Conference summary generation method

Abstract

The embodiment of the invention provides a conference summary generation method which can be applied to the technical field of artificial intelligence, in the method, speech content of participants in the conference process is collected, whether the speech content meets an event triggering mechanism or not is detected based on a dynamic knowledge graph constructed in the conference process, and if yes, the conference summary is generated. The event triggering mechanism represents the state change of the dynamic knowledge graph; if the event triggering mechanism is met, entity node and / or edge connection updating is carried out on the dynamic knowledge graph, and a conference summary is generated based on the dynamic knowledge graph at the end of the conference, so that the speaking content of each participant is compared with the constructed dynamic knowledge graph; according to the method, whether the new speaking content has the incidence relation with the previous speaking content is determined, the dynamic change relation of any viewpoint in the conference process from proposing to discussion of multiple participants to final decision is captured, the viewpoint evolution process is obtained, and therefore the accuracy of conference summary is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for generating meeting minutes. Background Technology

[0002] AI meeting assistants use natural language processing technology to understand the semantic information of meeting content and generate fluent and coherent text output. With the rapid development of large-scale language models (LLMs) technology, the new generation of AI meeting assistants has made significant improvements in text understanding and content generation.

[0003] In related technologies, AI meeting assistants perform batch processing and analysis of meeting audio transcripts, extracting key information such as summaries and tasks to be performed from the transcripts using large language model technology, and generating static meeting minutes for users. However, this approach treats the meeting content merely as an isolated and linear text (i.e., the meeting audio transcript), failing to capture and utilize the dynamic and changing contextual information within the meeting. For example, in unstructured meetings, it cannot capture the evolution of viewpoints, the influence of decisions on speakers, etc., resulting in low accuracy of the generated meeting minutes. Summary of the Invention

[0004] This invention provides a method for generating meeting minutes to improve the accuracy of meeting minutes.

[0005] On the one hand, embodiments of this application provide a method for generating meeting minutes, the method comprising: Collect the participants' remarks during the meeting; Based on the dynamic knowledge graph constructed during the meeting, it is detected whether the speech content satisfies the event triggering mechanism; the event triggering mechanism represents the state change of the dynamic knowledge graph. If the event triggering mechanism is met, the entity nodes and / or edge connections of the dynamic knowledge graph are updated. Meeting minutes are generated based on the dynamic knowledge graph at the end of the meeting.

[0006] Optionally, updating the entity nodes and / or edge connections of the dynamic knowledge graph if the event triggering mechanism is satisfied includes: Obtain first entity information from the speech content, and determine whether there is a first entity node corresponding to the first entity information in the dynamic knowledge graph; if not, create the first entity node in the dynamic knowledge graph. The semantic association degree between the first entity node and at least one second entity node in the dynamic knowledge graph is determined based on the attention mechanism, and the edge connection between the first entity node and the second entity node is updated based on the semantic association degree.

[0007] Optionally, creating the first entity node in the dynamic knowledge graph includes: If the content of the speech is related to the content of the speeches of other participants, an event node representing the first entity node is created; the content of the speeches of the other participants corresponds to the second entity node. Updating the edge connection between the first entity node and the second entity node based on the semantic relevance includes: Based on the semantic relevance of the speech content to the speeches content of other participants, an edge connection is established between the first entity node and the second entity node.

[0008] Optionally, the semantic relevance is characterized by one or more of the following: the state transition of the viewpoint node, the causal state between the viewpoint node and the decision node, and the directional state between the decision node and the task node; The semantic relevance conforms to one or more of the following constraints: temporal constraints, causal constraints, and role constraints.

[0009] Optionally, creating the first entity node in the dynamic knowledge graph includes: If the participant in the speech is speaking for the first time, then create a participant node corresponding to the speaker; If the content of the speech indicates a meeting topic, then a topic node is created; If the content of the statement is a viewpoint that is being presented for the first time, then a viewpoint node corresponding to the content of the statement is created.

[0010] Optionally, before generating the meeting minutes, the process further includes: Identify fuzzy nodes in the dynamic knowledge graph, where a fuzzy node is any entity node with ambiguous semantics; Calculate the correlation degree between the fuzzy node and any participant node, determine the candidate expert responsible for correcting the fuzzy node, and send a clarification request to the candidate expert.

[0011] Optionally, identifying fuzzy nodes in the dynamic knowledge graph includes: Based on the semantic clarity of any entity node other than the participant node in the dynamic knowledge graph, fuzzy nodes are identified. The calculation of the correlation degree between the fuzzy node and any participant node, and the determination of the candidate expert responsible for correcting the fuzzy node, includes: The influence of any participant node on the fuzzy node is calculated from multiple dimensions. Based on the influence of multiple dimensions and the weight values ​​of each dimension, the correlation degree between the fuzzy node and any participant node is obtained; The participant node with the highest correlation degree exceeding the preset threshold is identified as a candidate expert for correcting the fuzzy node.

[0012] Optionally, the multiple dimensions include the dynamic weight of the participant's role, the quality of the participant's speech, the participant's influence on the conference topic, the participant's authority in the field of the conference topic, the participant's trust in other participants, and the participant's innovative contribution to the conference topic.

[0013] Optionally, after determining the candidate experts for correcting the fuzzy nodes, the method further includes: Calculate the belonging distance between any entity node in the dynamic knowledge graph and the participant node identified as a candidate expert; the belonging distance is determined based on at least one of the following: the topological distance between the entity node and the participant node in the dynamic knowledge graph, the semantic distance between multiple speeches of the entity node and the participant node, the temporal distance between the time of generation of the entity node and the time of generation of the participant node, and the trust distance between the entity node and the participant node. At least one entity node whose distance does not exceed the first threshold is assigned to the associated region of the candidate expert's participant node.

[0014] Optionally, generating meeting minutes based on the dynamic knowledge graph at the end of the meeting includes: The verification agent checks whether there are logical conflicts in the edge connections of each entity node in the dynamic knowledge graph, and obtains the verification results. The evolution path of each viewpoint node is extracted by a summary-type agent, and the participant node corresponding to each task node in the dynamic knowledge graph is determined by an action-item agent. Based on the verification results, the evolution path, and the participant nodes corresponding to each task node, the dynamic knowledge graph is corrected; based on the corrected dynamic knowledge graph, meeting minutes are generated.

