Risk assessment device and risk assessment system

Abstract

To provide a risk assessment device that detects harassment and the like without hindering the free exchange of opinions and notifies the appropriate person of the risk assessment result. [Solution] A risk assessment device that performs risk assessment of statements contained in log information within a company, comprising: a log information acquisition unit that acquires the log information; a risk assessment unit that uses a large-scale language model to analyze the content of the statements and their contextual information contained in the log information, classifying the statements into at least one of three levels: high risk, medium risk, or low risk according to their compliance risk level, and generating risk assessment data by associating the risk level with the log information; a database that stores the risk assessment data and accumulates the stored risk assessment data as risk assessment history data organized by speaker; and a risk notification unit that outputs the risk assessment history data accumulated in the database as risk report data in a predetermined report format and sends it to a predetermined notification destination.

Description

[Technical Field]

[0001] The present disclosure relates to a risk assessment device and a risk assessment system for assessing compliance risks. [Background technology]

[0002] In recent years, companies have been required to take appropriate measures against harassment in internal communications (power harassment, sexual harassment, maternity harassment, etc.) as part of their corporate social responsibility and compliance requirements.

[0003] Patent document 1 discloses a video conference analysis program that converts speech made by participants in a video conference held over a network from audio data to text data, analyzes the text data, and obtains analysis results such as a harassment level that indicates the degree of harassment, and shares these analysis results with participants in the video conference. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2023-097789 Summary of the Invention [Problem to be solved by the invention]

[0005] The video conference analysis program disclosed in Patent Document 1 shares the harassment level of a participant who speaks with the speaker and other participants, which can cause participants to become too conscious of their own analysis results and hesitate to speak up, which can result in the free exchange of opinions being hindered. In view of the above-mentioned problems, an object of the present disclosure is to provide a risk assessment device that detects harassment and the like without hindering the free exchange of opinions and notifies the appropriate people of the detection results. [Means for solving the problem]

[0006] The risk assessment device disclosed herein is a risk assessment device that performs risk assessment of statements contained in log information within a company, and includes: a log information acquisition unit that acquires the log information; a risk assessment unit that uses a large-scale language model to analyze the content of the statements contained in the log information and their context information, classifies the statements into at least three levels of high risk, medium risk, and low risk according to their compliance risk level, and generates risk assessment data by associating the risk level with the log information; a database that stores the risk assessment data and accumulates the stored risk assessment data as risk assessment history data organized by speaker; and a risk notification unit that outputs the risk assessment history data accumulated in the database as risk report data in a predetermined report format and transmits the data to a predetermined notification destination. The large-scale language model is a model constructed by learning based on high risk, medium risk, and low risk labels assigned to past log information, and the predetermined notification destinations are set in advance to distinguish notification targets based on notification conditions set for each of the risk levels, and the notification conditions include at least one of a condition for sending a notification each time the risk assessment history data is generated and a condition for sending a notification when the generation of the risk assessment history data for the same speaker exceeds a predetermined number of times, and the risk notification unit selects a corresponding notification target from among the plurality of notification targets in accordance with the risk level related to the most recent comment content added to the risk assessment history data, and transmits the risk report data to the notification target. . [Effects of the Invention]

[0007] According to the risk assessment device of the present disclosure, harassment and the like can be detected without hindering the free exchange of opinions, and the appropriate people can be notified of the risk assessment results. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram showing a schematic configuration of a risk assessment system according to an embodiment. [Figure 2] FIG. 2 is a block diagram illustrating the functional configuration of the log information analysis server according to the embodiment. [Figure 3] FIG. 3 is a block diagram showing the functional configuration of the risk assessment server according to the embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of risk report data according to the embodiment. [Figure 5] FIG. 5 is a flowchart showing the flow of the risk assessment process of the risk assessment system according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. Note that the present disclosure is not limited to these examples, but is defined by the claims, and is intended to include all modifications within the meaning and scope of the claims.

