A Meeting Information Intelligent Minutes Generation System

Abstract

This invention belongs to the field of text processing technology, specifically relating to an intelligent meeting minutes generation system. The system includes: a semantic particle construction module, an interaction potential energy calculation module, a semantic entropy flow calculation module, and a meeting minutes generation module. It converts the meeting record text stream into a semantic particle sequence, identifies interaction events between speakers, calculates interaction potential energy based on the distribution characteristics of the semantic particle sequence within a time window and the interaction events, calculates semantic entropy flow based on the distribution disorder of semantic particles within the current time window, calculates an effective decision index based on the interaction potential energy and semantic entropy flow, and then filters key decision frames, extracting the semantic particles corresponding to the key decision frames to generate meeting minutes. This invention solves the problems of traditional static analysis's inability to capture dynamic decision-making processes and distinguish invalid monologues, effectively improving the accuracy and key information coverage of meeting minutes generation.

Description

Technical Field

[0001] This invention relates to the field of text processing technology. More specifically, this invention relates to an intelligent meeting minutes generation system. Background Technology

[0002] With the acceleration of enterprise digital transformation, online meetings have become a core scenario for cross-regional collaborative work and business decision-making. In order to retain and trace massive amounts of meeting content, existing meeting minutes generation systems usually adopt a standardized technical approach, which is to first use automatic speech recognition technology to transcribe the meeting speech into text data, and then apply natural language processing technology to generate meeting minutes by extracting keywords or summarizing algorithms.

[0003] However, in complex multi-person interactive meeting scenarios, the meeting process exhibits significant dynamic temporal characteristics, and existing summary generation methods based on word frequency statistics or static semantic analysis have significant technical limitations. Specifically, meeting discussions typically include two stages: divergent discussion and convergent decision-making. In the divergent discussion stage, participants often engage in heated debates, circular discussions, or ineffective agreement on unresolved issues. This process is accompanied by extremely high word frequency and interaction intensity, making it easy for traditional algorithms based on statistical features to misjudge and retain these as core conclusions. At the actual decision-making moment, when opinions tend to converge and semantics begin to converge, participants' expressions are often simplified to short affirmative pronouns such as "agree" or "no problem." In existing processing logic, these low-frequency words that carry the final decision are easily filtered out as stop words or low-weight information.

[0004] High engagement does not equate to high information value. Low-frequency words may contain contradictory characteristics of genuine decision-making, making it difficult for existing systems to accurately distinguish between divergent debate processes and convergent consensus results. The generated meeting minutes are often filled with redundant debate details but lose the final key decision conclusions, failing to meet users' needs for accurate review of meeting results. Summary of the Invention

[0005] To address the technical problem of existing meeting minutes generation technologies failing to distinguish between divergent arguments and convergent consensus, resulting in generated minutes filled with redundant argumentative details and losing the final decision conclusion, this invention provides an intelligent meeting minutes generation system, comprising the following modules:

[0006] The semantic particle construction module acquires the meeting transcript text stream, transforms the text stream into a semantic particle sequence using a sliding time window, and identifies interaction events between speakers. The interaction potential energy calculation module calculates the interaction potential energy of the current time window based on the distribution characteristics of the semantic particle sequence within the time window and the interaction events; this interaction potential energy characterizes the interaction tightness between speakers. The semantic entropy flow calculation module calculates the semantic entropy flow of the current time window based on the disorder of the semantic particle distribution within the current time window. The meeting minutes generation module calculates the effective decision index based on the interaction potential energy and the semantic entropy flow; it then filters decision keyframes based on the effective decision index and extracts the semantic particles corresponding to the decision keyframes to generate the meeting minutes.

[0007] This invention acquires meeting transcript text streams and transforms them into semantic particle sequences using a sliding time window while simultaneously identifying interaction events. This discretizes continuous unstructured text into temporally sequential analytical units, providing a refined data foundation for dynamic analysis. Furthermore, by calculating interaction potential energy to characterize the interaction tightness between speakers, it reflects the degree of coupling among members, effectively distinguishing between high-value interactive discussions and ineffective monologues. Further, by calculating semantic entropy streams to measure the disorder of semantic particle distribution, it can measure the dispersion and convergence of viewpoints, accurately identifying the process from disagreement to consensus. Finally, by integrating interaction potential energy and semantic entropy streams to construct an effective decision-making index to filter key decision frames, it can accurately locate the specific moment of decision generation by combining high-heat and high-consensus characteristics, extracting key information to generate meeting minutes, significantly improving the accuracy of minute generation and the coverage of key information.