[0015] Optionally, after generating the meeting minutes, the method further includes: Collect user ratings for the quality of the meeting minutes, and obtain ratings for the roles of different types of agents in generating the meeting minutes; A reward score is awarded based on the time the meeting minutes were generated and the adoption of clarification requests from each candidate expert. The meeting minutes system is optimized based on the quality score, the effectiveness score, and the reward score.

[0016] On one hand, embodiments of this application provide a meeting minutes generation apparatus, the apparatus comprising: The data collection module is used to collect the content of participants' speeches during the meeting; The detection module is used to detect whether the speech content satisfies the event triggering mechanism based on the dynamic knowledge graph constructed during the meeting; the event triggering mechanism represents the state change of the dynamic knowledge graph. An update module is used to update entity nodes and / or edge connections of the dynamic knowledge graph if the event triggering mechanism is met. The generation module is used to generate meeting minutes based on the dynamic knowledge graph at the end of the meeting.

[0017] Optionally, the update module is specifically used for: Obtain first entity information from the speech content, and determine whether there is a first entity node corresponding to the first entity information in the dynamic knowledge graph; if not, create the first entity node in the dynamic knowledge graph. The semantic association degree between the first entity node and at least one second entity node in the dynamic knowledge graph is determined based on the attention mechanism, and the edge connection between the first entity node and the second entity node is updated based on the semantic association degree.

[0018] Optionally, the update module is specifically used for: If the content of the speech is related to the content of the speeches of other participants, an event node representing the first entity node is created; the content of the speeches of the other participants corresponds to the second entity node. Updating the edge connection between the first entity node and the second entity node based on the semantic relevance includes: Based on the semantic relevance of the speech content to the speeches content of other participants, an edge connection is established between the first entity node and the second entity node.

[0019] Optionally, the semantic relevance is characterized by one or more of the following: the state transition of the viewpoint node, the causal state between the viewpoint node and the decision node, and the directional state between the decision node and the task node; The semantic relevance conforms to one or more of the following constraints: temporal constraints, causal constraints, and role constraints.

[0020] Optionally, the update module is specifically used for: If the participant in the speech is speaking for the first time, then create a participant node corresponding to the speaker; If the content of the speech indicates a meeting topic, then a topic node is created; If the content of the statement is a viewpoint that is being presented for the first time, then a viewpoint node corresponding to the content of the statement is created.

[0021] Optionally, the generation module is further configured to: Identify fuzzy nodes in the dynamic knowledge graph, where a fuzzy node is any entity node with ambiguous semantics; Calculate the correlation degree between the fuzzy node and any participant node, determine the candidate expert responsible for correcting the fuzzy node, and send a clarification request to the candidate expert.

[0022] Optionally, the generation module is specifically used for: Based on the semantic clarity of any entity node other than the participant node in the dynamic knowledge graph, fuzzy nodes are identified. The calculation of the correlation degree between the fuzzy node and any participant node, and the determination of the candidate expert responsible for correcting the fuzzy node, includes: The influence of any participant node on the fuzzy node is calculated from multiple dimensions. Based on the influence of multiple dimensions and the weight values ​​of each dimension, the correlation degree between the fuzzy node and any participant node is obtained; The participant node with the highest correlation degree exceeding the preset threshold is identified as a candidate expert for correcting the fuzzy node.

[0023] Optionally, the multiple dimensions include the dynamic weight of the participant's role, the quality of the participant's speech, the participant's influence on the conference topic, the participant's authority in the field of the conference topic, the participant's trust in other participants, and the participant's innovative contribution to the conference topic.

[0024] Optionally, the generation module is further configured to: Calculate the belonging distance between any entity node in the dynamic knowledge graph and the participant node identified as a candidate expert; the belonging distance is determined based on at least one of the following: the topological distance between the entity node and the participant node in the dynamic knowledge graph, the semantic distance between multiple speeches of the entity node and the participant node, the temporal distance between the time of generation of the entity node and the time of generation of the participant node, and the trust distance between the entity node and the participant node. At least one entity node whose distance does not exceed the first threshold is assigned to the associated region of the candidate expert's participant node.

[0025] Optionally, the generation module is specifically used for: The verification agent checks whether there are logical conflicts in the edge connections of each entity node in the dynamic knowledge graph, and obtains the verification results. The evolution path of each viewpoint node is extracted by a summary-type agent, and the participant node corresponding to each task node in the dynamic knowledge graph is determined by an action-item agent. Based on the verification results, the evolution path, and the participant nodes corresponding to each task node, the dynamic knowledge graph is corrected; based on the corrected dynamic knowledge graph, meeting minutes are generated.

[0026] Optionally, the generation module is further configured to: Collect user ratings for the quality of the meeting minutes, and obtain ratings for the roles of different types of agents in generating the meeting minutes; A reward score is awarded based on the time the meeting minutes were generated and the adoption of clarification requests from each candidate expert. The meeting minutes system is optimized based on the quality score, the effectiveness score, and the reward score.

[0027] On one hand, embodiments of this application provide a computer device, including: Memory, used to store program instructions; The processor is used to call the program instructions stored in the memory and execute the steps of the above-mentioned meeting minutes generation method according to the obtained program.

[0028] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer to perform the steps of the meeting minutes generation method described above.

[0029] On one hand, embodiments of this application provide a computer program product, including a computer program stored on a computer-readable storage medium, the computer program including program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described meeting minutes generation method.

