments.

Embodiment

[0015] Figure 1 is a block diagram of the information processing device according to Embodiment 1. The information processing device 1 acquires the details of an actual failure from the user terminal device 2, estimates the cause of the failure, and provides a response. As shown in Figure 1, the information processing device 1 includes a data storage unit 10, a target selection unit 11, a failure causal explanation text generation unit 12, a virtual failure knowledge graph generation unit 13, a coupling unit 14, and an LLM 15.

[0016] The data storage unit 10 pre-stores, for example, one or more actual failure knowledge graphs 101 and one or more non-fault case documents 102 that have been input from an external device.

[0017] The actual failure knowledge graph 101 is a knowledge graph generated from the failure case document 103. Figure 1 shows that the actual failure knowledge graph 101 is generated from the failure case document 103 by extending an arrow from the external failure case document 103 to the actual failure knowledge graph 101. The failure case document 103 is a report-style document such as a failure response report that describes the details of an actual failure that occurred and its cause.

[0018] Here, a knowledge graph is a graph that shows a knowledge network in which events related to an actual failure (an actual failure) are continuously linked by causal relationships that connect each event. An actual failure is also included as one event. Each event is represented as an entity in the knowledge graph. Furthermore, the causal relationships between events are represented as edges. The knowledge graph is represented in a form in which entities are connected by edges to show their mutual relationships. In the actual failure knowledge graph 101 generated from failure case document 103, the entity that serves as the starting point of the causal relationship represents the event that is the root cause of the failure, and the entity that serves as the ending point of the causal relationship represents the event that is the failure that ultimately occurs. Hereafter, the failure that ultimately occurs will be referred to as the "occurring failure".

[0019] Figure 2 shows an example of a failure case document and an actual failure knowledge graph. Failure case report 110 corresponds to an example of failure case document 103. The failure details described in failure case report 110 represent the actual failure. The event "User login impossible" described in this failure details is the event shown by entity 113, which corresponds to the actual failure in the actual failure knowledge graph 101. Furthermore, the event "Increased load on the authentication server" described in the details of the failure and the cause analysis is the event shown by entity 111, which corresponds to the cause of the failure in the actual failure knowledge graph 101. In addition, from the details of the failure and the cause analysis, network delay is identified as an event that occurs between the increased load on the authentication server, which is the cause of the failure, and the user login impossible, which is the resulting failure. Network delay is an event that exists in the causal relationship from the cause of the failure to the resulting failure, and is shown by entity 112. Hereafter, events that exist in the causal relationship from the cause of the failure to the resulting failure will be called intermediate events.

[0020] Entity 111 and Entity 113 are connected by a causal relationship via Entity 112, which represents an intermediate event. Therefore, in the actual failure knowledge graph 101, Entity 111 and Entity 112 are connected by an edge 114 that indicates a causal relationship, and Entity 112 and Entity 113 are connected by an edge 115 that indicates a causal relationship. Here, edges 114 and 115 are indicated by arrows pointing from cause to effect. Note that an event treated as a failure cause, intermediate event, or resulting failure in the actual failure knowledge graph 101 of a particular failure case document 103 may be treated as a different type in the actual failure knowledge graph 101 of other failure case documents 103.

[0021] Here, Figure 2 illustrates a case where the failure cause, intermediate events, and resulting failure are represented in a single row in the actual failure knowledge graph 101. However, in the actual failure knowledge graph 101, there may be multiple failure causes and intermediate events for a single resulting failure, and the edge connecting the resulting failure to the failure cause may branch at any intermediate event depending on the causal relationship.

[0022] In the following, entities that indicate the cause of a failure, such as entity 111 in the actual failure knowledge graph 101, will be called failure cause entities. Entities that indicate intermediate events, such as entity 112 in the actual failure knowledge graph 101, will be called intermediate event entities. Entities that indicate the occurrence of a failure, such as entity 113 in the actual failure knowledge graph 101, will be called failure entities. Furthermore, when failure cause entities, intermediate event entities, and failure entities are not distinguished from each other, they will simply be referred to as entities.

[0023] The actual failure knowledge graph 101 may be created manually from the failure case document 103, or it may be created automatically from the failure case document 103 using an external LLM or the like. In this embodiment, the actual failure knowledge graph 101 is described as pre-created information, but the information processing device 1 may also generate the actual failure knowledge graph 101 from the failure case document 103 using the LLM 15. Furthermore, one actual failure knowledge graph 101 may be generated from one failure case document 103, or multiple graphs may be generated.

[0024] Here, the actual failure knowledge graph 101 is conceptually represented as the graph in Figure 2, but the information processing device 1 actually grasps the information of the actual failure knowledge graph 101 through a data representation of combinations of entities and edges. For example, the data storage unit 10 holds a list of data that represents the combinations of entities and edges that make up the actual failure knowledge graph 101.

[0025] Figure 3 shows an example of data representation in a real-world fault knowledge graph. For example, let's describe a real-world fault knowledge graph 101 having entities 121-124 and edges 125-127. A-D described in entities 121-124 represent the events that each entity represents.

[0026] The data storage unit 10 holds an edge list 130 as data representing the actual failure knowledge graph 101 in Figure 3. The edge list 130 is a list of edge representations that show combinations of entities and edges. The edge representation (A,C) indicates that event A and event C have a causal relationship in which event A is the cause and event C is the result. That is, the edge representation (A,C) indicates that entity 121 representing event A and entity 123 representing event C are connected by an edge 125 that goes from entity 121 to entity 123. The edge representations (B,C) and (B,C) in the edge list 130 similarly indicate the relationship between each entity and edge. There are no restrictions on the order in which the edge representations are arranged in the edge list 130. In addition, events A to D themselves may be described, or identification information may be assigned to each event and this information may be stored, and the identification information may be used to represent each of them as events A to D.

[0027] Let's return to Figure 1 and continue the explanation. Non-fault case document 102 is a general document, such as a specification or technical paper. However, as explained below, in order to extract the content related to the target failure from the non-fault case document 102, it is preferable that the non-fault case document 102 is a document that includes descriptions related to the actual failure described in the failure case document 103.

[0028] Figure 4 shows an example of a non-fault case document and a virtual fault knowledge graph. The system operation management specification 140 shown in Figure 4 is an example of a non-fault case document 102. The system operation management specification 140 is a document that contains information about events that occur during system operation, and is highly likely to be a document that contains information about the target fault. The virtual fault knowledge graph 200 is data for generating a knowledge graph related to the actual fault knowledge graph 101, and is a knowledge graph generated from the non-fault case document 102, but the details will be explained later.

[0029] To make the following explanation easier to understand, we will briefly describe the overview of the processing performed by the information processing device 1 according to Example 1, referring to Figure 5. Figure 5 is a diagram showing the overview of the processing performed by the information processing device.

