Analysis device, analysis method, and program

Abstract

The present disclosure concerns an analysis device for analyzing the relevance between a first sentence to be analyzed and a second sentence to be analyzed. The analysis device comprises: a generation unit that executes a first process for generating a first causal graph by generating a first factor node indicating a factor and a first result node indicating the result in a causality indicated by the first sentence to be analyzed and connecting the first factor node to the first result node by a directed edge, and executes a second process for generating a second causal graph by generating a second factor node indicating a factor and a second result node indicating the result in a causality indicated by the second sentence to be analyzed and connecting the second factor node to the second result node by a directed edge; and a formation unit that forms a causality network based on the first sentence to be analyzed and the second sentence to be analyzed by connecting specific nodes to each other with non-directed edges if there are specific nodes within a prescribed similarity range in the first causal graph and the second causal graph.

Description

Analysis apparatus, analysis method and program

[0001] The present disclosure relates to a technique for analyzing relevance between two sentences to be analyzed.

[0002] In recent years, unpredictable situations such as climate change and regional conflicts have occurred, diversifying risks, impacts, and opportunities for corporations, and thus there is a concern that management may not be able to fully respond to unforeseen circumstances. Against this background, utilization of "scenario analysis" for predicting a plurality of uncertain future scenarios is required.

[0003] The elements and procedures required for this scenario analysis are as follows.

[0004] First, a focal question is set. A focal question is an important issue that should be focused on in relation to a company's management vision and objectives. It is set by the company's management. The content of this focal question may have a great impact in the future but involves uncertainty; that is, although it has an impact on the company, there are a plurality of options and possibilities, and various results can be expected.

[0005] Scenario analysis is to create a plurality of scenarios starting from this focal question. For example, assume a case where a question (important issue) "Does climate change affect our company?" is taken as a focal question. Since this is expected to have results that depend on various factors, it is a question involving uncertainty. For this question, there are factors or states that determine future trends, which are called "driving forces".

[0006] An example of a "driving force", which is a factor that greatly affects the above-mentioned focal question "Does climate change affect our company's business?", is as follows: (1) Spread of renewable energy (2) Increase in the number of high-temperature days (3) Increase in heavy rain For example, (1) "Spread of renewable energy" becomes a factor for an increase in electricity prices (assuming the current power transmission and distribution system remains unchanged), and becomes a significant burden on management. Conversely, if a company runs a business using renewable energy, the increased demand becomes an opportunity.

[0007] (2) The increase in the number of high-temperature days will increase the amount of electricity and equipment needed for air conditioning, which will be a burden on businesses. Conversely, if air conditioning is a business, it will be an opportunity.

[0008] (3) The increase in heavy rainfall increases the probability of equipment being damaged, which puts a burden on businesses. Conversely, if one is in the construction business, the demand for heavy rain countermeasures presents an opportunity.

[0009] Thus, in predicting the future based on the focal question, it is crucial to identify the "driving forces"—the major factors that will influence the future. Since these driving forces often lie outside the areas (and industries) in which the company is directly involved, it is desirable to comprehensively research, analyze, and extract information from each field.

[0010] Typically, this exploratory work (identifying related elements) is carried out by someone in the company's corporate planning department. Identifying the driving forces and related elements requires the person conducting the scenario analysis (the user) to read through a large amount of material and spend considerable time and effort researching, analyzing, or considering the matter themselves. Furthermore, this exploratory work must be comprehensive and cover a wide range of fields, without being biased towards any particular area.

[0011] However, it is difficult for users (humans) to find connections between content discussed from different fields or perspectives (for example, a microeconomic perspective and a macroeconomic perspective). For instance, it is difficult to find a multifaceted (multi-field) connection between a company's business vision from a micro perspective and climate change from a macro perspective. In the realm of scenario analysis, it is difficult to find a connection between external factors in a focal question and internal factors in a company's business vision and mission.

[0012] Thus, while a single user may be able to extract driving forces and their elements (nodes, node relationships) within the industry or field of knowledge they are familiar with, extracting driving forces from outside their area of ​​expertise is practically difficult. As a solution, it has been proposed to gather a large group of experts from various fields to explore driving forces and their elements (Non-Patent Literature 1).

[0013] Kuniko Urashima et al., "The Future Society Aimed for by Local Communities and Carbon Neutrality / Full Report," Research Center for Science and Technology Foresight and Policy Infrastructure, National Institute of Science and Technology Policy, Ministry of Education, Culture, Sports, Science and Technology, December 2023, <https: / / nistep.repo.nii.ac.jp / record / 2000061 / files / NISTEP-RM334-FullJ.pdf>

[0014] However, the method described in Non-Patent Document 1 requires gathering a large group of experts from various fields, which is costly and time-consuming. Therefore, it is difficult for individual companies, including small and medium-sized enterprises, to conduct uncertain future scenario analyses regarding their operations.

[0015] This disclosure aims to address the aforementioned circumstances by enabling relatively easy analysis of the relationships between multiple analysis texts when conducting tasks such as analyzing uncertain future scenarios related to business operations.

[0016] To achieve the above objective, the present disclosure provides an analysis device for analyzing the relationship between a first sentence to be analyzed and a second sentence to be analyzed, comprising: a generation unit that generates a first factor node indicating a factor and a first result node indicating a result in the causality shown by the first sentence to be analyzed, and performs a first process to generate a first causal graph by connecting the first factor node to the first result node with directed edges; a second process that generates a second factor node indicating a factor and a second result node indicating a result in the causality shown by the second sentence to be analyzed, and performs a second process to generate a second causal graph by connecting the second factor node to the second result node with directed edges; and a shaping unit that, if there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similar range, shapes a causal network based on the first sentence to be analyzed and the second sentence to be analyzed by connecting the specific nodes with undirected edges.

[0017] As explained above, this disclosure has the effect of making it possible to analyze the relationships between multiple sentences under analysis relatively easily.

[0018] This is an overall configuration diagram of the communication system according to the embodiment. This is an electrical hardware configuration diagram of the analysis device and input / output terminal. This is a functional configuration diagram of the analysis device according to the first embodiment. This is a conceptual diagram showing the minimum configuration of a causal graph. This is a conceptual diagram showing the state in which similarity is determined between objects of the same type included in predetermined nodes in two causal graphs. This is a flowchart showing the analysis method of the relationship between the analysis target sentence on the macro perspective and the analysis target sentence on the micro perspective according to the first embodiment. This is a flowchart showing the analysis method of the relationship between the analysis target sentence on the macro perspective and the analysis target sentence on the micro perspective according to the first embodiment. This is a flowchart showing the analysis method of the relationship between the analysis target sentence on the macro perspective and the analysis target sentence on the micro perspective according to the first embodiment. This is a conceptual diagram showing a causal network according to the first embodiment. This is a conceptual diagram showing the path of the causal network extended from Figure 9. This is a conceptual diagram showing the path of the causal network further extended from Figure 10. This is a diagram showing an example of an analysis result screen. This is a conceptual diagram showing the state in which intermediate variable nodes are generated in the summary sentence in addition to the first embodiment according to the second embodiment. This is a conceptual diagram showing a modified version of Figure 13 according to the second embodiment. This is a conceptual diagram relating to the third embodiment, showing a state in which, in addition to the first embodiment, objects for direct and indirect factors are created in the factor node. This is a conceptual diagram relating to the third embodiment, showing a state in which, in addition to the first embodiment, objects for short-term and long-term results are created in the result node. This is a diagram showing a specific example of a causal network.

[0019] Embodiments of the present invention will be described below with reference to the drawings. However, the present invention is not limited to the embodiments shown below, and various modifications are possible without departing from the technical spirit of the invention. Since the drawings are for conceptual explanation of the present invention, dimensions, ratios, or numbers may be exaggerated or simplified as necessary for ease of understanding.

[0020] ●The first embodiment will be described using Figures 1 to 12.

[0021] [System Configuration of the Embodiment] First, the overall configuration of the communication system according to the embodiment will be described using Figure 1. Figure 1 is an overall configuration diagram of the communication system according to the embodiment.

