Inter-event relationship graph construction device, inter-event relationship graph construction method, and recording medium

Abstract

In this inter-event relationship graph construction device, an acquiring means acquires event data from a text set. A graph construction means uses nodes and an edge to represent, as a graph, a pair of pieces of event data having a desired relationship from among the event data. A relationship generating means generates, for each node of the graph constructed by the graph construction means, a node corresponding to a consequent of each node by means of a machine learning model. A graph connecting means connects a pair of nodes that are determined to have the desired relationship from among pairs each formed of each node of the graph constructed by the graph construction means and each generated node. This inter-event relationship graph construction device constructs the graph to be able to assist decision making of a user.

Description

Inter-event relationship graph construction device, inter-event relationship graph construction method, and recording medium

[0001] The present disclosure relates to constructing an event relationship graph.

[0002] In recent years, knowledge graphs that represent knowledge in various fields have been developed and put to use. Non-Patent Document 1 proposes a language model that determines whether or not there is a causal relationship between any two sentences extracted from a document collection. Non-Patent Document 2 describes a method for generating a knowledge graph using a large language model (LLM).

[0003] Khetan Vivek, Ramnani Roshni, Anand Mayuresh, Sengupta Shubhashis, Fano Andrew E. "Causal BERT : Language models for causality detection between events expressed in text," 2020.West Peter, Bhagavatula Chandra, Hessel Jack, Hwang Jena D., Jiang Liwei, Bras Ronan Le, Lu Ximing, Welleck Sean, Choi Yejin. "Symbolic Knowledge Distillation: from General Language Models to Commonsense Models," 2021.

[0004] However, even with Non-Patent Document 1 and Non-Patent Document 2, there is a problem in that it is difficult to construct a consistent and effective graph.

[0005] One object of the present disclosure is to provide an event relationship graph construction device that can construct a consistent and effective graph structure using LLM.

[0006] In one aspect of the present disclosure, an event relationship graph construction device comprises: an acquisition means for acquiring event data from a text collection; a graph construction means for graphing pairs of event data having a desired relationship from among the event data using nodes and edges; a relationship generation means for generating, for each node of the graph constructed by the graph construction means, a node corresponding to the consequent of each node using a machine learning model; and a graph connection means for connecting each node of the graph constructed by the graph construction means with a pair of nodes determined to have a desired relationship from among pairs of each generated node.

[0007] In another aspect of the present disclosure, a method for constructing an event relationship graph executed by a computer includes: acquiring event data from a text collection; graphing pairs of event data having a desired relationship among the event data using nodes and edges; generating, for each node of the graph, a node corresponding to the consequent of each node using a machine learning model; and connecting pairs of nodes determined to have a desired relationship among pairs of each node of the graph and each generated node.

[0008] In yet another aspect of the present disclosure, a recording medium records a program that causes a computer to execute a process of acquiring event data from a text collection, graphing pairs of event data that have a desired relationship from the event data using nodes and edges, generating a node corresponding to the consequent of each node for each node of the graph using a machine learning model, and connecting pairs of nodes that are determined to have a desired relationship from among pairs of each node of the graph and each generated node.

[0009] According to the present disclosure, it is possible to construct a consistent and effective graph using LLM.

[0010] 1 shows the overall configuration of an inter-event relation graph construction system; FIG. 2 is a block diagram showing the hardware configuration of an inter-event relation graph construction device; FIG. 3 is a block diagram showing the functional configuration of an inter-event relation graph construction device; FIG. 4 is a diagram explaining a method for constructing a graph; FIG. 5 is a diagram explaining a method for constructing a graph; FIG. 6 is a flowchart of graph construction processing by an inter-event relation graph construction device; FIG. 7 is an example of input data; FIG. 8 is a display example of an inter-event relation graph; FIG. 9 is a block diagram showing the functional configuration of an inter-event relation graph construction device of a second embodiment; FIG. 10 is a flowchart of processing by an inter-event relation graph construction device of a second embodiment.

