Relevant information retrieval device, relevant information retrieval method, and relevant information retrieval program

Abstract

A relevant information retrieval device (100) comprises a static link unit (130) that connects a first node and a second node if a target similarity is equal to or greater than a similarity threshold, the target similarity being the similarity between a first text corresponding to a first node included in a first knowledge graph corresponding to a document and a second text corresponding to a second node included in a second knowledge graph corresponding to a document.

Description

Related information search device, related information search method, and related information search program

[0001] The present disclosure relates to a related information search device, a related information search method, and a related information search program.

[0002] There is a search technology based on a knowledge base. Patent Document 1 discloses a technology that, when a fault occurs in a device of equipment at a manufacturing site, realizes a function of presenting necessary knowledge as an appropriate measure to deal with the fault based on a knowledge base.

[0003] Patent No. 7242756

[0004] According to the technology disclosed in Patent Document 1, previously accumulated documents are converted into a knowledge graph, and relevant past knowledge is searched for according to the current situation. However, because this technology only extracts labels from individual documents, it has the problem of low accuracy in connecting documents (documents cannot be connected if their expressions are different, and documents are connected only if the words are the same), and it is not possible to search across different databases.

[0005] The present disclosure aims to provide a search technology based on knowledge graphs that achieves relatively high accuracy in the connections between documents and enables cross-sectional searches across different databases.

[0006] The related information search device according to the present disclosure includes a static link unit that connects a first node and a second node when an object similarity, which is the similarity between a first text corresponding to a first node included in a first knowledge graph corresponding to a document and a second text corresponding to a second node included in a second knowledge graph corresponding to the document, is equal to or greater than a similarity threshold.

[0007] According to the present disclosure, the static link unit determines whether to connect a first node and a second node based on the similarity between a first text corresponding to a first node included in a first knowledge graph and a second text corresponding to a second node included in a second knowledge graph. Therefore, according to the present disclosure, in a search technology based on a knowledge graph, it is possible to achieve a relatively high accuracy in the connection between documents and to enable cross-sectional searches across different databases.

[0008] 1 is a diagram for explaining an overview of the first embodiment. FIG. 1 is a diagram for explaining an overview of the first embodiment. FIG. 2 is a diagram showing an example of the configuration of a related information retrieval device 100 according to the first embodiment. FIG. 3 is a diagram for explaining a knowledge graph according to the first embodiment. FIG. 4 is a diagram for explaining a knowledge graph according to the first embodiment, where (a) is a table showing the meanings of node types and (b) is a table showing triples in the knowledge graph. FIG. 5 is a diagram for explaining the processing of a knowledge acquisition unit 110 according to the first embodiment. FIG. 6 is a diagram for explaining the processing of a static linking unit 130 according to the first embodiment. FIG. 7 is a diagram for explaining the processing of a static linking unit 130 according to the first embodiment. FIG. 8 is a diagram for explaining the processing of a related information inference unit 150 according to the first embodiment. FIG. 9 is a diagram for explaining the processing of a relevance calculation unit 160 according to the first embodiment, where (a) is a diagram for explaining each path and (b) is a table for explaining score calculation processing. FIG. 10 is a diagram for explaining the processing of a display data generation unit 170 according to the first embodiment. FIG. 11 is a diagram for explaining the processing of a display data generation unit 170 according to the first embodiment. FIG. 12 is a diagram for explaining the processing of a display data generation unit 170 according to the first embodiment. FIG. 13 is a flowchart showing the operation of the related information retrieval device 100 according to the first embodiment. FIG. 14 is a diagram for explaining an example of the hardware configuration of a related information retrieval device 100 according to a variation of the first embodiment. FIG. 1 is a diagram showing an example of the configuration of a related information search device 100 according to embodiment 2. FIG. 2 is a diagram explaining the processing of a connection condition specification unit 210 according to embodiment 2. FIG. 3 is a diagram explaining the processing of a static link unit 130 according to embodiment 2. FIG. 4 is a diagram explaining the processing of a static link unit 130 according to embodiment 2. A flowchart showing the operation of the related information search device 100 according to embodiment 2. FIG. 1 is a diagram showing an example of the configuration of a related information search device 100 according to embodiment 3. FIG. 2 is a diagram explaining the processing of a search condition specification unit 310 according to embodiment 3. A flowchart showing the operation of the related information search device 100 according to embodiment 3. FIG. 3 is a diagram showing an example of the configuration of a related information search device 100 according to embodiment 4. FIG. 4 is a diagram explaining the processing of a dynamic link unit 410 according to embodiment 4. A flowchart showing the operation of the related information search device 100 according to embodiment 4.

