A device for analyzing the time-series changes in emotional networks.

Abstract

Providing technical means to analyze the time-series changes in emotional networks. [Solution] The present invention provides an analysis device 1 for analyzing the time-series changes of an emotional network, comprising: a content data acquisition unit 111 that acquires conversation content data; an analysis unit 112 that derives a relationship index for each speaker pair based on emotional information obtained from the content data related to the analysis unit for each analysis unit of conversation content data along the time series; and a visual representation data generation unit 114 that generates visual representation data for the time-series changes of the graph of the emotional network constructed based on the relationship index. The graph generated by the visual representation data generation unit 114 is a graph in which each vertex corresponds to each speaker and each edge corresponds to the relationship index.

Description

【Technical Field】 【0001】 The present invention relates to an apparatus for analyzing the time-series changes of an emotion network. 【Background Art】 【0002】 In online meetings, consultations, conferences, symposiums, novels, scripts, or conversation games and other multi-person conversations, there is a need to analyze the emotions of speakers based on data recording multi-person conversations. However, when performing such analysis manually, difficulties may arise in ensuring the objectivity of the analysis results, reducing the required time for analysis, and reducing the labor. Therefore, technical means for automatically analyzing emotions are required. 【0003】 As a means for automatically analyzing emotions, Patent Document 1 discloses an apparatus for analyzing a sentence, which includes a sentence acquisition unit, a sentence segmentation unit, an emotion vector generation unit, an evaluation generation unit, a display command unit, and a first machine learning unit, and the evaluation generation unit can generate an evaluation corresponding to a correlation value regarding at least the correlation between the norm of the emotion vector and the moving average of the norm. 【0004】 The technique of Patent Document 1 can provide an analysis result that can easily grasp the characteristics of a sentence. 【Prior Art Documents】 【Patent Documents】 【0005】 【Patent Document 1】 Japanese Patent No. 7368684 【Summary of the Invention】 【Problems to be Solved by the Invention】 【0006】 Incidentally, in the analysis of data recorded from multi-person conversations, there is a demand for analysis of the time-series changes in the emotional relationships between speakers. Such analysis can help in understanding the emergence, changes, and disappearance of central figures in conversations, the formation, transformation, and dissolution of groups, and the identification of points of change in interpersonal relationships, as well as other time-series changes in relationships between speakers. 【0007】 While the technology described in Patent Document 1 can provide analysis results based on the analysis of emotions evoked by text or text accompanied by images in an unspecified number of viewers, there is room for further improvement in supporting the understanding of the time-series changes in emotion-based relationships between speakers. For this reason, there is a need for a technical means to analyze the time-series changes of an emotion network that summarizes the emotion-based relationships between speakers in order to understand the time-series changes in emotion-based relationships between speakers. 【0008】 This invention was made to solve the problems of the prior art described above, and aims to provide a technical means for analyzing the time-series changes of emotional networks. [Means for solving the problem] 【0009】 As a result of diligent research to solve the above problems, the inventors have found that the above problems can be solved by deriving a relationship index for each speaker pair using emotional information obtained from the content data, and by a configuration and other technical means that allows the time-series changes to be visually represented. The inventors have now completed the present invention. Specifically, the present invention provides the following: 【0010】 One aspect of the present invention provides an analysis device for analyzing the time-series changes of an emotional network, comprising: a content data acquisition unit that acquires content data including conversations; an analysis unit that derives a relationship index for each pair of speakers based on emotional information obtained from the content data relating to each analysis unit, for each analysis unit of the content data along the time series; and a visual representation data generation unit that generates visual representation data for the time-series changes in a graph of an emotional network constructed based on the relationship index, wherein the graph is a graph in which each vertex corresponds to each speaker and each edge corresponds to the relationship index. 【0011】 The analysis unit of the invention derives relationship indicators for each speaker pair using emotional information obtained from data related to the analysis unit of content data in a time series. The visual representation data generation unit of the invention then generates visual representation data of the time series changes of a graph with each speaker as a vertex and the relationship indicators for each speaker pair as edges, thereby providing the user with the analysis results of the time series changes of the emotional network. 【0012】 Therefore, the invention according to this aspect of the present invention can provide a technical means for analyzing the time-series changes of an emotional network. 【0013】 Furthermore, the present invention can take various forms, including: an embodiment that analyzes based on a high-granularity sentiment score that uses sentences, which are the basic units of conversation, and speaker pairs, which are the basic units of relationships, as processing units; an embodiment that assists in sentiment network analysis through indicators related to graph structure; an embodiment that distinguishes and expresses the feelings from one speaker to another and the feelings in the opposite direction; an embodiment that expresses the magnitude of emotions; an embodiment that adds automatically generated findings; and other various embodiments. Each of these embodiments of the present invention contributes to the analysis of time-series changes in sentiment networks through its own unique characteristics. [Effects of the Invention] 【0014】 This invention can provide a technical means for analyzing the time-series changes of emotional networks. [Brief explanation of the drawing] 【0015】 [Figure 1] Figure 1 is a block diagram showing an example of the hardware and software configuration of system S. [Figure 2] Figure 2 shows an example of the analytical information database 131. [Figure 3] Figure 3 is a continuation of the previous figure. [Figure 4] Figure 4 is a main flowchart showing an example of a preferred flow of the analysis process performed by the analysis device 1 of this embodiment. [Figure 5] Figure 5 is a continuation of the previous figure. [Figure 6] Figure 6 shows an example of a prompt. [Figure 7] Figure 7 shows an example of output corresponding to the prompt in the previous figure. [Figure 8] Figure 8 shows a detailed example of the processing flow from analysis to display in the analysis process. [Figure 9] Figure 9 shows an example of how analytical information is displayed. [Figure 10] Figure 10 is an example of a representation related to the emotional network analysis of a scene from a novel. [Modes for carrying out the invention] 【0016】 First, although the following disclosure, diagrams, and / or claims, etc. are described as being alone or in combination with one or more other aspects, the subject matter of the immediate disclosure is not intended to be limited as such. That is, the immediate disclosure, diagrams, and claims are intended to encompass the various aspects described herein, either alone or in one or more combinations with each other. For example, even if the immediate disclosure describes and illustrates a first embodiment, a second embodiment, and a third embodiment, and the first embodiment is described and illustrated particularly in relation to the second embodiment, or the second embodiment is described and illustrated only in relation to the third embodiment, the immediate disclosure and illustration are not limited as such, and may include only the first embodiment, only the second embodiment, only the third embodiment, or one or more combinations of the first, second, and / or third embodiments, such as the first embodiment and the second embodiment, the first embodiment and the third embodiment, the second embodiment and the third embodiment, or the first, second, and third embodiments. 【0017】 The use of the phrase "or" in the text shall mean a "non-exclusive" resolution unless otherwise specified. For example, in the case of "item x is A or B", it shall mean either (1) item x is only one of A or B, or (2) item x is both A and B. In other words, the word "or" is not used to define an "exclusive" resolution. 