A method for automatically clustering meeting minutes by using a topic model

Abstract

The application relates to the technical field of conference information processing and management decision analysis, and discloses a method for realizing automatic clustering of conference minutes by adopting a topic model, which comprises the following steps: acquiring an acoustic confidence sequence and an original topic entropy sequence by respectively analyzing a conference audio signal and corresponding transcribed text in parallel; then, forcibly weighting and correcting the original topic entropy sequence by using the acoustic confidence, identifying a decision anchor point in the obtained reliable entropy sequence, and performing reverse correlation clustering on the text with the decision anchor point as the core; the application avoids the limitation that the prior art can only perform content induction; by identifying the decision anchor point in the entropy sequence corrected by the acoustic confidence, the automatic reconstruction of a decision formation process is realized, the quality of a decision consensus can be further identified, and objective logical links are provided for management tracing and risk auditing.

Description

TECHNICAL FIELD

[0001] The application relates to a method for automatically clustering meeting minutes by using a topic model, and belongs to the technical field of meeting information processing and management decision analysis. BACKGROUND

[0002] In the current modern administrative, commercial or financial management practice, a meeting is a key link for forming decision distribution tasks and synchronizing information, and the voice or text record generated thereby is a core basis for subsequent management, supervision and review. At present, in order to improve information processing efficiency, the industry generally uses a natural language processing technology such as a topic model to automatically analyze meeting record texts, and by identifying and aggregating high-frequency words and semantic associations, the core issues discussed in the meeting are summarized. This approach indeed provides certain convenience in quickly extracting the key points of the meeting content.

[0003] However, when this commonly used technical approach is placed in the management scene of enterprise management and administrative supervision, which has strict requirements on process compliance and decision traceability, a constraint inherent in the approach is highlighted: the existing technical approach is rooted in a static document metaphor, and its core mechanism is to break the time sequence of the original text and reorganize the information based on content similarity as the only criterion. As a result, the most valuable process value of a meeting as a management activity, i.e., how group consensus is formed from differences to unified decision through dynamic logic in numerous discussions, is completely eliminated by this reorganization process. This loss of process information is not simply a lack of details, but it makes managers only see isolated conclusions and issues when reviewing, but cannot trace the complete argumentation path on which these conclusions are based. This constitutes a potential risk in serious management activities that require responsibility audits or process optimization.

[0004] To cope with this challenge, a seemingly direct improvement idea is to simply mark the time stamp information corresponding to each topic on the result of the existing topic clustering, but this way can only reflect when a topic is discussed, and cannot reveal the internal cause and evolution relationship between topics and between discussions and final decisions; The fundamental reason is that the internal mechanism and the logical goal of the reduction process of the foregoing technology for the purpose of content aggregation are in principle conflict, resulting in the following several aspects of the existing technology on this problem: 1. The existing method aims to summarize the content, and the processing process will actively destroy the original speech time sequence information which is crucial for reconstructing the decision logic, resulting in the formation process of the decision cannot be effectively restored; 2. The existing method outputs discrete and flat topic labels, which lack the ability to structure the presentation of how these topics interact with each other, progress and eventually converge to one or more binding resolutions, which is the core management process. Therefore, how to avoid the limitation of the existing technology that regards the meeting record as a static document for content analysis, and use a new information processing method to automatically identify all key decision formation nodes in the meeting process without destroying the original time sequence, and take this as the core to reversely reconstruct the complete discussion process with internal logic corresponding to each decision, becomes the technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a method for automatically clustering meeting minutes by using a topic model, which aims to solve the problem that the existing technology destroys the original speech time sequence due to its content aggregation mechanism, thereby failing to effectively restore the decision formation process and structurally present the logical context of the convergence of discussion topics to the final resolution.

[0006] To achieve the above purpose, the present application provides a method for automatically clustering meeting minutes by using a topic model, comprising the following steps:

[0007] Step a, obtaining the real-time audio signal of the meeting and the text unit sequence synchronized with the real-time audio signal in time;

[0008] Step b, processing the text unit sequence in turn with a sliding window, and running a topic model for the text units in each window to output a topic probability distribution vector, and then calculating an original topic entropy time sequence;

[0009] Step c, analyzing the real-time audio signal corresponding to each window in parallel to calculate the signal-to-noise ratio and the proportion of effective speech duration, and outputting an acoustic confidence time sequence based on the calculated signal-to-noise ratio and the proportion of effective speech duration;

[0010] Step d, weighting the original topic entropy time sequence point by point according to the acoustic confidence time sequence to generate a corrected topic entropy time sequence;

[0011] Step e: In the corrected topic entropy time series, identify the moment that satisfies the condition of taking a local minimum within the neighborhood radius and having its forward first difference continuously negative in the previous part of this neighborhood radius as the decision anchor point.

[0012] Step f: For any identified decision anchor point, analyze the corrected topic entropy time series within the 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.