[0030] As described above, the information processing device 1 holds an actual failure knowledge graph 101 generated from the failure case document 103. The information processing device 1 also generates a virtual failure knowledge graph 200 from the contents of the non-failure case document 102, showing the causal relationship between the failure cause and the resulting failure. Then, the information processing device 1 combines the actual failure knowledge graph 101 and the virtual failure knowledge graph 200 to generate a new combined knowledge graph 300. This new combined knowledge graph 300 includes, in addition to the actual failure knowledge graph 101, information on the causal relationship between the resulting failure and other failure causes obtained from the non-failure case document 102. Therefore, using this combined knowledge graph 300, the information processing device 1 can estimate the cause of a failure in response to user inquiries about failures, taking into account the failure case document 103 and the non-failure case document 102, thereby improving response performance. The details of the processing by the information processing device 1 are described below.

[0031] Returning to Figure 1, the explanation continues. When the data storage unit 10 holds multiple actual failure knowledge graphs 101, the target selection unit 11 performs the following processing for each actual failure knowledge graph 101.

[0032] The target selection unit 11 selects target entities one by one from the entities included in the actual failure knowledge graph 101 to be used as the source for joining. The source for joining is the entity that will be the joining point in the actual failure knowledge graph 101 when creating a combined knowledge graph 300 by combining the actual failure knowledge graph 101 and the virtual failure knowledge graph 200 created from the non-failure case document 102. Subsequently, the target selection unit 11 outputs the information of the selected target entities to the failure causal explanation statement generation unit 12.

[0033] Here, the target selection unit 11 may select target entities one by one from all entities included in the actual failure knowledge graph 101, or it may limit the types of entities and select one by one from within that limit. For example, by using an edge list, the target selection unit 11 can select target entities by type as follows.

[0034] Here, we will use the edge list 130 in Figure 3 for explanation. The target selection unit 11 can select a fault-causing entity as a target entity by selecting entities that appear only on the left side of the edge representations included in the edge list 130. For example, in the edge list 130, events A and B are events that appear only on the left side of the edge representations. Therefore, when the target selection unit 11 selects a fault-causing entity from the edge list 130, it selects event A and event B as target entities.

[0035] Furthermore, the target selection unit 11 can select a fault entity as a target entity by selecting an entity that appears only to the right of the edge representation included in the edge list 130. For example, in the edge list 130, the event that appears only to the right of the edge representation is event D. Therefore, when the target selection unit 11 selects a fault entity from the edge list 130, it selects event D as the target entity.

[0036] Furthermore, the target selection unit 11 can select intermediate event entities as target entities by selecting entities that appear on both sides of an edge representation included in the edge list 130. For example, in the edge list 130, the event that appears on both sides of an edge representation is event C. Therefore, when the target selection unit 11 selects an intermediate event entity from the edge list 130, it selects event C as the target entity. Here, the selection method is shown for each type, but by combining these selection methods, the target selection unit 11 can select target entities across multiple types.

[0037] This failure case document 103 is an example of a "first input document," the actual failure knowledge graph 101 is an example of a "first knowledge graph," and the entities included in the actual failure knowledge graph 101 are examples of "first entities." In other words, the target selection unit 11 selects a first entity from a first knowledge graph that includes multiple first entities representing events linked by cause-and-effect relationships based on the first input document. Alternatively, the target selection unit 11 may select a single entity included in the first knowledge graph.

[0038] Furthermore, the first entity includes a fault entity that shows the event of the actual fault occurring, a fault cause entity that shows the event that causes the fault, and an intermediate event entity that shows the event that occurs between the actual fault and the fault cause.The target selection unit 11 can then select one or a combination thereof of a fault entity, a fault cause entity, or an intermediate event entity from among the multiple first entities.

[0039] Returning to Figure 1, let's continue the explanation. The failure causal explanation unit 12 receives information about the target entity from the target selection unit 11. The failure causal explanation unit 12 then generates a failure causal explanation that explains the causal relationship between the cause of the failure and the resulting failure for the target entity, using the non-failure case document 102.

[0040] For example, the failure causal explanation generation unit 12 generates a prompt to generate a failure causal explanation for the target entity based on the non-failure case document 102, depending on whether the target entity is the cause of the failure or the target entity is the resulting failure. The failure causal explanation generation unit 12 then inputs the generated prompt to the LLM 15. Here, the LLM 15 receives the input prompt, which includes information about the target entity, instructions to generate a failure causal explanation for that target entity, and the non-failure case document 102, and can output a failure causal explanation that explains the causal relationship between the cause of the failure and the resulting failure. The LLM 15 will be described in detail later. The failure causal explanation generation unit 12 retrieves the failure causal explanation output from the LLM 15 according to the input prompt.

[0041] When generating a causal explanation for a failure where the target entity is the cause of the failure, the causal explanation generation unit 12 generates a prompt containing the following information. For example, the prompt may include a question such as, "If there is a description in the document below that allows us to estimate the failure caused by the following failure cause, please answer with a sentence that explains the mechanism of failure occurrence." The prompt may also include an example answer such as, "For example, please answer with, 'Event B occurs due to this failure cause, and as a result, failure C occurs.'" The prompt may also include information to specify the sentence that refers to the "Description of a Non-Failure Case Document." The prompt may also include information to specify the "Description of the Target Entity" as the cause of the failure.

[0042] Furthermore, when generating a causal explanation statement for a failure in the target entity, the failure causal explanation statement generation unit 12 generates a prompt containing the following information. For example, the prompt may include a question such as, "If there is a description in the document below that allows for the estimation of the cause of the failure below, please answer with a sentence that explains the mechanism of failure occurrence." The prompt may also include an example answer such as, "For example, please answer with, 'Event A causes Event B, and as a result, this failure occurs.'" The prompt may also include information that specifies the sentence to refer to in the "Description of a Non-Failure Case Document." The prompt may also include information that specifies the "Description of the Target Entity" as a failure.

[0043] Here, the causal explanation generation unit 12 may add a sentence to the prompt to exclude unreliable causal explanations. For example, the causal explanation generation unit 12 may add a sentence to the prompt such as, "Please answer with your confidence level regarding the answer. Please answer with a number between 0 and 100. A higher number means a higher confidence level." Then, if the confidence level included in the answer is below a threshold, the causal explanation generation unit 12 may regenerate the causal explanation using, for example, LLM 15. In this case, the causal explanation generation unit 12 may change the description of the prompt.

[0044] Furthermore, if the confidence level does not exceed a threshold even after recreating the causal explanation statement for a predetermined number of times, the causal explanation statement generation unit 12 may skip creating the causal explanation statement for that condition. This allows the causal explanation statement generation unit 12 to improve the reliability of the causal explanation statement for the failure and to improve the accuracy of the final estimation of the cause of the failure.