[0008] Preferably, the step of converting the text stream into a semantic particle sequence using a sliding time window and identifying interaction events between speakers includes: arranging the meeting record text stream by timestamp and extracting it using a sliding time window; extracting retained words within the time window and constructing a semantic particle set containing word vectors, speaker identifiers, and absolute timestamps; parsing the speaking sequence and content within the time window and classifying interaction events into replies, mentions, or interruptions; marking a mention in response to the explicit use of another person's name in the spoken text; marking a reply in response to a subsequent speaker speaking within a preset time threshold after the previous speaker has finished and having a semantic similarity greater than a preset similarity threshold; and marking an interruption in response to an overlap in the speaking time periods of the two speakers.

[0009] This invention utilizes a sliding time window to segment continuous meeting text streams and extracts retained words with actual referential meaning. This effectively preserves the temporal characteristics of meeting content evolution over time while filtering out meaningless noise interference. Furthermore, by analyzing the speaking sequence and content, interactive behaviors are subdivided into replying, mentioning, or interrupting. This provides fine-grained behavioral basis for subsequent accurate assessment of discussion heat and interaction patterns among members, thereby avoiding the one-sidedness of relying solely on speaking count statistics.

[0010] Preferably, the interactive potential energy satisfies the expression: In the formula, This represents the interaction potential energy within the current time window; This indicates the total number of people who participated in speaking within that time window; The spokesperson With the spokesperson The number of direct interactions within this time window; The spokesperson and The response delay coefficient between them; The spokesperson The total number of semantic particles output within this time window; The spokesperson The total number of semantic particles output within this time window.

[0011] This invention calculates interaction potential energy by comprehensively considering the number of participants, the number of direct interactions, the response delay coefficient, and the number of semantic particles output by both parties. It introduces a physics mapping index to quantify the coupling tightness between speakers. The response delay is used as the denominator, so that the faster the response, the higher the potential energy. At the same time, the geometric mean is used to weight the information content of both parties to ensure that the potential energy will increase significantly only when the interaction is frequent and both parties output substantial content. This effectively eliminates the interference of simple responses or monologues with low information content on the popularity assessment.

[0012] Preferably, the method for obtaining the response delay coefficient is as follows: obtaining the speakers within the time window. With the spokesperson The set of all interactive events; for each interaction in the set, in response to an interaction type of reply or mention, calculate the difference between the start timestamp of the next speaker and the end timestamp of the previous speaker to obtain the single response duration; in response to an interaction type of interruption, set the single response duration to 0; calculate the arithmetic mean of all single response durations in the set of interactive events as the average response delay, and use the ratio of the average response delay to a preset unit time as the response delay coefficient.

[0013] Preferably, the semantic entropy stream satisfies the expression: In the formula, This represents the semantic entropy stream of the current time window; This indicates the total number of different core keywords appearing within that time window; Indicates the first The percentage of frequency of each core keyword appearing within this time window; It is the natural logarithm function; Indicates the first The standard deviation of the frequency of use of the core keywords among different speakers; Indicates the first The average frequency of use of the core keywords among different speakers.

[0014] This invention introduces a correction factor based on the uniformity of core keywords among different speakers when calculating semantic entropy flow. It not only counts the word frequency ratio, but also combines the standard deviation and mean to measure the distribution of word usage. When a word is repeatedly mentioned by only one person, resulting in uneven distribution, the correction factor increases, thereby increasing the entropy value and identifying it as noise. When a word is mentioned by multiple people, the correction factor decreases, thereby reducing the entropy value, reflecting the orderliness of the system. This effectively distinguishes between two different semantic states: repeated opinions by one person and consensus reached by all members.

[0015] Preferably, the method for obtaining the core keywords is as follows: after calculating the importance of each semantic particle according to the word frequency inverse document frequency algorithm, a preset number of semantic particles are selected in descending order of importance as core keywords.