[0013] Step g involves taking any identified decision anchor point as the endpoint and merging all text units within the monotonically decreasing single entropy interval before that decision anchor point to form an issue cluster.

[0014] Preferably, in step c, the real-time audio signal corresponding to each window is analyzed in parallel and used to calculate the speaker overlap rate; furthermore, step c also includes, based on the effective speech duration ratio of the signal-to-noise ratio and the speaker overlap rate, outputting the acoustic confidence time series through a set normalization function.

[0015] Preferably, in step f, the entropy rebound factor The calculation follows these rules: ,in, To correct the topic entropy time series at time 1 The entropy value, The moment corresponding to the decision anchor point. This is a set minimum analysis time step.

[0016] Preferably, the method further 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 each text unit in the window, the influence of the text unit in the topic model calculation is weighted according to the speaker's identity 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.

[0017] Preferably, the method further includes: before the meeting begins, running the same topic model as in 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 consecutive windows in the deviation sequence is higher than a set deviation threshold, triggering a topic deviation warning.

[0018] Preferably, the method further comprises: storing each decision anchor and its corresponding window's theme probability distribution vector generated by the analysis of the previous meeting to form a historical decision anchor database; calculating the semantic similarity between the theme probability distribution vector of the window corresponding to any newly generated decision anchor of the current meeting and the theme probability distribution vector of the historical decision anchor stored in the historical decision anchor database; and when the semantic similarity is higher than a set link threshold, establishing a link between the current decision anchor and the corresponding historical decision anchor to represent the evolution relationship of the topic, so as to construct a cross-meeting topic evolution graph.

[0019] Preferably, the step e for identifying the decision anchor is further limited to: the condition that the forward first-order difference is negative in the previous part of the neighborhood radius is that, from the time point of the local minimum, the first time point of the forward first-order difference is traced back until the first time point of the forward first-order difference is converted from non-negative to negative.

[0020] Preferably, the step f of classifying the decision anchor as a stable anchor or a risk anchor is specifically: comparing the entropy rebound factor with a set risk threshold, and if the entropy rebound factor is higher than the risk threshold, the decision anchor is classified as a risk anchor, otherwise it is classified as a stable anchor.

[0021] 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.

[0022] Preferably, the method further comprises: extracting meeting type information from the meeting metadata before the analysis starts; selecting a matched entropy analysis mode from a mode library containing a decision mode and a divergence mode according to the extracted meeting type information; wherein when the decision mode is selected, the step e and the step f are executed; and when the divergence mode is selected, whether the entropy value of the corrected theme entropy time sequence is lower than a set creativity block threshold is monitored, and a creativity block point is output when the creativity block threshold is lower.

[0023] Compared with the prior art, the present application has the following advantages:

[0024] 1. Firstly, the conference text arranged in time sequence is converted into a continuous text window sequence by means of a sliding window, and the instant topic probability distribution of each independent window is calculated, and then the whole conference process is mapped into a time sequence of theme entropy values; on this basis, the system does not cluster according to content similarity, but locates the key moment of consensus formation or decision making through identifying the local minimum points in the entropy sequence, and marks them as decision anchor points; finally, with the decision anchor points as the terminal point, all text units in the corresponding whole entropy descending interval are attributed in reverse, so that an originally unstructured text recording only scattered speeches is reconstructed into a structured decision map with the final decision as the core and clearly showing all the logical context of the preceding discussion, so that the necessary decision process in administrative management is traced back and the responsibility audit is changed from manual interpretation of massive texts to direct examination of an objectively determined logical link.

[0025] 2. The application further utilizes the entropy sequence information formed after the decision anchor points, that is, for each identified decision anchor point, the system will continue to analyze the change trend of the theme entropy value in the subsequent preset time window; when the change trend, such as the initial stage of the entropy value growth slope, exceeds a preset risk threshold, the system will distinguish the decision anchor point as a risk anchor point; this mechanism enables the system to not only identify the fact that the decision has occurred, but also further identify the consensus quality of the decision, and distinguish between stable consensus accepted by all parties and fragile consensus involving disputes and possibly leading to rapid resurgence of discussions, providing a risk identification dimension for management and supervision activities.

[0026] 3. The method of the application further comprises: before the conference, generating a baseline theme vector according to a preset agenda text, and during the conference, taking the real-time topic probability distribution of each text window as a real-time topic vector, and monitoring the conference process by calculating the semantic deviation between the two; when the deviation continuously exceeds the threshold, a 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 instant topic modeling is performed on each window; this information reuse mode enables the system to perform the core task of reconstructing the decision link while providing a real-time automated process supervision capability for the conference host or manager with low additional resource consumption, ensuring that the decision process itself is effectively conducted within the preset topic boundary. BRIEF DESCRIPTION OF DRAWINGS

[0027] Fig. 1 the flow chart of the conference minutes automatic clustering method based on acoustic calibration theme entropy of the application;

[0028] Fig. 2This is a state transition diagram for consensus formation and quality determination based on topic entropy evolution, as presented in this invention.