[0022] As shown in Figure 1, the communication system 10 of this embodiment is constructed from an analysis device 30 and an input / output terminal 60.

[0023] The analysis device 30 and the input / output terminal 60 are composed of one or more computers. The analysis device 30 and the input / output terminal 60 are connected to each other via a communication network 100 such as the Internet. The connection method may be wired or wireless. The input / output terminal 60 is a terminal that transmits various input information to the analysis device 30 and receives various output information. If the analysis device 30 is composed of multiple computers, it may be referred to as "analysis device" or as "analysis system".

[0024] For example, when a company's management team presents a future scenario to stakeholders, the analysis device 30 automatically analyzes the relationship between macro-level analysis targets (such as future key issues and focal questions (FQs)) and micro-level analysis targets (such as the company's business vision), clarifying the relationship between the two and proposing it in the form shown in Figure 12. Therefore, the analysis device 30 may also be referred to as a "proposal device (proposal system)" that proposes the relationship. Alternatively, the analysis device 30 may also be referred to as an "exploration device (exploration system)" that explores the relationship.

[0025] The analysis device 30 is used in the following scenarios. It is highly likely that macro-level information and micro-level information, which at first glance seem unrelated, are connected through several elements. For example, smartphones and African children may seem unrelated at first glance, so it is unlikely that stakeholders will understand the connection between smartphones and African children. However, the raw materials for tantalum capacitors used in smartphones are mainly imported from Africa, and child labor at the mining sites for these raw materials is a social problem. In another example, customer centers and nuclear power plants may seem unrelated at first glance, but the generational AI (Artificial Intelligence) that is expected to be used in the automation of customer centers in the future consumes a huge amount of electricity, so the spread of small nuclear power plants is expected. In this way, by using different elements from several fields, it is possible to find the relationship between two things (the sentences to be analyzed) relatively easily.

[0026] Furthermore, a party that presents future scenarios for another company, rather than their own, may use the analysis device 30 to automatically analyze the relationship between macro-level analysis target statements (future key issues, FQs, etc.) and micro-level analysis target statements (other company's business vision, etc.) as part of the content of the future scenario.

[0027] [Hardware Configuration] Next, the electrical hardware configuration of the analysis device 30 will be explained using Figure 2. Figure 2 is an electrical hardware configuration diagram of the analysis device and input / output terminals.

[0028] For example, as shown in Figure 2, the analysis device 30 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a processor 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., which are all interconnected by a bus 1010.

[0029] The program that enables processing on the computer is executed using a recording medium 1001, such as a CD-ROM or memory card. When the recording medium 1001 containing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001; it may also be downloaded from another computer via a communication network such as the Internet. The auxiliary storage device 1002 stores the installed program as well as necessary files and data.

[0030] When a program startup command is received, the memory device 1003 reads the program from the auxiliary storage device 1002 and stores it. The processor 1004 implements the functions related to the memory device 1003 according to the program stored in the memory device 1003. The processor 1004 may include not only a CPU (Central Processing Unit) but also a GPU (Graphics Processing Unit).

[0031] The interface device 1005 is used as an interface for connecting to a communication network, etc. The display device 1006 displays a GUI (Graphical User Interface) or the like, programmed by the user. The input device 1007 consists of a keyboard and mouse, buttons, or a touch panel, and is used to input various operation instructions. The output device 1008 outputs the calculation results to an external device (input / output terminal 60, external display device, printer, etc.).

[0032] Since the input / output terminal 60 has the same configuration as the analysis device 30, a detailed explanation of these terminals will be omitted.

[0033] [Functional Configuration of the Embodiment] Next, the functional configuration of the analytical apparatus 30 of this embodiment will be explained using Figure 3.

[0034] As shown in Figure 3, the analysis device 30 includes an acquisition unit 31, a summarization unit 32, an object creation unit 33, a causal graph generation unit 34, a vectorization unit 35, a determination unit 36, a search unit 37, a formatting unit 39, a result creation unit 40, and an output unit 41. Each of these units is a function realized by instructions from the processor 1004 in Figure 2 based on a program. The analysis device 30 also has a storage unit 50 constructed from an auxiliary storage device 1002 or a memory device 1003. As shown in Figure 3, the storage unit 50 stores a related sentence management DB (Data Base) 51. The storage unit 50 also stores trained machine learning models m1 and m2.

[0035] <Related Sentence Management DB> The related sentence management DB 51 manages multiple related sentences acquired by the acquisition unit 31 via the communication network 100. Each related sentence contains information on multiple fields such as STEEPL (Sociological, Technological, Economic, Environmental, Political, Legal). In some cases, a single related sentence may contain information on multiple fields (for example, Sociological and Technological).

[0036] The related sentence management DB 51 may be built on an external device (such as a database server) separate from the analysis device 30, rather than within the analysis device 30. In this case, the output unit 41 of the analysis device 30 accesses the external device to store related sentences, or the search unit 37 of the analysis device accesses the external device to search for and retrieve related sentences.

[0037] <Machine Learning Model m1> Machine learning model m1 is a neural network that uses text summarization technology to generate and output a summary of a text or sentence (such as a section or chapter of a document) as input data. An example of this is the Large Language Model (LLM).

[0038] <Machine Learning Model m₂> The machine learning model m₂ generates feature vectors for each object included in each node in a causal graph as shown in Figure 4 as embeddings of text representations. For example, commonly used libraries such as Sentence-BERT can be mentioned.

[0039] <Each Functional Unit> Next, each functional unit included in the analysis device 30 will be described.

[0040] (Acquisition Unit) The acquisition unit 31 acquires related sentences (related texts) from a plurality of websites via the communication network 100, and stores them in the related sentence management DB 51. The acquisition unit 31 also acquires an analysis target sentence a₀ on the macro perspective side and an analysis target sentence b₀ on the micro perspective side from the input / output terminal 60.

[0041] (Summarization Unit) The summarization unit 32 uses the LLM m₁ to output a summary of a text or sentence based on the text or sentence. For example, the summarization unit 32 inputs the following prompt and format into the LLM m₁ to generate a summary sentence.

[0042] Example prompt: "Read the following text and summarize it based on the following format. Fill in appropriate content in the parentheses, and clearly distinguish the affected object (subject) A1 from the resulting object (subject) B1." Example format: The state (action) X1 of the object (subject) A1 is affected by the state (action) Z1 of the object (subject) C1, and the object (subject) B1 becomes the state (action) Y1.

[0043] For example, it is desirable to supplement the state (action) X1 with quantitative values and specific examples as much as possible, such as "cloud utilization increased by 25% year-on-year" or "the number of new entrants increased by 50% in the past five years".

[0044] Although the above prompt and format are only examples, summarization can be performed with such a format.

[0045] Furthermore, when LLMm1 is caused to perform causal inference itself, the occurrence of hallucinations becomes a problem; however, as in the present embodiment, by using LLMm1 only for generating summary sentences and further using a format that facilitates summarization, it is possible to minimize the occurrence of hallucinations. In addition, by organizing sentences (documents) through summarization, the amount of data to be processed can be reduced.

[0046] (Object Creation Unit) The object creation unit 33 creates each object (target (or subject), state (or action), and others) in a summary sentence. This "others" indicates time periods such as "recent years" and "the COVID-19 pandemic", regions such as "Middle East" and "Asia", and the like. In addition, "others" indicates quantitative information such as "temperature increased by 2°C" and "population fell below 100 million". Note that "others" is not necessarily required.

[0047] Furthermore, an example of creating each object of nodes from a summary sentence is described below.

