Method for realizing automatic clustering of conference summary by adopting topic model

Abstract

The invention relates to the technical field of conference information processing and management decision analysis, and discloses a method for realizing automatic clustering of conference summary by adopting a topic model, which comprises the following steps of: respectively acquiring an acoustic confidence sequence and an original topic entropy sequence by analyzing conference audio signals and corresponding transcripted texts in parallel; according to the method, the original topic entropy sequence is subjected to compulsory weighted correction by using the acoustic confidence coefficient, a decision anchor point is identified in the obtained reliable entropy sequence, and reverse association clustering is performed on the text by taking the decision anchor point as a core, so that the limitation that only content induction can be performed in the prior art is avoided; by recognizing the decision anchor points in the entropy sequence corrected by the acoustic confidence coefficient, automatic reconstruction of the decision forming process is achieved, the quality of decision consensus can be further discriminated, and an objective logic link is provided for management tracing and risk auditing.

Description

Technical Field

[0001] The present invention relates to a method for automatically clustering meeting minutes by adopting a topic model, and belongs to the technical field of meeting information processing and management decision analysis. Background Art

[0002] In modern administrative, commercial, or financial management practices, meetings are a key link in making decisions, assigning tasks, and synchronizing information. The voice or text records they produce are the core basis for subsequent management, supervision, and review. To improve information processing efficiency, the industry generally uses natural language processing technologies such as topic models to automatically analyze meeting record texts. By identifying and aggregating high-frequency words and semantic associations, they summarize the core topics discussed in the meeting. This method does provide certain convenience in quickly extracting the key points of the meeting content.

[0003] However, when we examine this widely adopted technical approach in the context of management scenarios such as corporate management and administrative supervision, which place strict demands on process compliance and decision traceability, an inherent constraint becomes apparent: the existing technical approach is rooted in a static document metaphor, and its core mechanism is to break the chronological order of the original text and reorganize information based solely on content similarity. As a result, the most valuable process value unique to meetings as a management activity—that is, the dynamic logic of how group consensus moves from disagreements to unified decisions and how they ultimately converge and form through numerous discussions—is completely eliminated by this reorganization process. This loss of process information is not simply a loss of details; it means that managers can only see isolated conclusions and issues during review, but cannot trace the complete argumentation path on which these conclusions are based. This poses a potential risk in serious management activities that require responsibility audits or process optimization.

[0004] To address this challenge, a seemingly straightforward improvement approach is to simply mark the timestamp information corresponding to each topic on the results of existing topic clustering. However, this approach can only reflect when a topic was discussed, but cannot reveal the intrinsic causal and evolutionary relationships between topics and between discussions and final decisions. The fundamental reason is that the aforementioned technologies aimed at content aggregation have a fundamental conflict between their internal mechanisms and the goal of restoring process logic, resulting in the following deficiencies in existing technologies on this issue: 1. Existing methods aim to summarize content, and their processing will actively destroy the original speech timing information that is crucial for reconstructing the decision logic, resulting in the inability to effectively restore the decision-making process. 2. Existing methods output discrete, flat topic labels, which lack the ability to structuredly present the core management process of how these topics influence each other progressively and ultimately converge to one or more binding resolutions. Therefore, how to avoid the limitation of existing technology that treats meeting minutes as static documents for content analysis, and use a new information processing method to automatically identify the formation nodes of all key decisions in the meeting process without destroying the original time series information, and based on this, reversely reconstruct the complete discussion process with internal logic corresponding to each decision, has become the technical problem to be solved by the present invention. Summary of the Invention

[0005] The present invention provides a method for automatically clustering meeting minutes using a topic model. Its main purpose is to solve the problem that the existing technology, due to its content aggregation mechanism, destroys the original speech sequence, making it impossible to effectively restore the decision-making process and unable to structuredly present the logical context of the discussion topics converging to the final resolution.

[0006] To achieve the above objectives, the present invention provides a method for automatically clustering meeting minutes using a topic model, comprising the following steps: Step a, obtaining a real-time audio signal of a conference and a sequence of text units synchronized with the real-time audio signal in time; Step b: Process the text unit sequence in sequence using a sliding window, and run the topic model on the text unit in each window to output the topic probability distribution vector, and then calculate the original topic entropy time series; Step c, analyzing the real-time audio signal corresponding to each window in parallel to calculate the signal-to-noise ratio and the effective speech duration ratio, and outputting an acoustic confidence time series based on the calculated signal-to-noise ratio and the effective speech duration ratio; Step d, weighting the original topic entropy time series point by point according to the acoustic confidence time series to generate a corrected topic entropy time series; Step e: in the corrected topic entropy time series, identify the moment that satisfies the local minimum within the neighborhood radius and whose forward first-order difference is continuously negative in the previous part of the neighborhood radius as the decision anchor point; Step f, for any identified decision anchor point, analyzing the corrected topic entropy time series within a subsequently set time window to calculate the entropy rebound factor and classify the decision anchor point as a stable anchor point or a risk anchor point; In step g, taking any identified decision anchor point as the end point, all text units in the single entropy monotonically decreasing interval before the decision anchor point are reversely merged to form an issue cluster.

[0007] Preferably, in step c, the real-time audio signal corresponding to each window is analyzed in parallel and is also used to calculate the speaker overlap rate; and step c also includes outputting an acoustic confidence time series based on the signal-to-noise ratio effective speech duration ratio and the speaker overlap rate through a set normalization function.

