Pre-conference data pushing method and system based on professional field and conference information of user

Abstract

The invention discloses a pre-conference data pushing method and system based on a user professional field and conference information, and the method comprises the steps: collecting and obtaining authorized multi-dimensional data which at least comprises participant data, conference data and external knowledge base data; processing the multi-dimensional data based on an open source semantic model, and constructing an associated knowledge network which is used for establishing a semantic association relationship among the conference issue, the core knowledge point, the participant role and the historical conference conclusion; when the target conference is created, a personalized pre-conference data recommendation list is generated for the participants based on the associated knowledge network and a preset recommendation strategy, and the recommendation strategy is at least based on the role matching degree, the historical preference similarity and the conference theme association degree; and pushing the pre-meeting data recommendation list to the corresponding participants, collecting interactive behavior data of the participants on the pushed data, and dynamically adjusting parameters of a recommendation strategy according to the interactive behavior data to optimize subsequent recommendation. According to the invention, precise and scenarized pushing of knowledge is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent meeting management and knowledge service technology, and in particular to a method and system for pushing pre-meeting materials based on the user's professional field and meeting information. Background Technology

[0002] In current mainstream conference systems, the core technical path for participants to obtain supplementary conference knowledge revolves around "goal-oriented proactive retrieval + basic material support," and the specific process includes:

[0003] 1. Direction Setting and Knowledge Supplementation: Participants should first clarify the core discussion direction based on the meeting notice, use the basic materials provided by the meeting to initially sort out the key topics, and finally supplement relevant knowledge content by combining external network resources, internal corporate knowledge reserves and personal knowledge areas.

[0004] 2. Knowledge Point Selection and Requirements List Formulation: By reading the meeting attachments, participants selected relevant knowledge points based on their roles (e.g., technical staff focused on "implementation details," business staff focused on "matching with the implementation scenario"); and manually recorded knowledge gaps related to key topics using local document and other note-taking tools to form a list to be searched (e.g., "whether the cost accounting method complies with the latest financial and tax policies," "security compliance basis for architecture upgrades," etc.).

[0005] 3. External resource retrieval: Based on general search engines such as Baidu and Google, or vertical databases such as industry report platforms and academic paper databases, external knowledge is obtained through keyword retrieval.

[0006] 4. Internal Knowledge Access: Utilizing the company's self-built knowledge base (such as employee handbooks or internal wikis) or document management platform (such as Alibaba Cloud Enterprise Edition), internal knowledge can be retrieved through keyword searches or directory navigation. However, search technology relies on literal matching and lacks semantic association capabilities. Participants need to input internal terms such as "Company Project Compliance Manual Article 3.2" or "Historical Cost Accounting Template" to retrieve past cases, standard processes, internal regulations, and other materials to ensure that the discussion content meets the company's actual requirements.

[0007] Current technologies rely heavily on a "human-led + tool-assisted" model, with the technical aspects focusing on basic file transfer and storage. The accuracy and efficiency of knowledge acquisition depend heavily on the participants' proactive actions and information filtering abilities. This process is cumbersome and inefficient, directly impacting meeting quality and often requiring multiple meetings to gather the necessary knowledge. Summary of the Invention

[0008] The purpose of this invention is to address the problems of low efficiency, lack of personalization, and missing context in pre-meeting knowledge delivery in existing technologies, and to provide a method and system for delivering pre-meeting materials based on the user's professional field and meeting information. By integrating participant profiles, meeting topic analysis, knowledge graph association, and real-time dynamic adjustments, it achieves precise and contextualized knowledge delivery.

[0009] The technical solution adopted in this invention is:

[0010] A method for pushing pre-conference materials based on users' professional fields and conference information includes the following steps:

[0011] Multidimensional data collection: Collect and acquire authorized multidimensional data, which includes at least participant data, meeting data, and external knowledge base data;

[0012] Knowledge organization: Based on open-source semantic models, multi-dimensional data is processed to construct a related knowledge network. The related knowledge network is used to establish semantic relationships between conference topics, core knowledge points, participant roles, and historical conference conclusions.

[0013] Personalized recommendations: When creating a target meeting, a personalized list of pre-meeting materials is generated for participants based on the associated knowledge network and preset recommendation strategies. The recommendation strategies are based at least on role matching, similarity of historical preferences, and relevance to the meeting topic.

[0014] Push and Feedback: A list of recommended materials will be pushed to the relevant participants before the meeting, and the interaction data of the participants with the pushed materials will be collected. The parameters of the recommendation strategy will be dynamically adjusted based on the interaction data to optimize subsequent recommendations.

[0015] Furthermore, the participant data collected through multidimensional data collection specifically includes role tags, historical meeting participation records, and knowledge preference data;

[0016] By constructing multi-level standardized role tags through role profiles, role tags include first-level tags for basic role categories and second-level tags for sub-fields that are sub-categories of the first-level tags. Specifically, the first-level tags include basic role categories such as technology research and development, product manager, and marketing operations, while the second-level tags are sub-fields of the first-level tags, such as front-end development and algorithm engineer under technology research and development. At the same time, a custom tag entry is set up to support users to add roles that are not covered.

[0017] The meeting system's API interface is used to obtain participants' historical meeting participation records, which include the meeting topics, agenda items, participation time, and interaction content. Specifically, the log capture module for meeting interaction records is used to collect user questions, speaking records, and interactive content.