[0045] Furthermore, if there are multiple non-fault case documents 102 or if the amount of text is large, one document may be split and a fault causal explanation document may be generated for each. In such cases, generating an explanation document for each split document may result in duplicate content. Therefore, the fault causal explanation document generation unit 12 may add a sentence to the prompt to exclude duplicate explanations from the fault causal explanation document. For example, when generating a fault causal explanation document in which the target entity is the cause of the fault, the fault causal explanation document generation unit 12 may add a sentence to the prompt such as, "Also, please include keywords that concisely indicate the estimated fault in your response." Similarly, when generating a fault causal explanation document in which the target entity is the actual fault, the fault causal explanation document generation unit 12 may add a sentence to the prompt such as, "Also, please include keywords that concisely indicate the estimated cause of the fault in your response."

[0046] The causal explanation unit 12 can then select one of the causal explanations of the disorder that contains overlapping keywords. In this case, the causal explanation unit 12 may limit keyword overlap to cases where the keywords are identical, or it may include cases where the keywords are similar. This allows the causal explanation unit 12 to reduce unnecessary causal explanations of the disorder, thereby reducing the following processing in the information processing device 1.

[0047] In addition, for example, when the failure causal explanation generation unit 12 generates a failure causality in which the target entity is the failure, it may add a description related to the target entity to the prompt in order to improve the response performance of the LLM 15. Figure 6 is a diagram illustrating an example of prompt options. In this case, as shown in Figure 6, the data storage unit 10 holds the failure case document 103.

[0048] The failure causal explanation generation unit 12 extracts descriptions of target entities, such as logs, from the failure case document 103 using the LLM 15. It is preferable to include the following information in the prompt for extracting descriptions of target entities. For example, the prompt may include a question such as, "Please provide a sentence that explains the following failure and the events observed in relation to the following failure, based on the document below." The prompt may also include an example answer such as, "For example, please answer as follows: 'Failure B occurred. The message Y was output to the log of X.'" The prompt may also include a condition such as, "Please do not include a description of the cause of this failure." The prompt may also include a condition such as, "Please do not include a description of failures other than this failure." However, this condition is optional and is an instruction to prevent the inclusion of descriptions of failures other than those of the target entity when the failure case document 103 contains multiple failures, or when the failure knowledge graph can contain multiple failures. The prompt may also include information that specifies a sentence that refers to "Description of the failure case document." The prompt also includes information specifying the "description of the target entity" as the fault.

[0049] The failure causal explanation unit 12 then specifies the description extracted using this prompt as a failure in the prompt for generating the failure causal explanation. In this case, it is preferable to include the following information in the prompt for generating the failure causal explanation. For example, the prompt may include a question such as, "If there is a description in the document below that allows for the estimation of the cause of the failure below, please answer with a sentence that explains the mechanism of failure occurrence." The prompt may also include an example answer such as, "For example, please answer as, 'Event A causes Event B, and as a result, this failure occurs.'" The prompt may also include information that specifies a sentence that refers to the "description in the non-failure case document." The prompt may also include information that specifies the "description concerning the target entity" obtained by the prompt for extracting a description concerning the target entity as a failure.

[0050] In this way, the causal explanation document generation unit 12 generates a causal explanation document from the second input document that explains the events related to the first entity in a cause-and-effect relationship. Furthermore, the causal explanation document generation unit 12 generates the causal explanation document using a large-scale language model. Specifically, the causal explanation document generation unit 12 generates a prompt to respond with a causal explanation document about the first entity based on the second input document, and inputs the generated prompt into the large-scale language model to generate the causal explanation document.

[0051] In this case, the causal explanation unit 12 may generate a prompt for the large-scale language model to respond with the confidence level of the causal explanation. Furthermore, the causal explanation unit 12 may exclude causal explanations with a confidence level below a threshold. In addition, the causal explanation unit 12 may generate a prompt for the large-scale language model to respond with keywords included in the causal explanation. In this case, the causal explanation unit 12 may select one of the causal explanations with overlapping keywords.

[0052] Furthermore, if the first entity is an event related to the occurrence of a specific failure, the failure causal explanation unit 12 can cause the large-scale language model to generate a specific failure explanation that describes the specific failure and the events observed regarding that specific failure, based on the second input document. A description related to the target entity to improve the response performance of LLM 15 is an example of a "specific failure explanation". The failure causal explanation unit 12 can then generate a prompt that prompts the user to respond with a failure causal explanation related to the specific failure, using the specific failure and the specific failure explanation as the occurring failure. In addition, the failure causal explanation unit 12 can generate a prompt for the specific failure explanation that instructs the system not to include descriptions of failures other than the specific failure in the specific failure explanation, and input this prompt to the large-scale language model to generate the specific failure explanation.

[0053] Returning to Figure 1, the explanation continues. The virtual failure knowledge graph generation unit 13 receives a failure causal explanation text based on the target entity as input from the failure causal explanation text generation unit 12. The virtual failure knowledge graph generation unit 13 then generates a virtual failure knowledge graph 200 from the failure causal explanation text, showing the causal relationship between the virtual failure cause and the resulting failure.

[0054] For example, the virtual failure knowledge graph generation unit 13 generates a prompt to generate a virtual failure knowledge graph 200 from the failure causal explanation text. The virtual failure knowledge graph generation unit 13 then inputs the generated prompt to the LLM 15. The LLM 15 receives a prompt instructing the generation of a virtual failure knowledge graph 200 based on the failure causal explanation text and generates and outputs the virtual failure knowledge graph 200 from the failure causal explanation text. The LLM 15 will be explained in detail later. Subsequently, the virtual failure knowledge graph generation unit 13 obtains the virtual failure knowledge graph 200 shown in Figure 4, which is output from the LLM 15 according to the input prompt. The virtual failure knowledge graph 200 shown in Figure 4 is generated based on the failure causal explanation text generated from the system operation management specification document 140, which is a non-failure case document 102.

[0055] For example, the virtual fault knowledge graph generation unit 13 generates a virtual fault knowledge graph 200 in which entity 201, which is the fault-causing entity, has the event of a QoS configuration error as shown in Figure 4, and entity 202, which is the fault-causing entity, has the event of network delay. The virtual fault knowledge graph 200 in Figure 4 shows the causal relationship between the fault cause and the fault that occurred, as indicated by description 203 included in the system operation management specification document 140.