[0011] Preferred embodiments of the present disclosure will be described below with reference to the drawings. First Embodiment To facilitate understanding of the first embodiment, the background and issues underlying the first embodiment will be described in detail. When a company plans a business strategy and makes decisions, it may make predictions using various information. For example, a strategic consultant may predict industry trends, future trends, and needs, and propose strategic directions. To make such predictions, the strategic consultant constructs megatrend hypotheses. Here, a "megatrend hypothesis" refers to a broad graph structure in which events (current events and potential future events) and needs in a certain industry are linked in a chain, primarily through causal relationships. This graph structure is constructed starting from trends in the industry. Furthermore, this graph structure must satisfy as many requirements as possible, such as the presence of many nodes between the start point and the end point, ensuring the reliability of the information, and including a variety of causal relationships. Constructing megatrend hypotheses is primarily performed manually, requiring time and effort.

[0012] Even if Non-Patent Document 1 or Non-Patent Document 2 is used, the distance from the starting point to the end point is long, and it is difficult to construct a consistent graph. Furthermore, even with Non-Patent Document 1 or Non-Patent Document 2, it is difficult to construct a graph that guarantees the reliability of causal relationships or a graph that includes a variety of causal relationships.

[0013] [Overall Configuration] Fig. 1 shows the overall configuration of an inter-event relation graph construction system according to the first embodiment. Note that an inter-event relation graph refers to a graph relating to the relations between events. In this embodiment, causal relationships will be described as examples of the relations between events.

[0014] The event relationship graph construction system 1 comprises a user's terminal device 5 and an event relationship graph construction device 100. The user operates the terminal device 5 to send an event relationship graph generation request to the event relationship graph construction device 100. The generation request includes keywords related to the event relationship graph to be generated. The event relationship graph construction device 100 generates an event relationship graph from a document set (text set) prepared in advance based on the user's generation request, and transmits the graph to the user's terminal device 5.

[0015] 2 is a block diagram showing the hardware configuration of the event relationship graph construction device 100. As shown in the figure, the event relationship graph construction device 100 includes an interface (IF) 12, a processor 13, a memory 14, a recording medium 15, and a database (DB) 16.

[0016] The IF12 communicates with an external device to acquire a request to generate an event relationship graph input by the user. The IF12 also transmits to the external device the event relationship graph generated by the event relationship graph construction device 100. Specifically, when a user uses the terminal device 5 to transmit a request to generate an event relationship graph to the event relationship graph construction device 100, the event relationship graph construction device 100 receives the generation request via the IF12 and transmits the event relationship graph generated based on the generation request to the user's terminal device 5 via the IF12.

[0017] The processor 13 is a computer such as a CPU (Central Processing Unit), and controls the entire event relationship graph construction device 100 by executing a prepared program. Specifically, the processor 13 may be a CPU, a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof. Specifically, the processor 13 executes the event relationship graph construction process.

[0018] The memory 14 is composed of a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The memory 14 stores various programs executed by the processor 13. The memory 14 is also used as a working memory while the processor 13 is executing various processes.

[0019] The recording medium 15 is a non-volatile, non-transitory recording medium such as a disk-shaped recording medium or semiconductor memory, and is configured to be detachable from the event relationship graph construction device 100. The recording medium 15 records various programs to be executed by the processor 13. When the event relationship graph construction device 100 executes various processes, the programs recorded on the recording medium 15 are loaded into the memory 14 and executed by the processor 13.

[0020] DB 16 stores a document set used when constructing an event relationship graph. For example, the event relationship graph construction device 100 may create a document set based on URLs or text data of new news or papers input by the user, and store the document set in DB 16. The event relationship graph construction device 100 may also crawl a website, create a document set from documents on the website, and store the document set in DB 16.

[0021] The DB 16 may also store a machine learning model used in the graph construction process described below. The DB 16 may also store graph data generated by the event relationship graph construction device 100 as a graph database.

[0022] In addition to the above, the event relationship graph construction device 100 may also be equipped with a display device such as a liquid crystal display or a projector, and an input device such as a keyboard or a mouse. These display devices and input devices are used by, for example, an administrator of the event relationship graph construction device 100 to perform the necessary management.