[0018] By using front-end tracking technology to record participants' actions on the conference document display page, knowledge preference data is generated. These actions include document opening time, scrolling depth, saving, sharing, and annotation.

[0019] Specifically, user actions are categorized as follows: 1) Time dimension: document opening and closing times, calculating effective reading time; 2) Operation dimension: recording document collection, sharing, and annotation behaviors, as well as page scrolling depth and click trajectories in key areas; 3) Document type: collecting the document types downloaded by attendees through the conference materials download interface.

[0020] Furthermore, during meeting data collection, the BiLSTM-CRF model was used to classify the meeting agenda into paragraph structures and extract agenda elements to construct an agenda element relationship graph, establishing dependencies between topics. A domain-adjusted BERT model was used, combined with TF-IDF features and contextual semantic features, to extract keywords for meeting topics through confidence screening (≥0.85) and clustering deduplication. The Sentence-BERT model was used to generate semantic vectors for project documents, and a multi-level association algorithm (keyword matching → semantic similarity → time decay factor) was used to calculate the association confidence between the semantic vectors and the current meeting topics to filter out (accuracy ≥0.90) related project documents.

[0021] Furthermore, the steps for constructing a related knowledge network are as follows:

[0022] Based on an open-source semantic model, entity recognition is performed on conference materials to extract core knowledge points, key figures, and time points.

[0023] Using the meeting topic text as input, the system associates the meeting topic with core knowledge points through rule matching and prompt word matching from an open-source semantic model.

[0024] A standard role library is established based on the enterprise's organizational structure, and an open-source semantic model is used to automatically label knowledge points with domain tags and content keywords (after manual review and correction of deviations) to establish the association between knowledge points and participant roles; specifically, a standardized role library containing "role ID + responsibility keyword" is formed based on the extracted core responsibility keywords.

[0025] Natural language processing technology is used to transform historical meeting conclusions and current topics into semantic vectors. Cosine similarity is used to calculate semantic relevance, and association rule algorithms are combined to mine high-frequency association rules to calculate association strength scores. Based on the association strength scores, the top N historical conclusions with the highest scores are selected. Knowledge graph storage technology is used to store the association relationships of the top N historical conclusions that have passed the verification in the form of triples in the graph, thereby establishing the association between historical meeting conclusions and current topics.

[0026] Furthermore, the knowledge network is dynamically updated. The dynamic update method includes: collecting user feedback on whether the knowledge points are relevant or irrelevant, using an incremental weight adjustment algorithm (if relevant, the weight is increased by 1; if irrelevant, it is decreased by 1) to update the knowledge point association weights, and automatically removing the corresponding knowledge points from the matching range when the weight is lower than the threshold (3).

[0027] Furthermore, the steps to generate a personalized pre-meeting materials recommendation list are as follows:

[0028] The current meeting topic is analyzed, and core knowledge points are extracted using an open-source semantic model to form a mapping list of "meeting topic - core knowledge points".

[0029] Query the related knowledge network to obtain a list of participant roles associated with the core knowledge points, as well as historical meeting conclusions and materials related to the core knowledge points;

[0030] Based on the participant's role information and data from the related knowledge network, the following tiered recommendations are made:

[0031] a) Push basic information that matches the role and responsibilities; for example, "Algorithm Engineer" corresponds to "NLP Model Tuning Technical Document" and "Historical Iteration Parameter Record Table", and "Product Manager" corresponds to "User Intent Recognition User Feedback Summary" and "Competitive Product Intent Recognition Function Analysis".

[0032] b) Push historical conference conclusions and materials related to core knowledge points as supplementary information for the historical context;

[0033] c) Filter out irrelevant information and remove duplicates to create a personalized pre-meeting information list for each role.

[0034] Furthermore, the recommendation strategy uses the following weighted formula to calculate the recommendation score of candidate knowledge items, and sorts and filters the candidate knowledge items according to the recommendation score (the same knowledge item is only retained once, and low-scoring items that are irrelevant to the user role / meeting topic, i.e., scores < 0.5), to form a pre-meeting material recommendation list; the specific steps are as follows:

[0035] Based on the knowledge association analysis results of the corresponding roles, the "role matching degree" is calculated (e.g., the matching degree between the manager and the "project resource coordination" knowledge item is 0.7).

[0036] The cosine similarity between the semantic vector of the candidate knowledge item and the keywords of the conference topic is calculated to obtain the "conference topic relevance".

[0037] By analyzing the behavioral preferences of similar users using collaborative filtering algorithms, we can obtain "historical preference similarity" (with a value of 0-1, such as user A having a preference similarity of 0.7 with a similar user group).

[0038] The recommended score is calculated using the following formula:

[0039] Recommendation score = α * Role matching degree + β * Historical preference similarity + γ * Meeting topic relevance + δ * Timeliness

[0040] Among them, α, β, γ, and δ are weight coefficients dynamically assigned based on the participant's role profile, historical behavior salience, the closeness of the association between knowledge items and the topic, and the creation time of knowledge items, respectively. Specifically, user role tags and historical preference features are combined with the current meeting scenario (such as "algorithm optimization meeting") to generate weighted user profiles (e.g., in the technical personnel profile, "technical details" has a weight of 0.6 and "project progress" has a weight of 0.2).