[0056] Returning to Figure 1, let's continue the explanation. The virtual failure knowledge graph generation unit 13 generates a prompt containing the following information. In this case, the prompt includes a question such as, "For the document below, represent the mechanism from the cause of the failure to the occurrence of the failure in a graph." The prompt also includes information specifying the graph generation conditions, such as, "Represent the events that occurred as entities, and the causally related events as edges." The prompt also includes an example answer, such as, "For example, if event B occurs due to event A, represent the edge as (A,B)." The prompt also includes information specifying the graph generation conditions, such as, "Extract all edges." The prompt also includes information to be specified as a sentence that refers to "Description of the failure case document."

[0057] Here, the virtual failure knowledge graph generation unit 13 may add a statement to the prompt to enable the virtual failure knowledge graph 200 to be properly linked with the actual failure knowledge graph 101. The virtual failure knowledge graph generation unit 13 can generate a virtual failure knowledge graph 200 that can be properly linked with the actual failure knowledge graph 101 by adding a statement that instructs the response to include an entity with the same content as the target entity. For example, the virtual failure knowledge graph generation unit 13 can add an instruction such as "Please be sure to include the following entity" to the prompt, and include information specifying the "target entity" as the entity in the prompt.

[0058] This non-fault case document 102 corresponds to the "second input document," the virtual fault knowledge graph 200 corresponds to an example of the "second knowledge graph," and the entities included in the virtual fault knowledge graph 200 correspond to an example of the "second entity." The entities included in the virtual fault knowledge graph 200 represent each event linked by cause and effect in the non-fault case document 102, which is the second input document related to the first entity, fault case document 103. The virtual fault knowledge graph generation unit 13 generates a second knowledge graph, which includes second entities representing each event linked by cause and effect in the second input document related to the first entity, based on the first entity and the second input document. More specifically, the virtual fault knowledge graph generation unit 13 generates a second knowledge graph based on the fault causal explanation document generated by the fault causal explanation document generation unit 12.

[0059] Furthermore, the virtual fault knowledge graph generation unit 13 may generate a second knowledge graph that connects the cause-and-effect relationships from the event of the occurrence of the fault cause of a specific fault to the event of the occurrence of the specific fault, wherein the first entity is the event of the occurrence of the specific fault.

[0060] The coupling unit 14 receives input for the virtual fault knowledge graph 200 from the virtual fault knowledge graph generation unit 13. The coupling unit 14 also obtains the actual fault knowledge graph 101 containing the target entity from the data storage unit 10. The coupling unit 14 then combines the actual fault knowledge graph 101 containing the target entity with the obtained virtual fault knowledge graph 200.

[0061] Figure 7 shows an example of how to combine a real-world fault knowledge graph and a virtual fault knowledge graph. For example, the case of combining a real-world fault knowledge graph 101 having entities 121 to 124 and a virtual fault knowledge graph 200 having entities 211 to 212 will be explained.

[0062] In this case, the actual fault knowledge graph 101 is represented by the edge list 131. The virtual fault knowledge graph 200 in this case is represented by the edge list 132.

[0063] The coupling unit 14 can simply combine the real failure knowledge graph 101 and the virtual failure knowledge graph 200, for example, by generating an edge list 133 by simply arranging the edge representations of edge list 131 and edge list 132 while eliminating duplicates. In this case, since entity 123 in the real failure knowledge graph 101 and entity 211 in the virtual failure knowledge graph 200 both represent the same event C, they are combined. That is, the coupling unit 14 has combined entity 123 in the real failure knowledge graph 101 and entity 211 in the virtual failure knowledge graph 200. As a result, the coupling unit 14 generates a combined knowledge graph 300 having entities 121 to 124 and entity 212 representing event K connected by an edge from entity 123.

[0064] In this way, the coupling unit 14 combines the actual failure knowledge graph 101 and the virtual failure knowledge graph 200 to generate a new combined knowledge graph 300. Subsequently, the coupling unit 14 outputs the generated combined knowledge graph 300 to the LLM 15.

[0065] This combined knowledge graph 300 is an example of a "third knowledge graph." That is, the combining unit 14 combines the first knowledge graph and the second knowledge graph to generate a third knowledge graph.

[0066] Figure 8 shows a concrete example of the connection between the actual failure knowledge graph and the virtual failure knowledge graph. For example, the actual failure knowledge graph 101 has an entity 141 representing the event "increased load on the authentication server," an entity 142 representing the event "network delay," and an entity 143 representing the event "user login failure." These have a causal relationship in which the event "increased load on the authentication server" causes the event "network delay," and the event "increased load on the authentication server" causes the event "increased load on the authentication server."

[0067] Furthermore, the virtual fault knowledge graph 200 includes entity 144, which represents the event "QoS configuration error," and entity 145, which represents the event "network delay." These have a causal relationship, where the event "QoS configuration error" causes the event "network delay."

[0068] Since entity 142 in the actual failure knowledge graph 101 and entity 145 in the virtual failure knowledge graph 200 represent the same event, the coupling unit 14 combines entity 142 and entity 145. As a result, the coupling unit 14 generates a combined knowledge graph 300 having entities 141 to 144. This combined knowledge graph 300 is a knowledge graph in which a causal relationship has been added to the original actual failure knowledge graph 101, showing that the event "network delay" occurs as a result of the event "QoS configuration error".

[0069] Here, we will explain several coupling patterns. In the virtual fault knowledge graph 200, the entities that are coupled are called the destination entities. The coupling pattern is determined by the types of the source entities and the destination entities.

[0070] The entities to be joined are the target entities selected by the target selection unit 11, and the types considered are three: failure cause entities, intermediate event entities, and failure entities. The entities to be joined are entities that represent the same event as the target entities in the virtual failure knowledge graph 200, and the types considered are two: failure cause entities and failure entities. There are two types of join patterns: one in which both the source and destination entities are failure entities, and another in which both the source and destination entities are failure entities. Furthermore, there are two types of join patterns: one in which the source entity is an intermediate event entity and the destination entity is either a failure cause entity or a failure entity.

[0071] Here, a join where the target entity is an intermediate event entity is included in the case where the target entity is either a failure cause entity or a failure entity. Therefore, here, a join where the target entity is an intermediate event entity is excluded from the join patterns. However, the join unit 14 may perform the join with the target entity as an intermediate event entity.

[0072] Figure 9 shows a coupling pattern when both the source and destination entities are fault entities. In this case, entity 151 in the actual fault knowledge graph 101 is the fault entity and the source entity. Also, entity 152 in the virtual fault knowledge graph 200 is the fault entity and the destination entity. The coupling unit 14 combines the actual fault knowledge graph 101 and the virtual fault knowledge graph 200 using entities 151 and 152 as coupling points to generate a combined knowledge graph 300. Entity 153 in the combined knowledge graph 300 is the coupling point. In this coupling pattern, the number of paths from other fault-causing entities connected to the fault entity increases compared to the original actual fault knowledge graph 101.