【0018】 Also, when the phrases "including at least one" or "including at least one of the following" used in the text are used in combination with a system or element, it means that the system or element includes one or more of the elements listed after the phrase. For example, if there are three types of elements, the first element to the third element, the phrases "including at least one" or "including at least one of the following" shall be interpreted as any of the following structural arrangements: a device including the first element, a device including the second element, a device including the third element, a device including the first and second elements, a device including the first and third elements, a device including the second and third elements, or a device including the first, second, and third elements. 【0019】 When the phrase "used in at least one of the following" is used in this document, the same interpretation is intended. Furthermore, "and / or" used in this document is used as a linguistic conjunction and is used to indicate that one or more of the described elements or conditions are included or occur. For example, a device including a first element, a second element, and / or a third element is interpreted as any of the following structural arrangements: a device including the first element, a device including the second element, a device including the third element, a device including the first and second elements, a device including the first and third elements, a device including the second and third elements, or a device including the first, second, and third elements. 【0020】 Note that the use of the phrase "and / or" in this document meaning a "non-exclusive" determination is also defined in "Style and Preparation Method of Standard Forms JIS Z 8301" of the Japanese Industrial Standards (JIS). 【0021】 Hereinafter, an example of an embodiment of the present invention will be described in detail with reference to the drawings. 【0022】 <System S> FIG. 1 is a block diagram showing an example of the hardware configuration and software configuration of System S. 【0023】 System S is configured to include an analyzer (Analyzer 1) for analyzing the time-series changes of the emotion network. Hereinafter, an aspect of a server-client in which Analyzer 1 cooperates with terminals T configured to be able to communicate with each other via network N will be described. Those skilled in the art will easily conceive of Analyzer 1 in a stand-alone aspect including input means and display means based on the following description. 【0024】 〔Analyzer 1〕 Analyzer 1 according to the present embodiment is a computer including a control unit 11, a storage unit 13, and a communication unit 14. Analyzer 1 performs analysis processing of the time-series changes of the emotion network by executing the program of the present embodiment. 【0025】 [Control Unit 11] The control unit 11 includes a Central Processing Unit (CPU), Random Access Memory (RAM), Read Only Memory (ROM), and other hardware components. 【0026】 The control unit 11 cooperates with at least one of the storage unit 13 and the communication unit 14 as needed. The control unit 11 then implements the content data acquisition unit 111, the analysis unit 112, the findings information generation unit 113, and the visual representation data generation unit 114, which are software components of the program of this embodiment executed in the analysis device 1. 【0027】 The analysis process performed by the program of this embodiment will be described later with reference to Figures 4 and 5. 【0028】 [Storage Unit 13] The storage unit 13 is a device on which data and / or files are stored, and has a storage unit that stores data non-temporarily using a hard disk, semiconductor memory, recording medium, memory card, or other storage material. Preferably, the storage unit 13 stores data for programs executed by the control unit 11, an analysis information database 131, recorded data, and other data. 【0029】 Furthermore, in order to achieve generation using a large-scale language model without transmitting content data to an external device that is different from the analysis device 1 and the managing entity, the storage unit 13 may store data for realizing the large-scale language model according to this embodiment (e.g., a set of parameters for inference, a tokenizer, vocabulary data, and executable code). 【0030】 (Analysis Information Database 131) The analysis information database 131 stores the analysis information generated by the analysis device 1. The analysis information stored in the analysis information database 131 includes the content data to be analyzed, data related to the time-series changes of the emotion network graph corresponding to the entire content data, and the analysis results, as well as content data, relationship indicators, emotion network graph data, and various analysis results corresponding to the analysis unit. 【0031】 The data relating to the time-series changes of the emotion network graph corresponding to the entire content data includes, for example, visual representation data for representing the graph corresponding to each analysis unit as a video. 【0032】 The analysis results corresponding to the entire content data include, for example, observations regarding the relationships between speakers, such as a graph of the emotion network or the temporal changes in said graph. Examples of such observations include text that outlines the relationships between speakers, text that shows the temporal changes in said outline, and text that provides advice for aiming for appropriate relationships depending on the type of conversation content data. 【0033】 The relationship indicators corresponding to the analysis units include, for example, the representative types and intensities of emotions between speakers, derived based on the content data corresponding to the analysis units. The types of emotions are, for example, the eight basic emotions in Plutchik's wheel of emotions. 【0034】 The data for the emotion network graph corresponding to the unit of analysis is data that shows a graph in which each vertex corresponds to each speaker and each edge corresponds to a relationship index between speakers. The format of this data can be, for example, represented in JSON format as a set of vertices and an edge set. Each edge in the edge set in this example is associated with a relationship index, such as a label indicating the type of emotion and a value indicating the strength of that emotion. 【0035】 The various analysis results corresponding to the analysis unit include, for example, indicators related to the structure of the graph (e.g., centrality index, network density). Furthermore, the analysis results may also include, for example, observational information regarding the relationships between speakers as shown by the content data corresponding to the analysis unit. Examples of such observational information include text outlining the relationships between speakers and text offering advice for achieving appropriate relationships according to the type of conversational content data. 【0036】 More detailed information about the analysis will be provided in the description of the analysis process flow below. In the following, the analysis information will be described as being stored in association with an identification ID (analysis information ID) used for storage, retrieval, and other data utilization. 【0037】 Figure 2 is an example of the analysis information database 131. Figure 3 is a continuation of the previous figure. The examples in Figures 2 and 3 show the first analysis information relating to the entire conversation content "Meeting Minutes 1" associated with analysis information ID "A0001," and the second analysis information relating to one analysis unit of the same content associated with analysis information ID "A1002." Note that analysis information relating to other analysis units has been omitted for clarity. 【0038】 The first analysis information for this example includes data from graphs summarizing each graph of the emotion network in a time-series manner, showing that there is a relationship index indicating a mutual moderate level of trust (trust: 0.5) between speaker A and speaker B and between speaker A and speaker C, a relationship index indicating a mutual strong level of trust (trust: 1.0) between speaker A and speaker D, no relationship index between speaker B and speaker C, a relationship index indicating a mutual weak level of trust (trust: 0.2) between speaker B and speaker D, and a relationship index indicating a one-sided weak level of trust (trust: 0.2) from speaker C towards speaker D. 【0039】 Furthermore, the first analysis information includes the following text as a result of the analysis of the emotional network graph: "Conclusion: This time-series network analysis shows that the conference has a structure that allows for efficient transitions from information sharing to decision-making while maintaining trust. However, there is a significant concentration of information on A, and promoting more decentralized communication is a challenge." 【0040】 Furthermore, the first analysis information includes the following text as advice (action plan) to aim for appropriate relationships according to the type of conversation content data: First advice: "1. Introduction of decentralized discussion segments · Set aside 10 minutes of 'all-hands discussion time' in the middle of the meeting to promote exchange of opinions without A intermediary. · Implementation method: Enforce an explicit pattern for a certain period of time, such as a triangular speaking order (A→B→C→A)" and Second advice: "2. Utilization of visualized feedback · Regularly provide feedback to the team on the results of sentiment network analysis like this one, and share an understanding of communication patterns..." 