[0048] (1) Summary sentence example 1: The state {facing soaring prices} of the target {low-income group} is affected by the action {tax rate increase} of {the government}, and the situation or state of the target {quality of life} becomes {further deteriorated}. <Object of factor node> Target A1: low-income group, State X1: facing soaring prices <Object of intervention node> Target C1: government, Action Z1: tax rate increase <Object of result node> Target B1: quality of life, State Y1: further deteriorated

[0049] (2) Summary sentence example 2: "A young tree growing in dry soil is affected by strong sunlight, and its leaves have begun to wilt." <Object of factor node> Target A1: young tree, State X1: growing in dry soil <Object of intervention node> Target C1: natural environment, Action Z1: strong sunlight <Object of result node> Target B1: leaves, State Y1: begun to wilt

[0050] (3) Example summary 3: "Immune cells that were not fully developed were affected by a viral infection, and the patient's condition worsened." <Objects in the Factor node> Target A1: Immune cell status X1: Not fully developed <Objects in the Intervention node> Target C1: Viral status Z1: Viral infection <Objects in the Result node> Target B1: Health condition Y1: Worsened

[0051] Through the above processing, the object creation unit 33 creates each object (subject, state, action, etc.) in the summary sentence s10 of the sentence to be analyzed a0, and creates each object (subject, state, action, etc.) in the summary sentence s20 of the sentence to be analyzed b0.

[0052] Furthermore, the object creation unit 33 creates each object (subject, state (action), etc.) in summary sentences such as related sentence a1 of the analysis target sentence a0 described later, and creates each object (subject, state (action), etc.) in summary sentences such as related sentence b1 of the analysis target sentence b0 described later.

[0053] (Causal Graph Generation Unit) As shown in Figure 4, the causal graph generation unit 34 creates each node (factor node, intervention node, and connection node) containing each object in the summary text, and connects each node with directional directed edges to create a causal graph. Here, each node will be explained in detail. Note that while Reference 1 shows a case where the causal graph is composed of two nodes, "factor" and "result," in this embodiment it is composed of three nodes: "factor," "intervention," and "result." <Reference 1> Kobayashi, Izumi et al. "Construction of a Causal Chain Presentation System Considering Output Diversity for the Discovery of Economic Ripple Effects," 2023 Natural Language Processing Society of Japan, 2L4-GS-3-01, March 2023.<https: / / www.anlp.jp / proceedings / annual_meeting / 2023 / pdf_dir / D11-3.pdf> Here, we will use Figure 4 to explain the minimum configuration of a causal graph. Figure 4 is a conceptual diagram showing the minimum configuration of a causal graph.

[0054] As shown in Figure 4, in the summary text s, the causal graph is composed of a factor node na, an intervention node nc, and an outcome node nb. Furthermore, the intervention node nc and the factor node na are connected by a directed edge that is directional from the intervention node nc to the factor node na. The factor node na and the outcome node nb are also connected by a directed edge that is directional from the factor node na to the outcome node nb.

[0055] A "factor node" is an element that indicates a cause in a causal relationship described by a sentence and has a causal influence on the "result node." "Factor" can be interpreted as "cause." Alternatively, it can be considered as the initial state in a causal relationship.

[0056] An "outcome node" is an element that represents the result in a causal relationship described by a sentence and is causally influenced by "factor nodes" and "intervention nodes."

[0057] An "intervention node" indicates an intervention in a causal relationship shown by a sentence, representing a factor that changes the outcome. In the causal relationship formed by the "factor node" and the "outcome node," it is an element that actively manipulates the target system, phenomenon, or other events or variables (actions, measures) and influences the outcome node.

[0058] Note that "summary sentence s" is a collective term for the summary sentences s10, s11, s12, s20, s21, and s22 described later. Also, "factor node na" is a collective term for the factor nodes na10, na11, na12, na20, na21, and na22 described later. "Result node nb" is a collective term for the result nodes nb10, nb11, nb12, nb20, nb21, and nb22 described later. "Intervention node nc" is a collective term for the result nodes nc10, nc11, nc12, nc20, nc21, and nc22 described later.

[0059] The essential objects of the factor node na are A1 (the subject) and X1 (the state / action), and the entire factor node is represented as A1(X1). The factor node na may also have another object, Ta1. Hereafter, the subject will be referred to as "subject(subject)" and the state (or action) as "state(action)".

[0060] The essential objects of an intervention node nc are C1 (the subject) and Z1 (the state / behavior), and the entire intervention node nc is represented as C1(Z1). In this case, for the state (behavior) Z1 of the intervention node nc, the options and the reason why a particular behavior was chosen are represented in the object. Note that the intervention node na may also have other objects Tc1.

[0061] The required objects of the result node nb are B1 as the subject and Y1 as the state (action), and the entire result node nb is represented as B1(Y1). The result node nb may also have other objects, such as Tb1.

[0062] And when other objects Ta1, Tb1, and Tc1 are included, the causal graph is represented as follows.

[0063] C1(Z1,Tc1)→A1(X1,Ta1)→B1(Y1,Tb1) Furthermore, the probability using a causal graph is expressed as follows:

[0064] P(B1=Y1,Tb1 | do(C1),A1=X1,Ta1) Furthermore, the causal graph can be expressed in natural language as follows: "At Ta1, A1, which is in state X1, is influenced by action Z1 at Tc1 by C1, and the state of B1 at Tb1 becomes Y1."

[0065] As described above, by representing each node in a causal graph by its object (subject) and state (action), it becomes possible to distinguish between the object (subject) of the affected "factor" and the object (subject) of the "result." This makes it easier to organize complex contexts and "results," and allows for the handling of various types of causal relationship sentences (summaries, etc.). For example, it can handle cases where the subject differs between the "factor" and the "result" of a causal relationship. It can also handle cases where the object (subject) is the same between the "factor" and the "result" (which can be represented as A1 = B1).

[0066] Furthermore, because the elements of each node (subject, state, etc.) are composed of phrases or collocations rather than single words, the meaning of the generated elements and causal graph is highly readable (see Figure 12). In addition, since the causal graph is a directed graph, analysis utilizing directionality is also possible.

[0067] Furthermore, the causal graph generation unit 34 can also understand the diversity of causality by performing a negativity analysis, that is, by determining whether each object in a node expresses a negative or positive sentiment. In this case, the causal graph generation unit 34 uses natural language processing (NLP Neuro-Linguistic Programming) to identify and classify only the state (behavior) elements in the subject (object) as either negative or positive expressions. For example, even if multiple summary sentences s produce slightly different results for the same factor and intervention, various sentences (summaries) can be handled by classifying the objects into two groups: negative or positive. This makes it possible to evaluate whether the causal development is negative or positive. Note that negativity analysis can also be performed using sentiment analysis with general-purpose modules such as NLTK (Natural Language Toolkit).

[0068] The intervention node nc may be included in any causal graph, and there may be causal graphs that do not include the intervention node nc. For example, the intervention node nc may be included in at least one of the causal graphs from both a macro and micro perspective.

[0069] (Vectorization Unit) The vectorization unit 35 uses the trained machine learning model m2 to derive the feature vectors of each object included in each node of the causal graph shown in Figure 4.

[0070] For example, since the factor node na is "object (subject) A1, state (action) X1", the vectorization unit 35 vectorizes A1 and X1 respectively, resulting in [A1, X1], thereby representing the node object as a matrix of element-wise vectors.

[0071] The vectorization unit 35 may derive feature vectors for each node rather than for each object. That is, the vectorization unit 35 may vectorize each node by combining the target (subject), state (action), and other objects into a single vector.

[0072] For example, in a node, if the subject A1 is "government", the state X1 is "expand subsidies", and other Ta1 is "starting next year", the vectorization unit 35 vectorizes "expand government subsidies starting next year", which combines each object into a single sentence separated by spaces.

[0073] Furthermore, instead of using spaces to separate objects, it is also possible to use natural language such as "ga" to group each object together, resulting in "The government will expand subsidies starting next year." The vectorization unit 35 may also vectorize "The government will expand subsidies starting next year."

[0074] In this way, by grouping and vectorizing each object within the same node, it can be treated as a rich feature that takes into account the history and realistic synthesis. Furthermore, if a causal network is created using only states (actions) without objects such as the subject or other objects (see Figures 9, 10, and 11), the resulting causal network will be difficult to understand (or meaningless). To avoid such situations, documents that include at least the objects of the subject and state (action) should be treated as feature vectors.