[0041] Furthermore, α (role weight): 0.4 for technical personnel and 0.3 for managers, dynamically assigned based on role tags in the user profile; β (historical preference weight): 0.2–0.3 based on collaborative filtering results (the more significant the user's historical behavior, the higher the weight); γ (topic relevance weight): 0.2–0.3 based on content-knowledge point matching results (the closer the knowledge item is to the meeting topic, the higher the weight); δ (timeliness weight): 0.2 for knowledge items from the past week and 0.05 for those older than 30 days, determined based on the knowledge item's creation time.

[0042] Furthermore, the steps for dynamically adjusting the recommendation strategy parameters based on interaction behavior data are as follows:

[0043] Participants' interest in the materials was categorized based on their interactive behavior, with levels of interest including high, medium, and low relevance. Interactive behavior included opening the document, reading time, and interactive actions. Materials with an opening time of >5 minutes, or those showing "favorites" or "marked as important" actions, were classified as high interest. Materials with an opening time of 2-5 minutes and no additional interaction were classified as medium interest. Materials that were not opened within a specified time (not clicked / downloaded within 24 hours) or with an opening time of <1 minute (i.e., classified as "invalid browsing") were classified as low relevance.

[0044] For data deemed low-relevance, the weight of historical preference similarity in the recommendation formula for the corresponding knowledge item is reduced, and the priority for future recommendations is lowered.

[0045] For data deemed of moderate interest, the weight remains unchanged, and it is recorded as "potentially relevant" for further recommendation and verification in conjunction with similar conference topics.

[0046] For materials deemed to have high interest, the weights of the role matching degree and conference topic relevance of the corresponding knowledge items are increased, and the recommendation priority is raised.

[0047] A pre-conference materials delivery system based on user expertise and conference information, the system includes:

[0048] The data acquisition module is used to collect and acquire authorized multi-dimensional data, which includes at least participant data, meeting data, and external knowledge base data.

[0049] The knowledge organization module is used to process multi-dimensional data based on an open-source semantic model and build a related knowledge network. The related knowledge network is used to establish semantic relationships between meeting topics, core knowledge points, participant roles, and historical meeting conclusions.

[0050] The personalized recommendation module is used to generate a personalized list of pre-meeting materials for participants when creating a target meeting, based on the associated knowledge network and preset recommendation strategies. The recommendation strategy is based at least on role matching, historical preference similarity and meeting topic relevance.

[0051] The push and feedback module is used to push a list of recommended materials to the corresponding participants before the meeting, and to collect the participants' interaction data with the pushed materials. Based on the interaction data, the parameters of the recommendation strategy are dynamically adjusted to optimize subsequent recommendations.

[0052] The present invention adopts the above technical solution and has the following beneficial effects compared with the prior art: 1) It accurately pushes personalized materials according to the participants' positions, professional fields and meeting topics, avoiding interference from irrelevant information; 2) It automatically links the conclusions of historical meetings and related knowledge points, so that participants can quickly understand the project background; 3) The user reading behavior feedback system can learn the participants' reading preferences, and the longer it is used, the more accurate the recommendations become, reducing the push of invalid information; 4) It automatically organizes the required materials before the meeting, shortens the preparation time, and reduces repeated questions during the meeting. Attached Figure Description

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0054] Figure 1 This is a flowchart illustrating the pre-meeting material delivery method based on the user's professional field and meeting information according to the present invention.

[0055] Figure 2 This is a schematic diagram illustrating the principle architecture of the pre-meeting material push system based on the user's professional field and meeting information according to the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0057] like Figure 1As shown in Figure 2, this invention discloses a method for pushing pre-conference materials based on the user's professional field and conference information. The method includes the following steps:

[0058] Multidimensional data collection: Collect and acquire authorized multidimensional data, which includes at least participant data, meeting data, and external knowledge base data;

[0059] Knowledge organization: Based on open-source semantic models, multi-dimensional data is processed to construct a related knowledge network. The related knowledge network is used to establish semantic relationships between conference topics, core knowledge points, participant roles, and historical conference conclusions.

[0060] Personalized recommendations: When creating a target meeting, a personalized list of pre-meeting materials is generated for participants based on the associated knowledge network and preset recommendation strategies. The recommendation strategies are based at least on role matching, similarity of historical preferences, and relevance to the meeting topic.

[0061] Push and Feedback: A list of recommended materials will be pushed to the relevant participants before the meeting, and the interaction data of the participants with the pushed materials will be collected. The parameters of the recommendation strategy will be dynamically adjusted based on the interaction data to optimize subsequent recommendations.

[0062] Furthermore, the participant data collected through multidimensional data collection specifically includes role tags, historical meeting participation records, and knowledge preference data;

[0063] By constructing multi-level standardized role tags through role profiles, role tags include first-level tags for basic role categories and second-level tags for sub-fields that are sub-categories of the first-level tags. Specifically, the first-level tags include basic role categories such as technology research and development, product manager, and marketing operations, while the second-level tags are sub-fields of the first-level tags, such as front-end development and algorithm engineer under technology research and development. At the same time, a custom tag entry is set up to support users to add roles that are not covered.