[0010] <Risk assessment system> FIG. 1 is a block diagram showing a schematic configuration of a risk assessment system S according to this embodiment. 1, the risk assessment system includes a log information acquisition server (log information acquisition device) 10, a log information analysis server (log information analysis device) 20, and a risk assessment server (risk assessment device) 30. These servers are connected to each other so that they can communicate with each other via a network N including the Internet, etc. The network N may include a LAN (Local Area Network) and / or a WAN (Wide Area Network).

[0011] Each server may be configured as a general-purpose computer such as a workstation or personal computer, or may be logically distributed in a cloud computing environment. That is, the configuration in this embodiment may be consolidated into one physical device or distributed across multiple physical devices.

[0012] Furthermore, various processing functions are realized as programs executed by a computer. These programs may be stored in a storage unit (e.g., ROM, RAM, SSD, HDD, etc.) and executed by a processing unit (e.g., CPU, GPU, AI accelerator, etc.).

[0013] <Log information acquisition server 10> The log information acquisition server 10 acquires log information related to chat messages, audio conferences, video conferences, etc. exchanged within a company, and transmits the acquired log information to the log information analysis server 20. The log information acquisition server 10 acquires log information using an API (Application Programming Interface) appropriate for the chat service, audio conference, or video conference being used. The log information includes data related to the sender of the chat message, the content sent, and the date and time of sending, the date and time of the audio or video conference, participant information, the content of comments, the content of chat messages sent during the conference, audio data, video data, transcription data, and comment section information for each speaker.

[0014] The log information acquisition server 10 may acquire, as text data, transcription data generated by a transcript (transcription function) on the audio conference or video conference application side, in addition to the audio information included in the log information.

[0015] The log information acquisition server 10 transmits log information including the various data and information described above to the log information analysis server 20. The log information acquisition server 10 may be installed within the company's network. By providing the log information acquisition server 10 on the same network as the various chat services, audio conferencing services, and video conferencing services operated by the company, log information can be acquired efficiently and safely.

[0016] <Log information analysis server 20> FIG. 2 is a block diagram showing the functional configuration of the log information analysis server 20 according to this embodiment. In Figure 2, the log information analysis server 20 includes a minutes data generation unit 201 that generates minutes data related to meetings and business activities based on the log information received from the log information acquisition server 10, and a minutes data transmission unit 202 that transmits the generated minutes data to the risk assessment server 30.

[0017] The minutes data generation unit 201 executes natural language processing tasks using a large language model (LLM) constructed using machine learning techniques including artificial intelligence (AI). An LLM is a language model that learns a large amount of text data in advance and can execute various language processing tasks by simply providing a few example tasks. Examples of such LLMs include GPT-4.5 and GPT-o3 developed by OpenAI.

[0018] The LLM used in the minutes data generation unit 201 has undergone pre-training in natural language processing, and has trained on large amounts of document data, for example, contained in minutes of meetings, meeting minutes, business chat exchanges, etc. This training enables the LLM to infer the meaning and intention of statements, taking into account differences in context and phrasing, and can execute processing to classify the content of statements into categories such as "agenda," "decisions," "unresolved matters," "issues," "to-do," and "concerns."

[0019] The minutes data generation unit 201 includes an agenda extraction unit, a decision item extraction unit, an unresolved item extraction unit, a problem extraction unit, a ToDo extraction unit, and a concern extraction unit.

[0020] The agenda extraction unit uses LLM to detect standard phrases such as "The purpose of this meeting is..." or "Today's theme is..." that appear during speech, as well as timing for topic changes at the beginning or middle of the meeting, and extracts the relevant content as agenda information.

[0021] The decision extraction unit uses LLM to detect language patterns that express decision-making, such as "It has been decided that..." or "It has been decided that..." in the utterances, and extracts the relevant content as decision information. Note that if the utterances span multiple utterances, the flow of the utterances is analyzed and integrated.

