Conference summary structured extraction and filing method

Abstract

The invention relates to the technical field of conference text processing, and discloses a conference summary structured extraction and filing method. The method comprises the following steps: acquiring conference record data, and determining conference content elements based on the conference record data; analyzing conference scene features according to the conference content elements; collecting the existing ambiguity information of the conference summary, and formulating a structured extraction parameter in combination with the conference scene features and the existing ambiguity information; scheduling text feature data and semantic structure data of the conference summary, and analyzing semantic performance representation corresponding to the text feature data; determining a signal propagation topology of the conference summary according to the semantic structure data; generating structured summary data based on the signal propagation topology; and finally, performing archiving processing based on the structured summary data. According to the method, through systematic element analysis, scene adaptation, ambiguity processing, semantic mining and topology construction, whole-process standardized processing of the conference summary from information extraction to structured filing is realized, and the information quality and management efficiency of the conference summary are improved.

Description

Technical Field

[0001] This invention relates to the field of meeting text processing technology, specifically a method for structured extraction and archiving of meeting minutes. Background Technology

[0002] In the daily operations of various organizations, meetings serve as a crucial vehicle for information exchange and decision-making. Meeting minutes play a central role in recording key information, tracing the decision-making process, and guiding subsequent work. However, current methods of processing meeting minutes generally suffer from numerous problems that urgently need to be addressed. In the traditional model, meeting minutes largely rely on manual compilation, a process heavily dependent on the recorder's personal experience and subjective judgment. This results in significant differences in the completeness and logical clarity of minutes produced by different recorders.

[0003] In existing technologies, some automated processing tools can only achieve simple text transcription or keyword extraction, and cannot deeply explore the inherent structure and semantic relationships of meeting content. In terms of extracting meeting content elements, due to the lack of unified standards and methods, key information is often omitted or there is too much redundant information. For example, it is difficult to systematically identify and extract core elements such as meeting topics, resolutions, and to-do items.

[0004] The diversity of meeting scenarios also presents challenges for minutes processing. Different types of meetings, such as project progress meetings and problem coordination meetings, have significantly different focuses and information presentation formats. Existing tools often adopt a general processing model and cannot adjust processing strategies according to specific scenario characteristics, resulting in a low degree of matching between extracted information and actual needs.

[0005] Meanwhile, meeting minutes contain a large amount of ambiguous information, including vague expressions, differences in technical terminology, and context-dependent statements. Existing processing methods are insufficient to identify and resolve this type of information, resulting in semantic ambiguity and unclear referencing in the compiled minutes. In the information structuring and archiving stages, due to the lack of effective semantic analysis and logical organization mechanisms, the minutes data is mostly stored in unstructured text form. This not only hinders subsequent retrieval, statistics, and reuse but also easily creates information silos, affecting the effective flow and value realization of meeting information. These problems collectively prevent the full release of the informational value of meeting minutes, hindering the organization's knowledge management and decision support. Summary of the Invention

[0006] The purpose of this invention is to provide a method for structured extraction and archiving of meeting minutes to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for structured extraction and archiving of meeting minutes, the method comprising: Obtain meeting record data, and based on the meeting record data, identify the key elements of the meeting content; Based on the aforementioned meeting content elements, analyze the characteristics of the meeting scenario; Collect existing ambiguous information from meeting minutes, and formulate structured extraction parameters based on the meeting scenario characteristics and the existing ambiguous information. The text feature data and semantic structure data of the scheduling meeting minutes are analyzed to determine the semantic performance representation corresponding to the text feature data. Based on the semantic structure data, determine the signal propagation topology of the meeting minutes; Based on the aforementioned signal propagation topology, structured summary data is generated; Based on the structured minutes data, archiving processing is performed.

[0008] Preferably, generating structured summary data based on the signal propagation topology includes: The synergistic coupling effect between the signal propagation topology and the semantic performance representation is evaluated, and the parameter priority sequence of the structured extraction parameters is determined based on the synergistic coupling effect. The core entity type and alternative entity type of the meeting minutes are queried. The meeting minutes are simulated and constructed by combining the core entity type and the alternative entity type to obtain the minutes prototype. Collect entity recognition data and topic classification data of the prototype minutes, and analyze the entity detection characteristics of the prototype minutes based on the entity recognition data; Based on the topic classification data, the topic consistency equivalent value of the minute prototype is calculated, and the semantic stability of the minute prototype is evaluated based on the topic consistency equivalent value. Determine the storage location requirements and access constraints of the meeting minutes in the archiving system, collect the adaptability parameters of the minutes prototype, and analyze the compatibility of the minutes prototype in the archiving system by combining the storage location requirements, the access constraints, and the adaptability parameters. Combining the entity detection characteristics, semantic stability, and adaptation compatibility, the optimal entity type is selected from the core entity types and the candidate entity types. Based on the structured extraction parameters, the parameter priority sequence, and the optimal entity type, structured summary data is generated.

[0009] Preferably, the step of combining the meeting scenario features and the existing ambiguous information to formulate structured extraction parameters includes: Feature extraction is performed on the meeting scene features to obtain scene feature factors; The existing ambiguous information is classified to obtain a set of ambiguous categories; Analyze the correlation between the scene feature factors and the ambiguous category set to obtain the correlation mapping matrix; Based on the aforementioned association mapping matrix, the key scenario features and key ambiguity categories of the meeting minutes are determined; Based on the key scenario features and key ambiguity categories, the extraction constraints of meeting minutes are analyzed. Based on the extraction constraints, the structured extraction parameters are formulated.

[0010] Preferably, the analysis of the semantic performance representation corresponding to the text feature data includes: The text feature data is standardized to obtain standardized text data; Extract the semantic attributes corresponding to the standardized text data, filter the semantic attributes, and obtain the key semantic attributes; Calculate the semantic performance index corresponding to the key semantic attributes, and generate the semantic performance representation corresponding to the text feature data based on the semantic performance index.