[0029] Fig. 3 This is a diagram of the system modules and operating architecture of the present invention. Detailed Implementation

[0030] To make the objectives, 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] This invention discloses a method for automatic clustering of meeting minutes using a topic model. The method is mainly implemented through the following steps: First, obtain time-aligned meeting audio streams and transcribed text unit sequences. Then, process the data in parallel, that is, convert the text sequence into an original topic entropy time series and the audio stream into an acoustic confidence time series. Next, use the confidence series to weight the original entropy sequence to generate a corrected topic entropy time series. Then, locate the key moment points representing the formation of consensus in the corrected sequence and evaluate the quality of the consensus. Finally, reconstruct the logical link from discussion to resolution based on the identified decision anchor points.

[0032] In a multi-departmental joint meeting scenario aimed at approving quarterly financial budgets, the audio during the conversion to text can be contaminated with noise or invalid text fragments due to overlapping of participants' speeches, low signal-to-noise ratio speeches in the far field, or prolonged silences. This distorted text input leads to biases in subsequent semantic analysis, making it difficult to accurately identify the true decision-making points. To address this issue, the method of this invention is configured to perform a quality assessment procedure for the original audio signal in parallel while performing text analysis. Specifically, after acquiring the real-time audio signal of the meeting and the sequence of text units synchronized with the real-time audio signal in time in step a, the system in step c analyzes the audio signal in parallel with each sliding window. The corresponding real-time audio signal is used to calculate the signal-to-noise ratio. Percentage of effective voice recording time ;in, Speech intelligibility is quantified by calculating the ratio of audio signal power to background noise power. The continuity of the discussion is determined by the ratio of valid audio frames within a statistical window to the total number of frames. For example, if a window has a total duration of 30 seconds, and the valid audio duration after audio activity detection is 24 seconds, then... 0.8; for further processing the audio track disorder caused by multiple speakers talking 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 , which can be set as a weighted model, for example , where all features are normalized to interval, so that the system obtains an acoustic confidence time series synchronized with the text window that can dynamically reflect the quality of the original signal .

[0033] Since the entropy sequence calculated based on the text stream is susceptible to signal quality, the system uses the aforementioned acoustic confidence time series to point-by-point weighted correction of the original topic entropy time series; in step b, the system processes the text unit sequence in a set sliding window , the size of the sliding window can be limited to adaptive adjustment according to the total length of the text unit sequence, for example, for a conference expected to last 2 hours, the window size can be set to cover an average of 5 minutes of speech to ensure the amount of text required for topic modeling while having the sensitivity to capture discussion focus switching; for each text unit in the window, the system runs the topic model to output a topic probability distribution vector , where is the preset number of topics, and then calculates the original topic entropy according to the Shannon entropy formula ; in step d, the system performs weighted operation to generate corrected topic entropy , for example, at time , the discussion is intense and the voice is clear, , , then ; while at time , there is a long period of silence and the speech recognition output is meaningless characters, resulting in a calculated of 0.4, but at this time is extremely low , and accordingly , in this way, the entropy value deviation caused by low signal quality is suppressed, thereby generating a corrected topic entropy time series that reduces the error caused by signal quality problems .

[0034] In the administrative activities, the consensus quality of different decisions is a technical link for the execution risk prediction; the confusion degree of the discussion, that is, the entropy value, immediately rises after a short decline, which usually indicates a decision with an unstable consensus foundation; to transform this phenomenon into a calculable index, the method of the present application performs stability analysis on the entropy dynamics in the neighborhood of the decision anchor point after identifying the decision anchor point; specifically, in step e, the time point satisfying the local minimum in the set neighborhood radius and the forward first-order difference being continuously negative in the previous part of the neighborhood radius is identified as the decision anchor point in the corrected theme entropy time series After that, step f analyzes the entropy sequence in the set time window after the decision anchor point to calculate an entropy rebound factor , which follows the rules: , wherein, is a set minimum analysis time step; has the same dimension as entropy, and its value represents the divergence rate of the discussion theme after the consensus is reached; then, the system compares the value of with a preset risk threshold ; if , the decision anchor point is classified as a risk anchor point, otherwise, it is a stable anchor point; the determination of the risk threshold can be achieved by offline analysis of historical meeting data, statistics of the distribution of the values of after all decision anchor points with smooth final execution, and taking the 95th percentile as a benchmark; in this way, the system can mark the decision matters that may need attention in a data-driven manner, whose consensus foundation is controversial or controversial; after obtaining reliable decision anchor points, the goal of the system is to reconstruct the argumentation process corresponding to each decision to form a topic cluster; for this purpose, the system performs reverse merging in step g with any identified decision anchor point as the terminal point; the procedure is as follows: backtracking from the time point of the decision anchor point to the time point when the first forward first-order difference changes from non-negative to negative, which is defined as the single entropy monotonicity 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 time flow recorded meeting text is reconstructed into multiple structured units with decisions as conclusions and discussions in the entropy reduction process as arguments, providing logical links for subsequent management review and responsibility audit.

