A conference agenda evaluation and real-time conference agenda dynamic correction system

Abstract

The application provides a conference agenda evaluation and real-time conference agenda dynamic correction system, and relates to the technical field of conference management, which comprises a pre-conference agenda evaluation system and a real-time dynamic correction system during the conference, wherein the pre-conference agenda evaluation system comprises an agenda multidimensional evaluator, which is used for evaluating a candidate agenda and outputting an optimal agenda based on a multidimensional evaluation model; the real-time dynamic correction system during the conference comprises a multi-modal perception layer, a state understanding layer, a decision engine layer and an execution and feedback layer, which are used for monitoring the conference state in real time, generating a correction strategy and executing feedback, and forming a closed-loop control; the multi-modal perception layer and the state understanding layer are used to quantitatively measure the conference state by means of data, so that the agenda optimization and correction decision is based on objective facts and is free from the dependence on the personal experience of the host; and the real-time monitoring mechanism is used to trigger an alarm when problems such as topic deviation and time overrun initially occur, and the decision engine automatically generates a correction strategy to seize the best correction opportunity.

Description

Technical Field

[0001] This invention relates to the field of meeting management technology, and in particular to a meeting agenda evaluation and real-time meeting agenda dynamic correction system. Background Technology

[0002] Currently, meeting management primarily relies on manually creating the agenda beforehand and the facilitator's experience-based guidance during the meeting. Existing technologies, online meeting software integrating agenda functionality, offer features such as static agenda display, manual timers, and auxiliary tools. This software allows users to upload their agendas before the meeting and display them to participants during the meeting, with the facilitator manually controlling the meeting's progress.

[0003] However, existing technologies have the following drawbacks: First, the system cannot proactively perceive the true state of the meeting, such as agenda deviation or progress, and lacks real-time data collection and analysis; second, it cannot quantify and understand the meeting status, such as participation enthusiasm or time deviation, leading to delayed intervention; third, due to the lack of perception and understanding, the system cannot intelligently generate corrective strategies; and finally, the entire management process relies on manual operation, failing to form a closed-loop control of "perception-understanding-decision-execution," resulting in low meeting efficiency and poor adaptability. Therefore, this invention proposes a meeting agenda evaluation and real-time dynamic meeting agenda correction system to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a meeting agenda evaluation and real-time dynamic meeting agenda correction system. This system uses a multimodal perception layer and a state understanding layer to quantify the meeting state with data, enabling agenda setting and correction decisions to be based on objective facts and eliminating reliance on the personal experience of the host.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a meeting agenda evaluation and real-time meeting agenda dynamic correction system, including a pre-meeting agenda evaluation system and a real-time dynamic correction system during the meeting, wherein the pre-meeting agenda evaluation system includes a multi-dimensional agenda evaluator, used to generate candidate agendas and output the optimal agenda based on a multi-dimensional evaluation model;

[0006] The real-time dynamic correction system during the meeting includes a multimodal perception layer, a state understanding layer, a decision engine layer, and an execution and feedback layer. It is used to monitor the meeting status in real time, generate correction strategies, and execute feedback to form a closed-loop control.

[0007] Further improvements are made in the following aspects: the evaluation dimensions of the agenda multi-dimensional evaluator include meeting efficiency, feasibility, goal achievement, and preference alignment; the meeting efficiency dimension is calculated using indicators such as time utilization rate, agenda completion rate, and time allocation rationality; the feasibility dimension is calculated using indicators such as time conflict detection and agenda dependency detection; the goal achievement dimension is calculated using indicators such as goal matching degree and key goal achievement rate; and the preference alignment dimension is calculated using indicators such as discussion session settings and voting resolution settings.

[0008] A further improvement is that the agenda multi-dimensional evaluator uses a weighted aggregation formula to calculate the total agenda score, as follows:

[0009] ,

[0010] in: Let the weight of the i-th dimension satisfy... ; The score for the i-th dimension is normalized to 0-1 or 0-100.

[0011] Further improvements include: the multimodal perception layer is used to collect voice, video, system logs, and manually input data in real time; voice data includes speech content, speaker ID, speech rate, and pauses; video data includes face detection, gaze direction, and facial expression analysis; system log data includes a list of participants, entry and exit records, raised hands, and chat messages.