[0011] Preferably, determining the signal propagation topology of the meeting minutes based on the semantic structure data includes: The semantic structure data is preprocessed to obtain the target semantic structure data; Extract the set of structural parameters for the meeting minutes from the target semantic structure data; The structural parameter set is subjected to semantic attribute association processing to obtain the attribute association parameter set; Based on the attribute-related parameter set, a numerical model of signal propagation corresponding to the meeting minutes is constructed; The signal propagation numerical model is simulated and processed to obtain dynamic data of signal propagation. The signal propagation dynamic data is subjected to topological abstraction processing to generate the signal propagation topology of the meeting minutes.

[0012] Preferably, the evaluation of the synergistic coupling effect between the signal propagation topology and the semantic performance representation includes: Extract the signal propagation features corresponding to the signal propagation topology, and perform dimensionality reduction processing on the signal propagation features to obtain dimensionality-reduced signal propagation features; Calculate the feature similarity index between the reduced-dimensional signal propagation features, and calculate the representation similarity index between the semantic performance representations; Calculate the correlation factor between the signal propagation topology and the semantic performance representation; By combining the correlation factor, the feature similarity index, and the representation similarity index, the degree of synergistic coupling between the signal propagation topology and the semantic performance representation is calculated. Based on the aforementioned cooperative coupling degree, the cooperative coupling effect between the signal propagation topology and the semantic performance representation is evaluated.

[0013] Preferably, calculating the correlation factor between the signal propagation topology and the semantic performance representation includes: The signal propagation topology and the semantic performance representation are vectorized respectively to obtain the propagation topology vector and the performance representation vector. Calculate the vector cosine between the propagation topology vector and the performance characterization vector; Calculate the vector mutual information between the propagation topology vector and the performance representation vector; By combining the vector cosine value and the vector mutual information, the correlation factor between the signal propagation topology and the semantic performance representation is calculated.

[0014] Preferably, the step of analyzing the entity detection characteristics of the prototype minutes based on the entity recognition data includes: The entity recognition data is cleaned to obtain cleaned entity recognition data; Analyze and extract the time-domain and frequency-domain features corresponding to the cleaned entity identification data; Based on the time-domain features and the frequency-domain features, an entity characteristic descriptor corresponding to the minutes prototype is generated; Based on the entity characteristic descriptor, the entity detection characteristics of the transcript prototype are analyzed.

[0015] Preferably, calculating the thematic consistency equivalent value of the minutes prototype based on the thematic classification data includes: Analyze the classification performance parameters corresponding to the topic classification data, and query the error patterns and error mechanisms corresponding to the meeting minutes; Variable analysis was performed on the error modes and error mechanisms to obtain the error variables corresponding to the meeting minutes; Based on the aforementioned error variables, thematic evaluation indicators corresponding to the meeting minutes are formulated. Based on the thematic evaluation indicators, thematic correlation parameters of the meeting minutes are extracted from the classification effectiveness parameters. Based on the topic classification data, the topic score value corresponding to the topic association parameter is calculated; Assign the parameter contribution degree corresponding to the topic association parameter, and calculate the topic consistency equivalent value of the minutes prototype by combining the topic score value and the parameter contribution degree.

[0016] Preferably, the step of analyzing the compatibility of the minutes prototype in the archiving system by combining the storage location requirements, the access condition constraints, and the adaptability parameters includes: Based on the storage location requirements, the storage area of ​​the prototype minutes in the archiving system is determined, and the area size parameters and minute size parameters corresponding to the storage area and the prototype minutes are measured. Based on the region size parameters and the minutes size parameters, the size fit of the minutes prototype in the archiving system is calculated; Analyze the adaptation factors corresponding to the adaptation parameters, and calculate the factor adaptation degree corresponding to the adaptation factors based on the access condition constraints and the adaptation parameters. Based on the aforementioned factor suitability, the access suitability of the minutes prototype in the archiving system is calculated. By combining the size adaptation and the access adaptation, the compatibility of the minutes prototype in the archiving system is analyzed. The archiving process based on the structured minutes data includes: Receive user interaction signals and determine the archiving operation type based on the user interaction signals; Based on the archiving operation type, the structured minutes data is dynamically corrected; Execute dynamically revised structured minutes data and collect archived feedback data; Based on the archived feedback data, update the text feature data and semantic structure data of the meeting minutes.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This method provides a complete and efficient solution for processing meeting minutes through a systematic step-by-step design. After acquiring the meeting record data, the process of clarifying the key elements of the meeting content allows for the identification of the core components of the meeting information from the source, avoiding information extraction biases caused by ambiguity in traditional processing and ensuring a clear objective for subsequent processing. Analyzing the characteristics of the meeting scenario based on content elements allows the processing to be adapted to the actual type of meeting and the focus of the agenda. Different meeting scenarios have different information needs; this targeted analysis ensures that the extracted information is more closely aligned with the actual needs of the scenario, reducing interference from irrelevant information.

[0018] By collecting existing ambiguous information and combining it with scenario characteristics to formulate structured extraction parameters, the problem of handling ambiguous content in meeting minutes was solved. The existence of ambiguous information often leads to misunderstandings. However, by considering scenario characteristics in conjunction with ambiguous information, a more comprehensive basis can be provided for parameter setting, making the extraction parameters more adaptable. This allows for the accurate identification and handling of vague expressions in different contexts, thereby improving the accuracy of information extraction.

[0019] By manipulating textual feature data and semantic structure data and analyzing semantic effectiveness representations, we can delve deeper into the intrinsic meaning of meeting texts. Textual features reflect the surface linguistic form, while semantic structure embodies the logical connections between information. Through comprehensive analysis of both, we can more fully capture the semantic connotation of meeting content and avoid semantic misunderstandings caused by relying solely on surface textual features.

[0020] Determining the signal propagation topology of meeting minutes based on semantic structure data provides a logical framework for the structured organization of information. The signal propagation topology clearly presents the flow path of information during the meeting, the relationships between topics, and the logical chain of decision-making, making the structured minutes more consistent with the actual logic of the meeting and facilitating users' quick understanding of the core message.

[0021] Structured meeting minutes data generated based on signal propagation topology transforms fragmented meeting information into standardized data with a clear structure and logical relationship. This structured data overcomes the shortcomings of unstructured text in terms of storage, retrieval, and reuse, making the presentation of meeting information more organized and facilitating users to quickly locate key content.