[0023] 3 is a block diagram showing the functional configuration of the event relationship graph construction device 100. The event relationship graph construction device 100 functionally comprises a first causal relationship determination unit 21, a causal graph construction unit 22, a causal relationship generation unit 23, a second causal relationship determination unit 24, and a graph connection unit 25.

[0024] The user inputs keywords, domain names, and the like related to the event relationship graph that the user wants to generate into the event relationship graph construction device 100 .

[0025] The first causal relationship determination unit 21 acquires a predetermined document set from the DB 16 based on keywords, domain names, etc. input by the user. The first causal relationship determination unit 21 then extracts events from each document included in the predetermined document set. Here, an event is text indicating an event that has already occurred or an event that may occur in the future. The text indicating an event may be a sentence, phrase, or paragraph. The first causal relationship determination unit 21 then determines whether or not a causal relationship exists for each of a plurality of pairs of events using a causal relationship extraction model. For example, the first causal relationship determination unit 21 determines that a pair of events has a causal relationship if the score output by the causal relationship extraction model is equal to or greater than a predetermined threshold. The first causal relationship determination unit 21 outputs the determination result to the causal graph construction unit 22.

[0026] The causal relationship extraction model is a model that determines whether or not there is a causal relationship between any two events extracted from a document set, and can be created using, for example, BERT (Bidirectional Encoder Representations from Transformers) described in Non-Patent Document 1.

[0027] Furthermore, the natural language processing technology for determining whether or not there is a causal relationship between any two events is not limited to the technology using a language model such as BERT, etc. The method for determining the causal relationship may be any rule-based algorithm, machine learning algorithm, etc.

[0028] The causal graph construction unit 22 obtains the determination result from the first causal relationship determination unit 21. The causal graph construction unit 22 extracts pairs of events having a causal relationship from the determination result and constructs a graph. Specifically, the causal graph construction unit 22 represents the extracted events as nodes and constructs a directed graph by connecting nodes having a causal relationship with directional edges. The causal graph construction unit 22 outputs the constructed graph to the causal relationship generation unit 23.

[0029] Fig. 4 is a diagram illustrating a method for constructing a graph. In Fig. 4, first, the first causal relationship determination unit 21 extracts pairs of events from the information source 1 and the information source 2 included in the document collection, and determines whether or not a causal relationship exists for each pair of events. The determination result 41 in Fig. 4 indicates the result of determining whether or not a causal relationship exists for each pair of events extracted from the information source 1 and the information source 2.

[0030] Next, the causal graph construction unit 22 constructs graphs 42a, 42b, and 42c based on the determination result 41. In the graphs 42a, 42b, and 42c, each event is represented by a rectangular node. In addition, in the graphs 42a, 42b, and 42c, the causal relationships between each node are expressed by representing edges with unidirectional arrows. In the graphs 42a, 42b, and 42c, the origin of the arrow corresponds to the cause (CAUSE), and the end of the arrow corresponds to the effect (EFFECT). Hereinafter, the node corresponding to CAUSE will also be referred to as the "antecedent," and the node corresponding to EFFECT will also be referred to as the "consequent." In addition, the causal graph construction unit 22 connects the antecedent and the consequent when they are synonymous. Whether the antecedent and the consequent are synonymous can be determined using a synonymity determination model that determines synonymity of document pairs that has been learned in advance.

[0031] The above method uses documents on websites and other sources as information sources, making the source of information clear and enabling comprehensive extraction of diverse causal relationships. However, with the above method, if the topics between information sources differ slightly, it is difficult to grasp the causal relationships, resulting in the construction of an unconnected graph with multiple components. For example, in Figure 4, the first causal relationship determination unit 21 determines that there is no causal relationship between the events extracted from information source 1 and the events extracted from information source 2. Therefore, in Figure 4, an unconnected graph including graphs 42a and 42b is generated.

[0032] 3 , the causal relationship generation unit 23 acquires a graph from the causal graph construction unit 22. Then, the causal relationship generation unit 23 uses a large-scale language model to generate new consequent nodes with each node included in the graph as antecedent. The causal relationship generation unit 23 outputs a graph including the new consequent nodes (hereinafter also referred to as a “node-added graph”) to the second causal relationship determination unit 24.