[0008] Preferably, in step f, the entropy rebound factor The calculation follows the following rules: ,in, is the corrected subject entropy time series at time The entropy value of is the moment corresponding to the decision anchor point, The minimum analysis time step is set.

[0009] Preferably, the method also includes: before the meeting begins, setting a decision weight coefficient for each participant based on the participant list and its organizational hierarchy; and, in step b, when outputting the topic probability distribution vector for the text unit in each window, weighting the influence of the text unit in the topic model operation based on the speaker identity of the text unit and its corresponding decision weight coefficient, so that the text unit corresponding to the speaker with a higher decision weight coefficient has a stronger influence on the formation of the topic probability distribution vector.

[0010] Preferably, the method also includes: before the meeting starts, running the same topic model as step b based on the meeting agenda text to generate at least one agenda baseline topic vector; during the meeting, using the topic probability distribution vector generated for each window in step b as the real-time topic vector, and calculating the semantic distance between the real-time topic vector and the agenda baseline topic vector in real time to generate a deviation sequence; when the deviation of more than a set number of windows in the deviation sequence is higher than the set deviation threshold, an agenda deviation warning is triggered.

[0011] Preferably, the method also includes: storing each decision anchor point generated by the analysis of previous meetings and the topic probability distribution vector of its corresponding window to form a historical decision anchor point database; for any newly generated decision anchor point in the current meeting, calculating the semantic similarity between the topic probability distribution vector of its corresponding window and the topic probability distribution vector of the historical decision anchor point stored in the historical decision anchor point database; and when the semantic similarity is higher than the set link threshold, establishing a link between the current decision anchor point and the corresponding historical decision anchor point to represent the topic evolution relationship, so as to construct a cross-meeting topic evolution graph.

[0012] Preferably, step e for identifying the decision anchor point is further defined as: the condition that the forward first-order difference is continuously negative in the previous part of this neighborhood radius refers to tracing back from the local minimum moment point until the moment point when the first forward first-order difference turns from non-negative to negative.

[0013] Preferably, in step f, the step of classifying the decision anchor point as a stable anchor point or a risk anchor point is specifically: comparing the entropy rebound factor with the set risk threshold; if the entropy rebound factor is higher than the risk threshold, classifying the decision anchor point as a risk anchor point; otherwise, classifying it as a stable anchor point.

[0014] Preferably, the size of the sliding window in step b is limited to be adaptively adjusted according to the total length of the text unit sequence or the total number of speakers.

[0015] Preferably, the method further includes: extracting meeting type information from the meeting metadata before the analysis begins; selecting a matching entropy analysis mode from a mode library containing decision-making modes and divergent modes based on the extracted meeting type information; wherein, when the decision-making mode is selected, executing steps e and f; when the divergent mode is selected, monitoring whether the entropy value of the corrected topic entropy time series is lower than the set creativity blocking threshold, and outputting the creativity blocking point when it is lower than the creativity blocking threshold.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention first converts the meeting text arranged in chronological order into a continuous text window sequence through a sliding window, and calculates the instant topic probability distribution for each independent window, and then maps the entire meeting process into a time series of topic entropy values; on this basis, the system does not cluster based on content similarity, but locates the key moments of consensus formation or decision-making by identifying local minimum points in the entropy sequence, and marks them as decision anchor points; finally, with these decision anchor points as the end point, all text units within the corresponding entire entropy reduction interval are reversely attributed. In this way, an originally unstructured text that only records scattered speeches is reconstructed into a structured decision map with the final decision as the core and clearly showing the logical context of all its previous discussions, so that the decision process tracing and responsibility auditing required in administrative management are transformed from manual interpretation of massive texts to direct review of an objectively determined logical link.

[0017] 2. The present invention further utilizes the entropy sequence information after the decision anchor point is formed. That is, for each identified decision anchor point, the system will continue to analyze the changing trend of the topic entropy value within a subsequent preset time window; when the changing trend, such as the entropy value growth slope in the initial stage, exceeds a preset risk threshold, the system will distinguish the decision anchor point as a risk anchor point; this mechanism enables the system to no longer just identify the fact that a decision has been made, but to further identify the consensus quality of the decision, distinguish between a stable consensus that is smoothly accepted by all parties and a fragile consensus that implies controversy and may lead to a rapid resurgence of discussion, providing a risk identification dimension for management and supervision activities.

[0018] 3. The method of the present invention also includes: before the meeting, generating a baseline topic vector based on the preset agenda text, and during the meeting, using the real-time topic probability distribution of each text window as the real-time topic vector, and monitoring the meeting progress by calculating the semantic deviation between the two; when the deviation continues to exceed the threshold, an early warning is triggered. This design is not an isolated monitoring function, but shares the same core calculation process with the aforementioned decision anchor point identification mechanism, that is, independent real-time topic modeling is performed on each window; this information reuse method enables the system to perform the core task of post-meeting decision link reconstruction while consuming lower additional resources, providing the meeting host or manager with an in-process automated process supervision capability, ensuring that the decision-making process itself is effectively carried out within the preset topic boundaries. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the automatic clustering method of meeting minutes based on acoustically corrected topic entropy of the present invention; Figure 2This is a state transition diagram of consensus formation and quality determination based on subject entropy evolution in the present invention; Figure 3 This is a diagram of the system modules and operation architecture of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0021] The present invention discloses a method for automatically clustering meeting minutes using a topic model. The method is mainly implemented by the following steps: first, obtaining a time-aligned conference audio stream and a sequence of transcribed text units, then processing the data in parallel, that is, on the one hand, converting the text sequence into an original topic entropy time series, and on the other hand, converting the audio stream into an acoustic confidence time series, then using the confidence sequence to weight the original entropy sequence to generate a corrected topic entropy time series, then locating the key moments representing the formation of consensus in the corrected sequence and evaluating its consensus quality, and finally, reconstructing the logical link from discussion to resolution with the identified decision anchor as the core.