[0009] In the description of the embodiments and the drawings, the same elements and corresponding elements are given the same reference numerals. The description of elements given the same reference numerals will be omitted or simplified as appropriate. Arrows in the drawings mainly indicate the flow of data or the flow of processing. Furthermore, "unit" may be read as "circuit," "step," "procedure," "process," or "circuitry" as appropriate.

[0010] Embodiment 1. This embodiment will be described in detail below with reference to the drawings. FIG. 1 is a diagram illustrating an overview of this embodiment. One of the objectives of this embodiment is to build a knowledge graph based on files that comprehensively indicate impacts, defects, and the like. More specifically, one of the objectives is to build a wide-ranging knowledge graph from internal documents that indicate designs, manufacturing processes, and market defects. Each knowledge graph corresponds to an information source. Files may also be referred to as documents. In this case, hidden relationships are predicted by utilizing document information, and multiple knowledge graphs are connected based on the predicted relationships. Note that using a language model specialized for a domain such as a company can reduce computational costs. Knowledge graph embedding technology incorporating document information is also utilized. As a specific example, the documents are business documents. As a specific example, this embodiment is utilized when it is desired to explore phenomena that may occur related to PPS (polyphenylene sulfide) when the material of a certain component is changed to PPS.

[0011] FIG. 2 is a diagram illustrating an overview of this embodiment. Knowledge graphs for each data type are shown in FIG. 2. Each node in each knowledge graph is determined according to the content of the knowledge. In this embodiment, as shown by the dashed lines in FIG. 2, nodes representing similar content in different knowledge graphs are linked. This allows each knowledge graph obtained from each of multiple databases to be semantically linked. Therefore, this embodiment enables cross-sectional searches by tracing semantic relationships between multiple databases.

[0012] ***Description of Configuration*** Fig. 3 shows an example of the configuration of a related information retrieval device 100 according to this embodiment. As shown in Fig. 3, the related information retrieval device 100 includes a knowledge acquisition unit 110, a knowledge registration unit 120, a static link unit 130, a DB operation unit 140, a related information inference unit 150, a relevance calculation unit 160, a display data generation unit 170, and a user interface unit 180. The related information retrieval device 100 also stores a knowledge graph DB 190. DB is an abbreviation for database.

[0013] Data representing files as a document group is input to the knowledge acquisition unit 110. The knowledge acquisition unit 110 acquires information on the nodes and edges of a knowledge graph from the input files. Here, edges correspond to the relationships between nodes. Generating a knowledge graph corresponds to acquiring knowledge.

[0014] FIG. 4 shows a specific example of a knowledge graph. The text in each node indicates the node type corresponding to that node. A text indicating the edge type corresponding to that edge is attached to each edge. Note that while the knowledge graph in FIG. 4 is a directed graph, the knowledge graph may also be an undirected graph. A knowledge graph acquired from one file is created as shown in FIG. 4. Such a knowledge graph is created for each data, and each created knowledge graph is stored in the knowledge graph DB 190.

[0015] FIG. 5 is a table corresponding to FIG. 4. (a) of FIG. 5 is a table showing the meaning of node types. (b) of FIG. 5 is a table showing triples in a knowledge graph. (b) of FIG. 5 shows "content" and the triples in the knowledge graph corresponding to "content". A triple in the knowledge graph is defined for each combination of node types corresponding to two nodes. In (b) of FIG. 5, for each combination of node types corresponding to two nodes, the edge type corresponding to the edge connecting the two nodes is shown.

[0016] As a specific example, if the input file is a tabular file, the knowledge acquisition unit 110 extracts text from each cell indicated by the tabular file and sets it as a node. In this case, a node type corresponding to each text may be predefined. Alternatively, the knowledge acquisition unit 110 may estimate a node type based on the content of each text and set the estimated node type as the node type corresponding to each text. Furthermore, the knowledge acquisition unit 110 determines relationships (edges) between multiple nodes using information such as whether corresponding cells for multiple nodes are included in the same row or column. Figure 6 is a diagram illustrating a process for generating a knowledge graph from a tabular file. In Figure 6, a node type corresponding to each column is predefined, and two nodes are appropriately connected by an edge having a corresponding edge type. Each node corresponds to the text in each cell. The knowledge acquisition unit 110 may determine whether to create an edge between two nodes based on the similarity between the text corresponding to the two nodes. Each edge type may be determined based on structural information of the tabular file.