[0064] The meeting system's API interface is used to obtain participants' historical meeting participation records, which include the meeting topics, agenda items, participation time, and interaction content. Specifically, the log capture module of the meeting interaction records (chat box, Q&A panel, meeting minutes) is used to collect user questions, speaking records, and interactive content.

[0065] By using front-end tracking technology to record participants' actions on the conference document display page, knowledge preference data is generated. These actions include document opening time, scrolling depth, saving, sharing, and annotation.

[0066] Specifically, user actions are categorized as follows: 1) Time dimension: document opening and closing times, calculating effective reading time; 2) Operation dimension: recording document collection, sharing, and annotation behaviors, as well as page scrolling depth and click trajectories in key areas; 3) Document type: collecting the document types downloaded by attendees through the conference materials download interface.

[0067] Furthermore, during meeting data collection, the BiLSTM-CRF model was used to classify the meeting agenda into paragraph structures and extract agenda elements to construct an agenda element relationship graph, establishing dependencies between topics. A domain-adjusted BERT model was used, combined with TF-IDF features and contextual semantic features, to extract keywords for meeting topics through confidence screening (≥0.85) and clustering deduplication. The Sentence-BERT model was used to generate semantic vectors for project documents, and a multi-level association algorithm (keyword matching → semantic similarity → time decay factor) was used to calculate the association confidence between the semantic vectors and the current meeting topics to filter out (accuracy ≥0.90) related project documents.

[0068] Specifically, a BiLSTM-CRF model is used for paragraph structure classification (accuracy ≥ 95%), combined with a rule engine to extract elements such as topics, sub-topics, time nodes, and participants, constructing an agenda element relationship graph and establishing dependencies between topics. Domain fine-tuning is performed on the BERT model (F1 score ≥ 0.88), fusing TF-IDF features and contextual semantic features (weight ratio 7:3). Deduplication is achieved through confidence screening (≥ 0.85) and clustering, extracting core keywords such as "AI algorithm optimization" and "customer requirement review." Sentence-BERT is used to generate document vectors, and a multi-level association algorithm (keyword matching → semantic similarity → time decay factor) is used to calculate association confidence, outputting the top 5 associated documents (precision ≥ 0.90), prioritizing association with recent project documents (such as research reports and decision lists).

[0069] Other data collection: 1) Connecting to enterprise knowledge bases: By deploying MCP server adapters and adopting a plug-in design to connect to internal enterprise document systems (such as employee handbooks, technical manuals, etc.), supporting PDF / DOCX / HTML format parsing, and achieving semantic retrieval of document content through a metadata index (Elasticsearch). Key steps include: configuring system monitoring services (inotify mechanism) to capture document changes in real time; uniformly converting structured / semi-structured data into the MCP standard JSON-LD format; and implementing fine-grained control of document access permissions based on the RBAC model. 2) Collecting third-party data sources: Standardizing access to third-party APIs (such as customer background databases, industry reporting platforms) through the MCP protocol adaptation layer. Core processes include: deploying Baidu / Google search MCP server adapters and submitting keyword search requests through standardized APIs; configuring search parameters (such as time range, number of results) and supporting multi-engine result aggregation and deduplication; deploying industry reporting platforms and academic paper database search MCP server adapters and submitting keyword searches through standardized APIs to obtain literature.

[0070] Furthermore, the steps for constructing a related knowledge network are as follows:

[0071] Based on an open-source semantic model, entity recognition is performed on conference materials to extract core knowledge points, key figures, and time points.

[0072] Specifically, when extracting core knowledge points, the BERT model is used to perform entity recognition on meeting materials (PPT, PDF, audio and video converted to text) to extract core knowledge points (such as technical terms "AIGC" and "variable cost accounting"), people (contacts), time nodes (project milestones), etc.

[0073] In the pre-meeting material preparation stage, enterprises across various business areas (R&D, administration, human resources, supply chain, etc.) generally face three common pain points: the materials received by participants are disconnected from their professional fields (e.g., R&D engineers receive administrative procurement process materials but lack technical specifications for reference); the materials are not related to the needs of the participant's role (e.g., when legal departments review procurement contracts, they only receive the contract text and lack historical compliance cases); and the materials do not integrate relevant historical conclusions (e.g., when HR formulates salary adjustment plans, they only receive current position salary data and do not incorporate past adjustment experience). The core purpose of establishing a knowledge network is to accurately filter and integrate information related to the participant's professional field, role responsibilities, and current topic through a full-link connection of "meeting agenda - core knowledge points - participant role - historical conclusions," thereby overcoming the bottleneck of "scattered searching and poor adaptability" of pre-meeting materials and shifting the pre-meeting materials for all business meetings from "general push" to "precise matching." For example, when R&D engineers discuss new product architecture, they automatically receive "microservice specifications + historical conclusions of similar architectures" before the meeting. When administrative staff prepare for procurement meetings, they obtain "supplier qualification standards and past procurement audit cases" before the meeting. Ultimately, this achieves the goal of "participants receiving precise information tailored to their own expertise and supporting the discussion of topics before the meeting," laying the foundation for subsequent meetings to shift from "disorderly discussions" to "precise decision-making." At the same time, it promotes the transformation of enterprise-wide knowledge from "passive storage" to "proactive adaptation to pre-meeting needs," improving the efficiency of knowledge reuse and the quality of meeting preparation.