[0022] The unresolved matter extraction unit uses LLM to detect statements containing reserved expressions such as "We will consider this in the future" or "We will reach a conclusion by next week," and extracts topics on which no conclusion has been reached as unresolved matter information.

[0023] The issue extraction section uses LLM to detect statements that include business problems, obstacles, and improvement suggestions (for example, "There is an issue with this point" or "We are not able to respond in time") and extracts them as issue information that needs to be addressed.

[0024] The ToDo extraction unit uses LLM to detect statements that specify the person in charge and specify an action (e.g., "Mr. / Ms. X, please prepare the materials by next week" or "I will check it") and extracts them as ToDo information. Note that the name of the person in charge and the deadline for action can also be added as metadata.

[0025] The concern extraction unit uses LLM to detect statements that contain negative expressions (for example, "There is a risk if things continue like this" or "This is a cause for concern"), or contexts such as raising doubts or sharing concerns, and extracts them as concern information.

[0026] The minutes data generating unit 201 organizes each piece of information extracted by each of the above-mentioned extracting units in accordance with a predetermined format to generate minutes data. The minutes data generator 201 classifies the extracted information by information type, organizes it by assigning corresponding category headings (for example, "agenda," "decisions," "unresolved matters," "issues," "to do," "concerns," etc.), and arranges the information in each category in chronological order according to the order in which the comments were made. This makes it possible to generate minutes data that is highly visible to the reader and has a unified information structure. Note that the above-mentioned category divisions and chronological arrangement are merely examples and are not limiting. The category structure and heading names may be changed as needed, or the information may be classified or arranged based on other criteria, such as by speaker or department.

[0027] The minutes data generation unit 201 adds metadata to the minutes data, such as the name of the person in charge, the name of the speaker, the deadline for response, and the date and time of the statement, and then converts the entire minutes into a specified data format (e.g., PDF format, HTML format, CSV format, etc.) and outputs it. The minutes data transmission unit 202 transmits the minutes data generated by the minutes generation unit to the risk assessment server 30. The minutes data transmission unit 202 also transmits the generated minutes data to a company terminal device (for example, an in-house server, a management PC, or a cloud management screen) via the network N. This allows the company to check, store, and reuse the minutes data.

[0028] <Risk Assessment Server 30> FIG. 3 is a block diagram showing the functional configuration of the risk assessment server 30 according to this embodiment. In Figure 3, the risk assessment server 30 includes a database 301 that stores minutes data received from the log information analysis server 20, a risk assessment unit 302 that evaluates statements that may pose compliance issues based on the stored minutes data, and a risk notification unit 303 that notifies specific individuals of the risks assessed by the risk assessment unit 302.

[0029] The database 301 organizes and stores various minutes data generated by the minutes data generator 201 in chronological order, and is capable of searching and extracting based on metadata (such as speaker, comment date and time, category, etc.). That is, each category of information in the minutes data (decisions, pending matters, to-do lists, concerns, etc.) is assigned metadata such as the corresponding speaker name, comment date and time, person in charge, and deadline for response, enabling conditional searches, narrowing down, and statistical processing based on the metadata. The database 301 may be a relational database system (RDBMS), or may be a document-based database, a time-series database with a timestamp, or a full-text search engine with an index structure (e.g., Elasticsearch).

[0030] The risk assessment unit 302 uses LLM to analyze the context of statements contained in the minutes data, automatically detect statements that pose compliance issues (particularly various types of harassment such as power harassment, sexual harassment, and maternity harassment), and evaluate the detected statements based on predetermined evaluation criteria according to their risk level (degree of risk). The LLM used in the risk assessment unit 302 is trained based on supervised data in which the content of statements contained in past meeting minutes, chat data, etc. are labeled with risk levels such as "high risk," "medium risk," and "low risk," and can classify risk levels according to the content and context of statements. In other words, the LLM functions as a classification model that comprehensively judges the content, tone, context, etc. of the input minutes data, evaluates whether the statement in question poses a problem from a compliance perspective, and classifies the risk level.