[0017] The knowledge registration unit 120 registers information indicating the nodes and edges acquired by the knowledge acquisition unit 110 in the knowledge graph DB 190 via the DB operation unit 140 .

[0018] The static linking unit 130 connects a first node and a second node when the object similarity is equal to or greater than a similarity threshold. The object similarity is the similarity between a first text and a second text. The first text corresponds to a first node included in a first knowledge graph corresponding to the document. The second text corresponds to a second node included in a second knowledge graph corresponding to the document. The object similarity may be the similarity between a vector corresponding to the first text and a vector corresponding to the second text. The static linking unit 130 may calculate the object similarity based on an edit distance. The similarity threshold may be determined in any manner. As a specific example, the static linking unit 130 converts text corresponding to all nodes registered in the knowledge graph DB 190 into vectors and calculates vector similarity between each two nodes using each converted vector. The static linking unit 130 then creates an edge between two nodes with high corresponding vector similarity. Note that the static linking unit 130 may link semantically similar nodes together using text embedding or graph embedding, and connect the linked nodes together with an edge. The two nodes to which an edge is to be established may be nodes in different knowledge graphs, or may be two nodes in the same knowledge graph.

[0019] The static linking unit 130 may vectorize text in any manner. Proposed methods for obtaining a vector representation corresponding to text include a method of aggregating vector representations of words and a method of vectorizing the text itself. The static linking unit 130 may vectorize each node using a graph. That is, when vectorizing a node, the static linking unit 130 may use text embedding (vectorizing a node using only text information) or graph embedding (vectorizing a node using information on a subgraph). Specific examples of methods for vectorizing text include BoW (Bag of Words), TF-IDF (term frequency-inverse document frequency), the average of Word2vec (word distributed representation), and Sentence BERT (Bidirectional Encoder Representations from Transformers). These methods use text information contained in nodes to obtain vectors. BoW is a method for calculating the sum of local representations of words that make up text. When TF-IDF is used, it is possible to increase the value of words that have characteristics that do not appear in other documents. The average of Word2vec is a method for calculating the average of word vectors of words that make up a sentence. Sentence BERT is a method for generating relatively high-quality sentence vectors by fine-tuning BERT.

[0020] After vectorizing each node, the static link unit 130 calculates the similarity between vectors, i.e., the similarity between nodes, by calculating the cosine similarity shown in Equation 1. Based on the calculated similarity, the static link unit 130 determines whether or not to create an edge between nodes.

[0021]

[0022] Fig. 7 shows a specific example in which an edge (SIMILAR_TO) is established between nodes whose corresponding similarity is equal to or greater than a threshold between knowledge graphs constructed from three data sets. In Fig. 7, between the knowledge graphs, the similarity between nodes with a SIMILAR_TO edge established is equal to or greater than a threshold, and the similarity between nodes without a SIMILAR_TO edge established is less than the threshold. A SIMILAR_TO edge is an edge established between similar nodes. Fig. 8 shows a specific example of a combination of node types corresponding to two nodes to which a SIMILAR_TO edge can be established.

[0023] The DB operation unit 140 has a function of operating the knowledge graph DB 190 .

[0024] The related information inference unit 150 selects a parent node from the first knowledge graph and the second knowledge graph based on the text indicated by the query, and searches for one or more nodes by tracing edges from the parent node in at least one of the first knowledge graph and the second knowledge graph. As a specific example, the related information inference unit 150 infers information related to a user's query roughly through the following two steps:

[0025] (Step 1: Node Acquisition) The related information inference unit 150 acquires a node in the knowledge graph that is most similar to the query entered by the user. In this case, the related information inference unit 150 may acquire one or more nodes corresponding to text that is relatively similar to the text indicated by the entered query by using natural language search, vector search, or the like. FIG. 9 is a diagram illustrating the processing of the related information inference unit 150. In the example shown in FIG. 9, the query indicates the text "The material of the blades was changed from metal to plastic." Nodes #123, #456, and #789 each correspond to a parent node. Here, node #123 is a node corresponding to the change "Blade material: metal ⇒ plastic." Node #456 is a node corresponding to the measure "The material was changed to Plastic A." Node #789 is a node corresponding to the measure "The material is hot-dip galvanized steel sheet...."