【0041】 In addition, the first analysis information includes, as centrality indicators among the indicators related to the structure of the graph, in-degree centrality of 0.667, out-degree centrality of 1.0, betweenness centrality of 0.0, adjacent centrality of 1.0, and eigenvector centrality of 0.653, and other indicators related to speaker A. Furthermore, the first analysis information includes a network density of 0.75 among the indicators related to the structure of the graph. 【0042】 Furthermore, the second analysis information for this example includes content data for the first 10 minutes, which constitutes the analysis unit, and data from a graph of the emotion network corresponding to that analysis unit. The graph data includes relationship indicators showing a moderate level of trust (trust: 0.5) from speaker A to speaker B, a moderate level of anger (anger: 0.5) from speaker B to speaker A, a weak level of surprise (surprise: 0.2) from speaker B to speaker C, and a weak level of expectation (expectation: 0.2) from speaker C to speaker B. 【0043】 In addition, the second analysis information includes, as centrality indicators among the indicators related to the structure of the graph, in-degree centrality of 0.5, out-degree centrality of 0.5, betweenness centrality of 0.0, closeness centrality of 0.8, and eigenvector centrality of 0.5, and other indicators related to speaker A. Furthermore, the second analysis information includes a network density of 0.667 among the indicators related to the structure of the graph. 【0044】 By storing this data in the analysis information database 131 of the example, the analysis device 1 can generate visual representation data based on this data and present the user with the time-series changes of the emotion network graph. 【0045】 [Communication Unit 14] The specific configuration of the communication unit 14 is not particularly limited, as long as it is for connecting the analysis device 1 to the network N and performing communication. The communication unit 14 may be configured using, for example, a network card compatible with the Ethernet standard, a wireless LAN compatible communication device, or other communication devices. 【0046】 [Network N] The type of network N is not particularly limited as long as it is used for communication between the analysis device 1 and other devices. Examples of network N include the internet, a mobile phone network, and a wireless LAN. 【0047】 [Terminal T] Terminal T is used by the user of the analysis device 1. Terminal T performs processes such as receiving and / or displaying data transmitted from the analysis device 1, transmitting user input to the analysis device 1, and other processes. The types of terminal T include, for example, portable terminals (e.g., laptop computers, smartphones, tablet computers) and stationary terminals (e.g., desktop computers). 【0048】 [Main Flowchart of Analysis Process] Figure 4 is a main flowchart showing an example of a preferred flow of the analysis process performed by the analysis device 1 of this embodiment. Figure 5 is a continuation of the previous figure. The following is a description of an example of a preferred flow of the analysis process performed by the analysis device 1, using Figures 4 to 5. 【0049】 [Step S1: Determine whether to acquire content data] The control unit 11 performs a process to determine whether to acquire content data including conversation using the content data acquisition unit 111 (content data acquisition determination step). If the control unit 11 determines to acquire the data, it moves the process to step S2; otherwise, it moves the process to step S3. The content data acquisition unit 111 is realized through the cooperation of the control unit 11, the storage unit 13, the communication unit 14, and other hardware components. 【0050】 The above determination in this step is achieved by a procedure that determines whether to acquire the data if it falls under a predetermined case (e.g., a case in which a user's instruction regarding the upload of content data is received, or a case in which a user's instruction to start analysis, including the specification of content data, is received). 【0051】 [Step S2: Acquire content data] The control unit 11 executes a process to acquire content data including conversation using the content data acquisition unit 111 (content data acquisition execution step). The control unit 11 then moves the process to step S3. 【0052】 The acquisition described above in this step is achieved, for example, by a procedure to receive content data transmitted by the user via terminal T, or by a procedure to acquire a resource specified using a Uniform Resource Identifier (URI) or other resource name identification means using the storage unit 13 or the communication unit 14. 【0053】 (Regarding conversation content data) The content data acquired by the content data acquisition unit 111 in this step is not particularly limited, as long as it includes data containing conversations between multiple people. The content data includes, for example, meeting transcripts or minutes, online meeting audio recordings, or audio-equipped video recordings. This allows the analysis device 1 to provide analysis results that are useful for examining problems within a meeting. 【0054】 Furthermore, the content data may include dialogue data within the story. Examples of "stories" include novels, comics, plays, videos, films, games, or stories related to interactive content. "Dialogue data within the story" is not limited to dialogue data contained in the work itself, but may also include data related to the production stage of the work (e.g., plot, dialogue, drafts, scenarios, scripts, or screenplay data). This allows the analysis device 1 to provide analysis results useful for examining problems in the story. 【0055】 After acquiring content data, the analysis device 1, through its analysis unit 112, performs a series of processes to derive relationship indicators for each speaker pair based on emotional information obtained from the content data related to the analysis unit, for each analysis unit of the content data in a time-series order. Steps S3 to S11 are an example of this process. The analysis unit 112 is realized, for example, through the cooperation of the control unit 11, the storage unit 13, the communication unit 14, and other hardware components. 【0056】 In this example, the analysis unit 112 performs a series of processes with respect to the sentiment information, including obtaining a sentiment score for each sentence based on the content data, extracting speaker pairs related to each sentence from the content data, associating the sentiment scores with the pairs, dividing the multiple sentences included in the content data chronologically into analysis units, and deriving a relationship index for the pairs based on the associations for each sentence included in the analysis units. 【0057】 (Sentiment score acquisition process) First, the analysis device 1 divides the content data into sentences and performs sentiment score acquisition processing to calculate sentiment scores corresponding to each sentence. Steps S3 to S4 are an example of this process. 【0058】 [Step S3: Content data is divided into sentences] The control unit 11 performs the process of dividing the acquired conversation content data into sentences using the analysis unit 112 (dividing step). The control unit 11 then moves the process to step S4. After that, the analysis unit 112 performs a series of processes from step S4 to step S6 for each divided sentence. 【0059】 [Step S4: Obtain Sentiment Score] The control unit 11 uses the analysis unit 112 to perform a process to obtain the sentiment score of the sentence to be processed (sentiment score acquisition step). The control unit 11 then moves the process to step S5. 【0060】 The process of obtaining the emotion score in this step is achieved, for example, by inputting the sentence to be processed into a converter, having the converter output an emotion score that includes the intensity of multiple types of emotions evoked by the sentence to be processed, and then obtaining the emotion score. 【0061】 The aforementioned "converter" is configured to analyze an input sentence and output an emotion score that includes the intensity of multiple types of emotions evoked by the sentence. These "multiple types of emotions" are, for example, the eight basic emotions in Plutchik's Wheel of Emotions. The eight basic emotions in Plutchik's Wheel of Emotions are anger, expectation, joy, trust, fear, surprise, sadness, and disgust. 【0062】 The emotion score is not particularly limited, as long as it includes the intensity of multiple types of emotions evoked by the sentence being processed. For example, it may include an emotion vector whose elements are the intensity of these multiple types of emotions. Hereafter, "the intensity of multiple types of emotions evoked by the sentence being processed" will also be simply referred to as "the emotion vector corresponding to the sentence being processed." 【0063】 The converter is configured using, for example, a pre-trained model obtained by pre-training with training data that associates sentences with corresponding emotion vectors. The converter may be implemented within the analysis device 1, or it may be implemented in an external device separate from the analysis device 1 included in system S. By implementing the converter in an external device, it is possible to reduce processing time through parallel processing between the analysis device 1 and the external device. 