[0012] Further improvements are made in the following ways: the state understanding layer is used to quantitatively analyze topic progress, topic deviation, participation intensity, and time deviation; topic progress is calculated using keyword coverage and semantic fragment similarity; topic deviation uses the Sentence-BERT model for dynamic topic consistency detection; participation intensity is calculated by weighting speaking frequency, average speaking duration, topic keyword frequency, and sentiment tendency; and time deviation is calculated by cumulative timeout value.

[0013] Further improvements include the following steps for topic deviation detection: pre-extracting a core keyword set for the topic; using Sentence-BERT encoding on the spoken text within the sliding window; calculating the maximum semantic similarity between the spoken text and the keyword set; and triggering a deviation alert when the similarity of multiple consecutive windows is below a threshold.

[0014] A further improvement is that the decision engine layer triggers a correction strategy based on the output of the state understanding layer, and the correction strategy includes extending the duration of the topic, skipping the topic, compressing the duration of subsequent topics, inserting new topics, reminding to return to the topic, inserting a break, and postponing topics.

[0015] Further improvements are made in the following ways: the strategy triggering logic of the decision engine layer is as follows: when the progress of the topic is consistently below the threshold, the extension or reminder focus strategy is triggered; when the topic deviates from the alarm, the reminder return to the topic strategy is triggered; when the participation enthusiasm is higher than the threshold and active, the insertion of new ideas or rest strategy is triggered; when the time deviation exceeds the threshold, the skip, postpone or compression strategy is triggered.

[0016] A further improvement is that the execution and feedback layer is used to push the correction strategy to the host and collect the host's confirmation or ignore feedback; the feedback results are recorded and used to re-perceive the meeting status to achieve closed-loop control.

[0017] Further improvements include the following corrective steps: generating candidate agendas, outputting the optimal agenda based on a multi-dimensional evaluation model; collecting meeting data in real time, quantitatively analyzing the meeting status, generating and executing corrective strategies, and collecting feedback to form a closed loop.

[0018] The beneficial effects of this invention are as follows:

[0019] 1. This invention uses a multimodal perception layer and a state understanding layer to quantify the state of a meeting, enabling agenda setting and corrective decisions to be based on objective facts and freeing them from dependence on the personal experience of the moderator.

[0020] 2. This invention uses a real-time monitoring mechanism to trigger alarms when issues such as topic deviation or timeouts first appear. The decision engine then proactively generates corrective strategies, transforming "post-event remediation" into "in-event intervention" and seizing the best opportunity for correction.

[0021] 3. This invention dynamically adjusts the agenda based on the real-time status of the meeting, transforming the agenda from a static document into a "dynamic plan" that adapts to changes on-site, thus avoiding the agenda becoming out of touch with the actual situation.

[0022] 4. The decision engine of this invention accurately matches correction strategies according to different problems, and combined with the human-machine collaborative execution mode, it avoids ineffective intervention and reduces ineffective discussions, significantly improving meeting efficiency and goal achievement rate. Attached Figure Description

[0023] Figure 1 This is a diagram of the pre-meeting agenda evaluation system architecture of the present invention;

[0024] Figure 2 This is a diagram of the real-time dynamic correction system architecture of the present invention. Detailed Implementation

[0025] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0026] Example 1

[0027] according to Figure 1 , 2 As shown, this embodiment proposes a meeting agenda evaluation and real-time meeting agenda dynamic correction system, including a pre-meeting agenda evaluation system and a real-time dynamic correction system during the meeting. The pre-meeting agenda evaluation system includes an agenda multi-dimensional evaluator, which is used to evaluate candidate agendas and output the optimal agenda based on the multi-dimensional evaluation model.

[0028] The real-time dynamic correction system during meetings comprises a multimodal perception layer, a state understanding layer, a decision engine layer, and an execution and feedback layer. It monitors meeting status in real time, generates correction strategies, and executes feedback to form a closed-loop control system. Through the collaborative design of pre-meeting assessment and in-meeting correction, it breaks down the traditional barriers between "static agendas" and "human facilitators," achieving intelligent control over the entire meeting management process. The closed-loop control mechanism ensures that meeting status feedback drives strategy optimization in real time, allowing agenda adjustments to closely follow meeting dynamics and significantly improving the accuracy of agenda adaptation to the actual scenario.