[0033] A large-scale language model (also simply referred to as a "language model") is a model that learns the relationships between words in a sentence and is a machine learning model that generates related strings from a target string. By using a language model that has been trained with sentences and passages from various contexts, it is possible to generate related strings with appropriate content that are related to the target string.

[0034] For example, a case where a language model is used in question answering will be described. The language model receives an input question such as "What kind of country is Japan?" as a target string. The language model generates a string such as "Japan is an island country in the Northern Hemisphere..." as an answer to the question.

[0035] The learning method of the language model is not particularly limited, but as an example, the language model may be trained to output at least one sentence including an input string. As a specific example, the language model is a Generative Pretrained Transformer (GPT) that outputs a sentence including an input string by predicting a string that is likely to follow the input string. Other examples of language models include T5 (Text-to-Text Transfer Transformer), BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly optimized BERT approach), and ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately).

[0036] Furthermore, the character strings generated by the language model are not limited to natural languages. For example, the language model may output an artificial language (such as a mathematical formula or program source code) in response to a character string input in a natural language. For example, the language model may accept an input question such as "How do I retrieve data containing a specific character string from a database?" as a target character string. The language model may output program source code for performing database processing. Alternatively, the language model may output a natural language corresponding to a character string input in an artificial language.

[0037] Furthermore, the content generated by the language model is not limited to character strings. For example, the language model may generate image data, video data, audio data, or other data formats corresponding to the input character string.

[0038] In this embodiment, the large-scale language model is a model that generates new events that have a causal relationship with existing events, and can be created using the GPT or Critic model described above. A user can generate a new consequent node by presenting an instruction sentence called a "prompt" to the large-scale language model. The prompt includes "instructions" that describe the desired information and, if necessary, a "sample" that serves as a model.

[0039] The second causal relationship determination unit 24 acquires the node addition graph from the causal relationship generation unit 23. The second causal relationship determination unit 24 determines the presence or absence of a causal relationship for each pair of a new consequent node and a node included in the node addition graph, using the causal relationship extraction model described above. Note that the second causal relationship determination unit 24 may determine the presence or absence of a causal relationship for all pairs of all nodes included in the node addition graph.

[0040] If the second causal relationship determination unit 24 determines that there is a pair of nodes having a causal relationship, it outputs the determination result to the graph connection unit 25. The graph connection unit 25 connects the pair of nodes having a causal relationship with an edge based on the determination result input from the second causal relationship determination unit 24.

[0041] On the other hand, if the second causal relationship determination unit 24 determines that there is no node pair having a causal relationship, it outputs the determination result to the causal relationship generation unit 23. The causal relationship generation unit 23 again uses the large-scale language model to generate new consequent nodes with each node included in the node addition graph as antecedent. The causal relationship generation unit 23 then outputs a graph including the new consequent nodes (hereinafter also referred to as a "node re-addition graph") to the second causal relationship determination unit 24. The second causal relationship determination unit 24 determines the presence or absence of a causal relationship for each pair of the new consequent node generated again by the causal relationship generation unit 23 and a node included in the node re-addition graph using the above-mentioned causal relationship extraction model. If the second causal relationship determination unit 24 determines that there is a node pair having a causal relationship, it outputs the determination result to the graph connection unit 25. If it determines that there is no node pair having a causal relationship, it repeats the above process. The above process is repeated until it is determined that there is a pair of nodes that have a causal relationship, or until the process has been repeated a predetermined number of times.

[0042] 5 is a diagram illustrating a method for constructing a graph. In FIG. 5, the causal relationship generation unit 23 generates new consequent nodes 44a, 44b, and 44c for antecedents of nodes 43a, 43b, and 43c, respectively. Note that the causal relationship generation unit 23 generates new consequent nodes for each node included in the graph, but for convenience of illustration, only a portion of them is shown in FIG. 5.