[0074] Using the conference topic text as input, the conference topics are associated with core knowledge points through rule matching and prompt word matching from an open-source semantic model; the specific steps are as follows:

[0075] The meeting agenda features were extracted, and the text was cleaned using regularization tools (such as the Python library re) to remove redundant symbols and repetitive expressions. Based on the enterprise's "business terminology dictionary", the Levenshtein edit distance algorithm was used to unify synonyms. Then, the TF-IDF and TextRank algorithms were used to extract weighted core keywords (weight ≥ 0.6), and the BERT-base Chinese model was called to generate semantic vectors. Finally, a structured feature vector containing "a list of core keywords + semantic vectors" was output.

[0076] The open-source semantic model prompts knowledge points. It calls open-source semantic models (such as Qwen2.5-7B) and guides the matching through preset prompts. The model calculates the semantic similarity of knowledge points based on the structured feature vector of "core keyword list + semantic vector" (using cosine similarity algorithm) and associates it with the knowledge point list (including association weights of 0-10).

[0077] A standard role library is established based on the enterprise's organizational structure, and an open-source semantic model is used to automatically label knowledge points with domain tags and content keywords (after manual review and correction of deviations) to establish the association between knowledge points and participant roles; specifically, a standardized role library containing "role ID + responsibility keyword" is formed based on the extracted core responsibility keywords.

[0078] Specifically, he established a standard role library. Based on the enterprise's organizational structure, roles were defined according to "responsibility area (e.g., R&D / HR) + job title (e.g., algorithm engineer / compensation specialist)". Core responsibility keywords were extracted (e.g., "algorithm engineer → model training, parameter tuning") to form a standardized role library containing "role ID + responsibility keywords". For role and knowledge attribute labeling, an open-source semantic model (e.g., Qwen-7B) was used to automatically label knowledge points (documents / data) with domain tags (e.g., "deep learning" "compensation system") and content keywords (e.g., "neural network" "salary adjustment range"). Manual review and correction of deviations resulted in a tagged knowledge point library.

[0079] Natural language processing technology is used to transform historical meeting conclusions and current topics into semantic vectors. Cosine similarity is used to calculate semantic relevance, and association rule algorithms are combined to mine high-frequency association rules to calculate association strength scores. Based on the association strength scores, the top N historical conclusions with the highest scores are selected. Knowledge graph storage technology is used to store the association relationships of the top N historical conclusions that have passed the verification in the form of triples in the graph, thereby establishing the association between historical meeting conclusions and current topics.

[0080] Specifically, Natural Language Processing (NLP) techniques (such as named entity recognition and key information extraction models) are used to structure the data of historical meeting conclusions (extracting fields such as meeting identifier, conclusion type, and core content) and current topics (extracting fields such as topic identifier and core theme) to clarify the key information of both. Combining domain knowledge to define feature dimensions such as domain matching degree, open-source semantic models (such as Qwen-7B) are used to transform the core content of historical conclusions and current topics into semantic vectors, and cosine similarity calculation is used to achieve semantic alignment. Association rule algorithms (such as Apriori) are used to mine high-frequency association rules of "historical conclusions - subsequent topics" from historical data. A correlation strength score (0-10 points, ≥6 points is considered strong association) is calculated by integrating multi-dimensional features such as domain matching degree and entity overlap. Based on the correlation strength score, the system automatically selects the top 5 strongly associated historical conclusions. Intervention is carried out through manual verification (such as adding / deleting associations), and reinforcement learning is used to optimize the association model, taking into account participant feedback (such as citations and marking irrelevant information). By leveraging knowledge graph storage technology, verified relationships are stored in the graph as triples (including metadata such as relationship strength and basis). Combined with a daily incremental update mechanism, the current meeting is converted into a historical meeting, and the relationship process is repeated for new topics, thus realizing graph iteration.

[0081] Furthermore, the knowledge network is dynamically updated. The dynamic update method includes: collecting user feedback on whether the knowledge points are relevant or irrelevant, using an incremental weight adjustment algorithm (if relevant, the weight is increased by 1; if irrelevant, it is decreased by 1) to update the knowledge point association weights, and automatically removing the corresponding knowledge points from the matching range when the weight is lower than the threshold (3).

[0082] Furthermore, the steps to generate a personalized pre-meeting materials recommendation list are as follows:

[0083] The process involves analyzing the current meeting topic and extracting core knowledge points using an open-source semantic model to create a mapping list of "meeting topic - core knowledge points". Specifically, natural language processing (NLP) techniques (such as keyword extraction and entity recognition models) are used to analyze the current meeting topic (e.g., "2024 Q4 Intelligent Customer Service Algorithm Optimization Meeting"). An open-source semantic model (e.g., Qwen-7B) is then used to perform semantic analysis on the topic text to extract core knowledge points, including technical terms (e.g., "NLP interaction model" and "intent recognition accuracy"), discussion focuses (e.g., "model response speed optimization" and "user intent misidentification problem"), etc., forming a mapping list of "meeting topic - core knowledge points" (e.g., topic-related knowledge points: [NLP model tuning, user intent recognition, response delay solutions]).