[0022] In a multi-departmental joint meeting scenario aimed at approving the quarterly financial budget, during the process of converting the conference audio into text, noise or invalid text fragments may be introduced due to overlapping speeches of participants, far-field speeches with low signal-to-noise ratio, or long periods of silence. These distorted text inputs will cause deviations in subsequent semantic analysis, making it difficult to accurately identify the real decision-making point. To address this problem, the method of the present invention is configured to execute a quality assessment procedure for the original audio signal in parallel with the text analysis. Specifically, after obtaining the real-time audio signal of the meeting and a sequence of text units synchronized with the real-time audio signal in time in step a, the system analyzes the text units in parallel with each sliding window in step c. The corresponding real-time audio signal is used to calculate the signal-to-noise ratio Ratio of effective voice duration ;in, Speech intelligibility is quantified by calculating the ratio of the audio signal power to the background noise power. The continuity of the discussion is determined by counting the ratio of the number of valid speech frames in the window to the total number of frames. For example, if the total length of a window is 30 seconds and the valid speech duration after voice activity detection is 24 seconds, then To further handle the confusion of audio tracks caused by multiple people speaking at the same time, the system can also calculate the speaker overlap rate. The system then integrates these acoustic features into a single acoustic confidence score through a preset normalization function , the function can be set as a weighted model, such as , where all features are normalized to In this way, the system obtains an acoustic confidence time series that is synchronized with the text window and can dynamically reflect the quality of the original signal. .

[0023] Given that the entropy sequence calculated based only on the text stream is easily affected by signal quality, the system uses the aforementioned acoustic confidence time series The original subject entropy time series is weighted and corrected point by point; in step b, the system uses a set sliding window The text unit sequence is processed in sequence. The size of the sliding window can be limited to be adaptively adjusted according to the total length of the text unit sequence. For example, for a meeting expected to last 2 hours, the window size can be set to cover an average of 5 minutes of speech, so as to ensure the amount of text required for topic modeling while being sensitive to the switching of discussion focus. For each text unit in the window, the system runs the topic model to output a topic probability distribution vector ,in is the preset number of topics, and then according to the Shannon entropy formula Calculate the original topic entropy ; In step d, the system performs a weighted operation to generate the corrected topic entropy , for example, in moment, the discussion was intense and the voice was clear, , ,but ; and in At this moment, there is a long period of silence, and the speech recognition outputs meaningless characters, resulting in the calculated is 0.4, but at this time Due to extremely low It also drops to 0.1 accordingly, and the final result is In this way, the entropy deviation caused by poor signal quality is suppressed, thereby generating a corrected topic entropy time series with reduced errors caused by signal quality problems. .

[0024] In administrative management activities, distinguishing the consensus quality of different decisions is a technical link for prejudging execution risks. The degree of confusion in the discussion, i.e., the entropy value, which rises immediately after a brief drop, usually indicates a decision whose consensus basis is not yet stable. In order to convert this phenomenon into a calculable indicator, the method of the present invention performs a stability analysis of the entropy dynamics in its neighborhood after identifying the decision anchor point. Specifically, in step e, the time series of the subject entropy after correction is The decision anchor point is identified as the moment when the local minimum is taken within the set neighborhood radius and the forward first-order difference is continuously negative in the previous part of this neighborhood radius. Then, step f will analyze the entropy sequence in the time window set after the decision anchor point to calculate an entropy rebound factor , which is calculated according to the rules: ,in, is a set minimum analysis time step; The dimension of is the same as entropy, and its value represents the divergence rate of discussion topics after consensus is reached; then, the system will Value and a preset risk threshold For comparison, if , then the decision anchor point is classified as a risk anchor point, otherwise it is a stable anchor point; the risk threshold The determination can be made by offline analysis of historical meeting data and counting all the decision anchor points that were successfully executed. The distribution of values ​​and its 95th percentile is taken as a benchmark. In this way, the system can, in a data-driven way, mark for managers those decision-making matters that may need attention, consensus-based or controversial. After obtaining reliable decision anchor points, the system's goal is to reconstruct the argumentation process corresponding to each decision and form an issue cluster. To this end, in step g, the system takes any identified decision anchor point as the end point and performs a reverse merging operation. The procedure is to start from the decision anchor point. The system traces back to the moment when the first forward first-order difference changes from non-negative to negative, and this interval is defined as the single entropy monotonically decreasing interval of the decision anchor point; the system merges all text units contained in this interval into a topic cluster belonging to the decision anchor point; in this way, a meeting text recorded in time flow is reconstructed into multiple structured units with decisions as conclusions and discussions in the entropy reduction process as arguments, providing a logical link for subsequent management review and responsibility audit.