[0035] It should be noted that in the management activities, the speeches of different roles have different driving effects on the formation of decisions; to take into account the influence difference of different participants in the entropy change analysis, the method of the present application further comprises setting a decision weight coefficient for each participant according to the participant list and its organizational hierarchy before the meeting​ For example, the project director is 2.0, the department manager is 1.5, and the engineer is 1.0; in step b, when calculating the topic probability distribution vector for each text unit within the window. At that time, the system will determine the speaker's identity and corresponding... One way to weight the influence of the speaker's text units in the topic model computation is to multiply the word frequency of each word in the speaker's text by its weight coefficient when constructing the bag-of-words model of the window. Thus, the text units corresponding to speakers with higher decision-making weight coefficients will have a greater impact on... The formation of this has a stronger impact; when it makes a concluding statement, it reduces the window's entropy more quickly. A deeper and more easily identifiable entropy valley is formed on the curve, thereby improving the accuracy of decision anchor point identification. Furthermore, to improve meeting management efficiency and prevent topic deviation, the method of this invention also provides a process monitoring 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 in 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 use the topics generated in step b for each window. This serves as a real-time topic vector, and the semantic distance between this real-time topic vector and the agenda baseline 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, a topic deviation warning is triggered. This method adds a real-time monitoring and guidance function to the system with low additional computational cost.

[0036] Furthermore, for topics spanning multiple meetings, this invention also provides a method for constructing their evolutionary trajectory; the core of this method is to associate each decision anchor point with its corresponding topic probability distribution vector. As a representation of the decision in the semantic space; the system will store each decision anchor and its representation vector generated by the analysis of the previous meeting, together with the metadata such as the meeting time, in a historical decision anchor database; when a new meeting analysis is completed and a new decision anchor is generated, the system will calculate the semantic similarity between its vector and all the vectors in the historical database; when the similarity between a new decision anchor and one or more decision anchors in the historical database is higher than a preset linking threshold, a link representing the evolution relationship of the topic is established between them; through continuous incremental linking, the system eventually constructs a cross-meeting topic evolution graph with decision anchors as nodes and semantic links as edges, enabling managers to trace the relevant decisions of any topic over time and providing coherent decision information for the supervision and review of long-cycle projects; Finally, to adapt to different types of management communication scenarios, the method of the present application also has the ability to adaptively switch between analysis modes based on the type of the meeting; before the analysis begins, the system extracts the type information of the meeting from the metadata of the meeting, such as a project decision meeting or a brainstorming meeting; according to the extracted type information, the system selects the matching entropy analysis mode from the mode library containing the decision mode and the divergence mode; when the decision mode is selected, the system executes the aforementioned decision anchor identification and grading process targeting the entropy valley; when the divergence mode is selected, the analysis logic of the system 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 block threshold, and output the creativity block point when it is lower than the threshold, to prompt the meeting organizer that the current discussion is insufficient in divergence.

[0037] Embodiment 1: In a multi-department joint meeting aimed at conducting a risk audit and path redefinition for a major new drug R&D project that has fallen into a bottleneck, participants from R&D, clinical, financial and marketing departments have long, unstructured discussions around multiple dimensions of the project, such as technical feasibility, budget overruns, clinical trial progress and market window risks. The text unit sequence generated in this stage, after time series entropy value calculation and processing, forms a corrected topic entropy time series Continuously fluctuating within a higher value range, representing the high divergence and lack of consensus of the topic; in the process, the chief financial officer accesses the system remotely, and due to the fluctuation of network transmission quality, the signal-to-noise ratio of the audio segment corresponding to his speech is low, and the system accordingly calculates the acoustic confidence score of his speech period to be a lower value, for example 0.4; at the same time, the chief operating officer, as the highest decision maker of the project, has a decision weight coefficient preset to 2.5, and when he makes a summary speech on the causes of the failure of the key technology node, his clear voice maintains above 0.9, and his weight coefficient is maintained above 0.9, and his weight coefficient By imposing influence on the topic model operation, the utterance content has a stronger effect on the formation of the topic probability distribution vector ; the acoustic confidence correction, according to the procedure , suppresses the entropy value of the CFO's utterance in the original topic entropy sequence , while the COO's utterance, due to the joint action of the high value and the high value, produces a more significant change in the entropy sequence . The synergy of the two mechanisms enables the change trend of the entropy sequence to more accurately reflect the flow of effective information in the meeting.