[0026] (Step 2: Graph Search) The related information inference unit 150 uses the node acquired in step 1 as the starting point (parent node) to acquire other nodes (child nodes) that can be traced from the parent node via a predetermined number of hops, as well as the paths from the parent node to each child node. Each child node corresponding to a parent node corresponds to an associated node corresponding to the parent node. By repeating step 2, the related information inference unit 150 can infer grandchild nodes from the child node as the starting point and trace associations. In the example shown in FIG. 9 , each node connected to each parent node is acquired. Here, node #062 corresponds to another node traced from each of node #123 and node #456, and is a node corresponding to the phenomenon "ultraviolet degradation." Node #554 corresponds to another node traced from node #789, and is a node corresponding to the phenomenon "material strength variation."

[0027] The relevance calculation unit 160 sets each of the one or more searched nodes as a target node and calculates the relevance corresponding to the target node based on the path from the parent node to the target node. The relevance calculation unit 160 calculates a score for each child node acquired by the related information inference unit 150 based on the path from the parent node to the child node. FIG. 10 is a diagram illustrating the processing of the relevance calculation unit 160. As a specific example, as shown in FIG. 10 , the relevance calculation unit 160 calculates a score corresponding to each node so that the closer the node is and the more paths that can be reached, the higher the relevance corresponding to the node. In the example shown in FIG. 10 , the relevance calculation unit 160 enumerates all possible paths from node S to node G, calculates the length of each enumerated path, calculates a score for each path based on the calculated length of each path, and aggregates the calculated scores to calculate the score corresponding to node G. FIG. 10 (a) is a diagram illustrating the path from node S to node G. FIG. 10 (b) is a table illustrating the processing for calculating the score corresponding to node G. In this example, the length of each edge is defined. The length of a path is the sum of the lengths of each edge on the path. The path score corresponding to each path is the inverse of the length of each path. This is because the longer the path (the farther the distance from node S to node G), the lower the path score. In Figure 10, the score of node G relative to node S is 1.66. The final score is also called the search score.

[0028] The display data generation unit 170 generates data displaying each node acquired by the related information inference unit 150 based on the score calculated by the relevance calculation unit 160. As a specific example, the display data generation unit 170 generates a screen displaying the acquired parent nodes and child nodes in descending order of their corresponding scores. FIGS. 11 and 12 show specific examples of screens generated by the display data generation unit 170. In FIG. 11, each parent node and each child node associated with each parent node are shown. As shown in FIG. 12, when a displayed child node is selected, a path from the parent node to the child node may be displayed. In FIG. 12, text corresponding to each node, a file containing each text, and the date on which each text was registered are shown.

[0029] The user interface unit 180 corresponds to a user interface and has functions such as accepting queries from users and outputting search results corresponding to the accepted queries.

[0030] The knowledge graph DB 190 is a database that stores data representing knowledge graphs.

[0031] 13 shows an example of the hardware configuration of the related information search device 100 according to this embodiment. The related information search device 100 is made up of a computer. The related information search device 100 may be made up of multiple computers.

[0032] As shown in the figure, the related information search device 100 is a computer that includes hardware such as a processor 11, a volatile storage device 12, a nonvolatile storage device 13, and an interface 14. These pieces of hardware are appropriately connected via signal lines.

[0033] The processor 11 is an integrated circuit (IC) that performs arithmetic processing and controls the hardware of a computer. Specific examples of the processor 11 include a central processing unit (CPU), a digital signal processor (DSP), or a graphics processing unit (GPU). The related information search device 100 may include multiple processors that replace the processor 11. The multiple processors share the role of the processor 11.

[0034] The volatile storage device 12 is also called a main storage device or a main memory, and a specific example thereof is a random access memory (RAM). The data stored in the volatile storage device 12 is saved in the non-volatile storage device 13 as needed.

[0035] The nonvolatile storage device 13 is also called an auxiliary storage device, and specific examples thereof include a ROM (Read Only Memory), an HDD (Hard Disk Drive), or a flash memory. Data stored in the nonvolatile storage device 13 is loaded into the volatile storage device 12 as needed. The volatile storage device 12 and the nonvolatile storage device 13 may be integrally configured.

[0036] The interface 14 has a function of communicating with other devices. The interface 14 may receive voice data and may transmit text data. Specific examples of the interface 14 include a communication chip or a NIC (Network Interface Card). Each unit of the related information search device 100 may use the interface 14 as appropriate when communicating with other devices.