[0043] Next, the second causal relationship determination unit 24 determines whether or not there is a causal relationship for each pair of the new consequent node and a node included in the graph, and generates a determination result 45. In FIG. 5 , the second causal relationship determination unit 24 determines that there is a causal relationship between node 44a and node 43d, and between node 44b and node 44c. Next, the graph connection unit 25 connects pairs of nodes having a causal relationship with edges. The graph connection unit 25 connects node 44a and node 43d with a dashed arrow 46a, and connects node 44b and node 44c with a dashed arrow 46b.

[0044] By using the above method, it is possible to complement the graphs 42a, 42b, and 42c constructed in FIG. 4, and to construct a consistent graph even when the distance from the starting node to the end node is long.

[0045] In the above configuration, the first causal relationship determination unit 21 is an example of an acquisition means, the causal graph construction unit 22 is an example of a graph construction means, the causal relationship generation unit 23 is an example of a relationship generation means, and the second causal relationship determination unit 24 and the graph connection unit 25 are examples of a graph connection means.

[0046] [Event Relation Graph Construction Process] Fig. 6 is a flowchart of the event relation graph construction process by the event relation graph construction device 100. This process is realized by the processor 13 shown in Fig. 2 executing a program prepared in advance and operating as each element shown in Fig. 3.

[0047] First, the event relationship graph construction device 100 acquires keywords, domain names, and the like related to the event relationship graph input by the user (step S21). Next, the first causal relationship determination unit 21 acquires a predetermined document set from the DB 16 based on the keywords, domain names, and the like. The first causal relationship determination unit 21 then extracts events from each document included in the predetermined document set and determines whether or not a causal relationship exists for each pair of events using a causal relationship extraction model (step S22). The first causal relationship determination unit 21 outputs the determination results to the causal graph construction unit 22. Next, the causal graph construction unit 22 extracts pairs of events having a causal relationship from the determination results and constructs a graph (step S23). The causal graph construction unit 22 outputs the constructed graph to the causal relationship generation unit 23.

[0048] Next, the causal relationship generation unit 23 uses a large-scale language model to generate new consequent nodes with each node included in the graph as antecedent (step S24). The causal relationship generation unit 23 outputs the node addition graph to the second causal relationship determination unit 24. Next, the second causal relationship determination unit 24 determines whether or not a causal relationship exists for each pair of the new consequent node and a node included in the node addition graph using the causal relationship extraction model (step S25). If the second causal relationship determination unit 24 determines that there is no node pair with a causal relationship (step S26: No), the process returns to step S24. On the other hand, if it determines that there is a node pair with a causal relationship (step S26: Yes), the second causal relationship determination unit 24 outputs the determination result to the graph connection unit 25. Based on the determination result, the graph connection unit 25 connects the node pairs with a causal relationship with an edge (step S27). Then, the event relationship graph construction process ends.

[0049] In the above embodiment, a graph is constructed based on the causal relationships between events, but the relationships between events used to construct the graph are not limited to causal relationships. For example, any relationship, such as an order relationship, in which the antecedent and the consequent cannot be switched may be used.

[0050] (Modification 2) The event relationship graph construction device of the above embodiment may be applied to construction of an event relationship graph related to medicine or healthcare. The event relationship graph construction device can organize the latest medical information by adding the latest medical information to an existing event relationship graph.

[0051] For example, a user inputs the URL or text data of new medical news or papers into the event-to-event relation graph construction device and requests the generation of a new event-to-event relation graph. FIG. 7 shows an example of medical input information. The event-to-event relation graph construction device generates a new event-to-event relation graph based on documents and text data from linked websites. The event-to-event relation graph construction device then determines whether or not there is a causal relationship between each node in the newly generated event-to-event relation graph and each node in an existing event-to-event relation graph stored in DB 16. The event-to-event relation graph construction device adds the newly generated event-to-event relation graph to the existing event-to-event relation graph by connecting pairs of nodes that have a causal relationship with an edge.

[0052] In addition, if there is no causal relationship between each node of the newly generated inter-event relationship graph and each node of the existing inter-event relationship graph, the inter-event relationship graph construction device may generate a new consequent node with each node as the antecedent, and determine the causal relationship again.