[0084] Query the related knowledge network to obtain a list of participant roles associated with the core knowledge points, as well as historical meeting conclusions and materials associated with the core knowledge points. Specifically, the association between knowledge points and roles is as follows: obtain the mapping relationship between "core knowledge points - participant roles" in the knowledge base (e.g., "NLP model tuning" is associated with "algorithm engineer" and "data scientist", "user intent recognition" is associated with "product manager" and "algorithm engineer"); the association between knowledge points and historical conclusions is as follows: obtain historical meeting conclusions related to the core knowledge points (e.g., the conclusion "intent recognition accuracy rate of 82% did not meet the standard" in the "2024Q3 Intelligent Customer Service Review Meeting", associated with the current knowledge point "user intent recognition"), and simultaneously extract the historical materials corresponding to the conclusions (e.g., "Q3 Intent Recognition Problem Review Report" and "Model Iteration Historical Data").

[0085] Based on the participant's role information and data from the related knowledge network, the following tiered recommendations are made:

[0086] a) Push basic information that matches the role and responsibilities; for example, "Algorithm Engineer" corresponds to "NLP Model Tuning Technical Document" and "Historical Iteration Parameter Record Table", and "Product Manager" corresponds to "User Intent Recognition User Feedback Summary" and "Competitive Product Intent Recognition Function Analysis".

[0087] b) Push historical meeting conclusions and materials related to core knowledge points as supplementary information for historical context; push historical meeting conclusions and corresponding materials (such as "Q3 Intent Recognition Problem Review Report") related to core knowledge points to the corresponding roles to supplement historical context (such as algorithm engineers need to refer to the reasons for historical problems, and product managers need to combine historical user feedback).

[0088] c) Filter out irrelevant information and remove duplicates to create a personalized pre-meeting materials list for each role. (For example, do not push technical parameter documents to "Administrative Specialist") Ensure that similar materials are not recommended repeatedly, and ultimately create a personalized pre-meeting materials list for each role.

[0089] Specifically, based on the logical association of "meeting theme - knowledge points - roles", pre-meeting materials are accurately pushed to participating users.

[0090] Furthermore, the recommendation strategy uses the following weighted formula to calculate the recommendation score of candidate knowledge items, and sorts and filters the candidate knowledge items according to the recommendation score (the same knowledge item is only retained once, and low-scoring items that are irrelevant to the user role / meeting topic, i.e., scores < 0.5), to form a pre-meeting material recommendation list; the specific steps are as follows:

[0091] Based on the knowledge association analysis results of the corresponding roles, the "role matching degree" is calculated (e.g., the matching degree between the manager and the "project resource coordination" knowledge item is 0.7).

[0092] The cosine similarity between the semantic vector of the candidate knowledge item and the keywords of the conference topic is calculated to obtain the "conference topic relevance".

[0093] By analyzing the behavioral preferences of similar users using collaborative filtering algorithms, we can obtain "historical preference similarity" (with a value of 0-1, such as user A having a preference similarity of 0.7 with a similar user group).

[0094] The recommended score is calculated using the following formula:

[0095] Recommendation score = α * Role matching degree + β * Historical preference similarity + γ * Meeting topic relevance + δ * Timeliness

[0096] Among them, α, β, γ, and δ are weight coefficients dynamically assigned based on the participant's role profile, historical behavior salience, the closeness of the association between knowledge items and the topic, and the creation time of knowledge items, respectively. Specifically, user role tags and historical preference features are combined with the current meeting scenario (such as "algorithm optimization meeting") to generate weighted user profiles (e.g., in the technical personnel profile, "technical details" has a weight of 0.6 and "project progress" has a weight of 0.2).

[0097] Furthermore, α (role weight): 0.4 for technical personnel and 0.3 for managers, dynamically assigned based on role tags in the user profile; β (historical preference weight): 0.2–0.3 based on collaborative filtering results (the more significant the user's historical behavior, the higher the weight); γ (topic relevance weight): 0.2–0.3 based on content-knowledge point matching results (the closer the knowledge item is to the meeting topic, the higher the weight); δ (timeliness weight): 0.2 for knowledge items from the past week and 0.05 for those older than 30 days, determined based on the knowledge item's creation time.

[0098] Furthermore, a multi-channel approach is adopted to push pre-meeting material recommendation lists, including: essential content (such as meeting agenda and decision-making list) and recommended pre-meeting materials (such as "detailed document of algorithm optimization solution" for technical personnel and "summary of customer needs" for managers).

[0099] Furthermore, the feedback module generates quantitative feedback data based on participants' interaction with the pushed content, iterating the recommendation weights in reverse to improve the accuracy of subsequent recommendations. The specific mechanism is as follows: Behavioral monitoring indicators and interest determination; Opening behavior: whether a document is opened (client click / email attachment download); Reading time: the duration from opening to closing (automatically recorded by the client, indirectly counted by email through "read receipts + attachment access logs"); Interaction actions: whether it is marked (e.g., client "favorites" "marked as important"), forwarded (e.g., forwarded to colleagues). High interest: opening time > 5 minutes, or "favorites" or "marked as important" actions exist; Medium interest: opening time 2-5 minutes, no additional interaction; Low relevance: not opened (no clicks / downloads within 24 hours), or opening time < 1 minute (judged as "invalid browsing").