[0037] The non-volatile storage device 13 stores a related information search program. The related information search program is a program that causes a computer to realize the functions of each unit included in the related information search device 100. The related information search program is loaded into the volatile storage device 12 and executed by the processor 11. The functions of each unit included in the related information search device 100 are realized by software.

[0038] Data used when executing the related information search program and data obtained by executing the related information search program are stored in a storage device as appropriate. Each part of the related information search device 100 uses a storage device as appropriate. Specific examples of the storage device include at least one of a volatile storage device 12, a nonvolatile storage device 13, a register in the processor 11, and a cache memory in the processor 11. Note that the terms "data" and "information" may have the same meaning. The storage device may be independent of the computer. The functions of the volatile storage device 12 and the nonvolatile storage device 13 may be realized by other storage devices.

[0039] The related information search program may be recorded on a computer-readable non-volatile recording medium. Specific examples of the non-volatile recording medium include an optical disk and a flash memory. The related information search program may be provided as a program product.

[0040] ***Explanation of Operation*** The operation procedure of the related information search device 100 corresponds to a related information search method. Also, the program that realizes the operation of the related information search device 100 corresponds to a related information search program.

[0041] 14 is a flowchart showing an example of the operation of the related information search device 100. The operation will be described with reference to FIG.

[0042] (Step S101) The knowledge acquisition unit 110 receives input of a group of documents and acquires knowledge from the received group of documents, i.e., generates each knowledge graph based on the received group of documents. The knowledge registration unit 120 registers each knowledge graph generated by the knowledge acquisition unit 110 in the knowledge graph DB 190 via the DB operation unit 140.

[0043] (Step S102) The static link unit 130 links semantically similar nodes together for each knowledge graph registered in the knowledge graph DB 190. The knowledge registration unit 120 updates the knowledge graph DB 190 via the DB operation unit 140 so that edges are established between the nodes linked by the static link unit 130.

[0044] (Step S103) The related information inference unit 150 receives a query input from the user and searches the knowledge graph DB 190 for a node that is most similar to the received query.

[0045] (Step S104) The related information inference unit 150 regards each node found in step S103 as a parent node, and acquires each related node corresponding to each parent node by tracing related nodes in order from each parent node.

[0046] (Step S105) The relevance calculation unit 160 calculates a score corresponding to each of the related nodes acquired by the related information inference unit 150 for each parent node.

[0047] (Step S106) The display data generation unit 170 displays the nodes on the screen in descending order of the scores calculated by the relevance calculation unit 160. At this time, the display data generation unit 170 may appropriately classify each node as either a parent node or a child node.

[0048] ***Description of the Effects of the First Embodiment*** In conventional technology, when a knowledge graph is constructed from each file, multiple nodes are treated as related only if there is no variation in notation between the multiple nodes or if the multiple nodes contain common keywords. However, in other cases, the multiple nodes are treated as unrelated nodes. In contrast, in this embodiment, a static link unit 130 is provided to calculate the similarity between nodes in multiple knowledge graphs constructed from each file, and link semantically similar nodes based on the calculated similarity. Note that this embodiment is not limited to in-house data, but can also be applied to public data. Furthermore, while this embodiment is described using data indicating designs and defects, this embodiment can also be applied to data other than designs and defects as long as the information can be represented by a knowledge graph.

[0049] ***Other Configurations*** <Variation 1> Fig. 15 shows an example of the hardware configuration of the related information retrieval device 100 according to this variation. The related information retrieval device 100 includes a processing circuit 18 instead of the processor 11, the processor 11 and the volatile storage device 12, the processor 11 and the non-volatile storage device 13, or the processor 11, the volatile storage device 12, and the non-volatile storage device 13. The processing circuit 18 is hardware that realizes at least a part of the components included in the related information retrieval device 100. The processing circuit 18 may be dedicated hardware, or may be a processor that executes a program stored in the volatile storage device 12.

[0050] When the processing circuitry 18 is dedicated hardware, the processing circuitry 18 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. The related information retrieval device 100 may include multiple processing circuits that replace the processing circuitry 18. The multiple processing circuits share the role of the processing circuitry 18.

[0051] In the related information search device 100, some of the functions may be realized by dedicated hardware, and the remaining functions may be realized by software or firmware.