[0025] It should be noted that in management activities, different roles have different effects on the promotion of decision-making. In order to take into account the differences in the influence of speeches by different participants in the entropy change analysis, the method of the present invention also includes setting a decision weight coefficient for each participant based on the list of participants and their organizational hierarchy before the meeting starts. For example, the project director is 2.0, the department manager is 1.5, and the engineer is 1.0; in step b, when the topic probability distribution vector is calculated for the text unit in each window When the system will be based on the identity of the speaker and his corresponding , weighting the influence of the speaker's text unit in the topic model operation. One way to achieve this is to multiply the frequency of each word in the speaker's text by its weight coefficient when building the bag-of-words model of the window ; In this way, the text unit corresponding to the speaker with a higher decision weight coefficient will The formation of has a stronger impact, and when it makes a summary statement, it will reduce the entropy of the window more quickly. A deeper and more easily identifiable entropy valley is formed on the curve, thereby improving the accuracy of decision anchor point identification; In addition, in order to improve the efficiency of meeting management and prevent the deviation of the agenda, the method of the present invention also provides a process supervision mechanism; This mechanism reuses the intermediate data generated in step b for calculating entropy, namely the topic probability distribution vector Before the meeting begins, the system first runs the same topic model as step b based on the meeting agenda text to generate one or more agenda baseline topic vectors representing the core topics of the meeting; during the meeting, the system will generate the same topic model as step b for each window. As a real-time topic vector, the semantic distance between the real-time topic vector and the agenda benchmark topic vector is calculated in real time, for example, by calculating the inverse of the cosine similarity between the two as the deviation The system continuously monitors the deviation sequence. When the deviation exceeds the set deviation threshold for a set number of consecutive windows, an agenda deviation warning is triggered. This method adds an in-process supervision and guidance function to the system at a lower additional computing cost.

[0026] Furthermore, for topics spanning multiple meetings, the present invention also provides a method for constructing their evolutionary context; the core of this method is to map each decision anchor point and its corresponding topic probability distribution vector As a representation of the decision in the semantic space; the system will store each decision anchor point and its representation vector generated by the analysis of previous meetings, together with metadata such as meeting time, in a historical decision anchor point database; when a new meeting analysis is completed and a new decision anchor point is generated, the system will calculate the semantic similarity of its vector with all vectors in the historical library; when the similarity between a new decision anchor point and one or more decision anchor points in the historical library is higher than a preset link threshold, a link representing the evolution relationship of the topic is established between them; through continuous incremental links, the system will eventually build a cross-meeting topic evolution graph with decision anchor points as nodes and semantic links as edges, so that managers can trace back all related decisions on any topic and provide coherent decision information for the supervision and review of long-term projects. information; finally, in order to adapt to different types of management communication scenarios, the method of the present invention also has an adaptive switching capability of analysis modes based on the type of meeting; before the analysis begins, the system extracts meeting type information from the metadata of the meeting, such as a project decision meeting or a brainstorming meeting; based on the extracted type information, the system selects a matching entropy analysis mode from a pattern library containing decision modes and divergent modes; when the decision mode is selected, the system executes the aforementioned decision anchor point identification and grading process with the goal of finding the entropy valley; and when the divergent mode is selected, the system's analysis logic is adjusted accordingly, and its main task is to monitor whether the entropy value of the corrected topic entropy time series is lower than a preset creativity blocking threshold, and output a creativity blocking point when it is lower than the threshold to prompt the meeting organizer that the divergence of the current discussion is insufficient.

[0027] Example 1: In a multi-departmental joint meeting aimed at conducting a risk audit and re-routing for a major new drug R&D project that had hit a bottleneck, participants from R&D, clinical, finance, and marketing departments engaged in a lengthy, unstructured discussion on multiple dimensions, including the project's technical feasibility, budget overruns, clinical trial progress, and market window risks. The text unit sequence generated during this phase was processed through temporal entropy calculation to form a corrected topic entropy time series. The continuous fluctuation within a relatively high value range indicates a high degree of divergence and lack of consensus on the topic. During this process, the CFO accessed the system through a remote system, and his speech fluctuated due to the fluctuation of network transmission quality, resulting in a high signal-to-noise ratio of the corresponding audio clip. The system calculates the acoustic confidence score of the speaker accordingly. is a lower value, such as 0.4; at the same time, the decision weight coefficient of the chief operating officer, who is the highest decision maker of the project, is It is preset to 2.5. When it makes a summary speech on the reasons for the failure of key technical nodes, its clear voice makes Maintained above 0.9, and its weight coefficient By influencing the topic model operation, the speech content affects the topic probability distribution vector The formation of a stronger effect; the acoustic confidence correction is based on The procedures for the CFO's speech in the original theme entropy sequence The entropy generated in the process was suppressed, and the COO's speech was Value and high The combined effect of values ​​in the entropy sequence The synergy of these two mechanisms enables the changing trend of the entropy sequence to more accurately reflect the flow of effective information in the meeting.