[0029] The multi-dimensional agenda evaluator evaluates four dimensions: meeting efficiency, feasibility, goal achievement, and preference alignment. Meeting efficiency is calculated using indicators such as time utilization, agenda completion rate, and time allocation rationality. Feasibility is calculated using indicators such as time conflict detection and inter-item dependency detection. Goal achievement is calculated using indicators such as goal matching degree and key goal achievement rate. Preference alignment is calculated using indicators such as discussion session settings and voting decision settings. This multi-dimensional and detailed indicator design comprehensively covers the core needs of meetings, ensuring both the efficiency and feasibility of the agenda while considering participant preferences and meeting goals, avoiding agenda imbalance caused by single-dimensional evaluation. All indicators are quantifiable and calculable objective parameters, overcoming the limitations of traditional agenda-setting relying on subjective experience, making the selection of the optimal agenda more scientific and practical.

[0030] The agenda multi-dimensional evaluator uses a weighted aggregation formula to calculate the total agenda score, as follows:

[0031] ,

[0032] in: Let the weight of the i-th dimension satisfy... ; The score for the i-th dimension is normalized to 0-1 or 0-100. The adjustable weight mechanism supports flexible adjustment of dimension priority according to meeting type, adapting to the meeting management needs of different scenarios; the score normalization process ensures that the evaluation results of each dimension are comparable horizontally, making the calculation logic of the total agenda score clearer and the selection criteria for the optimal agenda more accurate.

[0033] The multimodal perception layer is used to collect voice, video, system logs, and manually input data in real time. Voice data includes speech content, speaker ID, speech rate, and pauses; video data includes face detection, gaze direction, and facial expression analysis; system log data includes participant lists, entry and exit records, hand-raising records, and chat messages. Multimodal data collection covers the entire scenario of "voice + video + system + manual input," ensuring accurate and comprehensive meeting status data and providing comprehensive data support for subsequent quantitative analysis. The granular data collection items can accurately capture meeting details, avoiding misjudgments caused by a single data source and improving the accuracy of meeting status perception.

[0034] The state understanding layer is used to quantitatively analyze topic progress, topic deviation, participation intensity, and time deviation. Topic progress is calculated using keyword coverage and semantic fragment similarity. Topic deviation uses the Sentence-BERT model for dynamic topic consistency detection. Participation intensity is calculated by weighting speaking frequency, average speaking duration, topic keyword frequency, and sentiment tendency. Time deviation is calculated by cumulative timeout value. This transforms abstract concepts such as "topic progress" and "participation intensity" into calculable quantitative parameters, changing the meeting status from "subjective feeling" to "objective data," providing precise basis for intelligent decision-making. The application of the Sentence-BERT model overcomes the limitations of traditional keyword matching, accurately identifying semantic-level topic deviation and improving the robustness and accuracy of deviation detection.

[0035] The topic deviation detection includes the following steps: pre-extracting a core keyword set for the topic; using Sentence-BERT encoding on the spoken text within the sliding window; calculating the maximum semantic similarity between the spoken text and the keyword set; and triggering a deviation alert when the similarity of multiple consecutive windows is below a threshold. The sliding window + continuous multi-window verification logic effectively filters out interference from occasional speech deviations, reduces the false alarm rate, and avoids frequent interventions that could disrupt meeting flow. The semantic similarity-based detection method is adaptable to complex language scenarios such as synonyms and near-synonyms, and compared to traditional keyword matching, it can more accurately capture deviation trends from the core topic, making deviation correction more targeted.

[0036] The decision engine layer triggers corrective strategies based on the output of the state understanding layer. These strategies include extending the duration of each topic, skipping topics, compressing the duration of subsequent topics, inserting new topics, reminding participants to return to the main topic, inserting breaks, and postponing topics. These diverse corrective strategies cover common problems across all meeting scenarios (timeouts, going off-topic, insufficient discussion, unexpected topics), avoiding the limitations of single intervention methods. The strategy design balances "advancing meeting progress" and "ensuring discussion quality," such as extending popular topics and skipping consensus topics, achieving a balance between meeting efficiency and effectiveness.