[0016] Preferably, the effective decision index satisfies the expression: In the formula, This represents the effective decision-making index for the current time window; The interaction potential energy for the current time window; This is the potential energy amplification factor; To prevent small positive constants with a denominator of zero; This represents the rate of change of semantic entropy flow over time. It is a function for maximizing the value.

[0017] The effective decision index constructed in this invention integrates interactive potential energy and semantic entropy flow, and introduces the time change rate of entropy flow as a screening condition. It uses a mathematical model to capture the phase transition process from chaos to order. The effective decision index only reaches its peak when the interactive heat is high, the semantic entropy value is low, and the entropy value shows a rapid downward trend. This accurately locates the decision moment when a consensus is quickly reached after a heated discussion, effectively eliminating arguments that have heat but no conclusion or bland statements that have consensus but no discussion.

[0018] Preferably, the method for obtaining the temporal change rate of the semantic entropy stream is as follows: the difference between the semantic entropy stream of the current time window and the semantic entropy stream of the previous time window is taken as the temporal change rate of the semantic entropy stream.

[0019] Preferably, the step of selecting decision key frames based on the effective decision index includes: performing peak detection on the effective decision index sequence; in response to the existence of a peak exceeding an adaptive threshold, marking the time window where the peak is located as a decision key frame, wherein the adaptive threshold is a preset multiple of the mean of the effective decision index sequence; in response to the absence of a peak exceeding the adaptive threshold, selecting the top 5% of the time windows with the largest values ​​in the effective decision index sequence as decision key frames.

[0020] Preferably, the generation of meeting minutes further includes: identifying the speaker who contributes the most interaction potential energy within the decision keyframe as the decision initiator; identifying other speakers who explicitly agree or have the highest semantic particle overlap within the decision keyframe as the execution associates; and associating the summary content of the decision keyframe with the decision initiator and execution associates to generate a structured to-do task.

[0021] The beneficial effects of this invention are as follows: By constructing an interactive potential energy model, this invention uses response delay as a decay factor of interactive potential energy and combines it with geometric mean information content weighting to effectively filter out interference from lengthy speeches by a single person or low-frequency idle chatter, accurately locking in the high-intensity discussion interval of the meeting; This invention introduces an improved semantic entropy flow index, using the coefficient of variation to correct the traditional entropy value, taking into account the dispersion of words in the distribution of the population, successfully measuring the divergence and convergence of viewpoints, and effectively solving the technical problem that traditional word frequency analysis cannot distinguish between heated arguments and consensus; The effective decision index proposed in this invention, unlike the traditional approach of static feature matching, uses the rate of entropy decrease as a trigger condition, which can accurately capture the moment of decision-making from chaos to order in the meeting, greatly improving the accuracy and logic of automatically generated minutes, and ensuring that key conclusions are not missed. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating a system block diagram of an intelligent meeting information minutes generation system according to the present invention;

[0023] Figure 2 It schematically illustrates the curves showing the changes in interaction potential energy and semantic entropy flow;

[0024] Figure 3 It schematically illustrates the changing curve of the effective decision-making index. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] This invention provides an intelligent meeting minutes generation system. For example... Figure 1 As shown, a meeting information intelligent minutes generation system includes a semantic particle construction module 100, an interactive potential energy calculation module 200, a semantic entropy flow calculation module 300, and a meeting minutes generation module 400, which are described in detail below.

[0028] The semantic particle building module 100 is used to acquire the meeting record text stream, construct a semantic particle sequence based on a sliding time window, and classify and identify interactive events.

[0029] It should be noted that meeting minutes are not static text accumulations, but dynamic information flows over time. The decision-making information contained therein often depends on the evolution of the temporal context. Directly performing static analysis on the entire meeting text would lose the temporal dimension features of the topic from initiation, divergence to convergence, making it impossible to pinpoint the specific moment when the decision was generated. Furthermore, the original meeting text contains a large number of meaningless interjections, which would introduce significant noise if statistical analysis were performed directly. At the same time, a single speech record cannot reflect the intensity of interaction between members. Therefore, this invention uses this module to discretize the continuous text stream into analysis units with temporal features, and transforms unstructured text into high-density semantic particles. At the same time, it performs fine-grained classification of the interaction behaviors between members to facilitate subsequent quantitative calculations.