[0022] Archiving based on structured minutes data enables the orderly storage and management of meeting information, ensuring that minutes from different meetings maintain consistent standards and specifications during archiving. This facilitates subsequent statistical analysis, historical tracing, and cross-meeting information correlation queries, promoting the effective accumulation and value mining of meeting information. Attached Figure Description

[0023] Figure 1 This is a sequence diagram of the structured extraction and archiving method for meeting minutes described in this invention; Figure 2 A flowchart generated for structured minutes; Figure 3 A flowchart for generating a signal propagation topology from text processing; Figure 4 This is a flowchart of topological coupling analysis based on tensor decomposition. Detailed Implementation

[0024] 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 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.

[0025] Please see Figure 1 This invention provides a method for structured extraction and archiving of meeting minutes, the method comprising: Through multi-dimensional data processing and intelligent analysis technologies, the system achieves automated, structured processing and archiving management of meeting content. It acquires raw meeting minutes and uses natural language processing to identify key elements, including attendees, topics, and resolutions. Based on these identified elements, the system employs a deep learning model to analyze meeting scenario characteristics, distinguishing between different meeting types and discussion modes. By collecting ambiguous information from historical meeting minutes and combining it with the characteristics of the current meeting scenario, a structured extraction parameter system is established. The system simultaneously analyzes the lexical, syntactic, and discourse features of the meeting text to construct a semantic performance evaluation model. Based on the semantic structure analysis results, a logical network of connections between meeting content is established, forming a signal propagation topology. Finally, standardized structured minutes data is generated based on this topology and intelligently archived according to the enterprise document management system.

[0026] Example 1: See Figure 2 Upon receiving the signal propagation topology and semantic performance representation, the system initiates a tensor decomposition module to model the high-order relationship between the two. This module maps the node connections in the topology to the functional attributes in the semantic representation into a three-dimensional tensor space, decomposing them into a core tensor matrix using an alternating least squares algorithm. The system automatically parses the eigenvectors of the core matrix and extracts the parameter sequence by sorting them in descending order of eigenvalues. During execution, a dynamic threshold mechanism is set to retain only core parameters with a contribution rate higher than a preset threshold, while other parameters are assigned to an auxiliary queue. Simultaneously, the entity knowledge base retrieval function is activated. This knowledge base adopts a tree-like index structure, with the top layer containing core entity types such as resolution items and action items, and the sub-layers containing alternative entity types such as points of contention and reference bases. Guided by a priority sequence, the system dynamically matches entity combination schemes in the knowledge base, generating an independent simulation construction path for each scheme.

[0027] The construction of the minutes prototype adopts a phased iterative strategy. The first round builds a basic framework based on core entity types, and the system matches meeting segments with entity slots using a semantic alignment algorithm. The second round introduces alternative entity types to fill in supplementary information, and an attention weighting mechanism is used to adjust the presentation intensity of each entity. Upon completion, multiple parallel prototypes are generated, each retaining records of differentiated entity combinations. The system then enters a triple evaluation phase: the entity recognition module loads a Transformer-based hybrid model, which simultaneously processes the original minutes text and the prototype structured data. The model outputs dual results: entity boundary markers and type labels. Based on these, the system calculates the boundary consistency score for entity recognition and the confusion matrix for type discrimination. For topic classification data, the system deploys a hierarchical relevance propagation algorithm to analyze the logical dependencies between topics. This algorithm constructs a directed acyclic graph (DAG) to represent the topic hierarchy, assigning semantic association weights to each edge. Specifically, the algorithm first extracts topic nodes from the topic classification data, with each node representing a meeting topic or sub-topic. When constructing the DAG, nodes are arranged according to the hierarchical relationship of topics, with the top layer representing core topics and the sub-layers representing derived topics, ensuring an acyclic structure to avoid logical loops. The direction of the edges represents the dependencies between topics, such as from a parent topic to a child topic. Semantic association weights are assigned based on the semantic similarity calculation between topics, using a cosine similarity algorithm to measure the proximity of topic vectors in the semantic space. Specifically, for any two topic nodes, the system calculates the cosine value of their text feature vectors as the weight, using the following formula: in, This represents the edge weight from node i to node j. and Semantic vector representations for topics i and j are generated using a word embedding model such as Word2Vec. Weight values ​​range from -1 to 1, with positive values ​​indicating strong semantic associations and negative values ​​indicating conflicts. The system uses threshold filtering to retain only edges with weights higher than a preset value, ensuring the sparsity and interpretability of the graph. The entire process is integrated into the topic analysis module, which takes topic classification data as input and outputs a weighted directed acyclic graph for subsequent semantic propagation analysis.

[0028] The system traverses the semantic propagation paths in the graph, calculates the topic dispersion index for different paths, and finally outputs a standardized semantic stability score. Specifically, after constructing the directed acyclic graph, the system uses a depth-first search algorithm to traverse all semantic propagation paths, with each path extending from the root node to a leaf node. The topic dispersion index is calculated for each path, measuring the degree of variation in topic weights along the path, using the standard deviation formula: in, The topic dispersion index represents path p. The number of nodes on the path. Let the weight be the weight of the k-th edge. This represents the average weights along the path. A higher dispersion index indicates poorer thematic consistency along the path. Subsequently, the system standardizes the dispersion index of all paths, using a min-max normalization method to map the index to the [0,1] interval, as shown in the formula: here, and These are the minimum and maximum values ​​of the dispersion index for all paths, respectively. The standardized semantic stability score is ultimately calculated as the complement of the normalized dispersion index, i.e. A higher score indicates greater semantic stability of the topic. This process is implemented in the semantic analysis module, which takes a weighted directed acyclic graph as input and outputs a score to evaluate the semantic quality of the minutes prototype.

[0029] The archive compatibility assessment employs a multi-parameter mapping method, parsing the physical partitioning rules and logical naming rules in the storage location requirements, and performing a field-by-field matching check between the metadata structure in the compatibility parameters and the storage rules. The system establishes a virtual storage sandbox to simulate write operations and records the interaction logs between the prototype data and the storage interface.