[0052] The processing circuitry 18 is realized by, for example, hardware, software, firmware, or a combination of these. The processor 11, the volatile storage device 12, the non-volatile storage device 13, and the processing circuitry 18 are collectively referred to as the "processing circuitry." In other words, the functions of the functional components of the related information search device 100 are realized by the processing circuitry. Related information search devices 100 according to other embodiments may also have a configuration similar to that of this modified example.

[0053] Second Embodiment The following mainly describes the differences from the previous embodiment with reference to the drawings. The purpose of this embodiment is to appropriately remove noise caused by cross-sectional connections between multiple DBs.

[0054] 16 shows an example of the configuration of a related information search device 100 according to this embodiment. The related information search device 100 further includes a connection condition designation unit 210.

[0055] The connection condition specification unit 210 specifies connection conditions. The connection condition specification unit 210 is also referred to as a registration condition specification unit. A connection condition is a condition that two nodes to be connected must satisfy. The connection condition may be a condition regarding the combination of the node type of the first node and the node type of the second node. The connection condition specification unit 210 imposes constraints on the multiple nodes to be linked by the static link unit 130 according to a knowledge graph schema predefined by the user. These constraints correspond to connection conditions. The connection conditions are conditions for linking nodes that are semantically similar. As a specific example, in this embodiment, instead of calculating the similarity between all nodes and establishing edges, a restriction is imposed so that only nodes corresponding to node type combinations marked "SIMILAR_TO" in the table shown in FIG. 17 are candidates for establishing a SIMILAR_TO edge. FIG. 18 is a diagram corresponding to FIG. 17 and explains the processing of the static link unit 130 according to this embodiment. As a specific example, the static link unit 130 links nodes whose node type is Defect to each other according to the table shown in FIG.

[0056] The static link unit 130 according to this embodiment connects a first node and a second node when the first node and the second node satisfy a connection condition. As a specific example, when the nodes have similar or related node types, the static link unit 130 determines whether to establish an edge between the nodes by summarizing the similarities between the nodes, each of which has a similar or related node type. FIG. 19 is a diagram illustrating the processing of the static link unit 130. In FIG. 19 , each case corresponds to two nodes. In the example shown in FIG. 19 , "D (fault)" and "P (phenomenon)" are similar or related to each other. Therefore, for two nodes having node types "D (fault)" and "P (phenomenon)," the static link unit 130 establishes an edge only if both the faults and the phenomena are similar. In Cases 1 and 2 shown in FIG. 19 , the contents of the defects are not similar, so establishing an edge between the two nodes is likely to result in linking phenomena associated with completely different faults. In case 5, the contents of Phenomenon are not similar, so if an edge is established between the two nodes, there is a high possibility that faults associated with completely different phenomena will be linked. If edges are established between the two nodes corresponding to each of cases 1, 2, and 5, there is a high possibility that each established edge will become noise. Therefore, the static link unit 130 does not establish an edge between these two node pairs. In case 3 and case 4 shown in FIG. 19, the contents of Defect and Phenomenon are similar to each other, so the static link unit 130 establishes an edge between the two nodes.

[0057] ***Explanation of Operation*** Fig. 20 is a flowchart showing an example of the operation of the related information search device 100. This operation will be explained using Fig. 20.

[0058] (Step S201) The connection condition specification unit 210 reads a knowledge graph schema that has been defined in advance by the user. The read schema indicates connection conditions.

[0059] (Step S202) The static linking unit 130 links together semantically similar nodes that satisfy the connection conditions indicated by the loaded schema.

[0060] (Step S203) The knowledge registration unit 120 updates the knowledge graph DB 190 via the DB operation unit 140 so that edges are established between the nodes linked by the static link unit 130.

[0061] ***Explanation of the Effects of the Second Embodiment*** As described above, according to this embodiment, by imposing constraints on candidates for connecting edges, it is possible to exclude edges that should not be connected as use cases, even if there is a local semantic connection. Therefore, according to this embodiment, more effective searches become possible.

[0062] Embodiment 3. The following mainly describes the differences from the above-described embodiments, with reference to the drawings. Even when using a knowledge graph DB 190 consisting of knowledge graphs registered after noise removal, depending on the user's search conditions, the search may result in a mixture of information the user wants and information the user does not need. Therefore, the purpose of this embodiment is to enable the user to specify the node type or node type relationship they want to search.