[0038] As the meeting progresses, when the COO presents a decisive opinion on suspending the original technical route and switching to alternative solution B, the previously consistently high corrected topic entropy time sequence experiences a rapid one-way decline in a short period of time, and at the end of his utterance, at time , it reaches a local minimum. This point is marked as a decision anchor point by the decision anchor point recognition module because it satisfies the condition that the forward first-order difference is negative in the previous part of the neighborhood radius. This process does not rely on semantic understanding of the decision content, but rather maps the meeting process into entropy dynamics evolution, transforming the decision formation, a management behavior, into a signal feature identifiable in the time sequence, thereby positioning the key management nodes while preserving all original time sequence information, resolving the contradiction between comprehensive recording of process information and efficient recognition of decision nodes.

[0039] After the decision anchor point is identified, reverse correlation clustering is triggered, taking the anchor point as the endpoint, and merging all text units in the single entropy monotonically decreasing interval before this time point, including the COO's decision statement and the preceding related technical feasibility analysis, to form a topic cluster. At the same time, the decision anchor point grading module begins to analyze the entropy sequence within a preset time window after the anchor point, calculating the entropy rebound factor . Because there is no objection to the decision from the parties involved and they immediately move on to discuss resource allocation for the new solution, the entropy sequence presents a gentle climb, and the obtained value is lower than the preset risk threshold , the decision anchor is thus classified as a stable anchor; finally, in the structured minutes output by the system, the stable anchor is explicitly associated with its corresponding issue cluster and stored in the historical decision anchor database, providing a logical link from the converging of scattered discussions to the final resolution with a consensus quality evaluation attached; the method converts the speech stream text of the meeting into an entropy time series corrected by both acoustic signal quality and speaker organization weight, and identifies decision anchors on this sequence according to morphological features, and then reversely reconstructs the decision formation process with the decision anchors as the core, providing a technical path for objectively recording and structuring the dynamic evolution process of group consensus in administrative activities.

[0040] Example 2: To objectively verify the accuracy and robustness of the method of the present application in identifying decision anchors under complex acoustic environment and multi-level speech structure, the present test is specially set up; the purpose of the test is to quantitatively compare the performance difference between the method of the present application and an existing technology method based on content similarity in locating key decision points from meeting records in this core technical indicator; the test platform is based on a 100-hour standard business meeting corpus with artificial annotation, which covers three typical administrative management scenarios of project review, budget approval and strategic planning, wherein all key points of decision formation have been annotated by two independent annotators with cross-validation, and the corpus contains audio segments with signal-to-noise ratios from 5dB to 25dB to simulate real meeting environment; the test sets up a control group and an application group of the present application; the control group adopts a commonly used technical path in the industry, i.e. using a standard LDA topic model to perform topic clustering on the transcribed text of the entire meeting, and then associating each clustered topic with its timestamp set in the text; the application group of the present application fully deploys the technical solutions in the specific embodiments, including parallel acoustic confidence correction and decision weight coefficient weighting; the size of the sliding window is set to balance the stability of topic modeling and the ability to capture rapid switching of discussion focus, if the window is too small, the text amount in a single window is not enough to form a stable topic probability distribution, if the window is too large, a short but key decision moment may be smoothed out in time, therefore, its size is set to 5 times the average speaking duration of the participants in the corpus, i.e. 120 seconds; the risk threshold is set by statistically analyzing the entropy rebound factor of the anchors of the decisions that have been overturned in the corpus after the anchors, and taking the 80th percentile as the benchmark.

[0041] In the test execution phase, two groups of schemes are used to process all the meeting data in the corpus, and the control group outputs a series of topic labels and corresponding time stamps, while the inventive group outputs a structured text marked with decision anchor points and issue clusters; in terms of performance statistics, the accuracy of the decision anchor points recognized by the inventive group reaches 94.2%, while the accuracy of the control group is 28.6%; in terms of the recall rate of the decision anchor points, the inventive group is 91.5%, and the control group is 35.1%; in order to evaluate the robustness under poor acoustic conditions, special tests are conducted on the segments with a signal-to-noise ratio of 8dB in the corpus, and the accuracy of the decision anchor points of the inventive group is 90.3%, while the accuracy of the control group is reduced to 12.5%; in addition, two annotators evaluate the logical consistency of the clustering results generated by the two groups of schemes, and the logical consistency score of the issue clusters generated by the inventive group is 4.7, and the corresponding score of the control group is 1.8.

[0042] The difference analysis of the data shows that the low accuracy of the control group is due to the fact that its mechanism centered on content aggregation cannot distinguish between a discussion about a decision and the decision behavior itself; the high accuracy of the inventive group is directly derived from the technical path of associating the decision behavior with the signal feature of the local minimum of the topic entropy sequence; at the same time, under the condition of low signal-to-noise ratio, the performance degradation of the inventive group is much smaller than that of the control group, which verifies the role of acoustic confidence correction in suppressing signal noise interference and ensuring system robustness; the score of the logical consistency of the issue clusters further shows that the decision link formed by reverse association clustering has a high degree of consistency with the results of artificial annotation; the results of this test confirm that by introducing a topic entropy time sequence corrected by acoustic confidence and speaker weight, and identifying the decision anchor points based on morphological features on this sequence, the logical structure centered on decisions can be effectively and reliably reconstructed from unstructured meeting records; this technical fact confirms the application value of the method in providing objective and traceable decision-making process information for administrative supervision and prediction activities.