[0030] The 3D evaluation results are input into the entity selection engine, which employs a multi-objective optimization algorithm. The system defines entity detection characteristic values ​​as constraints, and semantic stability and fit compatibility as the optimization objective functions. Before performing a Pareto optimal solution search, the engine automatically loads a pre-trained scene decision model and sets weight preference coefficients based on the meeting scene characteristics. Meeting scene characteristics include meeting type (e.g., strategic decision-making or project execution meeting), participant roles, and agenda structure; these features are quantized into feature vectors. The scene decision model is a multilayer perceptron neural network. The input layer receives the scene feature vector, the hidden layer uses the ReLU activation function for non-linear transformation, and the output layer generates a weight preference coefficient vector. Model training uses historical meeting data and is optimized through backpropagation, with the mean squared error loss function to ensure that the coefficients reflect the impact of the scene on entity importance. Specifically, the weight preference coefficients... The calculation is as follows: in, For scene feature vectors, and For model weights and bias parameters, The sigmoid function is used to restrict the output to the range [0,1]. Coefficients are used to adjust the multi-objective optimization weights in the entity selection engine, such as weights that improve semantic stability in the strategy meeting and weights that enhance adaptability in the execution meeting. This system module is integrated into the entity selection phase to ensure that the weights dynamically adapt to scenario requirements.

[0031] After thousands of iterations, the optimal combination of entity types is output. The final structuring process performs a hierarchical fusion operation, mapping core parameters in the priority sequence to preset slots for the optimal entity type, while auxiliary parameters are stored in extended attribute fields. The system automatically verifies the logical closure of the structure and checks the completeness of the executable chain for each decision item using a source tracing detection algorithm. Upon completion, structured summary data is generated, and a metadata description file is output recording snapshots of key parameters in the decision path. An anomaly capture node is set up throughout the implementation process; when semantic gaps exceed the fault tolerance threshold, a backtracking mechanism is activated to self-correct from the entity type reorganization stage.

[0032] Example 2: See Figure 3 When text feature data is input, the system activates a preprocessing pipeline to perform multi-level normalization operations on the raw text. The terminology unification module maps vocabulary to a domain-specific terminology dictionary, converting colloquial expressions, industry abbreviations, and synonyms into standardized entries from the standard terminology database. The sentence regularization component parses complex sentence structures, breaks down nested clauses, and reconstructs standard subject-verb-object sentences, maintaining the original semantic role framework unchanged. The punctuation system is reconstructed according to formal document specifications, eliminating non-standard pause marks in spoken language. The standardized text data resulting from this process possesses grammatical structural uniformity and terminological consistency.

[0033] In the semantic attribute extraction stage, a hybrid analysis model is deployed, integrating a rule-based syntactic parser and a deep learning-based semantic parser. The syntactic parser employs an enhanced dependency grammar tree algorithm to identify the core predicates of a sentence and the structure of their governing argument roles. The semantic parser incorporates a multi-head attention mechanism to capture long-range contextual dependencies and identify implicit semantic relationships. The system establishes a dynamic filtering mechanism, implementing a three-layer filtering process on the extracted raw semantic attribute set: the first layer sorts the attributes according to their functional weight in the meeting scenario, retaining only attributes with decision relevance higher than a set threshold; the second layer uses a mutual exclusion detection algorithm to eliminate semantically repetitive or contradictory attributes; and the third layer applies fuzzy semantic resolution technology to refine the domain of attributes with unclear boundaries. The final output set of key semantic attributes forms a discrete yet complete set of semantic units.

[0034] The construction of semantic effectiveness representations employs a multi-dimensional quantitative evaluation model. The system assigns three core indicators to each key semantic attribute: information density, which measures the relevance of the attribute content to the core agenda of the meeting; functional completeness, which assesses the semantic integrity of the attribute; and timeliness, which records the applicable time period of the attribute. The system establishes an indicator weighting algorithm that automatically adjusts the weight coefficients based on the meeting type—increasing the weight of information density in strategic decision-making meetings and enhancing the weight of timeliness in project execution meetings. The indicator aggregation process uses a non-linear fusion function to avoid the indicator offsetting effect that may occur with simple weighted averaging. The generated semantic effectiveness representation forms a structured vector, with each dimension representing the quantitative value of the attribute at a specific effectiveness level.

[0035] The processing of semantic structure data begins with pre-processing. The system first applies a redundant information filter, using pattern matching algorithms to identify and remove non-substantive content such as formatted text paragraphs and ceremonial expressions. The referential parsing engine performs entity connections on cross-sentence pronouns, constructing referential chains for personal pronouns and association graphs for demonstrative pronouns. The semantic conflict detection component scans for logically contradictory expressions, initiating an agenda-priority-based resolution mechanism to selectively retain or structurally annotate conflicting content. The preprocessed target semantic structure data forms a semantically self-consistent and logically coherent sequence of text units.

[0036] The signal propagation topology construction process employs a hierarchical modeling strategy. The structural parameter extraction component parses four basic elements from the target semantic structure data: issue nodes contain the core discussion topic and its attributes; decision nodes record the final resolution; argumentation nodes store the chain of arguments supporting the resolution; and action nodes associate specific execution elements. The attribute association processor establishes relationship mappings between nodes, calculates the issue-decision association strength using a semantic similarity algorithm, verifies the causal association degree of argumentation-decision using logical reasoning rules, and establishes the sequence dependency relationship of action nodes based on a time-series model. The attribute association parameter set forms a weighted directed graph data model, with nodes carrying type and weight attributes, and edges labeled with association type and influence factors. The attribute association parameter set originates from the processing of semantic structure data. Nodes represent meeting elements such as issues, decisions, or action items. Node types are classified by the entity recognition module, and weight attributes are calculated based on the frequency and importance of node occurrence in the meeting, for example, using the TF-IDF algorithm. Edges represent associations between elements, with association types categorized as causal, temporal, or logical. Influence factors are derived through semantic similarity and logical reasoning. The calculation of influence factors uses an improved cosine similarity and incorporates a causal strength index, with the formula: in, Let i be the influence factor from node i to j. For semantic similarity function, This is a causal inference function (based on rule engine verification). This is a balancing parameter (usually set to 0.7). Similarity calculation is based on word vectors, and the causal function returns 0 or 1 to indicate the existence of a causal chain. The weighted directed graph model is constructed using a graph database, and node and edge attributes are used for subsequent signal propagation simulations.

