Meeting minutes generation method and device, storage medium and computer device

Abstract

The application provides a conference summary generation method and device, a storage medium and a computer device. A page-level evidence set is obtained by analyzing current conference materials, and historical conference summaries are structured and sliced. On this basis, query information is extracted from the current conference materials and multiple retrieval expressions are generated, and the corresponding types of retrieval are executed, overcoming the problem of incomplete recall caused by a single retrieval method. At the same time, the user-visible conference set is calculated before retrieval, and the same project conference is screened from the user-visible conference set and the corresponding retrieval slice set is obtained, so that the retrieval range is always limited within the historical conferences that the user has access to. Then, the evidence package is constructed based on the multi-path candidate evidence list and injected into the summary generation model, effectively avoiding the context overload and historical fact pollution of the current conclusion caused by full injection, so that the scheme can improve the content accuracy of the conference summary generation and enhance the usability of the generated conference summary under the premise of ensuring the authority security.

Description

Technical Field

[0001] This application relates to the field of data retrieval technology, and in particular to a method, apparatus, storage medium, and computer equipment for generating meeting minutes. Background Technology

[0002] Meeting minutes generation typically involves transcribing meeting audio to create summaries, agenda items, decisions, to-do lists, and risks. With the development of large language models, meeting minutes generation can now incorporate uploaded meeting materials, templates, and company knowledge to produce more comprehensive content.

[0003] In long-term enterprise projects, industry professionals often retrieve historical meeting minutes from the same project to supplement the context and ensure the minutes maintain project continuity. However, injecting all meeting materials can contaminate the current minutes, and the mainstream single search method struggles to balance precise matching and semantic relevance, thus reducing the usability of the generated meeting minutes. Summary of the Invention

[0004] The purpose of this application is to at least address one of the aforementioned technical deficiencies, particularly the technical deficiencies in the prior art where the full injection of meeting materials may contaminate the current minutes, and the mainstream single retrieval method also struggles to balance precise matching and semantic relevance, thereby reducing the usability of the generated meeting minutes.

[0005] Firstly, this application provides a method for generating meeting minutes, the method comprising:

[0006] Obtain current meeting materials and historical meeting minutes, parse the current meeting materials to obtain a page-level evidence set, and perform structured slicing on the historical meeting minutes to obtain a set of minute slices;

[0007] Extract query information from the current meeting materials, and generate various search expressions based on the query information;

[0008] Calculate the set of visible meetings for the current user, filter the meetings that belong to the same project as the current meeting in the set of visible meetings to obtain the target meeting set, and obtain the set of search slices corresponding to the target meeting set from the set of minutes slices;

[0009] Each search expression is used to perform a search of the corresponding type within the search scope formed by the page-level evidence set and the search slice set, resulting in multiple candidate evidence lists;

[0010] An evidence package is constructed based on each candidate evidence list, and the evidence package is injected into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting.

[0011] In one embodiment, the step of extracting query information from the current meeting materials and generating multiple retrieval expressions based on the query information includes:

[0012] Perform semantic analysis on the current meeting materials to extract query terms, query phrases, and topic tags;

[0013] Multiple pre-configured search types are obtained, and corresponding search expressions are constructed for each search type by combining the query terms, query phrases, and topic tags, resulting in multiple search expressions.

[0014] In one embodiment, calculating the current user's visible meeting set, and filtering meetings belonging to the same project as the current meeting from the visible meeting set to obtain a target meeting set, includes:

[0015] Retrieve the pre-set meeting prohibition and authorization condition sets, as well as the current user's status information;

[0016] For each of all currently existing meetings, if the status information does not match any prohibition condition in the meeting prohibition condition set, and the status information matches any authorization condition in the meeting authorization condition set, the meeting is determined to be a visible meeting.

[0017] Once all visible meetings are identified, a set of visible meetings is formed based on each visible meeting.

[0018] Obtain the project code of the current meeting, filter the meetings with the same project code as the current meeting from the set of visible meetings, and form a target meeting set based on the filtered meetings.

[0019] In one embodiment, the step of performing a corresponding type of search within the search scope comprised of the page-level evidence set and the search slice set based on each search expression, to obtain multiple candidate evidence lists, including:

[0020] The scope formed by the page-level evidence set and the search slice set is defined as the target search scope;

[0021] For each search expression, determine the search type of the search expression, and perform evidence retrieval within the target search range based on the search expression and its search type to obtain a list of candidate evidence;

[0022] Once all search expressions have been completed, multiple candidate evidence lists are obtained.

[0023] In one embodiment, the retrieval type includes a sparse retrieval type and a vector retrieval type. The step of performing evidence retrieval within the target retrieval range based on the retrieval expression and its retrieval type to obtain a candidate evidence list includes:

[0024] When the search type is a sparse search type, keyword matching search is performed based on the search expression within the target search range to obtain a candidate evidence list.

[0025] When the retrieval type is a vector retrieval type, the retrieval expression is converted into a query vector. Within the target retrieval range, semantic similarity retrieval is performed based on the query vector to obtain a candidate evidence list.