[0031] The LLM used in the risk assessment unit 302 performs risk assessment on the minutes data by sequentially executing processing such as dividing the minutes into statement units, adding context, classifying the minutes data, generating risk assessments, and storing assessment results.

[0032] The utterance unit division process divides a series of utterance logs included in the minutes data into speakers or utterance turns, and extracts each utterance as an individual evaluation unit. The extracted units are structured data that include information such as the utterance sentence, speaker, and utterance time.

[0033] The contextualization process refers to the content of previous and following statements and the flow of communication between speakers, and adds contextual information to supplement the meaning of each statement. The contextual information may include the topic of the conversation, repetition of phrases, negative tone, sarcastic connotation, etc.

[0034] The classification process involves the LLM receiving a statement and its contextual information as input, assessing the likelihood that the statement contains a compliance risk, and classifying it into three risk levels: "high risk," "medium risk," and "low risk" depending on the degree of risk. Here, LLM is a model constructed in advance through supervised learning, and is trained based on labels such as "high risk," "medium risk," and "low risk" assigned to past meeting logs, etc. LLM comprehensively analyzes the content, tone, emotional intensity, and social context of a statement to estimate the risk level of the statement and assign a classification label.

[0035] The risk assessment generation process is a process of structuring the risk assessment results obtained by the classification process as metadata corresponding to the comment. Specifically, in the classification process using LLM, serious problematic remarks that are likely to constitute harassment (e.g., "You're useless," "You still haven't gotten married," etc.) are classified as "high risk," problematic remarks that do not directly constitute harassment but may lead to psychological pressure or the imposition of prejudice if repeated (e.g., "You haven't done it yet?", "Have you got a boyfriend?", etc.) are classified as "medium risk," and problematic remarks that could be a sign of harassment (e.g., "Read the room," "You're doing a great job despite being a woman," etc.) are classified as "low risk."

[0036] Note that the above-mentioned example statements (such as "useless guy") are merely examples, and actual risk assessment is carried out after comprehensively analyzing the context, tone, and content of the statement based on a dataset trained by a large-scale language model (LLM). Therefore, the specific risk assessment results may vary depending on the situation in which the statement was used and its relevance to surrounding statements.

[0037] The risk assessment generation process generates assessment result data including information such as the utterance content, speaker identifier, utterance time, risk level, and risk category. Here, the utterance content is the text evaluated by LLM (e.g., "You're doing a great job, even though you're a woman"), the speaker identifier is the speaker's name, affiliation, or ID, the utterance time is the timestamp of the utterance, the risk level is the classification result assigned to the utterance (e.g., "low risk"), and the risk category is the type of risk (e.g., sexual harassment). The risk assessment data generated in this way is output in the following structure, for example.

[0038] (Statement: "You're doing a great job, even though you're a woman" / Speaker: "Employee A" / Time: "2025-05-01 14:35" / Risk level: "Low risk" / Category: "Sexual harassment")

[0039] The risk assessment accumulation process is a process of storing risk assessment data related to the most recent statements in the database 301. Each evaluation result data is accompanied by metadata such as the statement content, the name of the speaker, the time of statement, the risk level, and the risk category, and based on this information, risk assessment results for each speaker are organized in chronological order and accumulated to form risk assessment history data.

[0040] The database 301 is a storage device that accumulates risk assessment history data, and can perform conditional searches using metadata such as the content of statements, speaker information, time of statement, risk level, and risk category, range-specified extraction, statistical processing by category, time-series analysis, and speaker-specific aggregation. This makes it possible to track and analyze, on a speaker-by-speaker basis, whether or not the same speaker repeatedly makes problematic statements, risk trends in statements, or changes in the frequency of risk assessments over a specific period of time. In this way, based on the risk assessment history data of each speaker accumulated in database 301, it becomes possible to monitor and create reports using a variety of methods, such as extracting speech tendencies of "medium risk" or higher, visualizing the distribution of occurrences of specific categories (e.g., sexual harassment, power harassment, etc.), and even comparing risk distribution by department and time period.