[0037] The simulation of the signal propagation numerical model sets dynamic propagation rules. The information flow simulator defines three propagation modes: a diffusion propagation mode is enabled between issue nodes to reflect issue relevance; a directional propagation mode is used from argumentation to decision nodes to simulate the strength of argumentation support; and a sequential propagation mode is executed between action nodes to reflect time dependence. The system sets an attenuation factor function to describe the attenuation curve of information crossing nodes; a gain regulator is also configured to handle node aggregation effects. The attenuation factor function is used in the signal propagation numerical model to simulate the attenuation of information between nodes. The function is based on an exponential attenuation model, and the process involves fitting historical data: the system collects information propagation records from previous meetings and uses regression analysis to determine the attenuation constant. The specific function is as follows: here, This represents the attenuation factor when information travels a distance of d nodes. The decay constant is obtained by fitting historical data using the least squares method, and d is the topological distance between nodes. For example, the value of k is adjusted according to the type of meeting; k is larger for rapid decision-making meetings (faster decay) and smaller for detailed discussion meetings. The function is integrated into the propagation simulation module and is used to calculate the decay curve of information intensity.

[0038] The signal propagation dynamic simulator is run to perform multiple rounds of iterative calculations, recording the arrival strength, path delay, and stability coefficient of information at each node. The output signal propagation dynamic dataset includes a heatmap of node influence and a critical path topology map.

[0039] The topology abstractor receives dynamic data and performs structured compression. A multi-scale clustering algorithm aggregates nodes of similar function to form high-level functional groups. A critical path miner extracts information channels with propagation efficiency exceeding a set threshold, constructing a streamlined backbone topology. The abstraction process applies graph reduction techniques to remove inefficient and redundant branches, retaining high-influence nodes and their strong connections. The final generated signal propagation topology maps to the conference decision-making logic structure, with core node clusters corresponding to key decision regions and densely connected channels reflecting highly collaborative decision chains. This topology forms a logical relationship framework between semantic units, providing path guidance for subsequent structured processing.

[0040] Example 3: See Figure 4This paper employs tensor decomposition to process the signal propagation topology, transforming it into a dense representation in a low-dimensional vector space. The process begins by embedding nodes into the topological graph, generating node sequences using a random walk algorithm, and then learning the distributed representation of nodes in a continuous vector space using a Skip-gram model. Edge relationships are mapped to the same space through a bilinear transformation, forming a complete vectorized representation of the graph structure. The semantic performance representation, already in feature vector form, undergoes dimensional alignment to maintain the same dimensionality as the topological vectors.

[0041] The evaluation of collaborative coupling is based on a deep analysis of two vector spaces. The system designs a multi-scale similarity measurement framework, calculating the consistency of spatial distributions at the global level and analyzing the correspondence of feature clusters at the local level. Global similarity analysis employs an improved Wasserstein distance metric to measure the difference between the two distributions; this distance calculation is entropy regularized to enhance robustness to noise. Local similarity detection uses kernel density estimation to identify the degree of matching in the density distributions of corresponding feature regions in the two spaces. The system introduces a dynamic adjustment mechanism to automatically balance the contribution weights of global and local similarity based on data characteristics.

[0042] The correlation factor calculation employs a dual verification strategy, considering both linear correlation and nonlinear dependency. In the linear correlation analysis phase, the system centers the propagation topology vector and performance representation vector to eliminate bias caused by mean shift. The calculation process uses an improved cosine similarity metric, which is scale-normalized to eliminate the influence of vector magnitude differences. Nonlinear dependency analysis, based on mutual information theory, calculates the statistical dependency between the joint distribution and marginal distributions using kernel density estimation. An adaptive bandwidth selection algorithm is designed to ensure the accuracy of density estimation.

[0043] The final evaluation of the synergistic coupling effect employs a fusion analysis model that integrates linear and nonlinear analysis results. The system defines the coupling degree evaluation function as follows: in: Indicates the degree of cooperative coupling. and To adjust the parameters, This represents the improved cosine similarity function. and Let represent the topology vector and performance vector of the i-th dimension, respectively. This represents the mutual information between vectors V and W, where n is the vector dimension. The adjustment parameter in this function is dynamically adjusted based on the data type, increasing when a strong linear relationship is detected. The value increases when the nonlinear characteristic is significant. Weights.

[0044] The vectorization stage employs a deep feature extraction network. The signal propagation topology is input to a graph convolutional network for processing. This network has three convolutional layers, each containing node feature transformation and neighborhood information aggregation operations. A graph pooling layer is placed at the network's end to convert node-level representations into graph-level vector representations. Semantic performance representations are input to a multilayer perceptron, where feature refinement and dimensionality adjustment are achieved through nonlinear transformations. The output vectors from both processing channels are L2 normalized to ensure comparisons are performed on a uniform scale.

[0045] The similarity index calculation employs a multi-granularity analysis method. The feature similarity index is calculated using a sliding window technique, with the window size adaptively adjusted based on data characteristics. Within each window, the system calculates the KL divergence of the local feature distribution, and then obtains the overall similarity score through an aggregation function. The representation similarity index calculation uses a bidirectional matching strategy: first, the matching degree is calculated based on the topological vector, then a reverse calculation is performed based on the performance vector, and finally, the average of the two results is taken as the final index.

[0046] A quality monitoring mechanism is implemented during the correlation factor calculation process. The system monitors the sparsity and distribution characteristics of the input vector in real time, triggering a feature reconstruction process when an abnormal distribution is detected. Intermediate results generated during the calculation are stored in a cache for subsequent verification. The system establishes a result reliability evaluation model, assigning a confidence score to the values ​​generated at each calculation step, and finally outputting a reliability index for the correlation factor. The reliability evaluation model is constructed to assess the numerical reliability in collaborative coupling analysis. The model is based on a Bayesian inference framework, and the training data comes from error records of historical calculation steps. The confidence score for each calculation step is calculated using posterior probability, with the following formula: Where C is the confidence score, P(correct) is the prior correct probability, P(data|correct) is the likelihood function (based on the Gaussian distribution assumption), and P(data) is the evidence probability. The model input consists of the statistical characteristics of the calculated numerical values ​​(such as variance and bias), and the output is a score in the range [0,1]. The reliability index is ultimately output as metadata for the association factor; for example, a review process is triggered when the score falls below a threshold. The model is trained through cross-validation to ensure its robustness across various meeting scenarios.