[0075] By vectorizing as described above, it becomes possible to explore similar causal relationships using different subjects. For example, by performing a similarity search on subject A1 to find related subjects, a potentially applicable causal graph can be generated. Although this lacks scientific basis, it can be treated as a process that has a similar effect to the hypothesis formulation that humans typically perform in their brains. In this way, by vectorizing each object (using embedding representations), various hypotheses can be formulated as the state (subject) or state (action) of nodes in causal inference.

[0076] (Determination Unit) The determination unit 36 ​​determines whether there are specific nodes that are in a predetermined similarity range (relationship) by determining whether the similarity between the three types of objects at each node of the causal graph α at the end of the macro-viewpoint and at each node of the causal graph β at the end of the micro-viewpoint, as shown in Figure 5. The "predetermined range R" is a general term for the predetermined ranges R1, R2, and R3 described later. The "causal graph α" is a general term for the causal graphs α0, α1, and α2 described later (see Figure 11). The causal graph β is a general term for the causal graphs β0, β1, and β2 described later (see Figure 11).

[0077] Here, we will explain the determination process by the determination unit 36 ​​using Figure 5. Figure 5 is a conceptual diagram showing the state in which the similarity is determined between objects of the same type contained in predetermined nodes in two causal graphs.

[0078] Figure 5 shows the state when determining the similarity between two nodes of the same type, n1 representing a factor, intervention, or result, and n2 representing a factor, intervention, or result. Here, the determination unit 36 ​​determines whether the similarity between the subject object of node n1 and the subject object of node n2 falls within a predetermined range R1 (for example, between 50% and 100% (50% or more)).

[0079] Furthermore, the determination unit 36 ​​determines whether the similarity between the state (action) object of node n1 and the state (action) object of node n2 is within a predetermined range R2 (for example, within 70% to 100% (70% or more)).

[0080] Furthermore, the determination unit 36 ​​determines whether the similarity between the other objects of node n1 and the other objects of node n2 is, for example, within a predetermined range R3 (for example, within 90% to 100% (90% or more)).

[0081] Then, if the similarity between the three types of objects falls within a predetermined range R (Case 1: an example of a predetermined case), the determination unit 36 ​​determines that node n1 and node n2 are similar. On the other hand, if at least one of the similarity between the three types of objects is not within the predetermined range R (Case 2), the determination unit 36 ​​determines that node n1 and node n2 are not similar (dissimilar).

[0082] Furthermore, the upper limit of the specified range R does not necessarily have to be 100%. For example, the specified range R1 may be 50% to 90%.

[0083] On the other hand, when the vectorization unit 35 derives a feature vector for each node, the determination unit 36 ​​determines whether there are specific nodes in the two causal graphs that are within a predetermined similarity range.

[0084] (Search Unit) If the determination unit 36 ​​determines that at least one of the similarity levels between objects of the same type is not within a predetermined range (Case 2), the search unit 37 searches the related sentence management DB 51 using the constituent sentence of node n1 (natural sentence including the subject, state (action), and others) as the search key, and retrieves (acquires) the sentence as a related sentence (document). Similarly, the search unit 37 searches the related sentence management DB 51 using the constituent sentence of node n2 (natural sentence including the subject, state (action), and others) as the search key, and retrieves (acquires) the sentence as a related sentence (document). In other words, the search key is node-based, not object-based.

[0085] (Shaping Unit) If the determination unit 36 ​​determines that the similarity between the three types of objects is within a predetermined range R (Case 1), and that there are specific similar nodes, the shaping unit 39 connects each specific node with an undirected edge that signifies "similarity," thereby shaping each path based on the analyzed sentence (related sentence) and generating a causal network as shown in Figure 9, Figure 10, or Figure 11.

[0086] Furthermore, when the molding unit 39 has molded multiple paths, it selects a predetermined path from among the paths that satisfies predetermined criteria. The predetermined criteria are, for example, the shortest path (e.g., with three or fewer undirected edges) or a path with a predetermined number of edges or more (e.g., with five or more undirected edges) at a node connecting the macro viewpoint and the micro viewpoint.

[0087] (Result Creation Unit) Once the causal network is formed by the forming unit 39, the result creation unit 40 uses LLMm1 to summarize the data (information) used until the causal network was formed and creates a text as a story or scenario example. Then, as shown in Figure 12, the result creation unit 40 creates an analysis result that visually represents the causal network constructed by the selected predetermined path. This allows the user to understand the causal network both visually through graphs and through text. The analysis result screen shown in Figure 12 will be described later.

[0088] (Output Unit) The output unit 41 outputs the data of the analysis result screen created by the result creation unit 40 to a monitor or the like, as shown in Figure 12, to an input / output terminal 60 or the like.

[0089] [Processing of the Embodiment] Next, the analysis method of this embodiment will be explained using Figures 6 to 12. Figures 6 to 8 are flowcharts showing the method for analyzing the relationship between the target sentence from a macro perspective and the target sentence from a micro perspective. Here, we will explain an example of analyzing the relationship between the target sentence a0 from a macro perspective and the target sentence b0 from a micro perspective.

[0090] In the examples below, [] is used to explain the specific components of Figures 9 to 12, but this is just one example and is not intended to be the only one.

[0091] S11: The acquisition unit 31 acquires a first sentence to be analyzed from the macro perspective (an example of the first sentence) and a second sentence to be analyzed from the micro perspective (an example of the second sentence) from the input / output terminal 60.

[0092] S12: The summarization unit 32 uses the summarization function of LLMm1 to generate a summary of the text to be analyzed a0 (an example of a first summary) and a summary of the text to be analyzed b0 (an example of a second summary). Note that if the text obtained in process S11 is within a predetermined number of characters (for example, 200 Japanese characters) or has already been summarized, or if other predetermined conditions are met, process S12 may be omitted.

[0093] S13: The object creation unit 33 creates each object (subject, state, action, etc.) in the first summary sentence, and also creates each object (subject, state, action, etc.) in the second summary sentence.

[0094] S14: The causal graph generation unit 34 creates first nodes (factors, interventions, results) in the first summary sentence, each containing an object (subject, state (behavior), etc.), and connects these first nodes with directed edges to create a first causal graph. Similarly, the causal graph generation unit 34 creates second nodes (factors, interventions, results) in the second summary sentence, each containing an object (subject, state (behavior), etc.), and connects these second nodes with directed edges to create a second causal graph.

[0095] For example, as shown in Figure 9, the causal graph generation unit 34 generates a macro-perspective causal graph α0 in the summary sentence s10 of the analysis target sentence a0 (an example of a first summary sentence), and generates a micro-perspective causal graph β0 in the summary sentence s20 of the analysis target sentence b0 (an example of a second summary sentence). At this point, it is unclear whether there is any similarity between causal graphs α0 and β0, so causal graphs α0 and β0 are not connected by an undirected edge that signifies "similarity".

[0096] S15: The vectorization unit 35 derives the feature vectors of each object included in each node in each causal graph. In Figure 9, the vectorization unit 35 derives a total of nine feature vectors for three objects in each of the three types of nodes in causal graph α0, and also derives a total of nine feature vectors for three objects in each of the three types of nodes in causal graph β0.

[0097] S16: The determination unit 36 ​​determines whether to perform (start) a determination of the similarity between the causal graph α at the end of the macro-perspective and the causal graph β at the end of the micro-perspective. If the determination unit 36 ​​does not perform a determination of the similarity between the causal graph α0 at the end of the macro-perspective and the causal graph β0 at the end of the micro-perspective (NO), the process proceeds to the later process S36. That is, as shown in Figure 11, if the determination unit 36 ​​performs a determination of the similarity between, for example, the causal graph α0 at the end of the macro-perspective and the causal graph α1 relating to the summary sentence s11 of the retrieved related sentence a1, also on the macro-perspective side, the process proceeds to the later process S36. Alternatively, if the determination unit 36 ​​performs a determination of the similarity between, for example, the causal graph β0 at the end of the micro-perspective and the causal graph β1 relating to the summary sentence s21 of the retrieved related sentence b1, also on the micro-perspective side, the process proceeds to the later process S36.