[0028] As the meeting progressed, when the COO made a decisive suggestion to terminate the original technical route and switch to Alternative Plan B, the previously high corrected subject entropy time series There was a rapid one-way decline in a short period of time, and at the end of his speech, at time A local minimum is reached, and this point is marked as a decision anchor point by the decision anchor point identification module because it meets the condition that the forward first-order difference is continuously negative in the previous part of the neighborhood radius. This process does not rely on the semantic understanding of the decision content, but by mapping the meeting process into entropy dynamic evolution, the decision-making management behavior is converted into a signal feature that can be identified in the time series. Therefore, on the basis of retaining all the original time series information, the key management nodes are located, and the contradiction between comprehensive recording of process information and efficient identification of decision nodes is resolved.

[0029] At the decision anchor After being identified, reverse association clustering is triggered. Taking the anchor point as the end point, all text units in the single entropy monotonically decreasing interval before this moment, including the COO's decision statement and the preceding related technical feasibility analysis, are merged into a topic cluster. At the same time, the decision anchor point grading module begins to analyze Then the entropy rebound factor is calculated from the entropy sequence within a preset time window Since all parties involved have no objection to the decision and immediately turn to discuss the resource allocation of the new plan, the entropy sequence Showing a gentle climb, the income The value is lower than the preset risk threshold , the decision anchor point is therefore classified as a stable anchor point; finally, in the structured minutes output by the system, the stable anchor point and its corresponding topic cluster are clearly associated and stored together in the historical decision anchor point database, providing a logical link from decentralized discussion to final resolution with consensus quality assessment for possible future project review or audit; this method converts the speech flow text of the meeting into an entropy value time series that is doubly corrected by acoustic signal quality and speaker organization weight, and identifies decision anchor points based on morphological features in this sequence, and then reversely reconstructs the decision formation process based on this core, providing a technical path for objectively recording and structured presentation of the dynamic evolution process of group consensus in administrative management activities.

[0030] Example 2: In order to objectively verify the accuracy and robustness of the method of the present invention in identifying decision anchor points in complex acoustic environments and multi-level speech structures, this experiment is specially set up; the purpose of the experiment is to quantitatively compare the performance differences between the method of the present invention and an existing technical method based on content similarity in locating key decision points from meeting records, which is a core technical indicator; the experimental platform is constructed based on a corpus of standard business meeting recordings containing 100 hours of manually annotated data, which covers three typical administrative management scenarios: project review, budget approval, and strategic planning. Among them, all key moments in the formation of decisions have been cross-validated and annotated by two independent annotators. At the same time, the corpus contains audio clips with signal-to-noise ratios ranging from 5dB to 25dB to simulate a real meeting environment; Experimental Setting 1 The control group adopted an industry-standard technical approach, which is to first cluster the transcribed text of the entire meeting using the standard LDA topic model, and then associate each clustered topic with the timestamp set of its appearance in the text; the inventive sample group fully deployed the technical solution in the specific implementation method, including parallel acoustic confidence correction and decision weight coefficient weighting; the setting of the sliding window size is to balance the stability of topic modeling and the ability to capture rapid switching of discussion focus. If the window is too small, the amount of text in a single window is not enough to form a stable topic probability distribution. If the window is too large, a short but critical decision moment may be smoothed out in time. In view of this, its size is set to 5 times the average speaking time of the participants in the corpus, that is, 120 seconds; the risk threshold The setting is to use the entropy rebound factor after the anchor point marked as overturned decision in the corpus The values ​​are counted and the 80th percentile is taken as the benchmark.

[0031] During the experimental execution phase, the two sets of schemes were used to process all the meeting data in the corpus. The control group output a series of topic tags and their corresponding timestamps, while the sample group of the present invention output a structured text marked with decision anchors and topic clusters. In terms of performance statistics, the accuracy of decision anchors identified by the sample group of the present invention reached 94.2%, while the accuracy of the control group was 28.6%. In terms of the recall rate index of decision anchors, the sample group of the present invention was 91.5%, and the control group was 35.1%. In order to evaluate the robustness under harsh acoustic conditions, special tests were conducted on clips with a signal-to-noise ratio of 8dB in the corpus. The decision anchor accuracy of the sample group of the present invention was 90.3%, while the accuracy of the control group dropped to 12.5%. In addition, two annotators evaluated the logical consistency of the clustering results generated by the two sets of schemes. On a scale of 1 to 5, the logical consistency score of the topic clusters generated by the sample group of the present invention was 4.7 points, and the corresponding score of the control group was 1.8 points.

[0032] Data differential analysis shows that the low accuracy of the control group stems from its content-aggregation-based mechanism, which fails to distinguish between a discussion about a decision and the decision-making behavior itself. The high accuracy of the inventive group stems directly from its technical approach of associating decision-making behavior with the local minimum of the topic entropy sequence, a signal feature. Furthermore, in low signal-to-noise ratio environments, the performance degradation of the inventive group is much smaller than that of the control group. This difference verifies the role of acoustic confidence correction in suppressing signal noise interference and ensuring system robustness. The logical consistency score of the topic clusters further demonstrates that the internal logic of the decision chain formed through reverse association clustering is highly consistent with the results of manual annotation. The results of this experiment confirm that by introducing a topic entropy time series corrected by acoustic confidence and speaker weights and identifying decision anchors based on morphological features in this series, a logical structure centered on decision-making can be effectively and reliably reconstructed from unstructured meeting records. This technical fact confirms the application value of the inventive method in providing objective and traceable decision-making process information for administrative supervision and forecasting activities.