[0026] In one embodiment, constructing the evidence package based on each candidate evidence list includes:

[0027] Multiple candidate pieces of evidence are determined by the union of the lists of candidate evidence.

[0028] For each candidate piece of evidence, determine at least one list of candidate pieces of evidence in which the candidate piece of evidence exists, and calculate the overall score of the candidate piece of evidence based on the determined list of candidate pieces of evidence.

[0029] When the comprehensive score of each candidate piece of evidence is calculated, the candidate pieces of evidence are sorted from high to low according to the comprehensive score to form a target evidence list. The target evidence list is then subjected to diversity screening to obtain an evidence package.

[0030] In one embodiment, after injecting the evidence package into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting, the method further includes:

[0031] The meeting minutes are subject to citation verification. If an illegal citation is found in the meeting minutes, the minutes generation model is triggered to regenerate the minutes, or the illegal citation is downgraded to a no-citation minutes, until the meeting minutes pass the citation verification.

[0032] Secondly, this application provides a meeting minutes generation apparatus, the apparatus comprising:

[0033] The data acquisition module is used to acquire current meeting materials and historical meeting minutes, parse the current meeting materials to obtain a page-level evidence set, and perform structured slicing on the historical meeting minutes to obtain a set of minute slices;

[0034] The information extraction module is used to extract query information from the current meeting materials and generate various retrieval expressions based on the query information;

[0035] The scope determination module is used to calculate the set of visible meetings for the current user, filter the meetings belonging to the same project as the current meeting in the set of visible meetings to obtain the target meeting set, and obtain the retrieval slice set corresponding to the target meeting set from the set of minutes slices;

[0036] The evidence retrieval module is used to perform corresponding type of retrieval within the retrieval range consisting of the page-level evidence set and the retrieval slice set according to each retrieval expression, and obtain multiple candidate evidence lists;

[0037] The minutes generation module is used to construct evidence packages based on each candidate evidence list and inject the evidence packages into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting.

[0038] Thirdly, this application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the meeting minutes generation method as described in any of the above embodiments.

[0039] Fourthly, this application provides a computer device, including: one or more processors, and a memory;

[0040] The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the meeting minutes generation method as described in any of the above embodiments.

[0041] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0042] The meeting minutes generation method, apparatus, storage medium, and computer equipment provided in this application obtain a page-level evidence set by parsing current meeting materials and perform structured slicing of historical meeting minutes. Based on this, query information is extracted from the current meeting materials and multiple retrieval expressions are generated, each executing a corresponding type of retrieval. This approach balances precise keyword matching and semantically relevant recall, overcoming the incomplete recall problem caused by a single retrieval method. Simultaneously, before retrieval, the set of meetings visible to the user is calculated, and meetings from the same project are selected to obtain the corresponding retrieval slice set, ensuring that the retrieval scope is always limited to historical meetings accessible to the user. Then, an evidence package is constructed based on a multi-channel candidate evidence list and injected into the minutes generation model, effectively avoiding context overload and historical fact contamination of current conclusions caused by full injection. This allows the solution to improve the accuracy of generated meeting minutes and enhance their usability while ensuring access security. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a method for generating meeting minutes provided in an embodiment of this application;

[0045] Figure 2 An example diagram illustrating a meeting minutes generation method provided in this application embodiment;

[0046] Figure 3 A schematic diagram of a meeting minutes generation device provided in this application embodiment;

[0047] Figure 4 This is an internal structural diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] In one embodiment, this application provides a method for generating meeting minutes, and the following embodiments illustrate the application of this method to a server. It is understood that the method for generating meeting minutes can be performed by a single server or by a server cluster consisting of multiple servers; this application does not impose any specific limitations on this.

[0050] like Figure 1 As shown, this application provides a method for generating meeting minutes, the method comprising:

[0051] S101: Obtain current meeting materials and historical meeting minutes, parse the current meeting materials to obtain a page-level evidence set, and perform structured slicing of historical meeting minutes to obtain a set of minute slices.

[0052] The current meeting materials refer to files uploaded during the current meeting, including but not limited to PDF, DOCX, TXT, Markdown, and other formats, as well as files obtained by converting meeting audio to text. Historical meeting minutes refer to the minutes of past meetings for which minutes have already been generated. A page-level evidence set refers to a collection of evidence units formed by segmenting the current meeting materials according to their original page numbers or logical page dimensions. Each unit contains at least a meeting identifier, file name, page number or logical page sequence number, main text, and text hash, supporting page-by-page tracing and citation. A minutes slice set refers to a collection of retrieval units formed by segmenting historical meeting minutes according to structured field paths or chapters / paragraphs, with each slice serving as an independent retrieval granularity.