[0073] Figure 10 shows a join pattern when both the source and destination entities are fault-causing entities. In this case, entity 154 in the actual fault knowledge graph 101 is the fault-causing entity and the source entity. Also, entity 155 in the virtual fault knowledge graph 200 is the fault-causing entity and the destination entity. The join unit 14 joins the actual fault knowledge graph 101 and the virtual fault knowledge graph 200 using entities 154 and 155 as the join point to generate a combined knowledge graph 300. Entity 156 in the combined knowledge graph 300 is the join point. In this join pattern, the number of paths extending from fault-causing entities to other fault entities increases compared to the original actual fault knowledge graph 101.

[0074] Figure 11 shows a join pattern when the source entity is an intermediate event entity and the destination entity is a failure cause entity or a failure entity. In this case, entity 156 in the actual failure knowledge graph 101 is an intermediate event entity and is the source entity. Also, entity 157 in the virtual failure knowledge graph 200 is a failure entity and is the destination entity. The join unit 14 joins the actual failure knowledge graph 101 and the virtual failure knowledge graph 200 using entities 156 and 157 as the join point to generate a combined knowledge graph 300. Entity 158 in the combined knowledge graph 300 is the join point. In this join pattern, compared to the original actual failure knowledge graph 101, the number of paths connecting other failure cause entities to intermediate event entities or paths extending from intermediate event entities to other failure entities increases.

[0075] Here, a combined knowledge graph 300 is generated for the target entities sequentially selected by the target selection unit 11. For example, suppose one combined knowledge graph 300 is generated using one non-fault case document 102 for one entity included in one actual fault knowledge graph 101. In that case, assuming all entities in the actual fault knowledge graph 101 are used, a combined knowledge graph 300 is generated for the total number of entities included in each actual fault knowledge graph 101. The number of combined knowledge graphs 300 generated also increases depending on the number of non-fault case documents 102 used. Multiple new fault knowledge graphs generated in this way are sent to the LLM 15.

[0076] Furthermore, the above explanation used the example of a case where one virtual fault knowledge graph 200 is generated from one non-fault case document 102, but it is also possible that multiple virtual fault knowledge graphs 200 are generated from one non-fault case document 102. In addition, it is conceivable that multiple virtual fault knowledge graphs 200 can be combined with one actual fault knowledge graph 101. In such cases, the combining unit 14 can generate one combined knowledge graph 300 by sequentially single-connecting multiple virtual fault knowledge graphs 200 that can be connected to the actual fault knowledge graph 101.

[0077] Let's return to Figure 1 and continue the explanation. LLM15 is a large-scale machine learning model that performs natural language processing. For example, LLM15 receives prompts that include information about an entity and instructions to generate a causal explanation of the failure related to that entity, as well as a non-failure case document 102 from which information will be extracted, and outputs a causal explanation of the failure that describes the causal relationship between the cause of the failure and the resulting failure.

[0078] Furthermore, LLM15 receives prompts instructing it to generate a causal explanation for the failure and a virtual failure knowledge graph 200 based on that causal explanation, and generates and outputs the virtual failure knowledge graph 200 from the causal explanation for the failure.

[0079] Furthermore, LLM15 receives input of a fault from the user terminal device 2 and uses the combined knowledge graph 300 to estimate the cause of the specified fault. Then, LLM15 transmits the estimated cause of the specified fault to the user terminal device 2 and provides it to the user.

[0080] Figure 12 is a flowchart of the combined knowledge graph generation process by the information processing device according to Example 1. Next, the flow of the combined knowledge graph generation process by the information processing device 1 according to Example 1 will be explained with reference to Figure 12.

[0081] The target selection unit 11 selects target entities from the actual failure knowledge graph 101 to be used as the basis for generating the combined knowledge graph 300 (step S1).

[0082] The failure causal explanation document generation unit 12 generates a prompt to generate a failure causal explanation document for the target entity based on the non-failure case document 102, depending on whether the target entity is the cause of the failure or the target entity is the resulting failure. The failure causal explanation document generation unit 12 then generates the failure causal explanation document by inputting the generated prompt to the LLM 15 and obtaining its output (step S2).

[0083] The virtual fault knowledge graph generation unit 13 generates prompts from the fault causal explanation text to generate a virtual fault knowledge graph 200. The virtual fault knowledge graph generation unit 13 then inputs the generated prompts to the LLM 15 and obtains its output to generate the virtual fault knowledge graph 200 (step S3).

[0084] The coupling unit 14 simply combines the actual failure knowledge graph 101 and the virtual failure knowledge graph 200 to generate a combined knowledge graph 300 (step S4).

[0085] As described above, the information processing device 1 according to this embodiment generates a virtual fault knowledge graph 200 from the non-fault case document 102 with respect to the faults shown in the actual fault knowledge graph 101 based on the fault case document 103. Then, the information processing device 1 combines the actual fault knowledge graph 101 and the generated virtual fault knowledge graph 200 to generate a combined knowledge graph 300.

[0086] By using the combined knowledge graph 300 generated in this way to estimate the cause of a problem specified by the user, it becomes possible to perform estimations that take into account not only the content of the problem case document 103 but also the content of the non-problem case document 102, thereby improving the performance of answering questions.

[0087] Figure 13 illustrates the effects of generating a combined knowledge graph. Here, we will explain the case where two virtual failure knowledge graphs, 200A and 200B, are generated from a non-fault case document 102.

[0088] The information processing device 1 in this embodiment generates two virtual fault knowledge graphs 200A and 200B from the non-fault case document 102. The information processing device 1 then combines the fault case document 103 with the two virtual fault knowledge graphs 200A and 200B to generate a combined knowledge graph 300. The information processing device 1 then uses the combined knowledge graph 300 to estimate the cause of the fault specified by the user. As a result, the information processing device 1 can estimate and answer the cause of the fault based on the causal relationship of the fault, which cannot be grasped from either the single fault case document 103 or the non-fault case document 102, thus improving the answer performance.

[0089] For example, in a response based on a single failure case document 103, events A, B, and F are presented as the causes of event D, which is the failure that occurred. In a response based on a single non-failure case document 102, event K can be presented as the cause of event C, but its relationship to event D, which is the failure that occurred, is unclear.

[0090] In response to this, when asked what the cause of failure D is, the information processing device 1 according to this embodiment estimates the cause of failure by traversing the combined knowledge graph 300 from the entity of failure D towards the cause of failure. In this case, the information processing device 1 can answer that the cause of failure D is failure A, B, K, or F. That is, the information processing device 1 can present failure F, which would not be presented in an answer based on a single failure case document 103, and can provide a more appropriate answer to the question, thus improving the answer performance. [Examples]

[0091] Figure 14 is a block diagram of the information processing device according to Embodiment 2. The information processing device 1 according to this embodiment differs from Embodiment 1 in that it generates a combined knowledge graph 300 by considering multiple events that occur simultaneously and cause problems. In the following description, the operation of each part according to Embodiment 1 may be omitted.