[0098] S31: On the other hand, in processing S16, if the determination unit 36 ​​determines the similarity between the causal graph α0 at the end of the macro-viewpoint and the causal graph β0 at the end of the micro-viewpoint (YES), the process proceeds to Figure 7. Then, as shown in Figure 9, the determination unit 36 ​​determines whether the similarity between the three types of objects at each node of the causal graph α0 at the end of the macro-viewpoint and each node of the causal graph β0 at the end of the micro-viewpoint is within a predetermined range R, as shown in Figure 5, thereby determining whether there are specific nodes that are in a predetermined similarity range (relationship). That is, the factor node na10, result node nb10, and intervention node nc10 in the causal graph α0 at the end of the macro-viewpoint are subjected to a brute-force similarity determination with respect to the factor node na20, result node nb20, and intervention node nc20 in the causal graph β0 at the end of the micro-viewpoint, and at that time, the similarity between the three types of objects at each node is determined as shown in Figure 5.

[0099] S32: In processing S31, if the similarity between each of the same type of objects is within a predetermined range, and there are specific nodes that are in a predetermined similarity range (relationship) (YES), the shaping unit 39 connects each specific node [result node nb10, factor node na20] with an undirected edge indicating "similarity" in Figure 9, and shapes each path based on the analyzed sentences a0 and b0, thereby shaping a causal network as shown in Figure 9.

[0100] Furthermore, up to processing S32, multiple paths may be formed not only by the example shown in Figure 9, but also by connecting different types of nodes, such as by connecting result node nb10 and result node nb with an undirected edge, or by connecting factor node na10 and intervention node nc20 with an undirected edge. Alternatively, as shown in Figures 10 and 11, multiple paths may be formed by increasing the number of causal graphs α or β.

[0101] S33: When there are multiple paths, the molding unit 39 selects a predetermined causal network by selecting one or more predetermined paths that satisfy predetermined criteria, such as the shortest path, from among the multiple paths.

[0102] S34: Once a predetermined causal network is formed (selected) by the forming unit 39, the result creation unit 40 uses LLMm1 to summarize the data (information) used until the predetermined causal network was formed and creates a text as a story or scenario example. The result creation unit 40 then creates an analysis result screen as shown in Figure 12. Figure 12 shows the analysis result screen based on the causal network in Figure 11 and will be explained later.

[0103] S35: The output unit 41 outputs information of a predetermined causal network that visually represents the predetermined path selected in process S33 to the input / output terminal 60, etc., as shown in Figure 12.

[0104] As a result, it is possible to improve user visibility (enhance understanding), improve ease of overview, verify supporting information, or assist in the formation of knowledge chains and hypotheses.

[0105] S36: On the other hand, in processing S31, if the determination unit 36 ​​determines that there are no specific nodes in a predetermined similar range (relationship) (NO), the search unit 37 searches the related sentence management DB 51. That is, the search unit 37 searches the related sentence management DB 51 using each node included in the causal graph α at the end of the macro perspective as a search key to obtain each related sentence (related document) a1 (an example of a first related sentence), and also searches the related sentence management DB 51 using each node included in the causal graph β at the end of the micro perspective as a search key to obtain each related sentence (related document) b1 (an example of a second related sentence). Note that the search unit 37 may use at least one of the causal graph α at the end of the macro perspective and the causal graph β at the end of the micro perspective as a search key to search at least one of them.

[0106] After this, the process returns to S12, and the second (second round) of processing from S12 onward is executed, using related sentence b1 instead of the sentence a0 to be analyzed, and related sentence b2 instead of the sentence b0 to be analyzed.

[0107] S51: On the other hand, if in process S16 the determination unit 36 ​​does not determine the similarity between the causal graph α0 at the end of the macro-viewpoint and the causal graph β0 at the end of the micro-viewpoint (NO), the process proceeds to Figure 8. Then, as shown in Figure 10, the determination unit 36 ​​determines whether there are specific nodes that are in a predetermined similarity range (relationship) by determining whether the similarity between the three types of objects is within a predetermined range R at each node of the causal graph α0, which was the end of the macro-viewpoint, and at each node of the causal graph α1, which was also the end of the macro-viewpoint, as shown in Figure 5. That is, the factor node na10, result node nb10, and intervention node nc10 in the causal graph α0 are subjected to a brute-force similarity determination with respect to the factor node na11, result node nb11, and intervention node nc11 in the causal graph α1, and at that time, the similarity between the three types of objects at each node is determined as shown in Figure 5.

[0108] Similarly, as shown in Figure 10, the determination unit 36 ​​determines whether the similarity between the three types of objects in each node of the causal graph β0, which was the end of the micro-perspective side, and each node of the causal graph β1, which was also the end of the micro-perspective side, is within a predetermined range R, as shown in Figure 5, thereby determining whether there are specific nodes that are in a predetermined similarity range (relationship). That is, the factor node na20, the result node nb20, and the intervention node nc20 in the causal graph β0 are subjected to a brute-force similarity determination with respect to the factor node na21, the result node nb21, and the intervention node nc21 in the causal graph β1, and at that time, as shown in Figure 5, the similarity between the three types of objects in each node is determined.

[0109] S52: In processing S51, if it is determined that there are specific nodes in a predetermined similarity range (relationship) from at least one of the macro and micro viewpoints (YES), the molding unit 39 forms a path by connecting a specific node related to the search key [result node nb10] and a specific node related to the search result [factor node na11] with an undirected edge indicating "similarity" on the macro viewpoint side, as shown in Figure 10. Similarly, the molding unit 39 forms a path by connecting a specific node related to the search key [factor node na20] and a specific node related to the search result [result node nb21] with an undirected edge indicating "similarity" on the micro viewpoint side, as shown in Figure 10.

[0110] S53: On the other hand, in processing S51, if it is determined that there are no specific nodes in a predetermined similarity range (relationship) in at least one of the macro and micro perspectives (NO), the determination unit 36 ​​determines whether there is a predetermined range R that has not reached the predetermined range r for each node of causal graphs α and β that is the end of the perspective where there are no specific nodes in a predetermined similarity range (relationship). Specifically, as shown in Figure 5, it determines whether the predetermined range R1 of similarity between target (subject) objects has reached the predetermined range r1 [40% to 100%], whether the predetermined range R2 of similarity between state (action) objects has reached the predetermined range r2 [60% to 100%], and whether the predetermined range R3 of similarity between other objects has reached the predetermined range r3 [80% to 100%].

[0111] S54: If, in process S53, there is a predetermined range R that does not reach the specified range r (YES), the determination unit 36 ​​makes a change to expand at least one of the predetermined ranges R that does not reach the specified range r into the specified range r. Then, it returns to process S51 and makes a new determination.

[0112] S55: On the other hand, in process S53, if there is no predetermined range R that does not reach the specified range r (NO), that is, if all predetermined ranges R have reached the specified range r, the output unit 41 outputs information indicating that a causal network cannot be formed. This completes the processing of the analysis method.

[0113] After the second (second round) of processing S36 described above, the process returns to S12, and the third (third round) of processing starting from S12 is executed. As a result, the causal graph expands, as shown in Figure 11. Figure 11 shows the state in which the causal network is completed because, after the third (third round) of processing, the causal graph at the end of the macro-perspective [causal graph α2] and the causal graph on the micro-perspective side [causal graph β2] have become similar. Furthermore, when the processing from the fourth (fourth round) onwards is performed, the causal graph expands even further. In this way, during the process of expanding the causal network, result node nb10 may be connected to factor nodes na11 and na12 by undirected edges due to similar relationships, or factor node na11 may be connected to factor node na12 by undirected edges due to similar relationships.

[0114] <Analysis Results Screen> Next, we will explain the analysis results screen using the causal network shown in Figure 11, with reference to Figure 12. Figure 12 is a diagram showing an example of the analysis results screen according to the first embodiment.