[0043] Embodiment 3: This embodiment combines Figs. 1 to 3 a method for realizing automatic clustering of meeting minutes by using a topic model, as Fig. 1As shown, the process starts from the parallel acquisition of conference audio signals and the synchronized transcribed conference text unit sequence from the original speech data stream of the conference, where the conference text unit sequence is fed into the timing entropy value calculation, which outputs the original topic entropy time sequence through the sliding window and topic model processing according to the decision weight coefficient given by the participant's organizational level or role to affect the speaking weight, while the parallel acoustic confidence calculation module analyzes the conference audio signals to consider factors such as signal-to-noise ratio and effective speech duration, and outputs the acoustic confidence time sequence, which is then used to correct the original topic entropy sequence by weighting to suppress noise interference using acoustic confidence, generating the corrected topic entropy time sequence. On this corrected sequence, the system identifies the local minimum of the entropy sequence through the decision anchor point identification module to locate the key moment of consensus formation or decision knock, and 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 anchors and risk anchors, and on the other hand, reverse association clustering is performed to merge the text in the entropy decreasing interval before the decision anchor point, forming a topic cluster, and finally presenting a structured decision map with clear logical links centered on the decision. At the same time, the topic probability distribution vector generated in the timing entropy value calculation process is also used for topic deviation warning, which triggers an early warning when the topic deviates by calculating the deviation of the topic vector from the agenda benchmark in real time.

[0044] As shown, Fig. 2 At the beginning of the meeting, it is usually in the state of topic divergence, showing multiple viewpoints collision, open discussion and not yet focused, and the entropy value presents high fluctuation. When the key argument is put forward or the authoritative person makes a summary, the discussion process enters the topic convergence state, at which point the consensus is gradually formed and the discussion tends to be focused, and the entropy value presents a monotonic decrease until the consensus is reached or the discussion is over, and the entropy value reaches a local minimum, at which point the system identifies and marks this moment as a decision anchor point. After that, the system determines the quality of the consensus according to the comparison result of the entropy rebound factor and the preset risk threshold, if the entropy rebound factor is less than the risk threshold, the system enters the consensus stable state, showing a gentle rise in entropy and a smooth transition to a new topic, and the anchor point is determined as a stable anchor point, otherwise, if the entropy rebound factor is greater than the risk threshold, the system enters the consensus fragile state, showing a rapid rebound of entropy, indicating that there is a dispute or the topic may rekindle, and the anchor point is determined as a risk anchor point and may return to the topic divergence state due to the dispute. Similarly, the consensus stable state can also return to the topic divergence state by starting a new topic to start a new discussion cycle.

[0045] As shown, Fig. 3As shown, the architecture takes the application server cluster as the core, which receives audio and video collection and upload from the conference terminal and real-time audio-text data stream provided by the client. Inside the application server cluster, the data access and synchronization module is responsible for receiving data, the speech real-time transcription service converts it into text, then the parallel analysis engine schedules the acoustic analysis module and the timing entropy analysis module for processing, and the decision anchor analysis service and the real-time topic deviation early warning service perform the core analysis tasks. All analysis results and early warning information are finally sent to the visual analysis front end of the manager or analyst terminal for browser rendering. At the same time, the application server cluster interacts with the storage server in the back end through data read and write operations. The storage server contains a historical decision anchor database for cross-conference topic traceability, and a model and parameter configuration file library for adaptive switching of different conference modes.

[0046] Embodiment 4: The various operating parameters of the method of the present application can be determined by a systematic offline calibration procedure. The procedure first filters and constructs a calibration-specific corpus containing hundreds of conference recordings and their transcribed texts from the historical conference archives of a specific application scenario. Each conference in the corpus is accompanied by metadata annotations, including conference type, agenda, attendee job level, and key decision anchors and topic deviation events manually verified as the baseline truth. For the determination of the risk threshold on which the risk anchor classification depends, the procedure retrieves all project decision meetings in the corpus, runs the method of the present application to calculate the entropy rebound factor values corresponding to all baseline truth decision anchors, and divides these values into two sample sets of stable decisions and risk decisions according to the success or failure records of subsequent project execution. Through receiver operating characteristic analysis of the data distribution of the two sample sets, a value that can distinguish risk decisions with 90% accuracy is determined as 0.15 and used as the risk monitoring standard for this type of conference.