【0064】 (Speaker pair mapping process) Next, the analysis device 1 performs speaker pair mapping processing to associate the emotion score with the speaker. Steps S5 to S6 are an example of this processing. 【0065】 [Step S5: Identify speaker pairs] The control unit 11 uses the analysis unit 112 to perform a process to identify speaker pairs based on the content data (pair identification step). The control unit 11 then moves the process to step S6. 【0066】 Identifying speaker pairs in this step is achieved, for example, by identifying the speaker of the sentence to be processed and a different speaker who made an utterance immediately before or after the sentence, as a speaker pair related to the sentence to be processed. This allows the analysis device 1 to identify which speaker the emotions obtained based on the sentence to be processed are directed from the speaker of the sentence to. 【0067】 In this procedure, it is preferable that the analysis unit 112 identifies the speaker of the sentence to be processed and a different speaker who made an utterance immediately after the speaker's utterance, including the sentence to be processed, as a speaker pair. This allows the analysis unit 112 to identify pairs that correspond to the target of emotion in line with the normal flow of conversation, such as passing the conversation to the next speaker. It is preferable that the analysis unit 112 identifies the above pair as a directed pair that includes a direction from the speaker of the sentence to be processed to the aforementioned different speaker, so that the direction of emotion is reflected. 【0068】 [Step S6: Assigning Sentence Sentiment Scores to Speaker Pairs] The control unit 11 performs the process of assigning the sentiment scores of the sentences to be processed to the identified speaker pairs, using the analysis unit 112 (pair assignment step). The control unit 11 then moves the process to step S7. 【0069】 Through the pair identification step described above and this step, the analysis unit 112 can identify the correspondence between the sentiment score corresponding to the sentence to be processed and the speaker pair. Then, the analysis device 1 can analyze the sentiment network based on this correspondence. 【0070】 If a directed pair is identified in the pair identification step, the analysis unit 112 preferably assigns the sentiment score of the sentence to be processed to the directed pair. This allows the analysis device 1 to identify which speaker the sentiment score corresponding to the sentence to be processed is directed towards. Based on this correspondence, the analysis device 1 can then analyze the sentiment network, including the direction of the sentiment. 【0071】 In this step, it is preferable for the analysis unit 112 to identify the element with the largest absolute value of change from the emotion vector of the sentence immediately preceding the sentence to be processed to the emotion vector of the sentence to be processed as the emotion with the greatest change. This allows the analysis device 1 to utilize the emotion with the greatest change as a representative emotion for analysis. 【0072】 (Time-series chunking process) After calculating the sentiment score for each sentence and assigning the scores to speaker pairs, the analysis device 1 performs a series of processes (time-series chunking process) to identify analysis units along the time series. Step S7 is an example of this process. 【0073】 [Step S7: Identifying Analysis Units] The control unit 11, using the analysis unit 112, divides multiple sentences contained in the content data in chronological order and performs a process to identify the division results as analysis units (analysis unit identification step). The control unit 11 then moves the process to step S8. 【0074】 An example of the procedure for achieving the above identification in this step is given below. Preferably, the procedure for achieving the above identification is configured so that the procedure is selected according to the type of content. 【0075】 (Identifying units of analysis in meeting minutes content) Identifying units of analysis in meeting minutes or similar content can be achieved, for example, by a series of steps that divide the text medium based on the number of sentences. In this case, for example, a predetermined number of sentences may be identified as units of analysis. 【0076】 Meeting records contain statements made by multiple speakers, either consecutively or alternately. Each statement is typically represented as a single sentence in the written medium. However, written materials, particularly those obtained from audio recordings of meeting records using speech recognition, may not be properly divided into paragraphs. Procedures using segmentation based on the number of sentences can identify appropriate units of analysis even in such written materials. 【0077】 To analyze the emotional network at a granularity that allows for the interpretation of its time-series changes, the upper limit of the number in the "pre-specified number of sentences" is preferably 30 or less, and particularly preferably 25 or less. To achieve both reduced processing load and the ability to capture fine-grained changes, the lower limit of the number in the "pre-specified number of sentences" is preferably 15 or more, and particularly preferably 10 or more. 【0078】 (Identifying the unit of analysis in text content) Identifying the unit of analysis in text or similar content is achieved, for example, by a series of steps that involve obtaining the text medium from content data and dividing the text medium based on the constituent units of the text. In this case, each divided text is identified as a unit of analysis. Since the unit of analysis is a block of elements that make up the text, it is possible to analyze changes in emotion along the progression with an appropriate level of granularity. 【0079】 The division of a text medium based on its structural units is performed, for example, based on formal paragraphs or sentences. In the division of a text medium based on formal paragraphs, for example, each formal paragraph is identified as an analysis unit. In the division of a text medium based on sentences, for example, a predetermined number of sentences are identified as analysis units. The upper and lower limits of the number specified in this case may be the same as those for meeting minutes, for example. 【0080】 (Chunk-by-chunk cumulative processing) After identifying the analysis units in a time-series order, the analysis device 1 aggregates the emotion scores for each analysis unit for each speaker pair and performs a series of processes (chunk-by-chunk cumulative processing) to derive a relationship index for each pair. Steps S8 to S11 are an example of this process. 【0081】 [Step S8: Relationship Index Derivation Loop] The control unit 11 executes the process of starting the relationship index derivation loop using the analysis unit 112 (relationship index derivation loop start step). The control unit 11 then moves the process to step S9. 【0082】 [Step S9: Aggregate emotion scores for each speaker pair] The control unit 11 uses the analysis unit 112 to aggregate the emotion scores of each sentence included in the analysis unit for each speaker pair (emotion score aggregation step). The control unit 11 then moves the process to step S10. 【0083】 If the maximum change in emotion has been identified in the pair assignment step, it is preferable that the analysis unit 112 aggregates the maximum change in emotion values ​​for each pair in the analysis target in this step. 【0084】 In addition, in this step, the analysis unit 112 may achieve the aggregation described above by, for example, summing the emotion vectors for each sentence in the analysis unit that have the same speaker pair and associating them with that pair. As a result, the analysis device 1 can obtain an emotion score in the analysis target that includes emotion vectors associated with pairs, which serve as the basis for deriving relationship indicators for each speaker pair. 【0085】 [Step S10: Deriving relationship indicators for each speaker pair] The control unit 11, using the analysis unit 112, performs a process to derive relationship indicators for each speaker pair based on the aggregated sentiment scores described above (relationship indicator deriving step). The control unit 11 then moves the process to step S11. 【0086】 If the maximum change in emotion is identified in the pair assignment step, in this step, the analysis unit 112 preferably determines the type of relationship based on the frequency of occurrence of the maximum change in emotion for each pair in the analysis target. If it is not possible to determine the type of relationship in this step, it is preferable to derive a relationship index through a process that includes a step of determining the type of relationship using a large-scale language model. This allows the analysis unit 112 to determine the type of relationship based on the frequency of occurrence of the maximum change in emotion, even when this derivation is not possible, and to perform a comprehensive emotion network analysis. 【0087】 Determining the type of relationship using a large-scale language model is achieved, for example, by inputting prompts into the large-scale language model that include each sentence relating to the unit of analysis and instructions to estimate the type of relationship based on each sentence, and then determining the type of relationship based on the output of the large-scale language model. Cases where "determination cannot be made by this procedure" include, for example, deriving relationship indicators for pairs that were not assigned sentiment scores in the unit of analysis. 