[0047] The application phase of the synergistic coupling effect employs a hierarchical response strategy. Based on the numerical range of coupling degree, the system classifies the results into three levels: strong coupling, moderate coupling, and weak coupling. In the case of strong coupling, the current parameter settings are directly used; in the case of moderate coupling, a parameter fine-tuning mechanism is initiated; and in the case of weak coupling, a system self-check process is triggered to re-evaluate the quality of the input data. The system establishes a feedback learning mechanism, recording the operational parameters and result data of each coupling analysis, and continuously optimizing the analysis model through incremental learning.

[0048] The entire implementation process incorporates multiple verification nodes. During the data input phase, vector dimensions and numerical ranges are verified; during the intermediate computation phase, numerical stability and convergence are monitored; and during the result output phase, logical consistency is checked. The system employs a fault-tolerant design, automatically switching to a backup computation path when an anomaly is detected. All critical operations generate audit logs, recording decision points and parameter states during processing, supporting result traceability and process reproducibility. The implementation process emphasizes smooth transitions between stages; the output of the previous stage, after format conversion and caching, is seamlessly input into the next processing module, ensuring efficient data flow.

[0049] Example 4: Taking the minutes of a product decision-making meeting as an example, the system first receives entity identification data, which contains entity and attribute information extracted from the original text. The data cleaning module initiates multi-level filtering: the first layer removes obvious misidentifications, such as entities that mistakenly label "Q2 quarter" as a product model; the second layer merges duplicate entities, such as unifying scattered identifications of "marketing department" and "marketing department" into a standard expression; the third layer supplements missing attributes, such as supplementing the person in charge information for "project A". The cleaned entity data forms structured records, each record containing the entity name, type, location of occurrence, and contextual characteristics.

[0050] The temporal feature analysis module processes the cleaned entity data and generates a temporal distribution map of entity appearances. The system divides the meeting process into several logical segments and records the frequency and duration of each entity's appearance in each segment. Taking the "budget adjustment" entity as an example, the system records that it appears twice in the opening segment, eight times during the discussion phase, and three times during the decision-making phase. Frequency domain analysis uses a sliding window technique to calculate the distribution density of entities within fixed time intervals. The system identifies that key entities often exhibit specific frequency domain characteristics; for example, core decision-making entities show a bimodal distribution in the frequency domain, corresponding to the proposal and confirmation phases, respectively.

[0051] The generation of entity characteristic descriptors is combined with time-frequency analysis results. The system creates a feature vector for each entity, containing three dimensions: time-domain activity, frequency-domain concentration, and contextual diversity. The activity index reflects the sustained intensity of an entity's participation in discussions, the concentration measures the degree of clustering of an entity's appearance over time, and the diversity describes the contextual changes in an entity's appearance. These feature vectors constitute a quantitative expression of the entity detection characteristics, based on which the system establishes an entity importance ranking model. See Table 1 for characteristic description data of some entities.

[0052] Table 1: Characteristic description data of some entities.

[0053] Thematic classification data processing employs a hierarchical analysis approach. The system first establishes a thematic framework, dividing meeting content into three levels: core issues, supporting arguments, and implementation details. The classification performance evaluation module analyzes the discriminative characteristics of each thematic category, including dimensions such as vocabulary distribution, sentence patterns, and contextual relevance. The error analysis component identifies common classification biases, such as misclassifying implementation details as core issues or grouping cross-thematic discussions into a single category. The system establishes an error pattern library, recording the conditions under which various errors occur and methods for correction.

[0054] The extraction of topic relevance parameters is based on the classification quality assessment results. The system identifies four key parameters: category discrimination reflects the clarity of topic boundaries; topic coverage measures the scope of the classification system's coverage of the meeting content; semantic consistency assesses the internal coherence of similar topics; and decision relevance measures the strength of the association between the topic and the final decision. Each parameter is assigned a dynamic weight, which is automatically adjusted according to the meeting type and agenda structure. In the product decision-making meeting case, the decision relevance parameter receives a higher weight, while the semantic consistency weight is increased in regular progress meetings.

[0055] The topic score is calculated using a multi-level weighted method. The system defines scoring criteria for each topic's associated parameters; for example, the category discrimination score is based on the accuracy metric of the confusion matrix, and the topic coverage score is based on the proportion of unclassified content. Parameter contribution allocation considers the characteristics of the meeting scenario and is dynamically generated through a pre-trained weighted prediction model. The system implements a hierarchical scoring strategy, first calculating the weighted scores of parameters within each topic, and then aggregating them to obtain the overall topic consistency score. A smoothing process is incorporated into the scoring process to avoid excessive impact of local fluctuations on the overall result.

[0056] The generation of topic consistency equivalent values ​​undergoes a standardized transformation. The system establishes a score-equivalence mapping function to convert the original scores into standardized equivalent values. This function uses an S-curve to adjust the distribution, allowing the equivalent values ​​to intuitively reflect the relative level of topic consistency. A confidence interval is included with the output of the equivalent values, indicating the range of reliability of the assessment results. The system records all intermediate results during the equivalent value calculation process, forming a complete decision trajectory to support subsequent review and adjustments.

[0057] The implementation process emphasizes the collaborative operation of each stage. The output of the entity detection characteristic analysis module serves as supplementary information for topic classification, helping to identify the topic distribution of key entities. The topic consistency assessment results are fed back to the entity importance model to optimize the accuracy of entity ranking. The system incorporates a cross-validation mechanism to compare the consistency between the time-frequency analysis results and the topic classification data, identifying potential processing biases. The entire process adopts an incremental processing mode, with the output of each module being immediately passed to the next stage, while retaining a copy of the original data for retrospective analysis.

[0058] Quality control is implemented throughout the entire process. During the data input phase, the format integrity and content rationality of the entity recognition results are verified; during the feature extraction phase, the numerical stability of calculated indicators is monitored; and during the result generation phase, logical consistency and contextual relevance are checked. The system establishes an anomaly handling mechanism, automatically triggering a review process or switching to a backup algorithm when data anomalies or calculation deviations are detected. All processing steps generate detailed operation logs, recording key parameters and decision-making basis, forming a complete audit trail. The final output entity detection characteristic description and topic consistency equivalent value are accompanied by a quality label to guide trust assessment in subsequent applications.