[0047] For the two parameters of the deviating threshold and the number of continuous windows in the topic deviation early warning function, the procedure selects the meetings with clear agenda texts in the corpus, systematically calculates the semantic distance between the real-time topic vectors and the agenda benchmark theme vectors of all text windows manually labeled as on track and deviated, and determines a parameter combination that can achieve timely warning of 95% deviation events while keeping the false positive rate below 5% by traversing different combinations of thresholds and window numbers, where the deviation threshold is set to 0.65 and the number of continuous windows is set to 3. For the linking threshold used to construct the cross-meeting topic evolution graph, the procedure calculates the semantic similarity between the decision anchors within a series of meetings belonging to the same project but held in different sessions, as well as the semantic similarity between decision anchors of different projects, and selects a threshold of 0.8 that can maximize the distinction between the two types of similarity distributions as the basis for establishing the topic evolution relationship.

[0048] For the creativity blockage threshold used to judge creativity blockage in the divergent mode, the procedure focuses on brainstorming meetings in the corpus, and the participants score the fluency of the discussion throughout the meeting after the meeting. The procedure defines the period with a score below the preset value as the blockage period, and calculates the entropy value distribution of the corrected theme entropy time series in all non-blockage periods The 10th percentile of the distribution is 0.9, which is taken as a dynamic creativity lower limit. When the real-time entropy value in the meeting is lower than this threshold, a creativity blockage point prompt is triggered. Through the above series of calibration processes, a set of data-verified parameter configuration files matched with each typical meeting type in a specific application scenario is finally generated. This parameter calibration procedure converts the multiple key thresholds previously set by the operator into a closed and deterministic engineering process that can be executed before system deployment. In actual operation, the system first identifies the meeting type by reading its meeting metadata before processing any new meeting, 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 adapted to specific management contexts.

[0049] When the parameter profile is applied to the historical meeting archives of the financial institution for retrospective analysis, the system identifies a decision anchor point about approving the seed round investment of the Alpha project in the minutes of a first-quarter investment committee meeting and archives the corresponding topic probability distribution vector; when processing the minutes of a third-quarter strategic review meeting later, the system identifies a decision anchor point about suspending the follow-up investment of the Alpha project; the system then calculates the semantic similarity between the topic probability distribution vectors of the two decision anchor points belonging to different meetings, and the resulting score is 0.87, which is higher than the link threshold of 0.8 determined in the parameter profile of the institution, so the system automatically establishes a link between the two decision anchor points to represent the evolution of the topic, and when the manager reviews the suspension decision in the third quarter in the system, the original investment decision in the first quarter and its complete argumentation process will be presented as the direct and traceable decision precedent.

[0050] In a cross-departmental emergency response meeting held to respond to a sudden supply chain disruption, the initial agenda is set to discuss the screening of backup suppliers, but during the meeting, an emergency analysis of the force majeure clause in the existing contract becomes a new higher priority topic; in such a dynamic agenda adjustment scenario, to avoid continuous topic deviation alerts caused by topic switching, the method of the present invention provides an instant reset procedure for the agenda benchmark, which allows authorized meeting hosts to trigger the system to discard the agenda benchmark topic vector generated based on the original meeting agenda text and instantly regenerate a new agenda benchmark topic vector that reflects the current discussion focus by inputting the meeting text within a preset time window after the host announces the new topic, and using it as the benchmark for subsequent topic deviation calculations.

[0051] In the subsequent stage of the emergency response meeting, a legal consultant who is in the middle of the organizational hierarchy but is designated as the primary responsible person for legal affairs in this event handling, his speech has temporary influence higher than his job level on decision-making; to make the allocation of decision weight coefficients reflect such changes in authority based on task roles, the method of the present invention also provides a dynamic adjustment procedure for decision weight coefficients, which allows meeting hosts to assign a preset primary responsible person role template to designated attendees through instructions, and the system queries a preset role weight database to extract the decision weight coefficient corresponding to the role template after receiving the instruction, and uses this coefficient to replace the original decision weight coefficient of the legal consultant based on his organizational hierarchy for the remaining time of the meeting, so that his subsequent speech on contract risk analysis and response strategies can obtain a higher influence weight in the calculation of topic entropy.

[0052] Embodiment 6: In order to make the core algorithm of the method work in an optimized state that is suitable for the data characteristics of a specific scenario when the method is deployed in a specific business scenario, the system is configured to perform a set of off-line and on-line combined model configuration and parameter optimization procedures whose objective function is to maximize the analysis reliability before formal application; the procedures aim to determine the acoustic analysis model topic model and the key parameters in the decision point recognition algorithm in a data-driven manner, so as to configure the system as a verified engineering system.

[0053] The online part of the procedure performs a pre-check for each meeting with the built-in minimum admission confidence threshold as a constraint; within the first 60-second time window after the start of the meeting, the system calculates the continuous acoustic confidence score and obtains an average confidence value; if the value is lower than the admission threshold, it indicates that the current input signal quality cannot support the subsequent reliable analysis, and the system will abort the processing and issue a warning; the admission threshold itself is determined in the off-line stage of the procedure by analyzing the meeting segments that are manually annotated as unacceptable audio quality in the historical corpus, and taking the 95th percentile of the values.