[0030] In this embodiment, the participants' speeches during the meeting are collected. Based on the dynamic knowledge graph already constructed during the meeting, it is detected whether the speeches meet the event triggering mechanism, which represents the state change of the dynamic knowledge graph. If the event triggering mechanism is met, the entity nodes and / or edge connections of the dynamic knowledge graph are updated. Based on the dynamic knowledge graph at the end of the meeting, meeting minutes are generated. This allows for the comparison of each participant's speeches with the constructed dynamic knowledge graph to determine whether new speeches are related to previous speeches. It captures the dynamic changes of any viewpoint from its introduction to discussion among multiple participants and finally to the final decision during the meeting, thereby obtaining the evolution process of the viewpoint and improving the accuracy of the meeting minutes. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A schematic diagram of a system architecture provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for generating meeting minutes provided in an embodiment of this application; Figure 3 A flowchart illustrating a method for generating meeting minutes provided in an embodiment of this application; Figure 4 A flowchart illustrating a method for generating meeting minutes provided in an embodiment of this application; Figure 5 A flowchart illustrating a method for generating meeting minutes provided in an embodiment of this application; Figure 6 A flowchart illustrating a method for generating meeting minutes provided in an embodiment of this application; Figure 7 A schematic diagram of a meeting minutes generation device provided in this application embodiment; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0034] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0035] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0036] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0037] The term "module" refers to any known or subsequently developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware and / or software code that is capable of performing the functions associated with that element.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0039] The following is an explanation of the technical terms that appear in the implementation of this application: Large Language Model (LLM): This refers to an artificial intelligence model based on deep learning technology, designed to understand and generate natural language text. Trained on large amounts of text data, it can master complex language patterns, syntax, semantics, and contextual relationships, thereby generating coherent and logical text.

[0040] Multi-Agent System (MAS): A computing system consisting of multiple intelligent agents capable of autonomous decision-making, interaction, and cooperation, designed to solve complex problems that are difficult for a single intelligent agent to handle.

[0041] Dynamic Context Knowledge Graph (DCKG): A graph-structured data that is built and evolved in real time during collaborative tasks (such as meetings). It uses elements such as participants, topics, viewpoints, and decisions as nodes, and the dynamic relationships between them as edges, thereby providing a structured representation of the context of the collaborative process.

[0042] Multi-Agent Reinforcement Learning (MARL): A machine learning paradigm in which multiple agents learn optimal cooperative or competitive strategies by interacting with the environment and other agents and based on reward signals.

[0043] Network Voronoi Diagram: An algorithm that applies the concept of Vironoi diagram partitioning from computational geometry to graph structures (such as knowledge graphs). It divides the graph into regions based on the "distance" between nodes, with each region associated with a specific "center" node.

[0044] The following is a brief introduction to the system architecture diagram applicable to the technical solutions of the embodiments of this application. It should be noted that the process described below is only for illustrating the embodiments of this application and is not intended to limit the scope of the application.

[0045] refer to Figure 1 This is a system architecture diagram applicable to the embodiments of this application. The system architecture includes at least terminal device 101 and server 102. The number of terminal devices 101 can be one or more, and the number of servers 102 can also be one or more. This application does not specifically limit the number of terminal devices 101 and servers 102.

[0046] Terminal device 101 is pre-installed with an application or plugin that includes a meeting minutes system. This application can be a client application, a web application, a mini-program application, etc. Terminal device 101 can be a smartphone, tablet, laptop, desktop computer, smart home appliance, smart voice interaction device, smart in-vehicle device, etc., but is not limited to these.

[0047] Server 102 is the backend server of the application. Server 102 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, but it is not limited to these.

[0048] It should be noted that the method in this embodiment can be executed by the terminal device 101 or the server 102 alone, or by the terminal device 101 and the server 102 together.

[0049] In this embodiment, the terminal device 101 and the server 102 can communicate directly or indirectly through one or more networks. The network can be a wired network or a wireless network; for example, the wireless network can be a mobile cellular network or a Wireless-Fidelity (WIFI) network, or other possible networks. This embodiment does not limit the types of networks used.

[0050] The following is based on Figure 1The system architecture diagram shown illustrates the flowchart of a meeting minutes generation method provided in this application embodiment. This method's flowchart can be... Figure 1 The execution can be performed by the terminal device 101 shown, or by the server 102, or by interaction between the terminal device 101 and the server 102, as shown. Figure 2 As shown, it includes the following steps: Step 201: Collect the participants' remarks during the meeting.

[0051] Step 202: Based on the dynamic knowledge graph already constructed during the meeting, check whether the content of the speech meets the event triggering mechanism.

[0052] Specifically, the event triggering mechanism represents the state changes of the dynamic knowledge graph.

[0053] Furthermore, the triggering event could be identifying whether a participant's statement reflects a judgmental attitude towards the content in the constructed dynamic knowledge graph. The judgmental attitude could be agreement, partial agreement with suggestions for modification, or disagreement.

[0054] Step 203: If the event triggering mechanism is satisfied, then update the entity nodes and / or edge connections of the dynamic knowledge graph.

[0055] In some embodiments, if the participant speaking is speaking for the first time, a participant node corresponding to the speaker is created; if the speaking content indicates a meeting topic, a topic node is created; if the speaking content is a viewpoint presented for the first time, a viewpoint node corresponding to the speaking content is created.

[0056] Specifically, within a certain period after the start of the meeting, for any participant, if the participant speaks for the first time, a participant node corresponding to the participant is created, and it is identified whether the participant's speech content is related to the previous speech content of other participants. If it is not related, it is considered to be the first point of view raised by the participant, and a point of view node corresponding to the speech content is created.

[0057] In some embodiments, first entity information is obtained from the speech content, and it is determined whether a first entity node corresponding to the first entity information exists in the dynamic knowledge graph; if it does not exist, the first entity node is created in the dynamic knowledge graph; the semantic association degree between the first entity node and at least one second entity node in the dynamic knowledge graph is determined based on the attention mechanism, and the edge connection between the first entity node and the second entity node is updated based on the semantic association degree.

[0058] In some embodiments, if the content of a speech is related to the content of speeches of other participants, an event node representing the first entity node is created; the content of speeches of other participants corresponds to the second entity node; and an edge connection is established between the first entity node and the second entity node based on the semantic relevance of the content of the speech to the content of speeches of other participants.

[0059] Specifically, for any participant's speech during the meeting, the Dynamic Context-Aware Influence Network (D-CAIN) model in the meeting minutes system constructs a dynamic knowledge graph in real time. The D-CAIN model is a collaborative creation engine that integrates three major technologies: dynamic graph construction, influence network analysis, and real-time context awareness. It can understand, evolve dynamically, and make intelligent decisions in real time. It can identify changes in the influence of nodes in the network and understand how these changes are affected by the environmental context. In this embodiment, it mainly includes a dynamic knowledge graph construction module, an influence partitioning module, and an adaptive learning module.