[0037] The strategy triggering logic of the decision engine layer is as follows: when the progress of an issue consistently falls below a threshold, an extension or reminder-focused strategy is triggered; when an alert is triggered indicating a deviation from the topic, a reminder-return-to-topic strategy is triggered; when participation is above a threshold and active, a strategy of inserting new ideas or taking a break is triggered; when the time deviation exceeds a threshold, a skip, postpone, or compression strategy is triggered. This binding logic of "status threshold + corresponding strategy" automates and standardizes corrective decisions, reduces the manual judgment cost for the moderator, and improves meeting management efficiency. The triggering logic ensures both the timeliness of correction (e.g., reminders upon deviation) and avoids excessive intervention (e.g., flexibly inserting new ideas when participation is high), balancing the orderliness and flexibility of the meeting.

[0038] The execution and feedback layer is used to push the correction strategy to the host and collect the host's confirmation or ignore feedback. The feedback results are recorded and used to re-perceive the meeting status, achieving closed-loop control. The human-machine collaborative execution mode leverages the system's intelligent decision-making advantages while retaining the host's final decision-making authority, avoiding poor user experience caused by fully automated intervention. The feedback results are transmitted back to the perception layer to form a closed loop, allowing the system to dynamically adjust subsequent strategies based on the host's actions, improving the adaptive capability and practical application effect of the correction mechanism.

[0039] The corrective action process includes the following steps: generating candidate agendas, outputting the optimal agenda based on a multi-dimensional evaluation model; collecting meeting data in real time, quantitatively analyzing the meeting status, generating and executing corrective strategies, and collecting feedback to form a closed loop. This standardized corrective action process creates a standardized workflow for meeting management: "agenda optimization - status awareness - strategy execution - feedback optimization," reducing the learning cost for enterprise meeting organizations. Each step is interconnected and seamlessly integrated, ensuring efficient operation of the entire meeting management process, significantly reducing ineffective discussions and wasted time, and improving the meeting's goal achievement rate.

[0040] Example 2

[0041] according to Figure 1 , 2 As shown, this embodiment proposes a meeting agenda evaluation and real-time meeting agenda dynamic correction system, including a pre-meeting agenda evaluation system (based on an evaluator) and a real-time dynamic correction system during the meeting (based on a four-layer closed-loop model).

[0042] Pre-meeting agenda evaluation system:

[0043] Before the meeting begins, this component is used to evaluate and select the optimal meeting agenda. It mainly consists of an "Agenda Multidimensional Evaluator".

[0044] Agenda Multidimensional Estimator / Discriminator: Scores the N candidate agendas generated by the generator and selects the optimal solution.

[0045] Evaluation Dimensions and Indicators: Meeting Efficiency Dimension - Time Utilization (z1): Total time used for all agenda items / Total meeting time. Item Completion Rate (z2): Number of agenda items / Total planned agenda items. Time Allocation Rationality (z3): Matching degree between item duration and priority.

[0046] Feasibility Dimension - Time Conflict Detection (z4): Whether the agenda items overlap or exceed the total time limit. Agenda Dependency Detection (z5): Whether the agenda item dependencies are satisfied (e.g., agenda item B must follow agenda item A). Participant Time Compatibility: Comprehensively considering the available time of core participants. A group time compatibility algorithm is used to calculate the group compatibility index = Σ(individual compatibility × participant weight), ensuring that core personnel can participate in key agenda items.

[0047] Goal Achievement Dimension - Goal Matching Degree (z6): Semantic similarity between agenda content and meeting goals (e.g., calculated using BERT). Key Goal Achievement Rate (z7): Whether key topics were adequately addressed.

[0048] Preference Alignment Dimension - Discussion Session Setup (z8): Whether discussions have been arranged for the issues requiring them. Voting Decision Setup (z9): Whether voting has been arranged for the issues requiring them. Preference Alignment Score (AS): Using... The formula calculates the weighted cosine similarity between the agenda feature vector and the preference (e.g., efficiency, fairness, transparency) vector.

[0049] Overall assessment framework:

[0050] A weighted aggregation approach is used to calculate the total agenda score layer by layer, starting from the indicator, then moving to the dimension.