[0030] Specifically, the speech-to-text stream of the conference system is accessed in real time via a data interface, and the text stream is arranged in timestamp order, with a set length of [length missing]. A sliding time window with a set step size. This allows the sliding time window to overlap and slide on the text stream, with the sliding time window length... The empirical value is 30 seconds, step size The empirical value is 5 seconds. In other embodiments, implementers can adjust the sliding time window length and step size based on the average speaking speed and interaction density of the actual meeting.

[0031] For any given time window, extract all spoken content within it, filter out meaningless function words and modal particles using a pre-defined stop word list, and retain nouns, verbs, and key adjectives. Define each retained word as a semantic particle, and each semantic particle contains three-dimensional attributes: word vector, speaker identifier, and absolute timestamp, thereby constructing a set of semantic particles for that time window.

[0032] Furthermore, the sequence and content of speeches within this time window are analyzed to identify explicit and implicit interactions between speakers, and interactive events are marked as replies, mentions, or interruptions. An event is marked as a mention when the speech text explicitly includes a name or specific ID for another person; an event is marked as a reply when the time interval between the next speaker's speech and the previous speaker's speech is less than a preset time threshold, and the semantic particle sets of the two speakers overlap with a cosine similarity greater than a preset similarity threshold; and an event is marked as an interruption when the start time of the next speaker is earlier than the end time of the previous speaker, i.e., their speaking time periods overlap. The empirical value of the time threshold is 10 seconds, used to define the response time window for effective dialogue; the empirical value of the similarity threshold is 0.6, used to ensure that the reply content is semantically relevant to the previous speech. In other embodiments, implementers can adjust the above thresholds according to the formality of the meeting scenario, for example, appropriately shortening the time threshold and increasing the similarity threshold in a serious reporting meeting.

[0033] The interaction potential energy calculation module 200 is used to calculate the interaction potential energy of the current time window based on the distribution characteristics of the semantic particle set within the time window and the classified interaction events.

[0034] It should be noted that the frequency of speaking alone cannot objectively reflect the intensity of the discussion. Since a single person's long speech may be just an ineffective lecture or a monologue in a meeting, it cannot represent the attention of the group. The real heat comes from the high-frequency interaction and semantic collision between multiple people. At the same time, in intense discussions, the responses between speakers are usually very rapid, while in casual chats or loose discussions, the response delay is longer. Therefore, this invention introduces the physical mapping index of interaction potential energy to measure the coupling tightness and response speed between speakers, thereby assessing the activity level of the current topic.

[0035] Specifically, the interaction potential energy of the current time window satisfies the expression:

[0036]

[0037] In the formula, This represents the interaction potential energy within the current time window; This indicates the total number of people who participated in speaking within that time window; The spokesperson With the spokesperson The number of direct interactions within this time window is a weighted sum of the number of mentions, replies, and interruptions identified by the semantic particle building module. A higher number of interactions indicates a more enthusiastic discussion and greater interaction potential. The larger; The spokesperson and The response delay coefficient is the ratio of the average time difference between the start time of the latter's speech and the end time of the former's speech in each interaction within the time window to a preset unit time. For interaction events identified as interruptions in the semantic particle construction module, their time difference is considered zero, and the unit time is 1 second. The faster the response, the higher the response delay coefficient. The smaller, The larger the value, the closer the interaction, and the higher the corresponding interaction potential energy. The higher; The spokesperson The total number of semantic particles output within this time window; The spokesperson The total number of semantic particles output within this time window; The geometric mean weight is used to weight the amount of information exchanged between the two parties in the interaction. This term only applies when... and The interaction potential energy only increases significantly when both sides are relatively large, ensuring that the interaction potential energy only increases when there are frequent interactions, rapid responses, and a large amount of information on both sides. This will reach its peak, thus effectively eliminating interference from monologues or low-information chatter.

[0038] The semantic entropy flow calculation module 300 is used to calculate the distribution disorder of semantic particles within the current time window and obtain the semantic entropy flow index.