[0092] The data storage unit 10 in this embodiment holds information on concurrently occurring events 104. Concurrently occurring events refer to multiple events that would cause a problem if they occurred simultaneously.

[0093] Figure 15 shows an example of the information held by the information processing device according to Embodiment 2. The data storage unit 10 holds, for example, the edge list 401 showing the actual failure knowledge graph 101 shown in Figure 15, as well as the concurrently occurring event information 104 shown in Figure 15.

[0094] In the actual failure knowledge graph 101 in Figure 15, events A and B are the causes of the failure, event C is an intermediate event, and events D and E are the resulting failures. However, in this case, a failure does not occur unless events D and E occur simultaneously; that is, a failure does not occur if event D occurs alone or if event E occurs alone.

[0095] Therefore, in this embodiment, the data storage unit 10 holds simultaneous event information 104, which indicates that a failure occurs when events D and E occur simultaneously ((C,D),(C,E)). In other words, the set of edge list 401 and simultaneous event information 104 accurately represents the contents of the actual failure knowledge graph 101, including simultaneous events. That is, the simultaneous event information 104 allows multiple entities that are simultaneous events to be treated as an entity group.

[0096] A set of entities that indicate events that occur simultaneously and result in a specific event, as shown by this simultaneous event information 104, is an example of a "first entity group." In other words, the first knowledge graph, the actual failure knowledge graph 101, has a set of several first entities, each of which indicates an event that occurs simultaneously and results in a specific event, as a single first entity group.

[0097] The target selection unit 11 selects a target entity from the actual failure knowledge graph 101. However, for multiple entities that are designated as simultaneous events in the simultaneous event information 104, the target selection unit 11 selects those multiple entities as a single entity group. In other words, the target selection unit 11 can select one first entity group included in the first knowledge graph from the first knowledge graph, which has a set of first entities that cause a specific event to occur when the events indicated by each of the multiple first entities occur simultaneously, as a single first entity group.

[0098] The failure causal explanation generation unit 12 generates a prompt to generate a failure causal explanation for a target entity based on the non-failure case document 102, using the content obtained by combining each entity of the entity group with "and" as a single target entity. Then, the failure causal explanation generation unit 12 generates a failure causal explanation from the non-failure case document 102 using the LLM 15. The failure causal explanation generation unit 12 can generate a failure causal explanation for a failure that occurs when each event of the entity group occurs simultaneously.

[0099] In this case, the failure causal explanation text generation unit 12 can generate prompts that produce sentences that are expressed naturally. In this case, it is preferable for the failure causal explanation text generation unit 12 to generate prompts that include the following information. For example, the prompt may include a question such as, "Regarding the failure case document below, please provide a sentence that states that multiple failures shown in the failure list below occurred simultaneously." The prompt may also include information that specifies a sentence that refers to the "description of the failure case document." The prompt may also include information that specifies the description of an entity group as a failure.

[0100] The virtual fault knowledge graph generation unit 13 creates prompts to generate a virtual fault knowledge graph 200 from the fault causal explanation text. In these prompts, entity groups are treated as a single entity. The virtual fault knowledge graph generation unit 13 then inputs the generated prompts to the LLM 15 and obtains its output to generate the virtual fault knowledge graph 200. In the virtual fault knowledge graph 200 created at this stage, entity groups are represented as a single entity.

[0101] Next, if the target entity is an entity group, the virtual failure knowledge graph generation unit 13 decomposes the entity representing the entity group and replaces it with the original entity to update the virtual failure knowledge graph 200. This ensures that the representation of entities corresponding to entity groups in the virtual failure knowledge graph 200 matches the representation of entities corresponding to entity groups in the actual failure knowledge graph 101. The actual failure knowledge graph 101 and the virtual failure knowledge graph 200 become joinable. Furthermore, the virtual failure knowledge graph generation unit 13 generates simultaneous occurrence event information 104 indicating that multiple entities previously considered as an entity group are simultaneous occurrence events.

[0102] Figure 16 shows the generation of a virtual fault knowledge graph by the information processing device according to Embodiment 2. Here, we will explain the case where events D and E are simultaneous events. The virtual fault knowledge graph generation unit 13 generates a virtual fault knowledge graph 200 having entity 411 from the fault causal explanation text. Entity 411 is an entity group that indicates that events D and E are linked by "and" and that events D and E are simultaneous events. In practice, the virtual fault knowledge graph generation unit 13 generates an edge list 402 that represents this virtual fault knowledge graph 200.

[0103] Next, the virtual failure knowledge graph generation unit 13 removes the "and" from the entity group indicated by entity 411, decomposing it into entity 412, which represents event D, and entity 413, which represents event E, and updates the virtual failure knowledge graph 200. In practice, the virtual failure knowledge graph generation unit 13 divides the edge representation of the entity group (F,D and E) in the edge list 402 into two edge representations, (F,D) and (F,E), to create an edge list 403. Furthermore, the virtual failure knowledge graph generation unit 13 generates simultaneous event information 104, ((F,D),(F,E)), indicating that events D and E are simultaneous events, and adds it to the virtual failure knowledge graph 200.

[0104] In this way, the virtual fault knowledge graph generation unit 13 generates the virtual fault knowledge graph 200 by grouping a set of second entities that represent similar events to each of several first entities included in the first entity group into a single second entity group. For example, the pair of events D and E connected by "and" in Figure 16 is an example of a "second entity group".

[0105] The coupling unit 14 simply combines the actual failure knowledge graph 101 and the virtual failure knowledge graph 200 to generate a combined knowledge graph 300. In this case, the coupling unit 14 combines the simultaneous occurrence event information 104 attached to the actual failure knowledge graph 101 and the simultaneous occurrence event information 104 attached to the virtual failure knowledge graph 200 to form the simultaneous occurrence event information 104 for the combined knowledge graph 300.

[0106] As described above, the information processing device 1 according to this embodiment can generate a combined knowledge graph 300 even when there are events that cause problems when they occur simultaneously, by treating the simultaneous occurrence as the problem. Therefore, it is possible to generate a combined knowledge graph 300 that is more adapted to actual problems, and by estimating the cause of the problem using such a combined knowledge graph 300, the response performance can be improved. [Examples]

[0107] Next, Example 3 will be described. The information processing device 1 according to this example generates a single combined knowledge graph 300 from multiple actual failure knowledge graphs 101 and multiple non-fault case documents 102. In the following description, the operation of each part, which is the same as in Example 1, will be omitted.