【0088】 Furthermore, it is preferable that the intensity of emotion in the relationship index be calculated based on the intensity of emotion corresponding to the above-mentioned types in the emotion score aggregated for each pair of speakers. 【0089】 [Step S11: Generate graph data] The control unit 11 uses the analysis unit 112 to perform a process to generate data related to the graph of the emotion network, where each vertex corresponds to each speaker as described above, and each edge corresponds to the relationship index described above (graph data generation step). The control unit 11 then moves the process to step S12. 【0090】 By generating graph-related data, the analysis device 1 can present the emotional network shown by the relationship indicators for each analysis unit in a manner that is easy for the user to understand, thereby assisting in the analysis. By performing this generation in the relationship indicator derivation loop, the control unit 11 can efficiently execute various processes related to the time-series changes of the emotional network that are completed within the analysis unit. 【0091】 In order to indicate the direction of emotions, if the analysis unit 112 derives a relationship index based on the emotions (directed emotions) that one speaker directs towards the other speaker in a pair, it is preferable that the graph described above is a directed graph in which the direction of emotions directed from one speaker towards the other speaker is shown by directed edges. 【0092】 To show the type and intensity of emotions, the graph described above is preferably a graph with labels indicating the type of emotion related to the relationship indicator, labels assigning the intensity of the emotion to the edges, and weights. The type of emotion related to the relationship indicator is, for example, the representative emotion described above. 【0093】 [Step S12: Terminate if no analysis units remain] The control unit 11 uses the analysis unit 112 to determine whether or not there are any unprocessed analysis units remaining (relationship index derivation loop termination determination step). If the control unit 11 determines that there are no remaining units, it terminates the relationship index derivation loop and moves the process to step S13; otherwise, it continues the relationship index derivation loop and moves the process to step S8. 【0094】 (Display Processing) After deriving the relationship indicator, the analysis device 1 uses the visual representation data generation unit 114 to generate visual representation data of the time-series changes of the relationship indicator and performs a series of processes (display processing) to display it on a display device (e.g., the display means of terminal T). Steps S13 to S14 are an example of this process. 【0095】 (Utilization of large-scale language models) Regarding the display processing, it is preferable that the analysis device 1, using the findings information generation unit 113, inputs prompts including relationship indicators into a generative language model and performs a series of steps to generate findings information related to the relationships between speakers. Step S13 is an example of such steps. The findings information generation unit 113 is realized, for example, through the cooperation of the control unit 11, the storage unit 13, the communication unit 14 and other hardware components. 【0096】 Through this series of procedures, the analysis device 1 can cause a large-scale language model or other generative language model to generate various texts, including text that outlines the relationships between speakers, text that provides advice for aiming for appropriate relationships according to the type of conversation content data, and other such texts, and provide them to the user. 【0097】 [Step S13: Generate observation information] The control unit 11 uses the observation information generation unit 113 to input a prompt including the relationship indicators described above to the generative language model, and executes a process to cause the generative language model to generate observation information related to the relationship between speakers (observation information generation step). The control unit 11 then moves the process to step S14. 【0098】 In order to generate text based on insights related to emotional networks that are not included in general corpora, it is preferable that the large-scale language model is a pre-trained model obtained by pre-training using training data that associates emotional networks with insights related to those networks. 【0099】 (Indicators related to the structure of the graph) The findings information generation unit 113 preferably generates findings information that includes text based on the indicators related to the structure of the graph described above. This allows the analysis device 1 to generate text about the overall structure of the emotion network based on objective indicators and include it in the findings information. 【0100】 In order to perform graph analysis that reflects the presence or absence of emotions, it is preferable that the analysis device 1 calculates the above-mentioned index by excluding edges corresponding to pairs to which no emotion score was assigned in the analysis unit from the graph. 【0101】 Preferably, the above-mentioned index includes a network centrality index for each vertex corresponding to the speaker. This allows the analysis device 1 to analyze observational information indicating the importance of the speaker in the affect network based on an objective index called the network centrality index. Examples of network centrality indexes include betweenness centrality, degree centrality, closeness centrality, and eigenvector centrality. 【0102】 Furthermore, it is preferable that the above-mentioned indicators include network density. In calculating network density, the analysis device 1 calculates the network density by, for example, normalizing the number of edges included in the graph by the number of edges of a complete graph with the same number of vertices. This allows the analysis device 1 to generate graph structure analysis results using a network density that is independent of the number of speakers. 【0103】 The analysis device 1 may have the generative language model calculate the index, or the analysis unit 112 may calculate the index. By having the generative language model calculate the index, the process from calculating the relationship index to generating observation information based on the index can be realized in a single integrated process. By having the analysis unit 112 calculate the index, calculation errors based on the specifications of the generative language model can be suppressed. 【0104】 (Example of prompt and output) Figure 6 shows an example of a prompt. This figure shows an example of a prompt that is input to the generative language model in this step. 【0105】 The prompt includes instructions for generating findings information, such as a description of the network diagram (a graph of the emotion network), the colors of nodes (vertices) and edges in the visual representation, various assumptions in the graph representation of the emotion network, the procedure for deriving relationship indicators and other information, conversation content data related to the unit of analysis, graph data, graph indicator data, and a command statement instructing the generation of findings information. 【0106】 By including all the information from the network diagram description to the procedure for deriving relational indicators and other information in the prompt, the generative language model can generate observational information that aligns with the network diagram. By including content data and graph data in the prompt, the generative language model can perform a comprehensive analysis of the sentiment network, including not only the network diagram but also the conversation itself. By including graph indicator data, the generative language model can generate observational information based on the indicators. 【0107】 The prompt in the example shown in Figure 6, including the parts omitted in the figure, contains the following instruction: "Based on the above <conditions>, a network analysis was performed, and the <output> network diagram, network density, and results for each centrality index were obtained. Please provide your insights and conclusions that can be gleaned from these results, as well as specific action plans for making the meeting more meaningful, from multiple perspectives, in a concise, clear, and declarative tone. Also, if you are providing general methods when presenting specific action plans, please indicate that as well." This allows the generative language model to generate insights into the affective network, conclusions about the situation indicated by the affective network, and advice regarding that situation. 【0108】 Figure 7 shows an example of output corresponding to the prompt in the previous figure. In this example, Markdown text is output, including specifications for heading structure, highlighting, etc., for visually representing the findings information. As a result, the visual representation data generation unit 114 can provide the findings information in an easy-to-understand visual representation. 