[0060] In the dynamic knowledge graph construction module, the semantic understanding layer uses an improved BERT-based model to perform deep semantic analysis on the participants' statements, obtain the first entity information, and determine whether the first entity information is the participant's first opinion statement on the conference topic. If so, an opinion node is created in the dynamic knowledge graph. If the first entity information has a judgment attitude towards the opinion node or event node in the constructed dynamic knowledge graph, an event node or decision node is created in the dynamic knowledge graph. In the event detection layer, the event type is identified based on a hybrid architecture of sequence labeling and graph neural network. The event types include decision events, conflict events, and allocation events, etc. That is, the first entity information is monitored by sequence labeling, and the constructed dynamic knowledge graph is monitored by graph neural network. Then, in the relation reasoning layer, the semantic correlation between nodes in the constructed knowledge graph other than participant nodes and topic nodes is reasoned based on the attention mechanism. If the semantic correlation is greater than the semantic correlation threshold, the dynamic update layer creates a first entity node for the first entity information through a special graph structure incremental update algorithm. At this time, the first entity node can be either an event node or a decision node.

[0061] Furthermore, if the relational reasoning layer determines that the first entity information has a judgment attitude of disagreement or partial agreement with its previous viewpoint nodes or event nodes, then an event node is created based on the first entity information; if it determines that the first entity information has a judgment attitude of agreement with its previous viewpoint nodes or event nodes, then a decision node is created based on the first entity information.

[0062] In this embodiment, semantic relevance is characterized by one or more of the following: state transition of viewpoint node, causal state between viewpoint node and decision node, and directional state between decision node and task node; semantic relevance conforms to one or more of the following: temporal constraint, causal constraint, and role constraint.

[0063] Specifically, if an event node and an opinion node are connected by an edge in a rebuttal relationship, the opinion node's state transition is: proposing an opinion – disagreeing with the opinion. Subsequently, if an event node and an opinion node are connected by an edge in a correction relationship, the opinion node's state transition is updated to: proposing an opinion – disagreeing with the opinion – correcting the opinion. Then, if a decision node agrees with the corrected opinion, the opinion node's state transition is updated to: proposing an opinion – disagreeing with the opinion – correcting the opinion – agreeing with the corrected opinion. The causal relationship between opinion nodes and decision nodes is that a participant proposed a relatively balanced corrective suggestion during the meeting, so the corrected opinion is accepted. After a decision node appears, based on the participants' subsequent discussion of task allocation for the meeting topic, at least one task node is generated. The task node represents the subsequent execution work of each participant after the meeting, and based on the participants' statements, each task node is associated with a participant node to determine the executor of the task corresponding to that task node.

[0064] For example, refer to Figure 3 Regarding the meeting topic: budget determination; Participant A speaks first, with statement a: "I suggest cutting the budget by 10%", thus creating participant node A and opinion node a; Participant B follows Participant A, with statement b: "I disagree, that won't be effective for promotion". The event detection layer detects that statement b contradicts statement a, so it dynamically updates participant node B and event node b, and updates the edge connection between event node b and event node a to a rebuttal relationship; Participant C's statement c: "Then how about cutting it by 5%?" At this point, after participant C finishes speaking, the event detection layer can detect that the speech content c is a suggestion event based on the speech content a. Then, the dynamic update layer establishes participant node C and event node c, and updates the edge connection between event node c and event node a to a correction relationship. Participant D's speech content d is: "This is acceptable". At this point, after participant D finishes speaking, the event detection layer can detect that the speech content d is a decision event based on the speech content c. Then, the dynamic update layer establishes participant node D and event node d, and updates the edge connection between event node d and event node c to a decision relationship.

[0065] In this embodiment, the D-CAIN model captures the relationship between the participants' speech content and the constructed knowledge graph. At the same time, it introduces a consistency check of temporal logic (i.e., judging the order of speech time). Compared with the traditional knowledge graph construction method of extracting triple facts from static audio transcripts using offline batch processing mode, this application ensures the logical correctness of the construction of dynamic knowledge graph through temporal constraints, causal constraints (i.e., the evaluation attitude of the speech content towards the constructed dynamic knowledge graph) and role constraints.

[0066] Step 204: Generate meeting minutes based on the dynamic knowledge graph at the end of the meeting.

[0067] In this embodiment, the participants' speeches during the meeting are collected. Based on the dynamic knowledge graph already constructed during the meeting, it is detected whether the speeches meet the event triggering mechanism, which represents the state change of the dynamic knowledge graph. If the event triggering mechanism is met, the entity nodes and / or edge connections of the dynamic knowledge graph are updated. Based on the dynamic knowledge graph at the end of the meeting, meeting minutes are generated. This allows for the comparison of each participant's speeches with the constructed dynamic knowledge graph to determine whether new speeches are related to previous speeches. It captures the dynamic changes of any viewpoint from its introduction to discussion among multiple participants and finally to the final decision during the meeting, thereby obtaining the evolution process of the viewpoint and improving the accuracy of the meeting minutes.

[0068] In some embodiments, fuzzy nodes in a dynamic knowledge graph are identified, where a fuzzy node is any entity node with ambiguous semantics; the correlation degree between the fuzzy node and any participant node is calculated, a candidate expert responsible for correcting the fuzzy node is determined, and a clarification request is sent to the candidate expert.

[0069] In some embodiments, fuzzy nodes are determined based on the semantic clarity of any entity node other than the participant nodes in the dynamic knowledge graph; the influence of any participant node on the fuzzy node is calculated from multiple dimensions; the correlation between the fuzzy node and any participant node is obtained based on the influence of multiple dimensions and the weight values ​​of each dimension; and the participant node with the highest correlation exceeding a preset threshold is determined as a candidate expert for correcting the fuzzy node.

[0070] Specifically, for each node generated, a confidence score is calculated for that node. Based on the referential words, fluency, and coherence of any participant's speech, the node confidence scores for opinion nodes, decision nodes, event nodes, and task nodes are calculated. If the node confidence score of any node in the dynamic knowledge graph is less than the uncertain baseline, then that node is considered a fuzzy node.