[0051] Final Agenda Total Score Formula:

[0052] ,

[0053] Where: Wi: the weight of the i-th dimension (satisfying...) Di: Score of the i-th dimension (normalized to) ).

[0054] Score calculation for each dimension:

[0055] Meeting efficiency dimension D1:

[0056] ;

[0057] Indicator Explanation:

[0058] z1 = Total time used for all agenda items / Total meeting time × 100% (normalized to [0,1], the larger the value, the more reasonable the time utilization). The calculation method is as follows:

[0059]

[0060] Tuse is the effective usage time of the agenda (pure agenda duration excluding transition time and rest time between agendas).

[0061] Ttotal is the total meeting time planned in the agenda (the upper limit of the overall meeting duration preset).

[0062] α is the penalty coefficient for low utilization, and its possible value range is 0.8-1. The recommended default value is 1.

[0063] β is the timeout tolerance coefficient, which can take values ​​between 1 and 2, but cannot exceed 2. The recommended default value is 1.2.

[0064] z2 = Number of agenda items / Total planned agenda items × 100% (normalized to [0,1], the larger the value, the more reasonable the time utilization, same as z1);

[0065] z3: The matching degree between topic duration and priority (which can be calculated using cosine similarity or a custom matching function).

[0066] Feasibility dimension D2:

[0067] ;

[0068] Indicator Explanation: z4: Time conflict detection (combined overlap rate, participant compatibility); z5: Number of issues with non-compliance (needs to be normalized).

[0069] Goal Achievement Dimension D3:

[0070] ,

[0071] Metrics Explanation: z6: Target matching degree (BERT semantic similarity, 0~1); z7 = Number of key objectives achieved / Total number of key objectives × 100%.

[0072] Preference alignment dimension D4:

[0073] ,

[0074] Indicator Explanation: z8 = Number of discussions or time in the agenda / Total number of discussions or time required for all agenda items × 100%; z9 = Number of votes in the agenda / Total number of votes required for all agenda items × 100%.

[0075] Final total score:

[0076] ,

[0077] Final output:

[0078]

[0079] Evaluator feedback mechanism:

[0080] The evaluator not only outputs a score, but also provides interpretable feedback statements, such as: If A score >0.7 indicates a reasonable agenda; if D3 < 0.6, the objective may be deviated from; if D2 < 0.5, there is a serious time conflict; if z8 < 0.5, the discussion session is insufficient. Specific evaluation criteria and corresponding thresholds will be dynamically adjusted based on the type of meeting and its needs.

[0081] This feedback will guide adjustments to the agenda generator: topic duration allocation, interaction mode settings, topic ordering, and participant time compatibility.

[0082] Real-time dynamic correction system during the meeting:

[0083] After the meeting begins, this component monitors the meeting status in real time and performs dynamic corrections. It consists of a closed loop of four modules: "Multimodal Perception Layer," "State Understanding Layer," "Decision Engine Layer," and "Execution and Feedback Layer."

[0084] Multimodal Perception Layer (PL) - Function: Real-time, comprehensive data acquisition during meetings. Data Sources: Voice Modality: Real-time ASR (Automatic Speech Recognition) captures speech content, speaker ID, speech rate, and pauses. Video Modality: Camera-based face detection (speaker identification), gaze direction (attention assessment), and facial expressions (emotion analysis). System Modality: Meeting software logs, such as participant list, entry / exit records, raised hands, and chat messages. Manual Modality: Active input from the host or participants, such as "skip this topic" or "insert urgent topic."

[0085] State Understanding Layer (SUL) - Function: Analyzes the raw data from the perception layer to quantify the meeting state.

[0086] Topic Progress Identification: Determining the current topic progress. Method: Based on keyword coverage and semantic fragment similarity. Steps: Preprocessing the sliding window (Wt) of the current speech and the topic reference description (Ti, including keywords Ki and core sentence Si). Calculating keyword channel scores. Calculate semantic channel scores. Instantaneous progress fusion .