[0041] The risk notification unit 303 generates predetermined risk report data 310 based on the risky comment history data accumulated in the database 301 and transmits the generated risk report data 310 to pre-registered notification destinations. Here, a user (e.g., a company administrator) can pre-set notification recipients (notification recipients) corresponding to the risk levels included in the risk report data 310. Examples of notification recipients include a compliance officer (e.g., a compliance officer or chief compliance officer) who is notified when the risk assessment result is classified as "high risk," a manager in the department to which the commenter belongs who is notified when the risk assessment result is classified as "medium risk," and a monitor who regularly monitors risk assessments who is notified when the risk assessment result is classified as "low risk." The notification recipient setting information (notification recipient information) corresponding to each risk level is stored in an administrator setting database provided in the risk assessment server 30, and the user (company administrator) can register or change the setting information via a predetermined management screen (management interface). In addition, in order to enhance the linkage with the business systems of the company, the notification destination setting information may be registered or changed via an API provided by an external company system.

[0042] Here, risk assessment data is generated as a result of risk assessment processing for each statement, and is used primarily for storage in database 301 and analysis as intermediate data that holds the assessment details (risk level, category, etc.) for each individual statement, while risk report data 310, as shown in Figure 4, is generated for the purpose of providing information to notification recipients based on the most recent risk assessment results, and is output data for reporting that may include multiple risk assessment data as needed, past risk assessment history for the same speaker, chronological trend information, aggregated results, etc.

[0043] The risk report data 310 is structured and output in a predetermined data format such as PDF format, HTML format, CSV format, etc. Fig. 4 is a diagram showing an example of the risk report data 310 according to this embodiment, and various configurations and styles can be adopted for the display format of the risk assessment history, the method of compiling statistical information, the visualization format of trend analysis, etc. depending on the operational environment and purpose.

[0044] The risk notification unit 303 sends the risk report data 310 to the notification destination via a predetermined channel (email, chat tool, management dashboard, etc.), which enables prompt information transmission to the appropriate person in charge according to each risk level.

[0045] Furthermore, the risk notification unit 303 can set in advance conditions for the timing of notification. That is, a user (for example, a company administrator) can arbitrarily set notification conditions for each risk level. Specifically, a notification can be configured to be sent in stages, such as sending a notification in real time each time a "high risk" or "medium risk" utterance is detected, and sending a notification only when a "low risk" utterance is detected for that speaker a predetermined number of times (for example, three times) or more within a certain period of time. In this way, by making it possible to adjust notification conditions according to the severity and frequency of the risk, it is possible to promote rapid responses to comments that require a high level of urgency, while reducing the operational burden caused by excessive notifications.

[0046] Furthermore, the risk notification unit 303 according to this embodiment may be configured to automatically or semi-automatically request confirmation or review from an external expert (for example, a law firm, a corporate lawyer, an external compliance specialist organization, etc.) for risk report data 310 that satisfies predetermined conditions. Specifically, a user (for example, a company administrator) can pre-register an external expert as a notification destination in addition to the above-mentioned internal notification destinations, and the risk report data 310 may be sent to the external expert based on predetermined criteria, such as when a statement classified as "high risk" contains a specific keyword or when risk assessment data for the same speaker has been accumulated a certain number of times or more. Based on the received risk report data 310, the external expert organization can provide the company with a legal risk assessment of the person who made the comment, an evaluation of the appropriateness of the response policy, and the presentation of corrective measures and guidance as necessary. This allows for risk management that incorporates a highly objective third-party perspective, improving the accuracy and reliability of the company's compliance response.