[0053] FIG. 8 shows an example of the display of an event-to-event relation graph output by the event-to-event relation graph construction device. The display screen 50 includes a graph display section 51 and a search section 52. The graph display section 51 is an area for displaying the event-to-event relation graph. The search section 52 is an area for specifying search conditions. By specifying search conditions, a user can search for information related to a specific keyword or topic from the event-to-event relation graph. For example, in FIG. 8, the language model button 53 is selected. When the user selects the language model button 53, nodes generated by the large-scale language model are displayed in a different form (gray) from other nodes. Note that the display form of the search results is not limited to the above, and for example, only the graph portion generated by the large-scale language model may be displayed. Furthermore, the user may select a type of directed edge, such as a citation relationship or a reference relationship, to change the display form of the corresponding graph portion, or only the corresponding graph portion may be displayed.

[0054] As described above, by adding the latest medical information to an existing event relationship graph, users can organize the latest medical information and expand the relationships between events by continuously updating the medical information.

[0055] 9 is a block diagram showing the functional configuration of an event relationship graph construction device 70 according to a second embodiment. The event relationship graph construction device 70 includes an acquisition unit 71, a graph construction unit 72, a relationship generation unit 73, and a graph connection unit 74.

[0056] 10 is a flowchart of processing by the event relationship graph construction device 70 of the second embodiment. The acquisition means 71 acquires event data from a text collection (step S71). The graph construction means 72 graphs pairs of event data having a desired relationship among the event data using nodes and edges (step S72). The relationship generation means 73 generates nodes corresponding to the consequents of each node of the graph constructed by the graph construction means using a machine learning model (step S73). The graph connection means 74 connects pairs of nodes determined to have a desired relationship among the pairs of nodes of the graph constructed by the graph construction means and the generated nodes (step S74).

[0057] The event relationship graph construction device 70 of the second embodiment makes it possible to construct an effective graph that maintains consistency using the LLM. Therefore, the event relationship graph construction device 70 can support the user in making decisions based on the megatrend hypothesis.

[0058] A part or all of the above-described embodiments can be described as, but not limited to, the following supplementary notes.

[0059] (Supplementary Note 1) An event-to-event relationship graph construction device comprising: an acquisition means for acquiring event data from a text collection; a graph construction means for graphing pairs of event data having a desired relationship from among the event data using nodes and edges; a relationship generation means for generating, for each node of the graph constructed by the graph construction means, a node corresponding to the consequent of each node using a machine learning model; and a graph connection means for connecting pairs of nodes determined to have a desired relationship from among pairs of each node of the graph constructed by the graph construction means and each generated node.

[0060] (Supplementary Note 2) In the event that it is determined that there is no desired relationship between each node of the graph constructed by the graph construction means and each of the generated nodes, the relationship generation means further generates a node corresponding to the consequent of each node for each node of the graph constructed by the graph construction means or for each of the generated nodes, in the event relationship graph construction device described in Supplementary Note 1.

[0061] (Supplementary Note 3) The event relationship graph construction device according to Supplementary Note 1, wherein the relationship generation means additionally generates nodes corresponding to the consequents of each node of the graph constructed by the graph construction means or for each of the generated nodes, until the number of times that nodes have been generated exceeds a predetermined number of times or until it is determined that a desired relationship exists.

[0062] (Supplementary Note 4) An event relationship graph construction device according to Supplementary Note 1, comprising a storage means for storing an existing graph, wherein the graph connection means connects pairs of nodes determined to have a desired relationship among pairs of nodes of the existing graph and nodes of the graph constructed by the graph construction means, thereby expanding the existing graph.

[0063] (Supplementary Note 5) An event relationship graph construction device according to Supplementary Note 1, comprising: a display means for displaying a graph generated by the graph connection means by connecting pairs of nodes; and a receiving means for receiving instructions from a user regarding the display mode, wherein the display means switches the display mode of the graph based on the instructions regarding the display mode input by the user.

[0064] (Supplementary Note 6) The event relationship graph construction device according to Supplementary Note 5, wherein the instructions regarding the display mode include an instruction to visualize only specified nodes and specified edges, or an instruction to change the display mode of specified nodes and specified edges.