[0087] Topic Deviation Detection: Determines whether the current discussion has deviated from the topic. Method: Dynamic Topic Consistency Detection (DTCT), using Sentence-BERT for zero-shot detection. Steps: Pre-extract a core keyword set Ki from the topic. Encode the spoken text Tt in a 30-second sliding window using Sentence-BERT. Calculate the maximum semantic similarity between Tt and all words in Ki. If N consecutive windows (e.g., 5) all satisfy... If so, a deviation alarm rt=1 is triggered.

[0088] Popularity and Engagement Assessment: Judging the "Discussion Atmosphere". Method: Weighted calculation based on multiple indicators. Indicators: Frequency of speaking, average speaking duration, frequency of topic keywords, and sentiment (determined as positive / negative using the BERT model). Calculation: Discussion Popularity / Engagement = .

[0089] Meeting time deviation calculation: Determining whether "time is sufficient". Method: Calculating cumulative deviation. Calculation: Deviation for individual agenda items. Overall meeting cumulative deviation Trigger: If D > timeout threshold (e.g., 10 minutes), trigger a time deviation warning.

[0090] Decision Engine Layer (DEL) - Function: Generates corrective strategies based on the output of the State Understanding Layer. Strategy Trigger: IF Issue Progress If the value remains very low, then it is recommended to extend(i, (Extend) or REMIND (focus) (reminds you to stay focused). IF the topic deviates from rt=1 THEN trigger REMIND (topic) (reminds you to return to the topic). IF discussion is extremely hot and active THEN suggest INSERT (j,priority) (insert a new idea) or BREAK (…). (Halftime). IF time deviation D > threshold THEN, suggest SKIP(i), DEFER(i), or COMPRESS(i, (Compression follows).

[0091] Execution and Feedback Layer (EFL) - Function: Executes strategies and forms a closed loop. Execution: The system pushes the strategies generated by the decision engine to the meeting moderator as "intelligent reminders" (e.g., "Prompt: The current discussion has strayed from the topic [Return to the topic]?" or "Prompt: The current topic's time is up, but the discussion is still heated; [Extend for 5 minutes]?"). Feedback: The moderator confirms by clicking (e.g., "Agree to extend" or "Ignore reminder"), and the operation result is recorded by the system. Closed Loop: The moderator's action (or inaction) leads to a change in the meeting state, and the new state is recaptured by the "Multimodal Perception Layer," entering the next "perception-understanding-decision" cycle.

[0092] Regarding topic deviation detection: Traditional keyword-weighted matching (DTCT) methods can also be used. This involves pre-setting topic keywords and their weights, and calculating keyword scores for the spoken text within a sliding window. In comparison, this approach is simple to implement, but its drawback is poor robustness; it cannot handle synonyms, near-synonyms, or out-of-vocabulary words. The S-BERT scheme, preferred in this invention, achieves better results through semantic understanding and sliding window detection, which is one of the key aspects of this invention.

[0093] Regarding the understanding of states: It's also possible to implement only some functions, such as only "time deviation calculation." In contrast, while this simplified approach can partially solve the timeout problem, it cannot address core quality issues such as "going off-topic" and "insufficient discussion." This invention provides more comprehensive management capabilities by integrating the understanding of four states (progress, deviation, popularity, and time).

[0094] Regarding strategy execution: It can also be designed as "fully automatic execution," for example, if the system detects a timeout, it automatically compresses subsequent agenda items without the host's consent. In contrast, this "strong intervention" approach may negatively impact user experience. The "human-machine collaboration" mode adopted in this invention (i.e., the system provides suggestions, and the host confirms) is a superior implementation method, balancing intelligence with the host's or operator's control.

[0095] This meeting agenda evaluation and real-time dynamic agenda correction system utilizes a multimodal perception layer and a state understanding layer to quantify meeting status with data, ensuring that agenda setting and correction decisions are based on objective facts and free from reliance on the moderator's personal experience. Furthermore, through a real-time monitoring mechanism, the system triggers alarms as soon as issues such as agenda deviation or timeouts appear, and the decision engine proactively generates correction strategies, transforming "post-event remediation" into "in-event intervention" and seizing the optimal opportunity for correction. Simultaneously, the system dynamically adjusts the agenda based on the real-time meeting status, transforming the agenda from a static document into a "dynamic solution" that adapts to changes in the meeting environment, preventing the agenda from becoming out of touch with reality. In addition, the system's decision engine accurately matches correction strategies to different problems, combined with a human-machine collaborative execution mode, avoiding ineffective interventions and reducing ineffective discussions, significantly improving meeting efficiency and goal achievement rates.