[0071] In the embodiments of this application, multiple dimensions include the dynamic weight of the participant's role, the quality of the participant's speech, the participant's influence on the conference topic, the participant's authority in the field of the conference topic, the trust level between the participant and other participants, and the participant's innovative contribution to the conference topic.

[0072] Specifically, the influence partitioning module of the D-CAIN model calculates the influence degree of any participant node on the fuzzy node using the following formula (1): (1) in, This represents the degree of influence of participant node p on fuzzy node q; This indicates the dynamic weight of a participant node's role within the meeting topic; Let represent the sum of the quality of multiple speeches by participant p, u represent the content of each speech, and U represent the set of speeches by participant p. This represents the analytical influence of participant node p on fuzzy nodes; This represents the authority of participant p in the domain of the fuzzy node q; This represents the level of trust between participant p and other participants in their collaboration; This represents the degree of innovation contribution of participant p to the conference topic corresponding to fuzzy node q.

[0073] Based on the above formula (1), candidate experts who can answer fuzzy nodes are determined from multiple dimensions. The communication agent of the meeting minutes system can directly initiate voice Q&A with the candidate experts as a clarification request during the meeting, or send a private message to the candidate experts through a dialog box as a clarification request.

[0074] In this embodiment, fuzzy nodes are determined by multi-dimensional calculations. Traditional methods rely on statistical analysis of single speaking frequencies to identify human experts. This embodiment can improve the accuracy of expert positioning and effectively solve the problem of asking the wrong person in information clarification.

[0075] In some embodiments, the belonging distance between any entity node in the dynamic knowledge graph and the participant node identified as a candidate expert is calculated; the belonging distance is determined based on at least one of the topological distance between the entity node and the participant node in the dynamic knowledge graph, the semantic distance between multiple speeches of the entity node and the participant node, the temporal distance between the entity node's generation time and the participant node's generation time, and the trust distance between the entity node and the participant node; at least one entity node whose belonging distance does not exceed a first threshold is assigned to the associated region of the candidate expert's participant node.

[0076] Furthermore, in order to improve the accuracy of the dynamic knowledge graph, the entity nodes in the knowledge graph, excluding fuzzy nodes, participant nodes, and topic nodes, are divided into associated regions. The participant nodes of the candidate experts selected as fuzzy nodes are taken as the center of each associated region. The distance between any entity node and the participant node is calculated. The specific calculation formula is shown in the following formula (2): (2) in, This represents the distance between participating node p and any entity node v. This represents the topological distance between participant node p and any entity node v in the dynamic knowledge graph; This represents the semantic distance (i.e., semantic similarity) between the multiple statements made by any entity node v and participant node p. This represents the time distance between the generation time of any entity node v and the generation time of participant node p; This represents the trust distance between any entity node v and participant node p; , , , This represents the weight value.

[0077] The following formula (3) is used to determine whether any entity node can fall within the associated region of a participant node:

[0078] in, This represents the associated region of the participant node pi, which was selected as a candidate expert. V represents any entity node, and V represents the set of all entity nodes. This represents the distance between participating node pi and any entity node v; This represents the first threshold.

[0079] When an entity node falls into the associated region of a participant node, if the subsequent entity node becomes abnormal or is updated, the entity node is re-corrected in the dynamic knowledge graph based on any entity node related to the participant node, or a clarification request for the entity node is initiated to the participant node (i.e., candidate expert) through a communicative agent.

[0080] refer to Figure 4 When constructing a dynamic knowledge graph, the execution process of each module in the entire D-CAIN model is as follows: The input layer receives the transcribed text and segments each participant's speech according to timestamps using a discourse segmentation unit. The graph construction module performs entity recognition, event monitoring, and relation extraction (i.e., edge connections between different entities), and updates the graph structure through an attention mechanism to achieve real-time evolution of the knowledge graph. The resulting knowledge graph structure includes participant nodes, topic nodes, opinion nodes, decision nodes, and task nodes. The initial knowledge graph is then divided into influence categories, and fuzzy node identification is performed. Based on the fuzzy nodes, an influence score is calculated for each participant node. The participant influence score is calculated by comprehensively considering the participant's weight, speech quality, and the number of times the speech content is cited. The dynamic knowledge graph is then divided into network Veno graphs (i.e., each entity node is divided into associated regions). Based on the participant influence scores, candidate experts are located, and finally, candidate experts are output.

[0081] Furthermore, the adaptive learning module of the D-CAIN model jointly optimizes the model based on feedback from candidate experts and the collaborative work of multiple agents. Where α, β, γ, δ, ε, The equal weighting coefficients are automatically adjusted based on the quality of the final meeting minutes and the quality of feedback from candidate experts; the uncertain baseline is adaptively updated based on historical success rates for different meeting types and team characteristics. , , , Simultaneously, joint optimization is carried out considering expert positioning accuracy, response timeliness, and collaboration satisfaction.

[0082] In this embodiment, compared to traditional meeting assistants that either generate an equally vague result for ambiguous statements or simply ignore the information, leaving the burden of verifying meeting minutes entirely to human users, this application proactively identifies the uncertainties at each node and uses a communicative agent to initiate precise clarification requests from participants at appropriate times during the meeting, thereby improving the accuracy and efficiency of meeting minutes generation. Furthermore, the adaptive learning module ensures that the meeting minutes system maintains stable performance improvement over long-term use, avoiding the performance degradation problem of traditional systems.

[0083] In some embodiments, a verification agent checks whether there are logical conflicts in the edge connections of each entity node in the dynamic knowledge graph to obtain verification results; a summary agent extracts the evolution path of each viewpoint node, and an action agent determines the participant node corresponding to each task node in the dynamic knowledge graph; based on the verification results, evolution paths, and participant nodes corresponding to each task node, the dynamic knowledge graph is corrected; and based on the corrected dynamic knowledge graph, meeting minutes are generated.

[0084] Specifically, after generating a dynamic knowledge graph, various types of intelligent agents collaborate on the dynamic knowledge graph. Each type of intelligent agent undertakes different tasks. For example, a summarizing intelligent agent extracts the evolution path of each viewpoint node, that is, extracts the state transition of a viewpoint from its proposal to its decision (adoption).