[0059] Example 5: Compatibility Assessment First, storage area matching is performed. The system parses the partitioning rules of the archiving system, including physical storage capacity limits, file format constraints, and metadata specifications. For the target storage area, the system measures its maximum file size support capability, attribute field definition range, and access interface parameters. Simultaneously, the structural characteristics of the minutes prototype are extracted, and the number of logical storage units occupied by the prototype and the complexity of metadata fields are calculated. Size compatibility is calculated using a multi-dimensional matching algorithm, which comprehensively considers physical storage occupancy, index building efficiency, and retrieval response time. The core formula of the algorithm is as follows: in: Indicates size fit. Dimensional metrics representing the prototype of the minutes (including composite parameters such as file size and structural complexity). It is the maximum dimension threshold supported by the storage area. This is the metadata field set for the prototype of the minutes. It is the metadata model supported by the system. and This is a dynamically adjustable coefficient. The formula balances the dual impacts of physical storage limitations and logical structure compatibility.

[0060] Access compatibility assessment focuses on the alignment of permissions and security mechanisms. The system establishes a role-permission mapping model, mapping the information sensitivity tags in the minutes prototype to the access control levels of the archiving system. The compatibility factor analysis module identifies three key sets of factors: content security level corresponding to access permission level, update frequency affecting locking mechanism settings, and related retrieval needs determining index depth. The system designs a matching degree algorithm for each set of factors; for example, permission matching degree is calculated through policy compliance checks, and retrieval matching degree is evaluated based on query scenario coverage. The weighted fusion of factor compatibility results in the final access compatibility score. .

[0061] The compatibility and adaptation process employs a two-dimensional coupling model. System establishment... and The interaction matrix is ​​used to detect the impact of storage constraints on access efficiency. Compression optimization is automatically initiated when size adaptation is insufficient, and security encapsulation strategies are adjusted when access adaptation is inadequate. The synthesized result outputs a quantified adaptation compatibility score, along with a list of improvement suggestions, identifying key constraints affecting the score.

[0062] The archiving process begins with the parsing of user interaction signals. The system monitors three typical interactions: direct archiving commands triggering standard processing flows; archiving configuration modification requests initiating parameter adjustments; and archiving issue feedback activating the correction mechanism. The signal parsing engine identifies the operation type and its associated structured minutes data range, precisely locating data blocks for local correction needs.

[0063] The dynamic correction of structured data employs an incremental update strategy. The system establishes a version snapshot mechanism, saving the current valid version as a baseline. Correction operations are divided into two categories: content updates modifying entity attribute values, and structural updates adjusting relational topology. Each update generates an incremental change record, containing the old value, the new value, and the reason for the change. Conflict detection is performed before execution; when concurrent modifications are detected, they are handled according to the timestamp priority principle. The system maintains a change history graph, supporting retrospective reconstruction of any version.

[0064] The archiving execution phase employs a step-by-step commit strategy. First, a metadata index is written to ensure basic retrieval functions are immediately available; then, structured main data is stored; finally, relational links and attachment associations are added. The system monitors the storage node status in real time, automatically switching replica nodes in case of anomalies. After archiving is complete, an integrity verification procedure is activated, ensuring data is stored without loss through digest comparison.

[0065] The feedback data collection design incorporates a multi-dimensional monitoring system. Performance metrics include query response time and concurrent access throughput; quality metrics involve search result relevance scores; and functional metrics record API call success rates. The system includes a dedicated semantic search feedback channel to capture user feedback on the usability of the structured data. Data collection frequency is dynamically adjusted based on system load, automatically reducing the sampling rate during peak periods.

[0066] Data feedback-driven model updates employ a dual-channel mechanism. The text feature model update channel receives semantic retrieval feedback, adjusting keyword weight allocation and relation extraction rules; the semantic structure model update channel analyzes query path patterns and optimizes the signal propagation topology generation strategy. Model updates undergo A / B testing; new models are only deployed to the production environment after successful validation in a shadow environment. The system retains historical model versions and automatically rolls back to a stable version when performance degradation is detected.

[0067] The implementation process emphasizes operational atomicity and transaction consistency. Each processing step defines a clearly defined rollback point, and critical operations employ a two-phase commit protocol. The system establishes a resource isolation mechanism to ensure that high-priority archiving tasks are not disturbed. The entire process is monitored for health status; abnormal events trigger alarms and generate diagnostic reports, enabling operations personnel to quickly pinpoint the root cause of problems.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for structured extraction and archiving of meeting minutes, characterized in that, include: Obtain meeting record data, and based on the meeting record data, identify the key elements of the meeting content; Based on the aforementioned meeting content elements, analyze the characteristics of the meeting scenario; Collect existing ambiguous information from meeting minutes, and formulate structured extraction parameters based on the meeting scenario characteristics and the existing ambiguous information. The text feature data and semantic structure data of the scheduling meeting minutes are analyzed to determine the semantic performance representation corresponding to the text feature data. Based on the semantic structure data, determine the signal propagation topology of the meeting minutes; Based on the aforementioned signal propagation topology, structured summary data is generated; Based on the structured minutes data, archiving processing is performed.

2. The method for structured extraction and archiving of meeting minutes as described in claim 1, characterized in that, The generation of structured summary data based on the signal propagation topology includes: The synergistic coupling effect between the signal propagation topology and the semantic performance representation is evaluated, and the parameter priority sequence of the structured extraction parameters is determined based on the synergistic coupling effect. The core entity type and alternative entity type of the meeting minutes are queried. The meeting minutes are simulated and constructed by combining the core entity type and the alternative entity type to obtain the minutes prototype. Collect entity recognition data and topic classification data of the prototype minutes, and analyze the entity detection characteristics of the prototype minutes based on the entity recognition data; Based on the topic classification data, the topic consistency equivalent value of the minute prototype is calculated, and the semantic stability of the minute prototype is evaluated based on the topic consistency equivalent value. Determine the storage location requirements and access constraints of the meeting minutes in the archiving system, collect the adaptability parameters of the minutes prototype, and analyze the compatibility of the minutes prototype in the archiving system by combining the storage location requirements, the access constraints, and the adaptability parameters. Combining the entity detection characteristics, semantic stability, and adaptation compatibility, the optimal entity type is selected from the core entity types and the candidate entity types. Based on the structured extraction parameters, the parameter priority sequence, and the optimal entity type, structured summary data is generated.