[0108] The coupling unit 14 in this embodiment combines a virtual failure knowledge graph 200 based on one actual failure knowledge graph 101 and one non-failure case document 102 to generate one combined knowledge graph 300. The coupling unit 14 generates one combined knowledge graph 300 for each combination of multiple actual failure knowledge graphs 101 and multiple non-failure case documents 102. Then, the coupling unit 14 further combines the generated combined knowledge graphs 300 to generate one combined knowledge graph 300.

[0109] Figure 17 is a flowchart of the combined knowledge graph generation process by the information processing device according to Example 3. Next, referring to Figure 17, the flow of the combined knowledge graph generation process by the information processing device 1 according to this embodiment will be explained.

[0110] The target selection unit 11 selects one actual failure knowledge graph 101 from among multiple actual failure knowledge graphs 101 (step S11).

[0111] Furthermore, the failure causal explanation document generation unit 12 selects one non-fault case document 102 from among multiple non-fault case documents 102 (step S12).

[0112] Next, the target selection unit 11 selects target entities from the selected actual failure knowledge graph 101 to be used as the basis for generating the combined knowledge graph 300 (step S13).

[0113] Next, the failure causal explanation document generation unit 12 generates a prompt to generate a failure causal explanation document for the target entity based on the selected non-failure case document 102, depending on whether the target entity is the cause of the failure or the target entity is the failure that occurred. Then, the failure causal explanation document generation unit 12 generates the failure causal explanation document by inputting the generated prompt to the LLM 15 and obtaining its output (step S14).

[0114] The virtual fault knowledge graph generation unit 13 generates prompts from the fault causal explanation text to generate a virtual fault knowledge graph 200. The virtual fault knowledge graph generation unit 13 then inputs the generated prompts to the LLM 15 and obtains its output to generate the virtual fault knowledge graph 200 (step S15).

[0115] Next, the failure causal explanation unit 12 determines whether the generation of the virtual failure knowledge graph 200 has been completed for all non-failure case documents 102 (step S16). If there are still non-failure case documents 102 for which the virtual failure knowledge graph 200 has not been generated (step S16: negative), the combined knowledge graph generation process returns to step S12.

[0116] In contrast, if the generation of virtual failure knowledge graphs 200 is completed for all non-fault case documents 102 (step S16: affirmative), the coupling unit 14 simply combines the actual failure knowledge graph 101 with all the generated virtual failure knowledge graphs 200. As a result, the coupling unit 14 generates a single combined knowledge graph 300 (step S17).

[0117] Next, the target selection unit 11 determines whether the generation of virtual failure knowledge graphs 200 has been completed for all actual failure knowledge graphs 101 (step S18). If there are still actual failure knowledge graphs 101 for which virtual failure knowledge graphs 200 have not been generated (step S18: negative), the combined knowledge graph generation process returns to step S11.

[0118] In contrast, if the generation of virtual failure knowledge graphs 200 is completed for all actual failure knowledge graphs 101 (step S18: affirmative), the merging unit 14 merges all the generated merged knowledge graphs 300. As a result, the merging unit 14 generates one merged knowledge graph 300 (step S19).

[0119] As described above, the information processing device 1 according to this embodiment combines a plurality of actual failure knowledge graphs 101 and a plurality of virtual failure knowledge graphs 200 based on a plurality of non-failure case documents 102 to generate a single combined knowledge graph 300. That is, the generated combined knowledge graph 300 includes the causal relationship between the failures that occurred and the causes of the failures obtained from the plurality of failure case documents 103 and the plurality of non-failure case documents 102. Therefore, by using this combined knowledge graph 300 to estimate the cause of a failure specified by the user, it is possible to perform estimations based on the contents of the plurality of failure case documents 103 and the plurality of non-failure case documents 102, thereby improving the accuracy of the failure cause estimation.

[0120] In other words, if there are multiple actual failure knowledge graphs 101, one is an example of the first knowledge graph. The other is an example of a "fourth knowledge graph that includes multiple third entities, each representing an event linked by a cause-and-effect relationship based on the third input document." That is, the target selection unit 11 selects a third entity from the fourth knowledge graph that includes multiple third entities, each representing an event linked by a cause-and-effect relationship based on the third input document. The virtual failure knowledge graph generation unit 13 generates a fifth knowledge graph that includes fourth entities, each representing an event linked by a cause-and-effect relationship contained in the second input document, to which the third entity is related, based on the third entity and the second input document. The merging unit 14 further merges the merged knowledge graph 300, which is obtained by merging the fourth knowledge graph and the fifth knowledge graph, with the third knowledge graph generated based on the first knowledge graph and the second input document, to generate a single merged knowledge graph 300. In other words, the coupling unit 14 combines the third knowledge graph generated based on the first knowledge graph and the second input document with the newly generated fourth knowledge graph and fifth knowledge graph to generate a sixth knowledge graph.

[0121] Furthermore, if there are multiple non-fault case documents 102, one will be an example of a second input document, and another will be an example of a "third input document". The virtual fault knowledge graph generation unit 13 then generates a fourth knowledge graph, which includes a third entity representing each event linked by a cause-and-effect relationship contained in the third input document and related to the first entity, based on the first entity and the third input document. In other words, in this case, the third knowledge graph and the fourth knowledge graph are generated relative to the first knowledge graph. The merging unit 14 merges the first knowledge graph, the third knowledge graph, and the fourth knowledge graph to generate a fifth knowledge graph.

[0122] (Hardware configuration) Figure 18 is a hardware configuration diagram of the information processing device. Next, with reference to Figure 18, an example of a hardware configuration for realizing each function of the information processing device 1 will be described.

[0123] As shown in Figure 18, the information processing device 1 includes, for example, a CPU (Central Processing Unit) 91, memory 92, a hard disk 93, and a network interface 94. The CPU 91 is connected to the memory 92, hard disk 93, and network interface 94 via a bus.

[0124] The network interface 94 is an interface for communication between the information processing device 1 and an external device. For example, the network interface 94 relays communication between the user terminal device 2 and the CPU 91.

[0125] The hard disk 93 is an auxiliary storage device. The hard disk 93 implements the functions of the data storage unit 10 as illustrated in Figure 1. The hard disk 93 may also store the LLM 15. Furthermore, the hard disk 93 stores various programs, including programs for implementing the functions of the target selection unit 11, the fault causal explanation statement generation unit 12, the virtual fault knowledge graph generation unit 13, and the coupling unit 14 as illustrated in Figure 1.

[0126] Memory 92 is the main memory. Memory 92 can be, for example, DRAM (Dynamic Random Access Memory).