[0039] It should be noted that high interaction potential only indicates that the participants are having a heated discussion, but it cannot directly distinguish whether they are having a fierce disagreement or reaching a consensus. In a meeting scenario, when opinions diverge, the vocabulary used by different people varies greatly, resulting in high semantic entropy of the system. When opinions converge and a consensus is reached, the vocabulary of different people becomes similar, resulting in lower semantic entropy of the system. Furthermore, traditional Shannon entropy only considers word frequency probability and cannot distinguish between a person's high-frequency repetition and what everyone mentions. Therefore, this invention introduces a distribution uniformity correction and calculates the semantic entropy flow to measure the degree of dispersion of opinions.

[0040] Specifically, the semantic entropy flow of the current time window satisfies the expression:

[0041]

[0042] In the formula, This represents the semantic entropy stream of the current time window; This represents the total number of different core keywords appearing within the time window. The core keywords are semantic particles that rank among the most important within the time window, selected based on the term frequency inverse document frequency algorithm. Indicates the first The percentage of frequency of each core keyword appearing within this time window; It is the natural logarithm function; Indicates the first The standard deviation of the frequency of use of the core keywords among different speakers; Indicates the first The average frequency of use of the core keywords among different speakers; This is a correction factor for the coefficient of variation; when a word is repeatedly mentioned by only one person and not used by others, its distribution is highly uneven, with a standard deviation of [missing information]. A larger value leads to a significant increase in the correction factor, which in turn affects the calculated semantic entropy flow. An increase in standard deviation is considered high-entropy noise by the system; conversely, when a word is mentioned evenly by everyone, the standard deviation decreases. The value is relatively small, and the correction factor is close to 1. If the word frequency is high at this time, then the semantic entropy flow... The significant reduction reflects the overall orderliness of the system, which is beneficial for capturing the convergence moment when all members agree.

[0043] It should be noted that the method for obtaining core keywords is as follows: the importance of each semantic particle is calculated according to the term frequency inverse document frequency (TF-IDF) algorithm, and a preset number of semantic particles are selected in descending order of importance as core keywords. The empirical value of the preset number of core keywords is 10. In other embodiments, the implementers can adjust it according to the actual speaking speed and information density of the meeting.

[0044] The meeting minutes generation module 400 is used to construct an effective decision index, combine interactive potential energy and semantic entropy flow to filter decision fragments, and generate meeting minutes.

[0045] It should be noted that true meeting decision points often possess the dual characteristics of high engagement and low entropy, meaning that a linguistic consensus is quickly reached after heated discussions. Because static threshold judgments are prone to missing rapidly changing decision points and are difficult to adapt to the differences in benchmarks between different meeting styles, it is impossible to accurately pinpoint the moment a decision is generated. Therefore, this invention constructs a dynamic index based on the rate of entropy decrease, focusing on the phase transition process from chaos to order, in order to accurately screen decision segments.

[0046] Specifically, the effective decision index is calculated based on the interaction potential energy and semantic entropy flow:

[0047]

[0048] In the formula, This represents the effective decision-making index for the current time window; The interaction potential energy for the current time window; This is the potential energy amplification factor, used to highlight the effect of heat. The empirical value is between 1.5 and 2.0; in other embodiments, implementers may adjust this value depending on the type of meeting. For brainstorming meetings, the volume can be appropriately increased. ; The semantic entropy flow for the current time window, used as the denominator, signifies that the lower the semantic entropy flow, i.e., the higher the consensus, The larger the value; To prevent small constants with a denominator of zero; It represents the negative value of the temporal rate of change of semantic entropy flow, that is, the rate of decrease of entropy. In actual calculation, it is implemented by the difference between the entropy value of the current time window and the entropy value of the previous time window. This is a maximum value function used to filter the stages where the entropy value decreases. This ensures that the effective decision index is larger only when the entropy value is in a downward trend and the rate of decline is faster, thus eliminating the cold phase where the entropy value is low but remains unchanged for a long time.