[0054] In the off-line construction stage of the procedure, the weight coefficients of each acoustic feature in the normalization function used to calculate the acoustic confidence score are determined through an optimization process whose objective is to maximize the model prediction accuracy; in the process, the manually annotated speech intelligibility scores in the historical corpus are used as the dependent variable, and the normalized signal-to-noise ratio , the effective speech duration proportion and the speaker overlap rate are used as the independent variables to perform multiple regression analysis; the standardized regression coefficients obtained that can make the correlation between the model output and the manual scores reach the highest are fixed as the weights in the function.

[0055] After passing the initial signal quality check, the online part of the procedure will perform an instant optimization of the core hyperparameters of the lightweight topic model, i.e. the number of topics , for the current meeting to maximize the topic consistency of this analysis; the system uses the transcribed text of the first five minutes of the meeting to run the topic model in an iterative manner, where the value of is within a preset interval, the system calculates a standardized topic consistency score for the topics generated by each iteration, and finally selects the ​The value is the final number of topics in the subsequent text after processing the current meeting, and the neighborhood radius parameter used in the decision anchor identification step is dynamically set to be proportional to the average speech round length counted at the beginning of the current meeting, so that the judgment scale of local minimum can be adapted to the discussion pace of different meetings. The execution of this set of procedures makes the system complete the verification of the quality of the input data, the optimization configuration of the acoustic analysis model, and the dynamic adaptation of the topic model and the decision point identification algorithm before entering the core entropy sequence analysis. The whole process is guided by quantitative indicators, and finally generates a set of optimal operating parameter combinations that match the current specific working conditions.

[0056] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0057] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for automatic clustering of meeting minutes using a topic model, characterized in that, Includes the following steps: Step a: Acquire the real-time audio signal of the meeting and the sequence of text units that are time-synchronized with the real-time audio signal; Step b involves processing the text unit sequence sequentially using a sliding window, and running a topic model for each text unit in the window to output a topic probability distribution vector, thereby calculating the original topic entropy time series. Step c: Analyze the real-time audio signal corresponding to each window in parallel to calculate the signal-to-noise ratio and the proportion of effective speech duration, and output the acoustic confidence time series based on the calculated signal-to-noise ratio and the proportion of effective speech duration. Step d: The original topic entropy time series is weighted point by point according to the acoustic confidence time series to generate the corrected topic entropy time series; Step e: In the corrected topic entropy time series, identify the moment that satisfies the condition of taking a local minimum within the neighborhood radius and having its forward first difference continuously negative in the previous part of this neighborhood radius as the decision anchor point. Step f: For any identified decision anchor point, analyze the corrected topic entropy time series within the 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. Step g: Taking any identified decision anchor point as the endpoint, reversely merge all text units within the monotonically decreasing range of single entropy before that decision anchor point to form an issue cluster; In step c, the real-time audio signal corresponding to each window is analyzed in parallel and used to calculate the speaker overlap rate; furthermore, step c also includes outputting the acoustic confidence time series based on the effective speech duration ratio and speaker overlap rate through a set normalization function. In step f, the entropy rebound factor The calculation follows these rules: ,in, To correct the topic entropy time series at time 1 The entropy value, The moment corresponding to the decision anchor point. This is a set minimum analysis time step.

2. The method for automatic clustering of 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 list of participants and their organizational hierarchy; and in step b, when outputting the topic probability distribution vector for each text unit in the window, the influence of the text unit in the topic model calculation is weighted based on the speaker's identity and the corresponding decision weight coefficient of the text unit.

3. The method for automatic clustering of meeting minutes using a topic model according to claim 1, characterized in that, The method further includes: before the meeting starts, running the same topic model as in 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 consecutive windows in the deviation sequence is higher than the set deviation threshold, a topic deviation warning is triggered.

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

5. A method for automatic clustering of meeting minutes using a topic model according to claim 1, characterized in that, Step e for identifying the decision anchor point is further defined as follows: the condition that the forward first-order difference is continuously negative in the previous part of the neighborhood radius means tracing back from the local minimum point until the point where the first forward first-order difference changes from non-negative to negative.

6. The method for automatic clustering of 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 as follows: the entropy rebound factor is compared 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.

7. A method for automatic clustering of meeting minutes using a topic model according to claim 1, characterized in that, The method also includes: extracting meeting type information from meeting metadata before the analysis begins; selecting a matching entropy analysis mode from a mode library containing decision-making and divergent modes based on the extracted meeting type information; wherein, when the decision-making mode is selected, steps e and f are executed; when the divergent mode is selected, the entropy value of the corrected topic entropy time series is monitored to see if it is lower than the set creativity blocking threshold, and the creativity blocking point is output when it is lower than the creativity blocking threshold.

Images