[0047] In this way, in this embodiment, by analyzing the content of internal company comments and notifying predetermined recipients of the content according to the risk level, it is possible to grasp and respond to risks without disrupting the progress of the meeting. This allows participants in internal meetings, etc., to maintain free and open discussion without being intimidated, while ensuring a compliance risk monitoring system for the entire organization. Furthermore, since the assessed risk is notified only to recipients (e.g., compliance officers) according to the risk level, information is shared only with the minimum number of parties necessary, preventing unnecessary repercussions and the spread of misunderstandings.

[0048] <Risk assessment process flow> 5 is a flowchart showing the flow of the risk assessment process of the risk assessment system S according to this embodiment. Hereinafter, the flow of the risk assessment process of the risk assessment system S according to this embodiment will be described with reference to FIG.

[0049] In step S1, the log information acquisition server 10 acquires business log information such as chat messages, voice conferences, and video conferences within a company.

[0050] In step S2, the acquired log information is sent to the log information analysis server 20. The acquired log information includes the sender, the content of the speech, the time of speech, conference participant information, audio and video data, transcription data, and the like.

[0051] In step S3, the minutes generation unit of the log information analysis server 20 extracts information such as agenda items, decisions, pending matters, issues, tasks, concerns, etc. based on the log information using LLM, and generates minutes data organized by category. At this time, metadata such as the name of the speaker, the time of the comment, and the deadline for response are added to each piece of information.

[0052] In step S4, the minutes data transmission unit 202 of the log information analysis server 20 transmits the generated minutes data to the risk assessment server 30. The minutes data transmission unit 202 also transmits the generated minutes data to a terminal device on the company side via the network N.

[0053] In step S5, the risk assessment unit 302 of the risk assessment server 30 executes the following processes for the minutes data: dividing the minutes data into utterances, adding context, classifying the minutes data, generating evaluation results, and storing the evaluation results. Specifically, the utterances in the minutes data are divided into utterances, and context information is added to each utterance. Then, in the classification process, the LLM inputs the context-attached utterances, determines whether the utterances pose a compliance risk, determines their severity (high risk, medium risk, or low risk), and assigns a classification label (risk level). The classified risk assessment results are structured as risk assessment data including the utterance content, speaker identifier, utterance time, risk assessment, risk category, etc., by the evaluation result generation process, and stored in the database 301. Here, the risk assessment results for each speaker are organized in chronological order, and risk assessment history data is constructed. The risk notification unit 303 of the risk assessment server 30 generates risk report data 310 based on the risk assessment history data in the database 301.

[0054] In step S6, the risk notification unit 303 of the risk assessment server 30 sends the risk report data to the notification destination corresponding to the risk assessment (e.g., high risk: compliance officer, medium risk: manager of the department to which the commenter belongs, low risk: monitoring officer) via a predetermined channel (email, chat, etc.). The notification conditions (real-time notification, etc.) can be set in advance by the user. The risk notification unit 303 of the risk assessment server 30 may also send a risk report to an external expert organization (such as a legal advisor) if the risk report data 310 to be notified meets predetermined conditions (e.g., includes a specific keyword, or the cumulative number of risk cases exceeds a predetermined value). This allows for a review from a legal perspective.

[0055] In this way, a series of risk assessment processes for meeting minutes data can be automated in stages, and consistent management of recording, assessment, and notification regarding compliance risks can be achieved.