[0049] In the formula, the potential energy amplification factor is... One preferred value is 1.8, and the logic behind setting it is as follows: As an exponential term, it determines the model's nonlinear sensitivity to interaction heat, by setting... 1. This exponentially amplifies the weight of high-interaction potential ranges, thereby effectively suppressing interference from ordinary discussions of low to medium intensity when calculating the effective decision index. When implementing the settings, if the focus is on uncovering key conclusions after heated debates, the filtering threshold for intensity should be increased, and a high value close to 2.0 can be used. If the focus is on comprehensively recording consensus in the routine process, more information from mild discussions should be retained, and a low value close to 1.5 can be used. experience value Its setting principle is that the order of magnitude is much smaller than the semantic entropy stream. The normal fluctuation range is used to prevent the system from being in a state of absolute consensus, i.e. The computational overflow caused by a zero denominator is addressed, while ensuring efficient decision-making within the low-entropy range. It will not produce nonlinear distortion amplification. In other embodiments, implementers can set it according to the actual implementation situation, but in order to ensure the stability of numerical calculation, The range of values ​​is to .

[0050] For example, Figure 2 The curves represent the changes in interaction potential energy and semantic entropy flow. Figure 3 The curve showing the change in the effective decision-making index reveals that during the 30-80 second interval of the meeting, the interaction potential energy remains high at 1500-2000, while the semantic entropy flow is also high at 7-8. This indicates that although the interaction intensity is extremely high during this stage, the semantic particles are dispersed, and the viewpoints are in a highly divergent debate state. Subsequently, during the 80-85 second interval, the semantic entropy flow experiences a precipitous drop, rapidly falling below 1. At this time, although the interaction potential energy declines, it still maintains effective connectivity, indicating that the meeting has entered a convergence phase from chaos to order. Figure 3 It can be seen that the effective decision index shows a significant peak of over 18,000 around 82 seconds. In the preceding debate phase, from 30 to 80 seconds, although the interaction potential energy is high, the effective decision index remains at a low noise level due to the suppression of high entropy. This invention can effectively filter out ineffective, heated arguments by monitoring the coupling relationship between the rate of entropy decrease and potential energy, and accurately pinpoint the critical moment for decision generation and viewpoint convergence.

[0051] Furthermore, peak detection is performed on the calculated effective decision index sequence. An adaptive threshold is set, and the time window containing the peak exceeding this threshold is marked as a decision keyframe. The time range of this keyframe is expanded forward and backward to cover the complete semantic context. All semantic particles within this expanded range are extracted and restored to the original text. A pre-trained text summarization model is used to perform generative summarization on the extracted text to obtain the final decision minutes. Simultaneously, based on the interaction relationships such as replies, mentions, and interruptions identified in the semantic particle construction module, the speaker who contributes the most interaction potential energy within the decision keyframe is identified as the decision initiator. Other speakers who explicitly agree or have the highest semantic particle overlap are identified as executors. The decision content is automatically associated with the corresponding responsible parties, generating a structured task assignment. The adaptive threshold is set to 3 times the effective decision index sequence. In other embodiments, the implementer can adjust the multiple according to the need for conciseness of the minutes.

[0052] Furthermore, if there are no peaks exceeding the adaptive threshold in the effective decision index sequence, indicating a relatively smooth meeting process or a lack of significant abrupt decision-making transitions from chaos to order (e.g., a purely reporting meeting), the system automatically executes a rollback strategy, selecting the peak with the largest value in the effective decision index sequence. The time window is used as a keyframe, and adjacent keyframes are merged in chronological order. Then, the corresponding semantic particles are extracted to generate a summary, so as to ensure that minutes containing relatively high-value information can still be produced in low-adversarial or procedural meetings.

Claims

1. A meeting information intelligent minutes generation system, characterized in that, include: The semantic particle building module is used to acquire the meeting transcript text stream, convert the text stream into a semantic particle sequence using a sliding time window, and identify interaction events between speakers; The interaction potential energy calculation module is used to calculate the interaction potential energy of the current time window based on the distribution characteristics of the semantic particle sequence within the time window and the interaction events. The interaction potential energy is used to characterize the interaction tightness between speakers. The semantic entropy stream calculation module is used to calculate the semantic entropy stream of the current time window based on the disorder of the distribution of semantic particles within the current time window. The meeting minutes generation module is used to calculate the effective decision index based on the interaction potential energy and semantic entropy flow; filter decision key frames based on the effective decision index; extract the semantic particles corresponding to the decision key frames to generate meeting minutes. Semantic entropy flow satisfies the expression: In the formula, This represents the semantic entropy stream of the current time window; This indicates the total number of different core keywords appearing within that time window; Indicates the first The percentage of frequency of each core keyword appearing within this time window; It is the natural logarithm function; Indicates the first The standard deviation of the frequency of use of the core keywords among different speakers; Indicates the first The average frequency of use of the core keywords among different speakers.