【0109】 In the output for this example, following a heading with a visual representation specified in Markdown format, the overall structure of the emotion network, characteristics of the centrality index, changes in network density, edge characteristics (overall characteristics of the emotion types and colors associated with the edges of the graph), conclusions summarizing the analysis of the emotion network, and improvement suggestions based on the emotion network (text providing advice to aim for appropriate relationships according to the type of conversation content data). This allows the analysis device 1 to support the user in aiming for appropriate relationships based on the analysis results and advice of the emotion network. 【0110】 [Step S14: Generate Visual Representation Data] The control unit 11 uses the visual representation data generation unit 114 to perform the process of generating visual representation data for the time-series changes in the emotion network graph, as described above (visual representation data generation step). The control unit 11 returns the process to step S1 and repeats the process from step S1 to step S14. The visual representation data generation unit 114 is realized, for example, through the cooperation of the control unit 11, the storage unit 13, the communication unit 14, and other hardware components. 【0111】 If a centrality index has been calculated, it is preferable that the visual representation data generation unit 114 generates data that includes a representation that reflects the centrality index. One example of such a representation is a representation in which the size of the figure indicating the vertices corresponds to the size of the centrality index. The procedure for calculating this index may be the same as that for an index related to the structure of the graph. As a result, the analysis device 1 can show the importance of speakers in the emotion network based on an objective index called the network centrality index. 【0112】 The visual representation data generation unit 114 preferably generates data that includes representations that reflect the directed emotions described above. An example of such representation is a directed graph in which directed emotions are associated with directed edges. 【0113】 When the analysis unit 112 derives a relationship index based on the magnitude of the emotion scores assigned to a pair, it is preferable for the visual representation data generation unit 114 to generate data that includes a representation that reflects that magnitude. An example of such a representation is a weighted graph that shows the magnitude of emotion related to the relationship index by the thickness of its sides. 【0114】 To make the type and intensity of emotions easier to visualize, it is preferable that the visual representation data generation unit 114 generates data in which the magnitude of emotions related to relationship indicators is indicated by the thickness of the edges, and the type is indicated by attributes different from the thickness (e.g., color, line type). If observation information has been generated, it is preferable that the visual representation data related to this step includes data that displays the observation information as text. 【0115】 (Video Generation) When the content data includes video media, it is preferable that the visual representation data generation unit 114 superimposes a graph corresponding to the analysis unit onto the video in at least a portion of the playback range related to the analysis unit within the video media. This allows the analysis device 1 to provide a visual representation that makes it easy to grasp the relationship between conversational content data and the time-series changes of the emotion network. 【0116】 The visual representation data generation unit 114 preferably displays an animation in the playback range around the point where the analysis unit switches within the video medium, changing from a graph corresponding to the analysis unit before the switch to a graph corresponding to the analysis unit after the switch. This allows the analysis device 1 to clearly show the changes that occur when the analysis unit switches. 【0117】 [Detailed Example of Analysis Process] Figure 8 shows a detailed example of the processing flow from analysis to display in the analysis process. The following is a detailed explanation of a preferred flow of the analysis process performed by the analysis device 1, using the figure. Needless to say, embodiments of the present invention are not limited to the detailed example below. 【0118】 [Acquisition of Sentence-Based Emotion Scores] In the emotion score acquisition process (steps S3 to S4), the analysis device 1 performs sentence-based emotion score acquisition processing. Specifically, the analysis device 1 divides the conversation content data (work) into sentence units, extracts each sentence, analyzes the input sentence, and acquires eight types of emotion scores using a converter (e.g., STORYAI) configured to output emotion scores that include the intensity of multiple types of emotions evoked by the sentence. 【0119】 [Mapping Sentence Characters to Speaker Pairs] In the speaker pair mapping process (steps S5 to S6), the analysis device 1 performs a process of mapping sentence characters to speaker pairs. Specifically, the analysis device 1 pairs two adjacent statements, calculates the absolute value of the difference in emotion scores between the pairs, labels the maximum change in emotion identified based on this absolute value, aggregates representative emotions according to this label, and performs a series of processes to define the representative emotion corresponding to the sentence. 【0120】 [Time-series chunk splitting] In the time-series chunk splitting process (step S7), the analyzer 1 performs a process of splitting sentences into chunks (analysis units) of 15 to 20 sentences each. 【0121】 [Cumulative Processing for Each Chunk (Repeated)] In the cumulative processing for each chunk (steps S8 to S11), the analysis device 1 repeatedly performs cumulative processing for each chunk. Specifically, as long as there is a next chunk, the analysis device 1 aggregates the pairs + representative emotions related to the chunk and stores them in the memory unit 13, acquires and stores the structure of the emotion network based on the aggregated information, calculates and records relationship indicators based on the aggregated information or the output of the generative language model, creates a diagram showing the emotion network graph by color-coding and resizing the directed graph showing the emotion network, and converts the diagram into JSON format so that it can be transmitted as data for visual representation. This series of processes is executed in a loop. 【0122】 [Time-series animation generation] In the display process (steps S12 to S14), the analysis device 1 executes a process to generate a time-series animation. Specifically, the analysis device 1 sequentially takes the data of the above-mentioned figure generated by the program and executes a series of processes to combine it into an animation (e.g., video, slideshow). The analysis device 1 executes this series of processes, for example, by drawing within an HTML5 standard CANVAS element using the Echarts library related to the Javascript plugin. 【0123】 [Program for analyzing time-series changes in emotional networks] The above-described analysis process is performed, for example, by causing the analysis device 1 to execute a content data acquisition step to acquire content data including conversations; an analysis step to derive relationship indicators for each speaker pair based on emotional information obtained from the content data relating to the analysis unit for each analysis unit of the content data along the time series; and a visual representation data generation step to generate visual representation data for the time-series changes in the graph of the emotional network constructed based on the relationship indicators. The graph is executed by a program (program for analyzing time-series changes in emotional networks) in which each vertex corresponds to each speaker and each edge corresponds to the relationship indicators. 【0124】 [Effects of the Analysis Process] In the analysis process described above, the analysis device 1 of this embodiment performs the following: the content data acquisition unit 111 performs the process of acquiring content data including conversation (steps S1 to S2); the analysis unit 112 performs the process of deriving relationship indicators for each speaker pair based on the emotional information obtained from the content data related to the analysis unit for each analysis unit of the content data in a time series (steps S3 to S12); and the visual representation data generation unit 114 performs the process of generating visual representation data of the time series changes in the graph of the emotional network constructed based on the derived relationship indicators (steps S13 to S14). This graph is such that each vertex corresponds to each speaker as described above, and each edge corresponds to the relationship indicator as described above (step S11). 【0125】 In this analysis process, the analysis unit 112 uses emotional information obtained from data related to the analysis unit of content data in a time series to derive relationship indicators for each speaker pair. Then, the visual representation data generation unit 114 generates visual representation data of the time series changes of a graph with each speaker as a vertex and the relationship indicators for each speaker pair as edges, thereby providing the user with the analysis results of the time series changes of the emotional network. 