[0056] As described above, this embodiment is a risk assessment device that performs risk assessment of statements contained in log information within a company, and is equipped with a log information acquisition unit that acquires the log information, a risk assessment unit that uses a large-scale language model to analyze the content of the statements and their contextual information contained in the log information, and classifies the statements into at least one of three levels: high risk, medium risk, and low risk, depending on the compliance risk level, and associates the risk level with the log information to generate risk assessment data, a database that stores the risk assessment data and accumulates the stored risk assessment data as risk assessment history data organized by speaker, and a risk notification unit that outputs the risk assessment history data accumulated in the database as risk report data in a predetermined report format and sends it to a predetermined notification destination, thereby making it possible to detect harassment, etc. without hindering the free exchange of opinions, and notify the appropriate people of the risk assessment results. [Explanation of symbols]

[0057] S Risk Assessment System N Network 10 Log information acquisition server 20 Log information analysis server 201 Minutes Data Generation Department 202 Minutes Data Transmission Department 30 Risk Assessment Server 301 Database 302 Risk Assessment Department 303 Risk Notification Department 310 Risk Report Data

Claims

1. A risk assessment device that assesses the risk of statements contained in log information within a company, a log information acquisition unit that acquires the log information; a risk assessment unit that uses a large-scale language model to analyze the content of statements and their context information contained in the log information, classifies the statements into at least three levels of high risk, medium risk, and low risk according to their compliance risk level, and associates the risk level with the log information to generate risk assessment data; a database in which the risk assessment data is stored and which accumulates the stored risk assessment data as risk assessment history data organized by speaker; a risk notification unit that outputs the risk assessment history data accumulated in the database as risk report data in a predetermined report format and transmits the data to a predetermined notification destination; Equipped with the large-scale language model is a model constructed by learning based on high-risk, medium-risk, and low-risk labels assigned to past log information; The predetermined notification destinations are set in advance based on notification conditions set for each of the risk levels, with notification recipients being distinguished from one another, the notification condition includes at least one of a condition for issuing a notification each time the risk assessment history data is generated and a condition for issuing a notification when the risk assessment history data for the same speaker has been generated more than a predetermined number of times; the risk notification unit selects a corresponding notification destination from among the plurality of notification targets according to a risk level related to the most recent comment content added to the risk assessment history data, and transmits the risk report data to the corresponding notification destination; Risk assessment device.

2. The risk assessment data generated by the risk assessment unit is structured data including a risk level, a risk category, a statement content, a statement identification information, and a statement time for each statement unit, The risk report data generated by the risk notification unit includes a speech history, statistical information, and trend information according to the risk level of each speaker based on the accumulated risk assessment history data. The risk assessment device according to claim 1 .

3. The risk notification unit notifies the predetermined notification destination when at least one of the following conditions is met: when the risk level of the most recent comment content corresponds to high risk, when a specific keyword is included in the most recent comment content, and when the number of medium-risk comments by the same person exceeds a predetermined number. The risk assessment device according to claim 2 .

4. A risk assessment system that assesses the risk of comments included in a business log within a company, The system comprises a log information acquisition device, a log information analysis device, and a risk assessment device, which are connected to each other via a network, the log information acquisition device acquires log information of chats, voice conferences, etc. within the company, and transmits the log information to the log information analysis server; The log information analysis device generates minutes data in which the contents of comments are classified and structured based on the received log information, and transmits the generated minutes data to the risk assessment device; The risk assessment device analyzes the content of statements and their contexts based on the received minutes data using a large-scale language model constructed by learning based on the labels of high risk, medium risk, and low risk assigned to past log information, classifies the risk level of each statement into at least three levels of high risk, medium risk, and low risk to generate risk assessment data, stores the risk assessment data in a database as risk assessment history data organized by speaker, generates risk report data based on the risk assessment history data, and notifies the risk report data to a predetermined notification destination, the predetermined notification destinations being set in advance to distinguish notification targets based on notification conditions set for each risk level, the notification conditions including at least one of a condition for sending a notification each time the risk assessment history data is generated and a condition for sending a notification when the generation of risk assessment history data for the same speaker exceeds a predetermined number of times, and selects a corresponding notification destination from among a plurality of notification targets according to the risk level related to the most recent statement content added to the risk assessment history data, and transmits the risk report data to the notification destination. Risk assessment system.

Images