[0085] In some embodiments, user quality scores for meeting minutes are collected, as well as performance scores for different types of agents in generating meeting minutes; reward scores are obtained based on the time of meeting minutes generation and the adoption of clarification requests from each candidate expert; and the meeting minutes system is optimized based on the quality scores, performance scores, and reward scores.

[0086] Specifically, the Adaptive Multi-Agent Collaborative Policy Learning Algorithm (AMAC-Policy) is used to obtain quality assessments of the human-in-the-loop mechanism, multi-agent collaborative creation, and intelligent assessments of adaptive iterative learning. The human-in-the-loop mechanism refers to the liaison agent initiating clarification requests to candidate experts, transforming the feedback from candidate experts into adjustment signals for the dynamic knowledge graph, establishing collaborative trust with human experts, and thus dynamically adjusting human-machine collaboration. The specific calculation formula is shown in the following formula (4): (4) in, This indicates the validity of the feedback from the candidate experts; A score indicating the optimal timing of human-computer interaction; This indicates the accuracy of adopting human feedback. This indicates a change in the level of trust that human experts have in the system.

[0087] The quality assessment of multi-agent collaborative creation includes: establishing a multi-dimensional creation quality index system for multiple agents (such as content accuracy, language fluency, logical coherence, and innovation), realizing complementary learning of creation skills among agents, and calculating the role score of multiple agents in the process of generating meeting minutes through the following formula (5): (5) in, A score representing the collaborative creation efforts of multiple intelligent agents; Indicates the complementarity of agent skills; This represents a score indicating the innovativeness of content created by multiple intelligent agents. It indicates the coherence of multi-agent cooperation.

[0088] The intelligent evaluation of adaptive iterative learning includes: evaluation of adaptive adjustment of strategy parameters based on historical collaborative data, evaluation of multi-timescale learning mechanisms (short-term meeting minutes generation task optimization + long-term strategy evolution), evaluation of mechanisms for detecting and rapidly adapting to changes in meeting teams with different styles, and evaluation of strategies for balancing knowledge forgetting and updating. The reward score is calculated using the following formula (6), which reflects the learning and adaptive capabilities of the meeting minutes system: (6) in, This indicates the reward score for the meeting minutes system; This indicates the convergence speed of the adaptive learning module; This indicates how quickly the meeting minutes system adapts to changes in different meeting teams; This indicates the effective utilization rate of historical experience in the meeting minutes system; This represents the combined penalty for the time spent communicating during the meeting and the time it took to generate the meeting minutes.

[0089] And establish the four-dimensional collaborative reward function using the following formula (7): (7) in, This represents the four-dimensional collaborative reward function; Indicates the quality score of the meeting minutes; This indicates the validity of the feedback from the candidate experts; A score representing the collaborative creation efforts of multiple intelligent agents; This indicates the reward score for the meeting minutes system.

[0090] For example, refer to Figure 5 The AMAC-Policy learning algorithm process generally includes the following: The collaborative state of multiple agents during the meeting minutes generation process includes three modules: the feedback module (Human-in-the-Loop Feedback Integration, HILFI module), the evolution module (Collaborative CreationStrategy Evolution, CCSE module), and the learning module (Adaptive Iterative Learning Engine, AILE module).

[0091] The system comprises three modules: a feedback module, an evolution module, and a learning module. The feedback module assesses the creation quality of the summary agent, performs complementary skill learning, dynamically classifies roles, updates the experience base (agent experience), and optimizes the creation strategy. The learning module performs historical data analysis, multi-timescale learning, detects environmental changes, balances knowledge updates, and adjusts strategy parameters. Based on these modules, the system performs joint actions, including selecting debate topics (for multi-topic meetings), adjusting thresholds, and selecting editing strategies. A four-dimensional creation reward function is then calculated based on these joint actions, considering the quality score of the meeting minutes, the effectiveness of feedback from candidate experts, the role score of multiple agents (i.e., the reward score). The system updates its strategy based on this four-dimensional reward function and stores the experience generated from the meeting minutes for future iterations.

[0092] In this embodiment, the feedback module enhances human-computer interaction efficiency through intelligent candidate expert intervention, a dynamic trust model, and optimal timing selection. This allows the system to accurately seek help at critical junctures where human wisdom is most needed and to make autonomous decisions in areas where artificial intelligence can handle tasks independently, achieving true human-machine synergy. Furthermore, the evolution module, through multi-dimensional creation quality assessment and complementary learning of agent skills, enables the system to not only generate high-quality content but also leverage the advantages of multiple agents during the creation process, resulting in improved creation performance compared to a single AI model. The learning module and four-dimensional creation reward function design allow the system to optimize meeting minutes generation efficiency in the short term, continuously learn meeting minutes generation strategies in the long term, and quickly adapt to different meeting team styles. In summary, this solution aligns the dynamic knowledge graph before generating initial meeting minutes during the meeting process. Combined with a multi-dimensional reward mechanism, each human-computer interaction and collaborative creation by multiple agents has clear evaluation indicators and improvement directions, ensuring the predictability and controllability of the system's behavior.

[0093] To explain the embodiments of this application more simply and clearly, the following complete flowchart describes a method for taking meeting minutes provided by the embodiments of this application. The method consists of the following steps: Figure 1 The terminal device shown (i.e., the meeting minutes system) performs the following steps, as follows: Figure 6 As shown: Step 601, speech to text stream transcription.

[0094] Step 602: Activate the knowledge graph construction model.

[0095] Step 603: Construct the initial knowledge graph in real time.

[0096] Step 604, meeting adjourned.

[0097] Specifically, the meeting can end after the initial knowledge graph has been revised.

[0098] Step 605: Activation of the multi-agent system.

[0099] Step 606: Identify fuzzy nodes.

[0100] Step 607, Influence Analysis.

[0101] Step 608: Dispatch the contact agent.

[0102] Step 609: Select candidate experts at the appropriate time.

[0103] Step 610: Initiate a clarification request.

[0104] Step 611: Obtain feedback from candidate experts.

[0105] Step 612: Correct the initial knowledge graph.

[0106] Step 613, Multi-agent debate.

[0107] Step 6131, Summary Agent.

[0108] Step 6132: Verify the intelligent agent.