【0126】 Therefore, the analysis device 1 of this embodiment can provide a technical means for analyzing the time-series changes of an emotional network by performing the analysis process described above. 【0127】 [Configuration relating to the second feature] Furthermore, in the analysis process described above, the analysis unit 112 may be configured to obtain an emotional score for each sentence based on the sentences contained in the content data (step S4), extract speaker pairs related to each sentence from the content data (step S5), associate the emotional scores with the pairs (step S6), divide the multiple sentences contained in the content data in chronological order to form analysis units (step S7), and derive relationship indicators for the pairs based on the above-mentioned associations relating to each sentence contained in the analysis unit (steps S9 to S10) (configuration relating to the second feature). 【0128】 As a result, the analysis device 1 can estimate the object of emotion in each sentence with less computational processing than when estimating the object of emotion by syntactic analysis or other natural language processing, derive relationship indicators, and provide users with the results of an analysis of the time-series changes of the emotion network based on these indicators. 【0129】 Therefore, the analysis device 1 of this embodiment can provide a technical means for analyzing the time-series changes of an emotional network while suppressing excessive computational processing by performing the analysis processing configured as described above. 【0130】 [Configuration relating to the third feature] In the analysis processing of the configuration relating to the second feature, the visual representation data generation unit 114 may be configured to generate data that includes a visual representation of the indicators relating to the structure of the graph described above (configuration relating to the third feature). 【0131】 This allows the analysis device 1 to present the structure of the emotional network as a graph to the user through a visual representation based on objective indicators. For example, the analysis device 1 can show the importance of speakers in the emotional network based on an objective indicator called the network centrality index. Also, for example, the analysis device 1 can show the communication density in the emotional network based on an objective indicator called the network density. 【0132】 Therefore, the analysis device 1 of this embodiment can provide a technical means for analyzing the time-series changes of an emotional network, including structural analysis of the emotional network as a graph, by performing the analysis processing configured as described above. 【0133】 [Configuration relating to the fourth feature] In the analysis process relating to the second feature, the analysis unit 112 may be configured to derive relationship indicators based on the emotions (directed emotions) that one speaker directs towards the other speaker in the aforementioned pair, and the visual representation data generation unit 114 may be configured to generate data that includes expressions that reflect the directed emotions (configuration relating to the fourth feature). 【0134】 The emotions between speakers are not always symmetrical. For example, one speaker may feel strong trust towards another, while the other speaker feels strong anger towards the first speaker. Furthermore, in the time-series changes of the emotional network, such asymmetrical emotional relationships can arise, develop, converge, or change into symmetrical relationships. 【0135】 Therefore, in the analysis of emotion networks, the analysis of the direction of emotions is important. With the above configuration, the analysis device 1 can analyze the time-series changes in emotions between speakers through the representation of the emotion network, including its direction. 【0136】 Therefore, the analysis device 1 of this embodiment can provide a technical means for analyzing the time-series changes of an emotional network, including the analysis of the direction of emotions, by performing the analysis processing configured as described above. 【0137】 [Configuration relating to the fifth feature] In the analysis process relating to the second feature, the analysis unit 112 may be configured to derive a relationship index based on the magnitude of the emotion scores assigned to the above-mentioned pair, and the visual representation data generation unit 114 may be configured to generate data that includes a representation that reflects the magnitude (configuration relating to the fifth feature). 【0138】 The intensity of emotions is not always constant. Even if the type of emotion is the same, if its intensity differs, the emotional network between speakers can take on different characteristics. For example, an emotional network connected by relatively small feelings of anger, i.e., weak anger, may only show the beginnings of a conflict, while an emotional network connected by relatively large feelings of anger, i.e., strong anger, may indicate a conflict that is difficult to resolve. 【0139】 Furthermore, the results of the analysis of the time-series changes in the emotional network will be completely different depending on whether the time-series changes correspond to the progression of a situation or the resolution process. Therefore, the analysis of the intensity of emotions is important in the analysis of emotional networks. With the above configuration, the analysis device 1 can analyze the emotions between speakers through the representation of the emotional network, including the time-series changes in their intensity. 【0140】 Therefore, the analysis device 1 of this embodiment can provide a technical means for analyzing the time-series changes of an emotional network, including the analysis of the magnitude of emotions, by performing the analysis processing configured as described above. 【0141】 [Configuration relating to the sixth feature] The analysis process may be configured to include a process in which the findings information generation unit 113 inputs a prompt including relationship indicators to a generative language model and generates findings information relating to the relationships between speakers (configuration relating to the sixth feature). 【0142】 According to the above configuration, the analysis device 1 can provide the user with findings information including the overall structure of the emotion network, characteristics of the centrality index, changes in network density, edge characteristics (overall characteristics of the types and colors of emotions associated with the edges of the graph), conclusions summarizing the analysis of the emotion network, improvement suggestions based on the emotion network (text providing advice to aim for appropriate relationships according to the type of conversation content data), and other information. 【0143】 As a result, the analysis device 1 can provide the user with the results of the analysis of the time-series changes in the emotional network and advice based on that analysis. This allows the user to understand the emotional network, for example, by looking at the time-series changes in the emotional network graph and the results of its analysis, and to effectively utilize the advice based on that understanding. 【0144】 Therefore, the analysis device 1 of this embodiment can provide a technical means for analyzing the time-series changes of an emotional network, including the analysis of the magnitude of emotions, by performing the analysis processing configured as described above. 【0145】 <Example of Use> The following is a description of an example of use of the analytical apparatus 1 in this embodiment. 【0146】 [Action Time Series Network Analysis of Meeting Records] The following is an example of an action time series network analysis of meeting records. 【0147】 [Specifying the text to be analyzed] Users specify the content data of the conversation to be analyzed by uploading the content data, specifying a URL, or by other means. 【0148】 [Analysis Execution] Analysis device 1 acquires the specified content data and calculates an emotion score for each sentence. Then, analysis device 1 groups 20 to 25 sentences into analysis units (chunks), generates an emotion network graph for each, and generates visual representation data showing the time-series changes. 【0149】 Figure 9 shows an example of how the analysis information is displayed. In this example, the emotion network graph generated by the visual representation data generation unit 114 is displayed as sub-figures for each time segment: the whole, the first 10 minutes, the middle 10 minutes, and the last 8 minutes. The vertices of each sub-figure are associated with speakers A through D, and each side is associated with a relationship indicator. 【0150】 Each side of the diagram indicates directed emotions between speakers with an arrow, the type of emotion with a line type, and the intensity of the emotion with a line width. On the right side, a legend is displayed showing the correspondence between each emotion (anger, expectation, joy, trust, fear, surprise, sadness, and disgust) and the line type. This allows the user to see the structure of the emotion network based on relationship indicators and its temporal progression at a glance. 