[0065] (Supplementary Note 7) The event relationship graph construction device according to Supplementary Note 4 or 5, wherein the text set is a text set related to medical information, and the graph constructed by the graph construction means and the graph updated by the graph connection means include event relationships related to medical care.

[0066] (Supplementary Note 8) An event relationship graph construction method executed by a computer, comprising: acquiring event data from a text collection; graphing pairs of event data having a desired relationship among the event data using nodes and edges; generating, for each node of the graph, a node corresponding to the consequent of each node using a machine learning model; and connecting pairs of nodes determined to have the desired relationship among each node of the graph and each of the generated nodes.

[0067] (Supplementary Note 9) A recording medium having recorded thereon a program that causes a computer to execute the following process: acquire event data from a text collection; graph pairs of event data having a desired relationship among the event data using nodes and edges; generate, for each node of the graph, a node corresponding to the consequent of each node using a machine learning model; and connect pairs of nodes that are determined to have a desired relationship among the pairs of each node of the graph and each of the generated nodes.

[0068] Although the present disclosure has been described above with reference to the embodiments and examples, the present disclosure is not limited to the above-described embodiments and examples. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure.

[0069] REFERENCE SIGNS LIST 1 Event relation graph construction system 5 Terminal device 21 First causal relation determination unit 22 Causal graph construction unit 23 Causal relation generation unit 24 Second causal relation determination unit 25 Graph connection unit 100 Event relation graph construction device

Claims

1. An event relationship graph construction device comprising: an acquisition means for acquiring event data from a text collection; a graph construction means for graphing pairs of event data having a desired relationship from among the event data using nodes and edges; a relationship generation means for generating, for each node of the graph constructed by the graph construction means, a node corresponding to the consequent of each node using a machine learning model; and a graph connection means for connecting each node of the graph constructed by the graph construction means with a pair of nodes determined to have a desired relationship from among pairs of each node of the generated nodes.

2. The event relationship graph construction device according to claim 1, wherein, when it is determined that there is no desired relationship between each node of the graph constructed by the graph construction means and each of the generated nodes, the relationship generation means further generates additional nodes corresponding to the consequent of each node for each node of the graph constructed by the graph construction means or for each of the generated nodes.

3. The event relationship graph construction device according to claim 1, wherein the relationship generation means generates additional nodes corresponding to the consequent of each node for each node of the graph constructed by the graph construction means or for each of the generated nodes until the number of times the node has been generated exceeds a predetermined number of times or until it is determined that a desired relationship exists.

4. An event relationship graph construction device as described in claim 1, further comprising a storage means for storing an existing graph, wherein the graph connection means connects pairs of nodes determined to have a desired relationship among pairs of nodes of the existing graph and nodes of the graph constructed by the graph construction means, thereby expanding the existing graph.

5. An event relationship graph construction device as described in claim 1, comprising: a display means for displaying a graph generated by the graph connection means by connecting pairs of nodes; and a receiving means for receiving instructions from a user regarding the display mode, wherein the display means switches the display mode of the graph based on the instructions regarding the display mode input by the user.

6. The event relationship graph construction device according to claim 5, wherein the instructions regarding the display mode include an instruction to visualize only specified nodes and specified edges, or an instruction to vary the display mode of specified nodes and specified edges.

7. The event relationship graph construction device according to claim 4 or 5, wherein the set of texts is a set of texts relating to medical information, and the graph constructed by the graph construction means and the graph updated by the graph connection means include event relationship relating to medical care.

8. An event relationship graph construction method executed by a computer, comprising: acquiring event data from a text collection; graphing pairs of event data having a desired relationship among the event data using nodes and edges; generating, for each node of the graph, a node corresponding to the consequent of each node using a machine learning model; and connecting pairs of nodes determined to have the desired relationship among each node of the graph and each of the generated nodes.

9. A recording medium having recorded thereon a program that causes a computer to execute the process of acquiring event data from a text collection, graphing pairs of event data having a desired relationship from among the event data using nodes and edges, generating for each node of the graph a node that corresponds to the consequent of each node using a machine learning model, and connecting pairs of nodes that are determined to have the desired relationship from among pairs of each node of the graph and each of the generated nodes.

Images