[0063] 21 shows an example of the configuration of a related information search device 100 according to this embodiment. The related information search device 100 further includes a search condition designation unit 310.

[0064] The search condition specification unit 310 specifies search conditions. The search conditions are conditions that each node to be searched must satisfy. The search conditions may correspond to the node type of an intermediate node, and may correspond to the node type that the related information inference unit 150 should have after the intermediate node. The intermediate node is either a parent node or a node reached by tracing an edge from the parent node. As a specific example, the search condition specification unit 310 accepts, from the user, an instruction regarding the node type to be searched for or the relationship therebetween. As a specific example, by the search condition specification unit 310 accepting the specification of at least one of the node types of the parent node and the child node, relationships that are not necessary for the search can be explicitly excluded. The search condition specification unit 310 specifies the received instruction as a search condition. FIG. 22 is a diagram illustrating the processing of the search condition specification unit 310. The search condition specification unit 310 accepts the specification of the node type of the parent node and the node type of the child node.

[0065] The related information inference unit 150 according to this embodiment searches for each node that satisfies the search condition, that is, acquires each node according to the search condition specified by the search condition specification unit 310 .

[0066] ***Explanation of Operation*** Fig. 23 is a flowchart showing an example of the operation of the related information search device 100. The operation will be explained using Fig. 23.

[0067] (Step S301) The search condition designation unit 310 reads the node type of each node defined in advance by the user, and treats the read node type as a search condition.

[0068] (Step S302) This step is the same as step S103, except that the related information inference unit 150 searches for nodes according to the search conditions.

[0069] (Step S303) This step is the same as step S104, except that the related information inference unit 150 acquires each related node according to the search conditions.

[0070] ***Explanation of Effect of Third Embodiment*** As described above, according to this embodiment, the search targets are limited to nodes that meet the user's needs. Therefore, according to this embodiment, the user's desired information is more likely to be found.

[0071] Fourth Embodiment Hereinafter, differences from the above-described embodiments will be mainly described with reference to the drawings.

[0072] 24 shows an example of the configuration of the related information search device 100 according to this embodiment. The related information search device 100 further includes a dynamic link unit 410.

[0073] The dynamic linking unit 410 edits at least one of the first knowledge graph and the second knowledge graph based on at least a portion of the range traversed when searching for a target node. The dynamic linking unit 410 reads the paths traversed by the related information inference unit 150 and dynamically edits the knowledge graph based on the read paths. Specifically, the dynamic linking unit 410 filters each node or dynamically calculates the importance of each node. As another specific example, the dynamic linking unit 410 dynamically connects nodes or dynamically calculates the importance using information about the nodes and edges traversed by the related information inference unit 150 or the paths traversed by the related information inference unit 150 from a parent node to a child node or a grandchild node as a context, thereby thinning out some of the statically linked edges. In this case, the dynamic linking unit 410 may delete, from the edges linked by the static linking unit 130, edges other than those between nodes that have a similar relationship to the embedded subgraph including the paths traversed by the related information inference unit 150.

[0074] FIG. 25 is a diagram illustrating the processing of the dynamic linking unit 410. In the example shown in FIG. 25, it is assumed that the related information inference unit 150 traces from a parent node to a child node. Also, it is assumed that each node with a diagonal line drawn inside is a candidate for a grandchild node. At this time, the dynamic linking unit 410 extracts a subgraph including the periphery of each candidate grandchild node. Next, if a vector resulting from graph embedding corresponding to the extracted subgraph is similar to a vector resulting from graph embedding corresponding to the path traced by the related information inference unit 150, the dynamic linking unit 410 retains the grandchild node corresponding to the extracted subgraph. The processing of the dynamic linking unit 410 makes it possible to infer only measures that share common components or operations, as a specific example.

[0075] The related information inference unit 150 according to this embodiment may set a target reaching node as a search target if the similarity between the embedded representation of the subgraph corresponding to the target reaching node and the embedded representation of the subgraph corresponding to at least a part of the traced edge is equal to or greater than an expression similarity threshold. The target reaching node is each node that can be traced from a parent node. The expression similarity threshold may be determined in any manner.

[0076] When calculating the relevance, the relevance calculation unit 160 according to this embodiment uses the knowledge graph edited by the dynamic link unit 410. Specifically, when the first knowledge graph is edited, the relevance calculation unit 160 uses the edited first knowledge graph, and when the second knowledge graph is edited, the relevance calculation unit 160 uses the edited second knowledge graph.