[0096] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A meeting agenda evaluation and real-time dynamic agenda correction system, comprising a pre-meeting agenda evaluation system and a real-time dynamic agenda correction system during the meeting, characterized in that: The pre-meeting agenda evaluation system includes a multi-dimensional agenda evaluator, which generates candidate agendas and outputs the optimal agenda based on the multi-dimensional evaluation model. The real-time dynamic correction system during the meeting includes a multimodal perception layer, a state understanding layer, a decision engine layer, and an execution and feedback layer. It is used to monitor the meeting status in real time, generate correction strategies, and execute feedback to form a closed-loop control.

2. The meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 1, characterized in that: The evaluation dimensions of the agenda multi-dimensional evaluator include meeting efficiency dimension, feasibility dimension, goal achievement dimension, and preference alignment dimension; The meeting efficiency dimension is calculated using indicators such as time utilization rate, agenda completion rate, and time allocation rationality; the feasibility dimension is calculated using indicators such as time conflict detection and agenda dependency detection; the goal achievement dimension is calculated using indicators such as goal matching degree and key goal achievement rate; and the preference alignment dimension is calculated using indicators such as discussion session settings and voting resolution settings.

3. The meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 2, characterized in that: The agenda multi-dimensional evaluator uses a weighted aggregation formula to calculate the total agenda score, as follows: ， in: Let the weight of the i-th dimension satisfy... ; The score for the i-th dimension is normalized to 0-1 or 0-100.

4. The meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 1, characterized in that: The multimodal perception layer is used to collect voice, video, system logs and manually input data in real time; voice data includes speech content, speaker ID, speech rate and pauses; video data includes face detection, gaze direction and expression analysis; system log data includes participant list, entry and exit records, raised hands and chat messages.

5. A meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 1, characterized in that: The state understanding layer is used to quantitatively analyze topic progress, topic deviation, participation intensity, and time deviation. Topic progress is calculated using keyword coverage and semantic fragment similarity. Topic deviation uses the Sentence-BERT model for dynamic topic consistency detection. Participation intensity is calculated by weighting speaking frequency, average speaking duration, topic keyword frequency, and sentiment tendency. Time deviation is calculated by cumulative timeout value.

6. The meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 5, characterized in that: The topic deviation detection includes the following steps: pre-extracting a core keyword set for the topic; using Sentence-BERT encoding on the spoken text within the sliding window; calculating the maximum semantic similarity between the spoken text and the keyword set; and triggering a deviation alarm when the similarity of multiple consecutive windows is lower than a threshold.

7. The meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 1, characterized in that: The decision engine layer triggers correction strategies based on the output of the state understanding layer. These correction strategies include extending the duration of topics, skipping topics, compressing the duration of subsequent topics, inserting new topics, reminding users to return to the main topic, inserting breaks, and postponing topics.

8. A meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 7, characterized in that: The strategy triggering logic of the decision engine layer is as follows: when the progress of the topic is consistently below the threshold, the strategy of extension or reminder focus is triggered; when the topic deviates from the alarm, the strategy of reminder return to the topic is triggered; when the participation enthusiasm is higher than the threshold and active, the strategy of inserting new ideas or rest is triggered; when the time deviation exceeds the threshold, the strategy of skipping, delaying or compressing is triggered.

9. A meeting agenda evaluation and real-time meeting agenda dynamic correction system according to claim 1, characterized in that: The execution and feedback layer is used to push the correction strategy to the host and collect the host's confirmation or ignore feedback; Feedback results are recorded and used to re-perceive the meeting status, achieving closed-loop control.

10. A meeting agenda evaluation and real-time meeting agenda dynamic correction system according to any one of claims 1-9, characterized in that: The corrective action includes the following steps: generating candidate agendas, outputting the optimal agenda based on a multi-dimensional evaluation model; collecting meeting data in real time, quantitatively analyzing the meeting status, generating and executing corrective strategies, and collecting feedback to form a closed loop.

Images