[0033] Example 3: This example combines Figures 1 to 3 , a method for automatically clustering meeting minutes using a topic model is described, such as Figure 1As shown in the figure, the process starts with obtaining the conference audio signal and the synchronously transcribed conference text unit sequence in parallel from the original voice data stream of the conference, wherein the conference text unit sequence is sent to the time series entropy value calculation. This module outputs the original topic entropy time series through sliding window and topic model processing based on the decision weight coefficient given by the organizational hierarchy or role of the participants to influence the speech weight, and the parallel acoustic confidence calculation module analyzes the conference audio signal, considers factors such as signal-to-noise ratio and effective voice duration, and outputs the acoustic confidence time series. This acoustic confidence sequence is then used to perform weighted correction on the original topic entropy sequence to suppress noise interference by using acoustic confidence to generate a corrected topic entropy time series. In this correction, In the positive sequence, the system identifies the local minimum of the entropy sequence through the decision anchor point identification module to locate the key moment when consensus is formed or decision is finalized. Then, on the one hand, by analyzing the entropy change trend after the decision anchor point, the entropy rebound factor is calculated and the anchor point is graded to distinguish between stable anchor points and risk anchor points. On the other hand, reverse association clustering is performed, taking the decision anchor point as the end point, and reversely merging the text in the preceding entropy decrease interval to form an issue cluster, which is finally presented as a structured decision map with decision-making as the core and clear logical links. At the same time, the topic probability distribution vector generated in the process of calculating the time series entropy value is also used for issue deviation warning. By calculating the deviation between the issue vector and the agenda benchmark in real time, a warning is triggered when the issue deviates.

[0034] like Figure 2 As shown in the figure, at the beginning of the meeting, the topic is usually divergent, with multiple viewpoints colliding, the discussion open but not yet focused, and the entropy value fluctuating at a high level. When key arguments are put forward or an authoritative person summarizes, the discussion process turns to the topic convergence state. At this time, consensus is gradually formed, the discussion tends to be concentrated, and the entropy value decreases monotonically until consensus is reached or the discussion ends. The entropy value reaches a local minimum, and the system recognizes and marks this moment as the decision anchor point. After that, the system adjusts the decision according to the entropy rebound factor. The quality of consensus is determined by comparing the result with the preset risk threshold. If the entropy rebound factor is less than the risk threshold, the system enters a stable consensus state, which is manifested as a gentle recovery of entropy and a smooth transition to a new topic. This anchor point is determined to be a stable anchor point. Conversely, if the entropy rebound factor is greater than the risk threshold, the system enters a fragile consensus state, which is manifested as a rapid rebound of entropy, indicating that there is a dispute or the topic may resurface. This anchor point is determined to be a risk anchor point and may return to the topic divergence state due to the resurgence of the dispute. Similarly, the stable consensus state can also return to the topic divergence state by opening a new topic to start a new discussion cycle.

[0035] like Figure 3As shown in the figure, the architecture is centered on the application server cluster, which receives real-time audio and video data streams provided by the audio and video acquisition and upload client from the conference room terminal. Within the application server cluster, the data access and synchronization module is responsible for receiving the data, and the real-time voice transcription service converts it into text. The parallel analysis engine then schedules the acoustic analysis module and the time series entropy analysis module for processing, and the decision anchor analysis service and the real-time topic deviation warning service perform the core analysis tasks. All analysis results and warning information are eventually sent to the visual analysis front-end of the manager or analyst terminal for browser presentation. At the same time, the application server cluster interacts with the back-end storage server through data reading and writing operations. The storage server contains a historical decision anchor database for tracing across conference topics, as well as a model and parameter configuration file library for adaptive switching of different conference modes.

[0036] Example 4: The various operating parameters of the method of the present invention can be determined through a systematic offline calibration procedure; the procedure first selects and constructs a calibration-specific corpus containing hundreds of meeting recordings and their transcriptions from the historical meeting archives of a specific application scenario. Each meeting in this corpus is annotated with metadata, including the meeting type, agenda, position level of the participants, and key decision anchors and agenda deviation events that are manually verified as the baseline truth; the risk thresholds relied on for risk anchor classification The procedure retrieves all project decision-making meetings in the corpus and runs the method of the present invention to calculate the entropy rebound factor corresponding to all benchmark truth decision anchor points. value, and based on the success and failure records of subsequent project execution, these The values ​​are divided into two sample sets: stable decision and risk decision. By performing receiver operating characteristic analysis on the data distribution of these two sample sets, a The value was determined to be 0.15 and was used as the risk monitoring criterion for this type of meeting.

[0037] Regarding the two parameters of the deviation threshold and the number of consecutive windows in the agenda deviation warning function, the program selected meetings with clear agenda texts in the corpus and systematically calculated the semantic distance between the real-time agenda vectors and the agenda baseline topic vectors of all text windows manually marked as on-track and deviated. By traversing different combinations of thresholds and number of windows, a parameter combination was determined that can achieve timely warning of 95% of deviation events while maintaining a false alarm rate below 5%. The deviation threshold was set to 0.65 and the number of consecutive windows was set to 3. For the link threshold used to construct the cross-conference agenda evolution graph, the program calculated the semantic similarity between the internal decision anchors of a series of meetings in the corpus that belong to the same project but are held in batches, as well as the semantic similarity between the decision anchors of different projects, and selected a threshold of 0.8 that can maximize the distinction between these two types of similarity distributions as the basis for establishing the judgment basis for establishing the agenda evolution relationship.