[0077] ***Explanation of Operation*** Fig. 26 is a flowchart showing an example of the operation of the related information search device 100. The operation will be explained using Fig. 26.

[0078] (Step S401) The dynamic link unit 410 reads the path followed by the related information inference unit 150, and edits the knowledge graph based on the read path.

[0079] ***Explanation of Effects of Fourth Embodiment*** As described above, according to this embodiment, by dynamically editing the knowledge graph, it is possible to display search results with higher accuracy.

[0080] ***Other Embodiments*** The above-described embodiments can be freely combined, or any of the components of each embodiment can be modified, or any of the components can be omitted from each embodiment. Furthermore, the embodiments are not limited to those shown in embodiments 1 to 4, and various modifications are possible as needed. The procedures described using flowcharts, etc., can be modified as appropriate.

[0081] 11 Processor, 12 Volatile storage device, 13 Non-volatile storage device, 14 Interface, 18 Processing circuit, 100 Related information search device, 110 Knowledge acquisition unit, 120 Knowledge registration unit, 130 Static link unit, 140 DB operation unit, 150 Related information inference unit, 160 Relatedness calculation unit, 170 Display data generation unit, 180 User interface unit, 190 Knowledge graph DB, 210 Connection condition designation unit, 310 Search condition designation unit, 410 Dynamic link unit.

Claims

1. A related information search device comprising a static link unit that connects a first text corresponding to a first node included in a first knowledge graph corresponding to a document and a second text corresponding to a second node included in a second knowledge graph corresponding to the document when the object similarity between the first text and the second text is equal to or greater than a similarity threshold.

2. The related information search device according to claim 1, wherein the object similarity is a similarity between a vector corresponding to the first text and a vector corresponding to the second text.

3. The related information search device according to claim 1 or 2, wherein the static linking unit calculates the object similarity based on an edit distance.

4. The related information search device according to any one of claims 1 to 3 further comprises a connection condition specification unit that specifies connection conditions that must be satisfied by two connected nodes, and the static link unit connects the first node and the second node when the first node and the second node satisfy the connection conditions.

5. The related information search device according to claim 4, wherein the connection condition is a condition relating to a combination of a node type possessed by the first node and a node type possessed by the second node.

6. The related information search device according to any one of claims 1 to 5, further comprising: a related information inference unit that selects a parent node from the first knowledge graph and the second knowledge graph based on text indicated by a query, and searches for one or more nodes by tracing edges from the parent node in at least one of the first knowledge graph and the second knowledge graph; and a relevance calculation unit that sets each of the one or more searched nodes as a target node and calculates the relevance corresponding to the target node based on the path from the parent node to the target node.

7. The related information search device according to claim 6, further comprising a search condition specification unit that specifies search conditions that must be satisfied by each node to be searched, and the related information inference unit searches for each node that satisfies the search conditions.

8. The related information search device described in claim 7, wherein the search condition is a condition corresponding to the node type possessed by an intermediate node, which is either the parent node or a node reached by tracing an edge from the parent node, and is a condition corresponding to the node type that the related information inference unit should possess after tracing the intermediate node.

9. The related information search device further comprises a dynamic link unit that edits at least one of the first knowledge graph and the second knowledge graph based on at least a portion of the range traversed when searching for the target node, and the relevance calculation unit uses the edited first knowledge graph when the first knowledge graph is edited, and uses the edited second knowledge graph when the second knowledge graph is edited. The related information search device described in any one of claims 6 to 8.

10. A related information search device as described in claim 9, wherein when each node that can be traced from the parent node is defined as a target reaching node, the related information inference unit sets the target reaching node as a search target if the similarity between the embedded representation of the subgraph corresponding to the target reaching node and the embedded representation of the subgraph corresponding to at least a portion of the traced edge is equal to or greater than an expression similarity threshold.

11. A related information search method in which a computer connects a first node and a second node when an object similarity, which is the similarity between a first text corresponding to a first node included in a first knowledge graph corresponding to a document and a second text corresponding to a second node included in a second knowledge graph corresponding to the document, is equal to or greater than a similarity threshold.

12. A related information search program that causes a computer related information search device to execute static link processing that connects a first node and a second node when the object similarity, which is the similarity between a first text corresponding to a first node included in a first knowledge graph corresponding to a document and a second text corresponding to a second node included in a second knowledge graph corresponding to the document, is equal to or greater than a similarity threshold.

Images