3. The method for structured extraction and archiving of meeting minutes as described in claim 1, characterized in that, The step of combining the meeting scenario features and the existing ambiguous information to formulate structured extraction parameters includes: Feature extraction is performed on the meeting scene features to obtain scene feature factors; The existing ambiguous information is classified to obtain a set of ambiguous categories; Analyze the correlation between the scene feature factors and the ambiguous category set to obtain the correlation mapping matrix; Based on the aforementioned association mapping matrix, the key scenario features and key ambiguity categories of the meeting minutes are determined; Based on the key scenario features and key ambiguity categories, the extraction constraints of meeting minutes are analyzed. Based on the extraction constraints, the structured extraction parameters are formulated.

4. The method for structured extraction and archiving of meeting minutes as described in claim 1, characterized in that, The analysis of the semantic performance representation corresponding to the text feature data includes: The text feature data is standardized to obtain standardized text data; Extract the semantic attributes corresponding to the standardized text data, filter the semantic attributes, and obtain the key semantic attributes; Calculate the semantic performance index corresponding to the key semantic attributes, and generate the semantic performance representation corresponding to the text feature data based on the semantic performance index.

5. The method for structured extraction and archiving of meeting minutes as described in claim 1, characterized in that, The step of determining the signal propagation topology of the meeting minutes based on the semantic structure data includes: The semantic structure data is preprocessed to obtain the target semantic structure data; Extract the set of structural parameters for the meeting minutes from the target semantic structure data; The structural parameter set is subjected to semantic attribute association processing to obtain the attribute association parameter set; Based on the attribute-related parameter set, a numerical model of signal propagation corresponding to the meeting minutes is constructed; The signal propagation numerical model is simulated and processed to obtain dynamic data of signal propagation. The signal propagation dynamic data is subjected to topological abstraction processing to generate the signal propagation topology of the meeting minutes.

6. The method for structured extraction and archiving of meeting minutes as described in claim 2, characterized in that, The evaluation of the synergistic coupling effect between the signal propagation topology and the semantic performance representation includes: Extract the signal propagation features corresponding to the signal propagation topology, and perform dimensionality reduction processing on the signal propagation features to obtain dimensionality-reduced signal propagation features; Calculate the feature similarity index between the reduced-dimensional signal propagation features, and calculate the representation similarity index between the semantic performance representations; Calculate the correlation factor between the signal propagation topology and the semantic performance representation; By combining the correlation factor, the feature similarity index, and the representation similarity index, the degree of synergistic coupling between the signal propagation topology and the semantic performance representation is calculated. Based on the aforementioned cooperative coupling degree, the cooperative coupling effect between the signal propagation topology and the semantic performance representation is evaluated.

7. The method for structured extraction and archiving of meeting minutes as described in claim 6, characterized in that, The calculation of the correlation factor between the signal propagation topology and the semantic performance representation includes: The signal propagation topology and the semantic performance representation are vectorized respectively to obtain the propagation topology vector and the performance representation vector. Calculate the vector cosine between the propagation topology vector and the performance characterization vector; Calculate the vector mutual information between the propagation topology vector and the performance representation vector; By combining the vector cosine value and the vector mutual information, the correlation factor between the signal propagation topology and the semantic performance representation is calculated.

8. The method for structured extraction and archiving of meeting minutes as described in claim 2, characterized in that, The step of analyzing the entity detection characteristics of the prototype minutes based on the entity recognition data includes: The entity recognition data is cleaned to obtain cleaned entity recognition data; Analyze and extract the time-domain and frequency-domain features corresponding to the cleaned entity identification data; Based on the time-domain features and the frequency-domain features, an entity characteristic descriptor corresponding to the minutes prototype is generated; Based on the entity characteristic descriptor, the entity detection characteristics of the transcript prototype are analyzed.

9. A method for structured extraction and archiving of meeting minutes as described in claim 2, characterized in that, The step of calculating the thematic consistency equivalent value of the minutes prototype based on the thematic classification data includes: Analyze the classification performance parameters corresponding to the topic classification data, and query the error patterns and error mechanisms corresponding to the meeting minutes; Variable analysis was performed on the error modes and error mechanisms to obtain the error variables corresponding to the meeting minutes; Based on the aforementioned error variables, thematic evaluation indicators corresponding to the meeting minutes are formulated. Based on the thematic evaluation indicators, thematic correlation parameters of the meeting minutes are extracted from the classification effectiveness parameters. Based on the topic classification data, the topic score value corresponding to the topic association parameter is calculated; Assign the parameter contribution degree corresponding to the topic association parameter, and calculate the topic consistency equivalent value of the minutes prototype by combining the topic score value and the parameter contribution degree.

10. A method for structured extraction and archiving of meeting minutes as described in claim 2, characterized in that, The analysis of the compatibility of the minutes prototype in the archiving system, combining the storage location requirements, access constraints, and adaptability parameters, includes: Based on the storage location requirements, the storage area of ​​the prototype minutes in the archiving system is determined, and the area size parameters and minute size parameters corresponding to the storage area and the prototype minutes are measured. Based on the region size parameters and the minutes size parameters, the size fit of the minutes prototype in the archiving system is calculated; Analyze the adaptation factors corresponding to the adaptation parameters, and calculate the factor adaptation degree corresponding to the adaptation factors based on the access condition constraints and the adaptation parameters. Based on the aforementioned factor suitability, the access suitability of the minutes prototype in the archiving system is calculated. By combining the size adaptation and the access adaptation, the compatibility of the minutes prototype in the archiving system is analyzed. The archiving process based on the structured minutes data includes: Receive user interaction signals and determine the archiving operation type based on the user interaction signals; Based on the archiving operation type, the structured minutes data is dynamically corrected; Execute dynamically revised structured minutes data and collect archived feedback data; Based on the archived feedback data, update the text feature data and semantic structure data of the meeting minutes.

Images