2. The intelligent meeting minutes generation system according to claim 1, characterized in that, The process of converting a text stream into a sequence of semantic particles using a sliding time window and identifying interaction events between speakers includes: The meeting transcript text stream is arranged by timestamp and truncated using a sliding time window. Retained words within the time window are extracted to construct a semantic particle set containing word vectors, speaker identifiers, and absolute timestamps. The speaking sequence and content within the time window are analyzed, and interaction events are categorized as replies, mentions, or interruptions. A mention is marked when the spoken text explicitly contains a name of another speaker. A reply is marked when a subsequent speaker speaks within a preset time threshold after the previous speaker's speech and the semantic similarity is greater than a preset similarity threshold. An interruption is marked when the speaking time periods of two speakers overlap.

3. The intelligent meeting minutes generation system according to claim 2, characterized in that, The interaction potential energy satisfies the expression: ； In the formula, This represents the interaction potential energy within the current time window; This indicates the total number of people who participated in speaking within that time window; The spokesperson With the spokesperson The number of direct interactions within this time window; The spokesperson and The response delay coefficient between them; The spokesperson The total number of semantic particles output within this time window; The spokesperson The total number of semantic particles output within this time window.

4. The intelligent meeting minutes generation system according to claim 3, characterized in that, The method for obtaining the response delay coefficient is as follows: Get speakers within the time window With the spokesperson The set of all interactive events; for each interaction in the set, in response to an interaction type of reply or mention, calculate the difference between the start timestamp of the next speaker and the end timestamp of the previous speaker to obtain the single response duration; in response to an interaction type of interruption, set the single response duration to 0; calculate the arithmetic mean of all single response durations in the set of interactive events as the average response delay, and use the ratio of the average response delay to a preset unit time as the response delay coefficient.

5. The intelligent meeting minutes generation system according to claim 1, characterized in that, The method for obtaining the core keywords is as follows: After calculating the importance of each semantic particle using the inverse document frequency algorithm, a preset number of semantic particles are selected in descending order of importance as core keywords.

6. The intelligent meeting minutes generation system according to claim 1, characterized in that, The effective decision index satisfies the expression: ； In the formula, This represents the effective decision-making index for the current time window; The interaction potential energy for the current time window; This is the potential energy amplification factor; To prevent small positive constants with a denominator of zero; This represents the rate of change of semantic entropy flow over time. It is a function for maximizing the value; Represents the semantic entropy stream of the current time window.

7. The intelligent meeting minutes generation system according to claim 6, characterized in that, The method for obtaining the time change rate of the semantic entropy stream is as follows: The difference between the semantic entropy stream of the current time window and the semantic entropy stream of the previous time window is used as the temporal rate of change of the semantic entropy stream.

8. The intelligent meeting minutes generation system according to claim 1, characterized in that, The step of filtering decision key frames based on the effective decision index includes: Peak detection is performed on the effective decision index sequence; in response to the existence of a peak exceeding an adaptive threshold, the time window containing the peak is marked as a decision key frame, where the adaptive threshold is a preset multiple of the mean of the effective decision index sequence; in response to the absence of a peak exceeding the adaptive threshold, the top 5% of the time windows with the largest values ​​in the effective decision index sequence are selected as decision key frames.

9. A meeting information intelligent minutes generation system according to claim 8, characterized in that, The generation of meeting minutes also includes: The speaker who contributes the most interaction potential energy within the decision keyframe is identified as the decision initiator; other speakers who explicitly agree or have the highest semantic particle overlap within the decision keyframe are identified as the execution associates; the summary content of the decision keyframe is associated with the decision initiator and the execution associates to generate a structured to-do task.

Images