[0038] For the creativity blocking threshold used to judge creativity blocking in the divergent mode, the protocol focuses on the brainstorming sessions in the corpus, and the participants score the discussion fluency of the entire session afterwards. The protocol defines the period when the score is lower than the preset value as the blocking period, and calculates the time series of the corrected topic entropy in all non-blocking periods. The entropy distribution is taken as the 10th percentile of 0.9 as a dynamic lower limit of creativity. When the real-time entropy value in the meeting is lower than this threshold, the creativity blocking point prompt is triggered. By executing the above series of calibration processes, a set of matching parameter configuration files that have been verified by data is generated for each typical meeting type in a specific application scenario. This set of parameter calibration procedures transforms the multiple key thresholds that originally relied on operator settings into a closed, deterministic engineering process that can be executed before system deployment. In actual operation, before processing any new meeting, the system first identifies the meeting type by reading its meeting metadata and automatically loads the corresponding parameter configuration file. This mechanism enables the system's analysis conclusions to be based on statistically verified quantitative benchmarks that are suitable for specific management scenarios.

[0039] When this set of parameter configuration files was applied to the financial institution's historical meeting archives for retrospective analysis, the system identified a decision anchor point regarding the approval of the Alpha Project's seed round investment in the minutes of a first-quarter investment committee meeting and archived its corresponding topic probability distribution vector. When subsequently processing the minutes of a third-quarter strategic review meeting, the system identified another decision anchor point regarding the termination of subsequent additional investment in the Alpha Project. The system then calculated the semantic similarity between the topic probability distribution vectors of these two decision anchor points, belonging to different meetings, and obtained a score of 0.87, which was higher than the link threshold of 0.8 determined in the institution's parameter configuration file. Therefore, the system automatically established a link between the two decision anchor points representing the evolutionary relationship of the topic. When managers reviewed the termination decision in the third quarter in the system, the original investment decision in the first quarter and its complete argumentation process would be presented as its direct and traceable decision predecessor.

[0040] Example 5: In an inter-departmental emergency response meeting convened to address a sudden supply chain disruption, the initial agenda was set to discuss the screening of backup suppliers. However, during the meeting, an urgent analysis of force majeure clauses in existing contracts became a new, higher-priority topic. In such conditions where the agenda is dynamically adjusted, to avoid continuous agenda deviation warnings caused by topic switching, the method of the present invention provides an instant reset procedure for the agenda baseline. This procedure allows an authorized meeting host to trigger the system to discard the agenda baseline topic vector generated based on the original meeting agenda text through specific instructions, and to immediately re-run the topic model using the meeting text within a preset time window after the host announces the new topic as input to generate a new agenda baseline topic vector that can reflect the current focus of the discussion, and use this as the benchmark for subsequent topic deviation calculations.

[0041] In the subsequent stage of the emergency response meeting, a legal advisor who is in the middle level of the organizational hierarchy but is designated as the main person in charge of legal affairs in handling this incident, has a temporary influence on the decision-making process that is higher than his job grade. In order to enable the allocation of decision-making weight coefficients to reflect this change in authority based on task roles, the method of the present invention also provides a dynamic adjustment procedure for decision-making weight coefficients. This procedure allows the meeting host to assign a preset main person in charge role template to a designated participant through instructions. After receiving the instruction, the system queries a preset role weight database, extracts the decision weight coefficient corresponding to the role template, and uses the coefficient to replace the legal advisor's original decision weight coefficient based on his organizational hierarchy during the remaining time of the meeting, so that his subsequent speeches on contract risk analysis and response strategies can obtain higher influence weights in the calculation of topic entropy.

[0042] Example 6: In order to enable the core algorithm of the method of the present invention to operate in an optimized state adapted to the data characteristics of the specific scenario when deployed in a specific business scenario, the system is configured to execute a set of offline and online combined model configuration and parameter optimization procedures with maximizing analysis reliability as the objective function before formal application; this procedure aims to determine the key parameters in the acoustic analysis model topic model and decision point identification algorithm in a data-driven manner, thereby configuring the system into a verified engineering system.

[0043] The online part of the procedure performs a pre-verification for each meeting, using a built-in minimum admission confidence threshold as a constraint. During the first 60-second window after the meeting begins, the system calculates a continuous acoustic confidence score. And get an average confidence value. If the value is lower than the admission threshold, it means that the current input signal quality cannot support subsequent reliable analysis, and the system will stop processing and issue an early warning. The admission threshold itself is set during the offline stage of the procedure by analyzing the conference clips in the historical corpus that have been manually marked as having unacceptable audio quality, and taking The 95th percentile of the value is determined.

[0044] During the offline construction phase of the protocol, it is used to calculate the acoustic confidence score The weight coefficients of each acoustic feature in the normalization function are determined by an optimization process with the goal of maximizing the model prediction accuracy; in this process, the manually annotated speech intelligibility scores in the historical corpus are used as the dependent variable, and the normalized signal-to-noise ratio is used as the , effective voice duration ratio Speaker overlap rate A multiple regression analysis is performed with ∑ i = 1 i + ∑ i = 1 i + ∑ i = 2 ...