[0109] Step 6133, Action Item Agent.

[0110] Step 614: Reach a consensus.

[0111] Step 615: Generate the first draft of the collaborative agent.

[0112] Step 616: Review the agent's marked modification points.

[0113] Step 617: Refine the intelligent agent.

[0114] Specifically, the editing agent refines the initial draft with revisions.

[0115] Step 618: Determine if the quality meets the requirements. If yes, proceed to step 619; otherwise, return to step 616.

[0116] Step 619: Generate meeting minutes.

[0117] Step 620: Collect user feedback.

[0118] Step 621, Quality Evaluation.

[0119] Step 622, optimization and update of multi-agent system.

[0120] Based on the same technical concept, this application provides a schematic diagram of the structure of a meeting minutes generation device, such as... Figure 7 As shown, the meeting minutes generation device 700 includes: The data acquisition module 701 is used to collect the content of participants' speeches during the meeting; The detection module 702 is used to detect whether the speech content satisfies the event triggering mechanism based on the dynamic knowledge graph constructed during the meeting; the event triggering mechanism represents the state change of the dynamic knowledge graph. The update module 703 is used to update the entity nodes and / or edge connections of the dynamic knowledge graph if the event triggering mechanism is satisfied. Module 704 is used to generate meeting minutes based on the dynamic knowledge graph at the end of the meeting.

[0121] Optionally, the update module 703 is specifically used for: Obtain first entity information from the speech content, and determine whether there is a first entity node corresponding to the first entity information in the dynamic knowledge graph; if not, create the first entity node in the dynamic knowledge graph. The semantic association degree between the first entity node and at least one second entity node in the dynamic knowledge graph is determined based on the attention mechanism, and the edge connection between the first entity node and the second entity node is updated based on the semantic association degree.

[0122] Optionally, the update module 703 is specifically used for: If the content of the speech is related to the content of the speeches of other participants, an event node representing the first entity node is created; the content of the speeches of the other participants corresponds to the second entity node. Updating the edge connection between the first entity node and the second entity node based on the semantic relevance includes: Based on the semantic relevance of the speech content to the speeches content of other participants, an edge connection is established between the first entity node and the second entity node.

[0123] Optionally, the semantic relevance is characterized by one or more of the following: the state transition of the viewpoint node, the causal state between the viewpoint node and the decision node, and the directional state between the decision node and the task node; The semantic relevance conforms to one or more of the following constraints: temporal constraints, causal constraints, and role constraints.

[0124] Optionally, the update module 703 is specifically used for: If the participant in the speech is speaking for the first time, then create a participant node corresponding to the speaker; If the content of the speech indicates a meeting topic, then a topic node is created; If the content of the statement is a viewpoint that is being presented for the first time, then a viewpoint node corresponding to the content of the statement is created.

[0125] Optionally, the generation module 704 is further configured to: Identify fuzzy nodes in the dynamic knowledge graph, where a fuzzy node is any entity node with ambiguous semantics; Calculate the correlation degree between the fuzzy node and any participant node, determine the candidate expert responsible for correcting the fuzzy node, and send a clarification request to the candidate expert.

[0126] Optionally, the generation module 704 is specifically used for: Based on the semantic clarity of any entity node other than the participant node in the dynamic knowledge graph, fuzzy nodes are identified. The calculation of the correlation degree between the fuzzy node and any participant node, and the determination of the candidate expert responsible for correcting the fuzzy node, includes: The influence of any participant node on the fuzzy node is calculated from multiple dimensions. Based on the influence of multiple dimensions and the weight values ​​of each dimension, the correlation degree between the fuzzy node and any participant node is obtained; The participant node with the highest correlation degree exceeding the preset threshold is identified as a candidate expert for correcting the fuzzy node.

[0127] Optionally, the multiple dimensions include the dynamic weight of the participant's role, the quality of the participant's speech, the participant's influence on the conference topic, the participant's authority in the field of the conference topic, the participant's trust in other participants, and the participant's innovative contribution to the conference topic.

[0128] Optionally, the generation module 704 is further configured to: Calculate the belonging distance between any entity node in the dynamic knowledge graph and the participant node identified as a candidate expert; the belonging distance is determined based on at least one of the following: the topological distance between the entity node and the participant node in the dynamic knowledge graph, the semantic distance between multiple speeches of the entity node and the participant node, the temporal distance between the time of generation of the entity node and the time of generation of the participant node, and the trust distance between the entity node and the participant node. At least one entity node whose distance does not exceed the first threshold is assigned to the associated region of the candidate expert's participant node.

[0129] Optionally, the generation module 704 is specifically used for: The verification agent checks whether there are logical conflicts in the edge connections of each entity node in the dynamic knowledge graph, and obtains the verification results. The evolution path of each viewpoint node is extracted by a summary-type agent, and the participant node corresponding to each task node in the dynamic knowledge graph is determined by an action-item agent. Based on the verification results, the evolution path, and the participant nodes corresponding to each task node, the dynamic knowledge graph is corrected; based on the corrected dynamic knowledge graph, meeting minutes are generated.

[0130] Optionally, the generation module 704 is further configured to: Collect user ratings for the quality of the meeting minutes, and obtain ratings for the roles of different types of agents in generating the meeting minutes; A reward score is awarded based on the time the meeting minutes were generated and the adoption of clarification requests from each candidate expert. The meeting minutes system is optimized based on the quality score, the effectiveness score, and the reward score.

[0131] In this embodiment, the participants' speeches during the meeting are collected. Based on the dynamic knowledge graph already constructed during the meeting, it is detected whether the speeches meet the event triggering mechanism, which represents the state change of the dynamic knowledge graph. If the event triggering mechanism is met, the entity nodes and / or edge connections of the dynamic knowledge graph are updated. Based on the dynamic knowledge graph at the end of the meeting, meeting minutes are generated. This allows for the comparison of each participant's speeches with the constructed dynamic knowledge graph to determine whether new speeches are related to previous speeches. It captures the dynamic changes of any viewpoint from its introduction to discussion among multiple participants and finally to the final decision during the meeting, thereby obtaining the evolution process of the viewpoint and improving the accuracy of the meeting minutes.