[0115] Figure 12 shows various causal graphs, such as α0. For example, causal graph α0 shows the factor node NA10, the outcome node NB10, the intervention node NC10, and directed edges.

[0116] Note that the factor node NA10, the outcome node NB10, and the intervention node NC10 correspond to factor node na10, outcome node nb10, and intervention node nc10 in Figure 11, respectively. Furthermore, factor node NA10, outcome node NB10, and intervention node NC10 show either the object (subject) or a sentence (natural sentence) that includes the object (subject), state (behavior), and others.

[0117] Furthermore, undirected edges indicating "similarity" relationships are shown between each causal graph. Note that in Figure 12, due to size constraints, only some nodes of causal graph α0 and causal graphs α1 and α2 are shown; however, in reality, all nodes are shown for all causal graphs.

[0118] Furthermore, important nodes, such as nodes NB10, NA11, and NA12, have a different display configuration from other nodes. In this case, the criteria for extracting important nodes by the result generation unit 40 include nodes with a large number of branches due to undirected edges showing similar relationships, nodes with many connections, and nodes with a high frequency of occurrence. Important nodes may be treated as a driving force in scenario planning.

[0119] Furthermore, because nodes linked by similarity have almost the same meaning, they are displayed in a visually identifiable group, such as in the group-indicating frame G1, to show that they belong to the same group.

[0120] For example, if the result node NB10 and the factor node NA11 have similar meanings, the molding unit 39 may use a machine learning model m1 or the like to summarize the natural language in the result node NB10 and the factor node NA11 and combine them into a single node.

[0121] Here, the result creation unit 40 may create an analysis result screen so that when the user uses cursor c1 to select a desired node (factor node NA10 in Figure 12), a screen displaying specific object information is shown in a pop-up or sub-window. The sub-window may be displayed when the desired node is selected with cursor c1, or it may be displayed beforehand. The sub-window may also include other potentially related nodes or their object information.

[0122] Furthermore, the results generation unit 40 may include basic information about the causal network in the analysis results screen, such as path length (distance), number of nodes, number of branches, number of edges, and other evaluation values ​​related to the network (e.g., similarity between nodes or number of identical documents). In this case, the results generation unit 40 may also include object information such as document information, document title, author, year, publisher, issuer, page number, or chapter (section) title in the analysis results screen.

[0123] Furthermore, the result generation unit 40 provides a simplified representation of the object information for each node, and expresses the object information in natural language. Alternatively, it may be expressed in a summary sentence s instead of natural language. For example, the subject of the node may be included within the node's frame, while other important object information, such as the state (action), may be displayed around the node's frame.

[0124] [Effects of the First Embodiment] As described above, according to this embodiment, it is possible to mechanically extract and analyze complex causal relationships involving changes and relationships between objects (subjects) and states (actions). Furthermore, since each node contains multiple objects to form a chain, it is possible to create a causal network from a causal graph even in situations where the causal relationships are complex and difficult to understand. Moreover, since the analysis can be automated, by increasing the number of materials, it is possible to explore relationships from a wide range of fields (realization of comprehensive investigation and exploration). As a result, if the user inputs a0, which is the target sentence for analysis from a macro perspective, and b0, which is the target sentence for analysis from a micro perspective, it is possible to relatively easily analyze and clarify the relationship between macro-perspective information such as FQ and micro-perspective information such as the company's business vision.

[0125] For example, even if a company's management team tells stakeholders that FQ and business vision appear unrelated at first glance, they can clearly communicate the relationship between FQ and business vision by presenting a causal network as an analysis result, as shown in Figure 12.

[0126] Furthermore, stakeholders can visually survey the chain of causal relationships through the causal network, allowing them to understand how the desired state can be achieved through various causal processes.

[0127] Furthermore, by putting the contents of the causal network into written form ("natural sentences" in Figure 12), the chain of causal relationships can be understood as a story or scenario.

[0128] Furthermore, by changing the content of the input object, it can be used for problem-solving methods, scenario analysis (scenario planning), future predictions, and more.

[0129] ●Second Embodiment The second embodiment will be described using Figures 13 and 14. The second embodiment implements the same functional configuration and processing as the first embodiment, so the differences will be explained below. Figure 13 is a conceptual diagram relating to the second embodiment, showing a state in which intermediate variable nodes are generated in the summary text in addition to the first embodiment. Figure 14 is a conceptual diagram relating to the second embodiment, showing a modified version of Figure 13.

[0130] Regarding the minimum configuration of the causal graph, in the first embodiment, there were three types of nodes na, nb, and nc as shown in Figure 4, but in the second embodiment, there are four types of nodes na, nb, nc, and nm as shown in Figure 13.

[0131] As shown in Figure 13, the causal graph generation unit 34 inserts an intermediate variable node nm between the factor node na and the result node nb. In this case, the factor node na and the intermediate variable node nm are connected by a directed edge from the factor node na to the intermediate variable node nm. Also, the intermediate variable node nm and the result node nb are connected by a directed edge from the intermediate variable node nm to the result node nb.

[0132] An "intermediate variable node" is located between the factor node and the result node in the causal relationship shown by the sentence, and clearly indicates the influence of an intervention on a factor that changes the result. In other words, it shows the influence that the intervention node has on the result node in order to reach the state (behavior) object in the result node. For example, the causal graph generation unit 34 inserts an intermediate variable node when the target (subject) object B1 and the state (behavior) object Y1 in the result node nb are closely related but are not independent as a result of the intervention node nc.

[0133] For example, the intermediate variable node nm shows "how intervention node nc C1 (Z1) affects the subject object B1 in the result node nb, ultimately leading to the state (behavior) Y1 in the result node nb." For instance, it can represent the causal relationship "subject B1: customer satisfaction → intermediate variable node nm: decrease in customer churn rate → state (behavior) Y1: improvement in profit margin" step by step.

[0134] Furthermore, as shown in Figure 14, the causal graph may be constructed by connecting the intervention node nc to the intermediate variable node nm with a directed edge. In this case, since the intervention node nc is connected to the intermediate variable node nm, even if the result node nb and the factor node na, which have similar meanings, are combined into one, the causal network can be formed by branching the patterns of the intervention elements.

[0135] Although Figures 13 and 14 show one intermediate variable node nm, there may be multiple intermediate variable nodes.

[0136] Furthermore, the insertion of the intermediate variable node nm may be performed in any causal graph, and there may be causal graphs in which the insertion does not occur. For example, the insertion of the intermediate variable node nm may be performed in at least one of the causal graphs on the macroscopic and microscopic perspectives.

[0137] [Effects of the Second Embodiment] According to the second embodiment, in addition to the effects of the first embodiment, Figure 12 shows the influence that the intervention node has on the result node in order to reach the state (action) object in the result node.

[0138] Furthermore, in the case of Figure 14, since the intervention node nc is connected to the intermediate variable node nm, even if the result node nb and the factor node na, which have similar meanings, are combined into one, the pattern of intervention elements can be branched to form a causal network.

[0139] ●The third embodiment will be described using Figures 15 and 17. The third embodiment has the same functional configuration and processing as the first embodiment, so the differences will be explained below.

[0140] <Other Examples of Factor Nodes> First, let's explain other examples of factor nodes using Figure 15. Figure 15 relates to the third embodiment and is a conceptual diagram showing a state in which direct factor and indirect factor objects have been created in the factor node, in addition to the first embodiment.

[0141] In the first embodiment, the factor node na includes three types of objects A1, X1, and Ta1, as shown in Figure 4. However, in the third embodiment, as shown in Figure 15, a reason (direct factor) object R1d and a reason (indirect factor) object R1m are added, resulting in a total of five types of objects A1, X1, Ta1, R1d, and R1m.

[0142] The "Reason (Direct Cause) Object R1d" indicates the reason why state (action) X1 occurs in the same node, and represents the direct cause.