[0100] Furthermore, the steps for dynamically adjusting the recommendation strategy parameters based on interaction behavior data are as follows:

[0101] Participants' interest in the materials was categorized based on their interactive behavior, with levels of interest including high, medium, and low relevance. Interactive behavior included opening the document, reading time, and interactive actions. Materials with an opening time of >5 minutes, or those showing "favorites" or "marked as important" actions, were classified as high interest. Materials with an opening time of 2-5 minutes and no additional interaction were classified as medium interest. Materials that were not opened within a specified time (not clicked / downloaded within 24 hours) or with an opening time of <1 minute (i.e., classified as "invalid browsing") were classified as low relevance.

[0102] For data deemed low-relevance, the weight of historical preference similarity in the recommendation formula for the corresponding knowledge item is reduced, and the priority for future recommendations is lowered.

[0103] For data deemed of moderate interest, the weight remains unchanged, and it is recorded as "potentially relevant" for further recommendation and verification in conjunction with similar conference topics.

[0104] For materials deemed to have high interest, the weights of the role matching degree and conference topic relevance of the corresponding knowledge items are increased, and the recommendation priority is raised.

[0105] Furthermore, this invention employs a feedback closed-loop mechanism: feedback data is automatically aggregated every morning and updated to the user database and knowledge modules: at the user profile level, the weight of user "interest tags" is updated (e.g., the weight of the highly interested "algorithm optimization case" tag is increased by 0.2); at the knowledge association level, the association strength of "knowledge item-role" is updated (e.g., the association strength between the low-relevance "customer needs summary" and technical personnel decreases from 0.3 to 0.2); ultimately achieving a closed loop of "push-feedback-adjustment-re-push", ensuring that the recommendation model continuously meets the actual needs of the participants.

[0106] A pre-conference materials delivery system based on user expertise and conference information, the system includes:

[0107] The data acquisition module is used to collect and acquire authorized multi-dimensional data, which includes at least participant data, meeting data, and external knowledge base data.

[0108] The knowledge organization module is used to process multi-dimensional data based on an open-source semantic model and build a related knowledge network. The related knowledge network is used to establish semantic relationships between meeting topics, core knowledge points, participant roles, and historical meeting conclusions.

[0109] The personalized recommendation module is used to generate a personalized list of pre-meeting materials for participants when creating a target meeting, based on the associated knowledge network and preset recommendation strategies. The recommendation strategy is based at least on role matching, historical preference similarity and meeting topic relevance.

[0110] The push and feedback module is used to push a list of recommended materials to the corresponding participants before the meeting, and to collect the participants' interaction data with the pushed materials. Based on the interaction data, the parameters of the recommendation strategy are dynamically adjusted to optimize subsequent recommendations.

[0111] The present invention adopts the above technical solution and has the following beneficial effects compared with the prior art: 1) It accurately pushes personalized materials according to the participants' positions, professional fields and meeting topics, avoiding interference from irrelevant information; 2) It automatically links the conclusions of historical meetings and related knowledge points, so that participants can quickly understand the project background; 3) The user reading behavior feedback system can learn the participants' reading preferences, and the longer it is used, the more accurate the recommendations become, reducing the push of invalid information; 4) It automatically organizes the required materials before the meeting, shortens the preparation time, and reduces repeated questions during the meeting.

[0112] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. A method for pushing pre-conference materials based on users' professional fields and conference information, characterized by: The method includes the following steps: multidimensional Data Acquisition: Collect and acquire authorized multi-dimensional data, which includes at least participant data, meeting data, and external knowledge base data; Knowledge organization: Based on open-source semantic models, multi-dimensional data is processed to construct a related knowledge network. The related knowledge network is used to establish semantic relationships between conference topics, core knowledge points, participant roles, and historical conference conclusions. Personalized recommendations: When creating a target meeting, a personalized list of pre-meeting materials is generated for participants based on the associated knowledge network and preset recommendation strategies. The recommendation strategies are based at least on role matching, similarity of historical preferences, and relevance to the meeting topic. Push and Feedback: A list of recommended materials will be pushed to the relevant participants before the meeting, and the interaction data of the participants with the pushed materials will be collected. The parameters of the recommendation strategy will be dynamically adjusted based on the interaction data to optimize subsequent recommendations.

2. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: The participant data collected through multidimensional data collection specifically includes role tags, historical meeting participation records, and knowledge preference data; Multi-level standardized character tags are constructed through character profiles. Character tags include first-level tags for basic character categories and second-level tags for subdivided fields that serve as first-level tags. The meeting system's API interface is used to obtain the participants' historical meeting participation records, which include the historical meeting topics, issues, participation time, and interaction content. By using front-end tracking technology to record participants' actions on the conference document display page, knowledge preference data is generated. These actions include document opening time, scrolling depth, saving, sharing, and annotation.

3. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: During conference data collection, the BiLSTM-CRF model was used to classify the conference agenda into paragraph structures and extract agenda elements to construct an agenda element relationship graph and establish dependencies between topics. A domain-adjusted BERT model is used in conjunction with TF-IDF features and contextual semantic features. Confidence-based filtering and clustering are used to remove duplicates in order to extract keywords for the conference topics. The Sentence-BERT model is used to generate semantic vectors for project documents, and a multi-level association algorithm is used to calculate the association confidence between the semantic vectors and the current conference topics in order to filter out related project documents.

4. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: The steps to construct an interconnected knowledge network are as follows: Based on an open-source semantic model, entity recognition is performed on conference materials to extract core knowledge points, key figures, and time points. Using the meeting topic text as input, the system associates the meeting topic with core knowledge points through rule matching and prompt word matching from an open-source semantic model. Based on the core responsibility keywords extracted from the enterprise's organizational structure, a standardized role library containing "role ID + responsibility keywords" is formed. An open-source semantic model is used to automatically label knowledge points with domain tags and content keywords, establishing a connection between knowledge points and participant roles. Natural language processing technology is used to transform historical meeting conclusions and current topics into semantic vectors. Cosine similarity is used to calculate semantic relevance, and association rule algorithms are combined to mine high-frequency association rules to calculate association strength scores. Based on the association strength scores, the top N historical conclusions with the highest scores are selected. Knowledge graph storage technology is used to store the association relationships of the top N historical conclusions that have passed the verification in the form of triples in the graph, thereby establishing the association between historical meeting conclusions and current topics.

5. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: The knowledge network is dynamically updated. The methods for dynamic updating include: collecting user feedback on whether the knowledge points are relevant or irrelevant, using an incremental weight adjustment algorithm to update the knowledge point association weights, and automatically removing the corresponding knowledge points from the matching range when the weight is lower than the threshold.

6. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: The steps to generate a personalized list of pre-meeting materials recommendations are as follows: The current meeting topic is analyzed, and core knowledge points are extracted using an open-source semantic model to form a mapping list of "meeting topic - core knowledge points". Query the related knowledge network to obtain a list of participant roles associated with the core knowledge points, as well as historical meeting conclusions and materials related to the core knowledge points; Based on the participant's role information and data from the related knowledge network, the following tiered recommendations are made: a) Basic information on push notifications and role / responsibilities; b) Push historical conference conclusions and materials related to core knowledge points as supplementary information for the historical context; c) Filter out irrelevant information and remove duplicates to create a personalized pre-meeting information list for each role.

7. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: The recommendation strategy uses the following weighted formula to calculate the recommendation score of candidate knowledge items, and sorts and filters the candidate knowledge items according to the recommendation score to form a pre-meeting material recommendation list; The steps for calculating the recommendation score are as follows: Based on the knowledge association analysis results of the corresponding roles, the "role matching degree" is calculated; The cosine similarity between the semantic vector of the candidate knowledge item and the keywords of the conference topic is calculated to obtain the "conference topic relevance". By analyzing the behavioral preferences of similar users using a collaborative filtering algorithm, "historical preference similarity" is obtained. The recommended score is calculated using the following formula: Recommendation Score = α * Role Matching Degree + β * Historical Preference Similarity + γ * Meeting Topic Relevance + δ * Timeliness Where α, β, γ, and δ are the weighting coefficients for role matching degree, historical preference similarity, meeting topic relevance, and timeliness, respectively; weighting coefficient α is dynamically assigned based on role tags in the user profile; weighting coefficient β is based on collaborative filtering results, with a value range of 0.2 to 0.3; weighting coefficient γ is based on content-knowledge point matching results, with a value range of 0.2 to 0.3; and weighting coefficient δ is based on the knowledge item creation time.

8. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 1, characterized in that: The steps for dynamically adjusting recommendation strategy parameters based on interaction behavior data are as follows: Participants' interest in the materials was categorized based on their interactive behaviors, with interest levels classified as high, medium, and low relevance. Interactive behaviors included opening actions, reading time, and interactive actions. For data deemed low-relevance, the weight of historical preference similarity in the recommendation formula for the corresponding knowledge item is reduced, and the priority for future recommendations is lowered. For data deemed of moderate interest, the weight remains unchanged, and it is recorded as "potentially relevant". It will be recommended and verified again in conjunction with similar conference topics. For materials deemed to have high interest, the weights of the role matching degree and conference topic relevance of the corresponding knowledge items are increased, and the recommendation priority is raised.

9. The method for pushing pre-meeting materials based on user's professional field and meeting information according to claim 8, characterized in that: Materials that are opened for more than 5 minutes, or that have been "favorited" or "marked" are considered highly relevant; materials that are opened for 2-5 minutes and have no additional interaction are considered moderately relevant; materials that are not opened within a specified time or are opened for less than 1 minute are considered lowly relevant.

10. A pre-meeting materials push system based on user's professional field and meeting information, employing the pre-meeting materials push method based on user's professional field and meeting information as described in any one of claims 1 to 9, characterized in that: The system includes: The data acquisition module is used to collect and acquire authorized multi-dimensional data, which includes at least participant data, meeting data, and external knowledge base data. The knowledge organization module is used to process multi-dimensional data based on an open-source semantic model and build an associated knowledge network. The associated knowledge network is used to establish semantic relationships between meeting topics, core knowledge points, participant roles, and historical meeting conclusions. The personalized recommendation module is used to generate a personalized list of pre-meeting materials for participants based on the associated knowledge network and preset recommendation strategies when creating a target meeting. The recommendation strategy is based at least on role matching degree, historical preference similarity, and meeting topic relevance. The push and feedback module is used to push a list of recommended materials to the corresponding participants before the meeting, collect the participants' interaction data with the pushed materials, and dynamically adjust the parameters of the recommendation strategy based on the interaction data to optimize subsequent recommendations.

Images