[0127] The CPU 91 reads various programs from the hard disk 93, loads them into memory 92, and executes them. In this way, the CPU 91 realizes the functions of the target selection unit 11, the fault causal explanation text generation unit 12, the virtual fault knowledge graph generation unit 13, and the coupling unit 14, as illustrated in Figure 1. [Explanation of Symbols]

[0128] 1. Information Processing Device 2. User terminal device 10 Data storage unit 11. Target Selection Section 12. Disability causal explanation text generation unit 13. Virtual Fault Knowledge Graph Generation Unit 14 Joint 15 LLM 101 Real-World Problem Knowledge Graph 102 Non-Fault Case Documents 103 Incident Case Documents 104 Information on simultaneous events 200 Virtual Failure Knowledge Graph 300 Combined Knowledge Graphs

Claims

1. Information processing device, Select a first entity from a first knowledge graph that includes multiple first entities representing events linked by cause-and-effect relationships based on the first input document. Based on the selected first entity and second input document, a second knowledge graph is generated which includes a second entity representing each event linked by a cause-and-effect relationship contained in the second input document, relating to the first entity. The first and second knowledge graphs are combined to generate a third knowledge graph. A generation method characterized by performing a process.

2. The generation method according to claim 1, characterized in that the process of selecting the first entity includes the process of selecting one entity included in the first knowledge graph.

3. The generation method according to claim 1, characterized in that the first input document is a document describing the details of the actual failure and the cause of the failure.

4. The generation method according to claim 3, characterized in that the second input document is a document that includes a description relating to the actual failure described in the first input document.

5. The generation method according to claim 3, characterized in that the process of selecting the first entity includes a process of selecting one or a combination thereof of the fault entity, the fault cause entity, or the intermediate event entity from among the plurality of first entities included in the first knowledge graph, which include a fault entity indicating the event of the occurrence of the actual fault, a fault cause entity indicating the event that causes the fault, and an intermediate event entity indicating an event that occurs between the actual fault and the fault cause.

6. The generation method according to claim 1, characterized in that the generation process of the second knowledge graph includes a process of generating the second knowledge graph, which is a second knowledge graph linked by cause-and-effect relationships from the event of occurrence of the cause of a specific failure to the event of occurrence of the specific failure, wherein the first entity is the event of occurrence of the specific failure.

7. The generation method according to claim 1, characterized in that the generation process of the second knowledge graph is a second knowledge graph linked by cause-and-effect relationships from the event of occurrence of the cause of a specific failure to the event of occurrence of the specific failure, wherein the first entity is the event of occurrence of the cause of the failure.

8. The process for generating the second knowledge graph described above is: From the second input document, a causal explanation document is generated that describes the events related to the first entity in a cause-and-effect relationship. Based on the aforementioned explanation of the cause and effect of the failure, the second knowledge graph is generated. The generation method according to claim 1, characterized by including processing.

9. The generation method according to claim 8, characterized in that the process for generating the causal explanation of the disorder includes the process for generating the causal explanation of the disorder using a large-scale language model.

10. The process for generating the aforementioned causal explanation statement for the disorder is as follows: Based on the second input document, a prompt is generated to provide the causal explanation statement for the first entity, The large-scale language model is given the prompt to generate the causal explanation statement for the failure. The generation method according to claim 9, characterized by including processing.

11. The generation method according to claim 10, characterized in that the prompt generation process includes a prompt that causes the large-scale language model to respond with the degree of confidence of the causal explanation statement of the failure.

12. The generation method according to claim 11, characterized in that the generation process for the causal explanation of the failure includes a process for excluding the causal explanation of the failure whose degree of confidence is below a threshold.

13. The generation method according to claim 10, characterized in that the prompt generation process includes prompting the large-scale language model to respond with keywords included in the causal explanation of the failure.

14. The generation method according to claim 13, characterized in that the process for generating the causal explanation of the impairment includes a process for selecting one of the causal explanations of the impairment in which the keywords overlap.

15. The process for generating the aforementioned prompt is as follows: If the first entity is an event of a specific failure occurring, the large-scale language model generates a specific failure description that describes the specific failure and the events observed regarding the specific failure, based on the second input document. The aforementioned specific failure and the description of the specific failure are treated as the resulting failure, and a prompt is generated to provide a causal explanation of the failure related to the said failure. The generation method according to claim 10, characterized by including processing.

16. The process for generating the aforementioned prompt is as follows: A prompt for the specific defect description is generated, which instructs that the specific defect description not include any description of defects other than the specific defect. The aforementioned prompt for the specific defect description is input into a large-scale language model to generate the aforementioned specific defect description. The generation method according to claim 15, characterized by including processing.

17. The generation method according to claim 1, characterized in that the process of selecting the first entity includes a process of selecting one first entity group from a first knowledge graph having a set of first entities in which a specific event occurs when the events indicated by each of the plurality of first entities occur simultaneously.

18. The first knowledge graph has a group of first entities in which a specific event occurs when the events represented by each of the plurality of first entities occur simultaneously, as a single first entity group. The process for generating the second knowledge graph includes generating the second knowledge graph by grouping sets of second entities that represent events similar to each of the sets of first entities included in the first entity group into a single second entity group. The method for generating according to feature 1.

19. The aforementioned information processing device is Select a third entity from a fourth knowledge graph that includes multiple third entities representing events linked by cause-and-effect relationships based on the third input document. Based on the selected third entity and the second input document, a fifth knowledge graph is generated, which includes a fourth entity representing each event linked by a cause-and-effect relationship contained in the second input document, relating to the third entity. The third, fourth, and fifth knowledge graphs are combined to generate a sixth knowledge graph. The generation method according to claim 1, characterized by further processing.

20. The aforementioned information processing device is Based on the selected first entity and third input document, a fourth knowledge graph is generated, which includes a third entity representing each event linked by a cause-and-effect relationship contained in the third input document that is related to the first entity. The first, third, and fourth knowledge graphs are combined to generate a fifth knowledge graph. The generation method according to claim 1, characterized by further processing.

21. Select a first entity from a first knowledge graph, which includes multiple first entities representing events linked by cause-and-effect relationships based on a first input document. Based on the selected first entity and second input document, a second knowledge graph is generated which includes a second entity representing each event linked by a cause-and-effect relationship contained in the second input document, relating to the first entity. The first and second knowledge graphs are combined to generate a third knowledge graph. A generation program characterized by having a computer perform the processing.

22. A target selection unit that selects a first entity from a first knowledge graph, which includes a plurality of first entities representing events linked by cause-and-effect relationships based on a first input document, A knowledge graph generation unit generates a second knowledge graph, which includes a second entity representing each event linked by a cause-and-effect relationship in the second input document that is related to the first entity, based on the first entity and the second input document selected by the target selection unit. A coupling unit that combines the first knowledge graph and the second knowledge graph to generate a third knowledge graph. An information processing device characterized by having the following features.

Images