[0045] After passing the initial signal quality check, the online part of the procedure will adjust the core hyperparameter of the lightweight topic model, namely the number of topics , performs an instant optimization for the current meeting to maximize the theme consistency of this analysis; the system uses the transcript of the first five minutes of the meeting to iteratively run the theme model, where The value of is within a preset range. The system calculates a standardized topic consistency score for each topic generated in each iteration and finally selects the topic that can maximize the score. value, which serves as the final number of topics for processing subsequent texts of this meeting; at the same time, the neighborhood radius parameter used in the decision anchor point identification step is dynamically set to be proportional to the average speaking turn duration counted at the beginning of this meeting, so that the judgment scale of the local minimum can adapt to the discussion rhythm of different meetings; the implementation of this set of procedures enables the system to complete the verification of input data quality, the optimal configuration of the acoustic analysis model, and the dynamic adaptation of the topic model and decision point identification algorithm before entering the core entropy sequence analysis. The entire process is guided by quantitative indicators and ultimately generates a set of optimal operating parameter combinations that match the current specific working conditions.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatically clustering meeting minutes using a topic model, characterized in that: The following steps are involved: Step a, obtaining a real-time audio signal of a conference and a sequence of text units synchronized with the real-time audio signal in time; Step b: Process the text unit sequence in sequence using a sliding window, and run the topic model on the text unit in each window to output the topic probability distribution vector, and then calculate the original topic entropy time series; Step c, analyzing the real-time audio signal corresponding to each window in parallel to calculate the signal-to-noise ratio and the effective speech duration ratio, and outputting an acoustic confidence time series based on the calculated signal-to-noise ratio and the effective speech duration ratio; Step d, weighting the original topic entropy time series point by point according to the acoustic confidence time series to generate a corrected topic entropy time series; Step e: in the corrected topic entropy time series, identify the moment that satisfies the local minimum within the neighborhood radius and whose forward first-order difference is continuously negative in the previous part of the neighborhood radius as the decision anchor point; Step f, for any identified decision anchor point, analyzing the corrected topic entropy time series within a subsequently set time window to calculate the entropy rebound factor and classify the decision anchor point as a stable anchor point or a risk anchor point; In step g, taking any identified decision anchor point as the end point, all text units in the single entropy monotonically decreasing interval before the decision anchor point are reversely merged to form an issue cluster.

2. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: In step c, the real-time audio signal corresponding to each window is analyzed in parallel and used to calculate the speaker overlap rate. In addition, step c also includes outputting an acoustic confidence time series based on the signal-to-noise ratio effective speech duration ratio and the speaker overlap rate through a set normalization function.

3. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: In step f, the entropy rebound factor The calculation follows the following rules: ,in, is the corrected subject entropy time series at time The entropy value of is the moment corresponding to the decision anchor point, The minimum analysis time step is set.

4. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: The method also includes: before the meeting begins, setting a decision weight coefficient for each participant based on the participant list and its organizational hierarchy; and in step b, when outputting the topic probability distribution vector for the text unit in each window, weighting the influence of the text unit in the topic model operation based on the speaker identity of the text unit and its corresponding decision weight coefficient.

5. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: The method also includes: before the meeting begins, running the same topic model as step b based on the meeting agenda text to generate at least one agenda baseline topic vector; during the meeting, using the topic probability distribution vector generated for each window in step b as a real-time topic vector, and calculating the semantic distance between the real-time topic vector and the agenda baseline topic vector in real time to generate a deviation sequence; when the deviation of more than a set number of windows in the deviation sequence is higher than a set deviation threshold, triggering an agenda deviation warning.

6. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: The method also includes: storing each decision anchor point generated by the analysis of previous meetings and the topic probability distribution vector of its corresponding window to form a historical decision anchor point database; for any new decision anchor point generated in the current meeting, calculating the semantic similarity between the topic probability distribution vector of its corresponding window and the topic probability distribution vector of the historical decision anchor points stored in the historical decision anchor point database; and when the semantic similarity is higher than a set link threshold, establishing a link between the current decision anchor point and the corresponding historical decision anchor point to represent the topic evolution relationship, so as to construct a cross-meeting topic evolution graph.

7. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: The identification of the decision anchor point in step e is further defined as follows: the condition that the forward first-order difference is continuously negative in the previous part of this neighborhood radius refers to tracing back from the local minimum moment until the moment when the first forward first-order difference turns from non-negative to negative.

8. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: In step f, the step of classifying the decision anchor point as a stable anchor point or a risk anchor point is specifically as follows: comparing the entropy rebound factor with the set risk threshold; if the entropy rebound factor is higher than the risk threshold, the decision anchor point is classified as a risk anchor point; otherwise, it is classified as a stable anchor point.

9. The method for automatically clustering meeting minutes using a topic model according to claim 1, characterized in that: The method also includes: before the analysis begins, extracting meeting type information from the meeting metadata; based on the extracted meeting type information, selecting a matching entropy analysis mode from a mode library containing a decision mode and a divergence mode; wherein, when the decision mode is selected, executing steps e and f; when the divergence mode is selected, monitoring whether the entropy value of the corrected topic entropy time series is lower than a set creativity blocking threshold, and outputting a creativity blocking point when it is lower than the creativity blocking threshold.

Images