[0132] Based on the same technical concept, embodiments of this application provide a computer device, which can be... Figure 1 The server shown is as follows: Figure 8 As shown, it includes at least one processor 801 and a memory 802 connected to at least one processor. In this embodiment, the specific linking medium between the processor 801 and the memory 802 is not limited. Figure 8 Taking the connection between the processor 801 and the memory 802 via a bus as an example, the bus can be divided into address bus, data bus, control bus, etc.

[0133] In this embodiment of the application, the memory 802 stores instructions executed by at least one processor 801. By executing the instructions stored in the memory 802, at least one processor 801 can perform the steps of the above-described meeting minutes generation method.

[0134] The processor 801 is the control center of the computer device. It can connect to various parts of the computer device using various interfaces and lines. By running or executing instructions stored in the memory 802 and calling data stored in the memory 802, it generates meeting minutes. Optionally, the processor 801 may include one or more processing modules. The processor 801 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 801. In some embodiments, the processor 801 and the memory 802 may be implemented on the same chip; in some embodiments, they may be implemented on separate chips.

[0135] The processor 801 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0136] Memory 802, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 802 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 802 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer device, but is not limited thereto. In the embodiments of this application, memory 802 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0137] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the above-described meeting minutes generation method.

[0138] Based on the same inventive concept, embodiments of this application provide a computer program product, including a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described meeting minutes generation method.

[0139] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for generating meeting minutes, characterized in that, Suitable for meeting minutes systems, including: Collect the participants' remarks during the meeting; Based on the dynamic knowledge graph constructed during the meeting, it is detected whether the speech content satisfies the event triggering mechanism; the event triggering mechanism represents the state change of the dynamic knowledge graph. If the event triggering mechanism is met, the entity nodes and / or edge connections of the dynamic knowledge graph are updated. Meeting minutes are generated based on the dynamic knowledge graph at the end of the meeting.

2. The method as described in claim 1, characterized in that, If the event triggering mechanism is satisfied, the dynamic knowledge graph is updated with entity nodes and / or edge connections, including: Obtain first entity information from the speech content, and determine whether there is a first entity node corresponding to the first entity information in the dynamic knowledge graph; if not, create the first entity node in the dynamic knowledge graph. The semantic association degree between the first entity node and at least one second entity node in the dynamic knowledge graph is determined based on the attention mechanism, and the edge connection between the first entity node and the second entity node is updated based on the semantic association degree.

3. The method as described in claim 2, characterized in that, Creating the first entity node in the dynamic knowledge graph includes: If the content of the speech is related to the content of the speeches of other participants, an event node representing the first entity node is created; the content of the speeches of the other participants corresponds to the second entity node. Updating the edge connection between the first entity node and the second entity node based on the semantic relevance includes: Based on the semantic relevance of the speech content to the speeches content of other participants, an edge connection is established between the first entity node and the second entity node.

4. The method as described in claim 2, characterized in that, The semantic relevance is characterized by one or more of the following: the state transition of the viewpoint node, the causal state between the viewpoint node and the decision node, and the directional state between the decision node and the task node. The semantic relevance conforms to one or more of the following constraints: temporal constraints, causal constraints, and role constraints.

5. The method as described in claim 2, characterized in that, Creating the first entity node in the dynamic knowledge graph includes: If the participant in the speech is speaking for the first time, then create a participant node corresponding to the speaker; If the content of the speech indicates a meeting topic, then a topic node is created; If the content of the statement is a viewpoint that is being presented for the first time, then a viewpoint node corresponding to the content of the statement is created.

6. The method according to any one of claims 1 to 5, characterized in that, Before generating the meeting minutes, the following steps are also included: Identify fuzzy nodes in the dynamic knowledge graph, where a fuzzy node is any entity node with ambiguous semantics; Calculate the correlation degree between the fuzzy node and any participant node, determine the candidate expert responsible for correcting the fuzzy node, and send a clarification request to the candidate expert.

7. The method as described in claim 6, characterized in that, The identification of fuzzy nodes in the dynamic knowledge graph includes: Based on the semantic clarity of any entity node other than the participant node in the dynamic knowledge graph, fuzzy nodes are identified. The calculation of the correlation degree between the fuzzy node and any participant node, and the determination of the candidate expert responsible for correcting the fuzzy node, includes: The influence of any participant node on the fuzzy node is calculated from multiple dimensions. Based on the influence of multiple dimensions and the weight values ​​of each dimension, the correlation degree between the fuzzy node and any participant node is obtained; The participant node with the highest correlation degree exceeding the preset threshold is identified as a candidate expert for correcting the fuzzy node.

8. The method as described in claim 7, characterized in that, After determining the candidate experts for correcting the fuzzy nodes, the method further includes: Calculate the belonging distance between any entity node in the dynamic knowledge graph and the participant node identified as a candidate expert; the belonging distance is determined based on at least one of the topological distance between the entity node and the participant node in the dynamic knowledge graph, the semantic distance between multiple speeches of the entity node and the participant node, the temporal distance between the time of generation of the entity node and the time of generation of the participant node, and the trust distance between the entity node and the participant node. At least one entity node whose distance does not exceed the first threshold is assigned to the associated region of the candidate expert's participant node.

9. The method as described in any one of claims 1 to 5, characterized in that, The process of generating meeting minutes based on the dynamic knowledge graph at the end of the meeting includes: The verification agent checks whether there are logical conflicts in the edge connections of each entity node in the dynamic knowledge graph, and obtains the verification results. The evolution path of each viewpoint node is extracted by a summary-type agent, and the participant node corresponding to each task node in the dynamic knowledge graph is determined by an action-item agent. Based on the verification results, the evolution path, and the participant nodes corresponding to each task node, the dynamic knowledge graph is corrected; based on the corrected dynamic knowledge graph, meeting minutes are generated.

10. The method as described in claim 9, characterized in that, After generating the meeting minutes, the process also includes: Collect user ratings for the quality of the meeting minutes, and obtain ratings for the roles of different types of agents in generating the meeting minutes; A reward score is awarded based on the time the meeting minutes were generated and the adoption of clarification requests from each candidate expert. The meeting minutes system is optimized based on the quality score, the effectiveness score, and the reward score.

Images