[0143] The "Reason (Indirect Factor) Object R1m" represents the reason why state (behavior) X1 occurs in the same node, and indicates an indirect factor. Since multiple factors can be involved in the cause of state (behavior) X1, the analysis becomes easier by hierarchically dividing this reason into "Primary Cause" and "Contextual Cause". For example, for "increased competition", it can be divided into "market saturation (direct factor)" and "increase in new entrants (contextual factor)", and each can be represented by a separate object.

[0144] In this case, the determination unit 36 ​​determines that node n1 and node n2 are similar not only when the similarity between the three types of objects described above falls within each predetermined range R (Case 1), but also when the similarity between the reason (direct cause) objects R1d and the reason (indirect cause) objects R1m falls within each predetermined range R at each node. The predetermined range R41 for the similarity between reason (direct cause) objects R1d is, for example, 80% to 100%, and the predetermined range R42 for the similarity between reason (direct cause) objects R1d is, for example, 60% to 100%.

[0145] <Other Examples of Result Nodes> Next, we will explain other examples of result nodes using Figure 16. Figure 16 relates to a third embodiment and is a conceptual diagram showing a state in which, in addition to the first embodiment, short-term result and long-term result objects have been created in the result node.

[0146] In the first embodiment, the result node nb contains three types of objects B1, Y1, and Tb1 as shown in Figure 4. However, in the third embodiment, as shown in Figure 16, a short-term result object R1S and a long-term result object R1L are added, resulting in a total of five types of objects A1, X1, Ta1, R1S, and R1L.

[0147] The "Short-Term Outcome Object R1S" represents the short-term results that are causally influenced by the factor node.

[0148] The "Long-Term Outcome Object R1L" represents the causal influence from the factor nodes and the resulting outcomes observed over the long term.

[0149] The result node nb may also include medium-term result objects and very-long-term result objects.

[0150] The "Medium-Term Outcome Object R1M" represents the results observed over the medium term, causally influenced by the factor nodes.

[0151] The "Ultra-Long-Term Result Object R1SL" represents the results observed over the ultra-long term, causally influenced by factor nodes.

[0152] Note that short-term, medium-term, long-term, and very long-term refer to, for example, 1 year, 1 to 5 years, 5 to 10 years, and 10 years or longer, respectively.

[0153] In this case, the determination unit 36 ​​determines that node n1 and node n2 are similar not only when the similarity between the three types of objects described above is within each predetermined range R (Case 1), but also when the similarity between similar objects, such as short-term result objects R1S, is within each predetermined range R at each node.

[0154] The predetermined range R for the similarity between objects in the short-term, medium-term, long-term, and very long-term results is, for example, 90% to 100%, 80% to 100%, 70% to 100%, and 60% to 100%, respectively.

[0155] Here, we will describe examples of each object, including the intermediate variable node nm shown in the second embodiment.

[0156] (Examples of each object) [1] Global business development <Object of the factor node> Target A1: Our own ICT business Reason (direct factor) R1d: Intensifying competition in overseas markets, Reason (background factor) R1m: Deregulation status in each country X1: Regional unit integration is underway <Object of the intervention node> Subject C1: Our own ICT business Action Z1: Business integration and service expansion <Object of the intermediate variable node> Target M1: Service supply status W1: Cost reduction <Object of the result node> Target B1: Overseas market result Y1: Improvement of economic profitability in the short term, establishment of competitive advantage in the region in the long term

[0157] [2] Climate Change Response <Objects of Factor Nodes> Target A1: Environmental policy reasons (direct factors) R1d: Reasons for global carbon neutrality targets (indirect (background) factors) R1m: State of strengthened government regulations X1: Target achievement is required <Objects of Intervention Nodes> Stakeholder C1: Company actions Z1: Research and development of environmental technologies <Objects of Intermediate Variable Nodes> Target M1: Company state W1: Emission reduction <Objects of Outcome Nodes> Target B1: Environmental impact Outcome Y1: Emission reduction in the short term, establishment of sustainable operations in the long term

[0158] [3] Work Style Reform <Object of Factor Node> Target A1: Reasons for employees' work style (direct factor) R1d: Expansion of demand for hybrid work (direct factor), Reasons (indirect (background) factor) R1m: Issues with commuting environment State X1: Flexibility is required <Object of Intervention Node> Subject C1: Company actions Z1: Development of remote work environment, strengthening of support systems <Object of Intermediate Variable Node> Target M1: Employee state W1: Improvement of satisfaction <Object of Result Node> Target B1: Labor productivity Result Y1: Productivity improvement in the short term, realization of lifestyle diversification in the long term

[0159] [4] Cybersecurity <Objects in the Factor Node> Target A1: Corporate information protection system Reason (Direct Factor) R1d: Sophistication of cyberattacks (Direct Factor) Reason (Indirect (Background) Factor) R1m: Expansion of digital dependence (Indirect (Background) Factor) State X1: Exposed to increasing threats <Objects in the Intervention Node> Subject C1: IT company Action Z1: Provision of AI-powered security solutions <Objects in the Intermediate Variable Node> Target M1: Customer State W1: Reduction of risk response time <Objects in the Result Node> Target B1: Customer information assets Result Y1: Reduction of information leakage risk, improvement of reliability

[0160] [5] Smart Life Business <Object of Factor Node> Target A1: Consumer-facing service Reason (Direct Factor) R1d: Increased needs in an aging society (Direct Factor) Reason (Indirect (Background) Factor) R1m: Spread of IoT devices (Indirect (Background) Factor) State X1: Service expansion is needed <Object of Intervention Node> Subject C1: IT company Action Z1: Strengthening investment in the medical and financial fields <Object of Intermediate Variable Node> Target M1: Service content State W1: Promotion of personalization <Object of Result Node> Target B1: Consumer experience Result Y1: Improved user convenience and increased satisfaction The above is an example of the third embodiment.

[0161] (Other 1) The reason why state (behavior) X1 occurs may be represented by a single object without hierarchical structuring. In this case, the factor node na contains four types of objects A1, X1, Ta1, and R1dm. An example of this case is shown below.

[0162] [6] Digital Economy Market Subject A1: Digital Economy Market State X1: Rapid growth continues (e.g., cloud usage rate increased by 25% year-on-year) Reason R1: Spread of cloud and AI technology (direct factor), spread of remote work (background factor)

[0163] In this case, the determination unit 36 ​​determines that nodes n1 and n2 are similar not only when the similarity between the three types of objects described above falls within each predetermined range R (Case 1), but also when the similarity between the reason objects R1dm at each node falls within each predetermined range R (an example of a third similarity). The predetermined range R4 related to the similarity between reason objects R1dm is, for example, 70% to 100%.

[0164] Note that the object reason R1 may be made into a separate reason node.

[0165] (Other 2) Objects R1d and R1m may each be divided into two, and the direct cause of the object (subject) A1 and the direct cause of the state (action) X1 may be included in the factor node na. In this case, the factor node na will contain seven types of objects.

[0166] <Specific Example of a Causal Network> Here, using Figure 17, we will explain a specific example of the causal network in Figure 9 that includes reason (direct cause) objects R10d, R20d, reason (indirect (background) cause) objects R10m, R20m, short-term result objects R10S, R20S, and long-term result objects R10L, R20S. Note that this example shows the case where other nodes of each node are not included.

[0167] First, the target sentence a0 from the macro perspective and the target sentence b0 from the micro perspective in this specific example are as follows:

[0168] (Analysis target sentence a0 from a macro perspective) The use of AI is gradually increasing in general companies. The direct cause of this is the need for increased efficiency in business operations, but this is not only due to the need to respond to the direct cause of cost reduction demands, but also because of the underlying cause of labor shortages. In response to these needs, low-cost AI services have recently been offered by major IT companies, and their use is expanding even further. Along with the expansion of demand, the amount of server usage is increasing, and in recent years, the use of AI has been expanding rapidly in data centers.