【0151】 In this example, by displaying a graph of the "overall" period, which aggregates the entire timeframe, and graphs for each interval on the same screen, it is possible to comparatively grasp the points of change, the emergence of relationships, increases and decreases in intensity, asymmetry in direction, and other characteristics for each analysis unit (chunk). Furthermore, although omitted in this figure, if a centrality index is calculated, the size of the vertex figures can be displayed corresponding to the size of that index. 【0152】 In this example, the findings information generation unit 113 presents the "conclusion" and "action plan" as text side-by-side, and the graph analysis results and advice based on those results are displayed on the same screen. This allows the user to continuously refer to the visualization of the structure and progression of the emotional network and the implications based on that visualization. 【0153】 [Analysis of Emotional Time-Series Networks in Novels] The following is an example of an analysis of emotional time-series networks in novels. 【0154】 [Specifying the text to be analyzed] The user specifies the content data of the conversation to be analyzed by uploading the content data, specifying a URL, or by other means. The user specifies content data related to the beginning of Dostoevsky's novel "Crime and Punishment". 【0155】 [Analysis Execution] Analysis device 1 acquires the specified content data and calculates an emotion score for each sentence. Then, analysis device 1 groups 20 to 25 sentences into analysis units (chunks), generates an emotion network graph for each, and generates visual representation data showing the time-series changes. 【0156】 Figure 10 is an example of a representation of an emotional network analysis for the first scene of a novel. In this example, the following emotional network is depicted for an early scene in Crime and Punishment. 【0157】 For Raskolnikov (symbol A), the filled-in vertex corresponding to him indicates feelings of disgust. Furthermore, the type and thickness of the directed edge from him to Sonya (Sofia Semyonovna; symbol B) indicates weak disgust towards her, the type and thickness of the directed edge from him to Katerina Ivanovna (symbol C) indicates strong sadness towards her, the type and thickness of the directed edge from him to Marmeladov (symbol D) indicates strong sadness towards her, and the type and thickness of the directed edge from him to Alyona Ivanovna (symbol E) indicates moderate disgust towards her. 【0158】 For Sonya (symbol B), the filled-in vertex corresponding to her indicates a feeling of trust. A strong mutual trust is depicted between Sonya and Marmeladov. 【0159】 For Katerina Ivanovna (symbol C), the filled-in vertex corresponding to her indicates anger. The directed edge pattern and thickness from Katerina Ivanovna toward Marmeladov indicates strong anger directed toward the other person, and furthermore, the directed edge pattern and thickness from her toward Sonya indicates strong sadness directed toward the other person. 【0160】 For Marmeladov (symbol D), the filled-in vertex corresponding to him indicates sadness. The directed edge pattern and thickness from Marmeladov to Raskolnikov indicates strong trust in the other, the directed edge pattern and thickness from him to Sonya indicates strong trust in the other, and the directed edge pattern and thickness from him to Katerina Ivanovna indicates strong fear in the other. 【0161】 Regarding Alyona Ivanovna (symbol E), the filled-in vertex corresponding to her indicates feelings of disgust. The pattern and thickness of the directed edge from Alyona Ivanovna to Raskolnikov indicate a weak fear directed towards the other. 【0162】 Thus, the graph provides a visual representation that allows viewers to see at a glance each character's emotions and the emotions they direct towards others, and to see both asymmetrical emotional relationships (e.g., the interplay of trust and sadness between Raskolnikov and Marmeladov) and symmetrical emotional relationships (e.g., mutual trust between Marmeladov and Sonya). 【0163】 Furthermore, the example includes the following commentary on the network analysis: "This network graph visualizes the complex relationships and emotional dynamics of the characters in the early part of Dostoevsky's novel 'Crime and Punishment.' At the center of the graph is the protagonist, Raskolnikov, and two different worlds are linked together: his isolated relationship with Alyona Ivanovna, the target of his murderous intent, and the tragic relationship he forms with the Marmeladov family, whom he encounters by chance." This demonstrates that analysis device 1 not only displays a graph but also provides supplementary explanations to help users easily understand the meaning of the illustrated structure by interpreting the relationships between the characters. 【0164】 Furthermore, the example includes a user interface for adjusting vertex size (node ​​size) and edge thickness, along with a button to command printing. In addition, this example displays a character list sorted by degree, indicating the number of times each character interacts with other characters, with symbols accompanied by degree and a filled-in legend. Thus, the analysis device 1 provides a visual representation and user interface that allows the user to intuitively grasp the strength of relationships and the relative importance of each character, adjust the display as needed, and print, record, and share the information. 【0165】 In this example, the analysis device 1 provides a visual representation that allows for an overview of the complex emotional network between characters in the opening section of Dostoevsky's novel "Crime and Punishment." By providing a similar visual representation for each analysis unit, the analysis device 1 can provide a visual representation that allows for an overview of the changes in the characters and the chronological changes in the complex emotional network between them. 【0166】 <Note> Within the scope of the concept of the present invention, a person skilled in the art can conceive of various modifications and alterations, and it is understood that such modifications and alterations also fall within the scope of the present invention. For example, any addition, deletion, or design change of components, or addition of processes or changes in conditions, made by a person skilled in the art to the above-described embodiment, is also included within the scope of the present invention, as long as it retains the gist of the present invention. [Explanation of Symbols] 【0167】 S System 1 Analyzer 11 Control Unit 111 Content Data Acquisition Unit 112 Analysis Department 113 Findings information generation unit 114 Visual representation data generation unit 13 Storage section 131 Analysis Information Database 14 Communications Department N Network T terminal

Claims

[Claim 1] A content data acquisition unit that acquires content data including conversations, For each analysis unit of the content data arranged in chronological order, an analysis unit derives a relationship index for each speaker pair based on the emotional information obtained from the content data related to that analysis unit, A visual representation data generation unit generates data for visually representing time-series changes in an emotion network graph constructed based on the aforementioned relationship indicators. An observation information generation unit inputs a prompt including the aforementioned relationship indicator into a generative language model and generates observation information related to the relationship between speakers, including the results of an analysis of the time-series changes in the affect network and text showing advice based on the analysis. Equipped with, The graph described above is a graph in which each vertex corresponds to each speaker and each edge corresponds to the relationship index. A device for analyzing the time-series changes in emotional networks. [Claim 2] The aforementioned analysis unit is With respect to the aforementioned sentiment information, a sentiment score is obtained for each sentence based on the content data, From the aforementioned content data, speaker pairs related to each sentence are extracted, and the sentiment scores are associated with those pairs. Multiple sentences contained in the aforementioned content data are divided chronologically and used as the analysis units. Based on the correspondence for each sentence included in the aforementioned unit of analysis, a relationship index for the pair is derived. The analytical apparatus according to claim 1. [Claim 3] The aforementioned visual representation data generation unit generates data including a visual representation of the indicators related to the structure of the graph. The analytical apparatus according to claim 2. [Claim 4] The analysis unit derives the relationship index based on the emotions (directed emotions) that one speaker directs towards the other speaker in the pair. The aforementioned visual representation data generation unit generates data that includes representations that reflect the directed emotion. The analytical apparatus according to claim 2. [Claim 5] The analysis unit derives the relationship index based on the magnitude of the emotion scores assigned to the pair, The aforementioned visual representation data generation unit generates data that includes a representation that reflects the size. The analytical apparatus according to claim 2.

Images