[0169] (Analysis of the micro-perspective, sentence b0) In our own data center business, the demand for data centers is on the rise. This is due to the use of enormous computing resources in various companies' AI services, but behind this is the rapid increase in AI service users due to the expansion of low-cost AI services. Furthermore, realizing sustainable business is not only a corporate responsibility, but also something that we have received many requests for from our customers. For this reason, in recent years we have been continuously increasing the amount of electricity derived from renewable energy. Through this initiative, we are able to switch a portion of the power consumption of our data centers to renewable energy generation, thereby providing zero-emission services to customers who request it. In the future, we will continue our activities with the aim of decarbonizing all of our businesses.

[0170] (Causal Network) Each of the above analyzed sentences a0 and b0 is summarized, and in the end, a causal network like the one shown in Figure 17 is formed. Figure 17 is a diagram showing a specific example of a causal network.

[0171] [Effects of the Third Embodiment] According to the third embodiment, in addition to the effects of the first and second embodiments, the addition of reason objects R1d, R1m, etc., makes it possible to output analysis results using a causal network that is even easier to understand.

[0172] [Supplement] The present invention is not limited to the embodiments described above, and may also have configurations or processes (operations) as shown below.

[0173] (1) The analysis device 30 can be realized by a computer and a program, but it is also possible to record and provide this program on a (non-temporary) recording medium, and it is also possible to provide the program via a communication network 100 such as the Internet.

[0174] (2) The hardware processor 1004 may be single or multiple.

[0175] 10 Communication system 30 Analysis device 31 Acquisition unit 32 Summarization unit 33 Object creation unit 34 Causal graph generation unit (example of generation unit) 35 Vectorization unit 36 ​​Judgment unit 37 Search unit 39 Formatting unit 40 Result creation unit 41 Output unit 50 Storage unit 51 Related sentence management DB (example of related sentence management unit) 60 Input / Output terminal a0 Sentence to be analyzed (example of the first sentence to be analyzed) bo Sentence to be analyzed (example of the second sentence to be analyzed) a1 Related sentence (example of the first related sentence) a2 Related sentence b1 Related sentence (example of the second related sentence) b2 Related sentence na10 Factor node (example of the first factor node) nb10 Result node (example of the first result node, example of a specific node) nc10 Intervention node (example of the first intervention node) na20 Factor node (example of the second factor node, example of a specific node) nb20 Result node (example of a second result node) nc10 Intervention node (example of a second intervention node) nm Intermediate variable node (example of a first intermediate variable node, example of a second intermediate variable node) A10, B10, C10 Target or subject object (example of a first object) X10, Y10, Z10 State or action object (example of a second object) A20, B20, C20 Target or subject object (example of a third object) X20, Y20, Z20 State or action object (example of a fourth object) R1d Reason (direct cause) object R1m Reason (indirect cause) object R1dm Reason object (example of a first reason object, example of a second reason object) α0 Causal graph (example of a first causal graph) α1 Causal graph α2 Causal graph β0 Causal graph (example of a second causal graph) β1 Causal graph β2 Causal graph

Claims

1. An analysis device for analyzing the relationship between a first sentence to be analyzed and a second sentence to be analyzed, comprising: a generation unit that performs a first process of generating a first factor node indicating a factor and a first result node indicating a result in the causality shown by the first sentence to be analyzed, and generating a first causal graph by connecting the first factor node to the first result node with directed edges; and a second process of generating a second factor node indicating a factor and a second result node indicating a result in the causality shown by the second sentence to be analyzed, and generating a second causal graph by connecting the second factor node to the second result node with directed edges; and a shaping unit that, if there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similar range, shapes a causal network based on the first sentence to be analyzed and the second sentence to be analyzed by connecting the specific nodes with undirected edges.

2. The analysis apparatus according to claim 1, wherein the generation unit generates a first intervention node indicating an intervention for a factor that changes the result in the causal relationship shown by the first sentence to be analyzed, and generates the first causal graph by connecting the first intervention node to the first factor node with a directed edge, or generates a second intervention node indicating an intervention for a factor that changes the result in the causal relationship shown by the second sentence to be analyzed, and generates the second causal graph by connecting the second intervention node to the second factor node with a directed edge.

3. The analysis apparatus according to claim 2, wherein the generation unit generates a first intermediate variable node that clearly indicates the influence of intervention on factors that change the result in the causal relationship shown by the first sentence to be analyzed, inserts the first intermediate variable node between the first factor node and the first result node, connects the first factor node to the first intermediate variable node with a directed edge, and connects the first intermediate variable node to the first result node with a directed edge to generate the first causal graph, or generates a second intermediate variable node that clearly indicates the influence of intervention on factors that change the result in the causal relationship shown by the second sentence to be analyzed, inserts the second intermediate variable node between the second factor node and the second result node, connects the second factor node to the second intermediate variable node with a directed edge, and connects the second intermediate variable node to the second result node with a directed edge to generate the second causal graph.

4. The analysis apparatus according to any one of claims 1 to 3, comprising: a vectorization unit for deriving a feature vector for each node; and a determination unit for determining whether there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similarity range.

5. The analysis apparatus according to claim 4, wherein each of the first factor node and the first result node has an object creation unit that creates a first object indicating an object or subject and a second object indicating a state or action for the first sentence to be analyzed, and each of the second factor node and the second result node has an object creation unit that creates a third object indicating an object or subject and a fourth object indicating a state or action for the second sentence to be analyzed, the vectorization unit derives a feature vector for each object, and the determination unit determines that if the similarity of each feature vector of the first object and the third object is within a first predetermined range and the similarity of each feature vector of the second object and the fourth object is within a second predetermined range, then there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similarity range.

6. The analysis apparatus according to claim 5, wherein the object creation unit creates a first reason object at the first factor node that indicates the reason why a state or action occurs, and creates a second reason object at the second factor node that indicates the reason why a state or action occurs; the vectorization unit derives feature vectors of the first reason object and the second reason object; and the determination unit determines that, in the predetermined case and when the similarity of each feature vector of the first reason object and the second reason object is within a third predetermined range, there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similarity range.

7. The object creation unit creates a first short-term result object at the first result node that shows results observed in the short term due to causal influence from the first factor node, and a first long-term result object that shows results observed in the long term due to causal influence from the first factor node, and at the second result node, it creates a second short-term result object at the second result node that shows results observed in the short term due to causal influence from the second factor node, and a second long-term result object that shows results observed in the long term due to causal influence from the second factor node, and the vectorization unit derives feature vectors for the first short-term result object, the first long-term result object, the second short-term result object, and the second long-term result object, The analysis apparatus according to claim 5, wherein the determination unit determines that in the predetermined case, and the similarity of the feature vectors of the first short-term result object and the second short-term result object is within a fourth predetermined range, and the similarity of the feature vectors of the first long-term result object and the second long-term result object is within a fifth predetermined range, there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similarity range.

8. The analysis apparatus according to claim 1, comprising a search unit that, when there are no specific nodes in a predetermined similar range in the first causal graph and the second causal graph, retrieves a first related sentence of the first sentence to be analyzed by searching using a search key based on the first factor node and a search key based on the first result node, or retrieves a second related sentence of the second sentence to be analyzed by searching using a search key based on the second factor node and a search key based on the second result node, wherein the generation unit, when the first related sentence is retrieved, performs the first processing on the first related sentence instead of the first sentence to be analyzed, and when the second related sentence is retrieved, performs the second processing on the second related sentence instead of the second sentence to be analyzed.

9. A computer analysis method for analyzing the relationship between a first sentence to be analyzed and a second sentence to be analyzed, wherein the computer performs a generation process which generates a first factor node indicating a factor and a first result node indicating a result in the causality shown by the first sentence to be analyzed, and generates a first causal graph by connecting the first factor node to the first result node with directed edges, and performs a second process which generates a second factor node indicating a factor and a second result node indicating a result in the causality shown by the second sentence to be analyzed, and generates a second causal graph by connecting the second factor node to the second result node with directed edges, and performs a shaping process which, if there are specific nodes in the first causal graph and the second causal graph that are within a predetermined similar range, connects the specific nodes with undirected edges to form a causal network based on the first sentence to be analyzed and the second sentence to be analyzed.

10. A program that causes a computer to perform the method described in claim 9.

Images