[0053] In this step, when the server receives a minutes generation instruction, it determines the meeting for which minutes are to be generated based on the instruction and retrieves the meeting materials and historical meeting minutes. When parsing the current meeting materials, the file type is first identified, and parsing is performed in an appropriate manner based on the file type. Specifically, for PDF files, the text content of each page can be extracted according to the original page numbers to generate page-level evidence; for formats such as DOCX, TXT, and Markdown, logical page evidence can be divided according to logical pages, chapters, paragraphs, or length thresholds. Each page-level evidence, in addition to preserving the main text, also includes information such as the filename, page number, and text hash to ensure that subsequent citations can accurately trace back to a specific page in the original file.

[0054] Meanwhile, when structuring and slicing historical meeting minutes, the server can prioritize identifying the structured fields of the minutes. If the minutes contain explicit fields such as key decisions, to-do items, risk issues, and work deployments, then each minute is sliced ​​according to the field path, and each slice independently saves its associated meeting, minutes version, slice type, and main text content. If the minutes are in Markdown or unstructured format, then they are sliced ​​according to heading level and paragraph boundaries, with excessively long paragraphs further split according to semantic boundaries. The final result is a set of minutes slices.

[0055] S102: Extract query information from the current meeting materials and generate various search expressions based on the query information.

[0056] The query information refers to structured data extracted from the current conference materials to characterize the retrieval needs of this conference, including but not limited to query terms, query phrases, and topic tags. The retrieval expression refers to an expression or input constructed based on the query information to perform a specific type of retrieval.

[0057] In this step, when extracting query information from the current meeting materials, the server can first gather multi-source input from the current meeting materials, including the current meeting title, speech-to-text subtitles, module names in the minutes template, project name snapshots, project hot keywords, filenames of uploaded materials, and additional content specified by the user. Then, these inputs are cleaned to filter out noise such as colloquial words, low-value function words, speaker tags, timestamp residue, and meaningless numbers, retaining words with substantial search value such as client names, project names, version numbers, technical terms, risk items, to-do keywords, and filename keywords.

[0058] Subsequently, when generating search expressions based on query information, the extracted vocabulary can be organized into various search expressions to adapt to different search paths. For example, for sparse search, query terms and phrases can be organized into keyword expressions, and each word can be weighted according to factors such as word origin, word frequency, and whether it is a proper noun, to strengthen the matching priority of important words. For vector search, query terms, query phrases, and topic tags can be combined into query text, and then converted into dense vectors through a vector model as input for semantic search. The same query information can simultaneously generate both sparse index-oriented and vector index-oriented expressions. In addition, other types of search expressions can be generated based on query information to ensure parallel execution of multi-path searches.

[0059] S103: Calculate the set of visible meetings for the current user, filter the meetings that belong to the same project as the current meeting in the set of visible meetings to obtain the target set of meetings, and obtain the set of search slices corresponding to the target set of meetings in the set of minutes slices.

[0060] The visible meeting set refers to the set of all meetings that the current user has the right to view. The retrieval slice set refers to a subset of the minutes slice set, which consists of the minutes slices of historical meetings on the same project as the current meeting.

[0061] In this step, when calculating the set of visible meetings for the current user, access permissions for each meeting can be determined session by session based on the user's status. The determination logic can consist of both prohibition conditions and authorization branches. For example, deleted meetings are first excluded, and then it is determined whether the user meets any authorization condition, such as whether the user is the creator of the meeting, or whether the meeting is public and completed, and the user is in the list of participants or recipients of the meeting. It is important to note that even if a user is listed as a member of a project, it does not mean that they automatically have read access to all meetings under that project.

[0062] Based on the established set of visible meetings, further filtering is performed to identify meetings belonging to the same project as the current meeting, resulting in a target set of meetings. In other words, each meeting in this target set meets two conditions: first, the user has the right to view it; and second, it belongs to the same project / business category as the current meeting. Subsequently, multiple slices corresponding to each meeting in the target set are located in the set of minutes slices, forming a search slice set. It can be understood that since the minutes slices have already saved their respective meeting identifiers during the indexing phase, the meeting identifiers in the target set are used as filtering conditions to extract minutes slices belonging to these meetings and representing the currently selected completed versions from the shared minutes slice index, thus forming the search slice set.

[0063] Before a search is performed, the scope of historical meeting searches is limited to valid minutes slices that the current user has the right to view and that are from the same project as the current meeting. This completely decouples meeting access permissions from project member identities, ensuring that users cannot access any meeting content they are not authorized to, thus eliminating the risk of permission overreach at the source. At the same time, by converging within the same project, subsequent searches are only performed within the business-related historical context, effectively avoiding cross-project noise interference.

[0064] S104: Perform the corresponding type of search within the search scope consisting of the page-level evidence set and the search slice set according to each search expression to obtain multiple candidate evidence lists.

[0065] The candidate evidence list refers to the set of evidence obtained after performing a search of the type corresponding to the search expression within a specified search range based on the corresponding search expression. Each evidence record in the list includes at least an identifier, source type, conference or document to which it belongs, evidence text fragment, and relevance score.

[0066] In this step, before executing the search, the search scope is first determined. This scope consists of the union of the current meeting's page-level evidence set and the search slice set, covering both currently uploaded materials and authorized historical meeting data. Then, each previously generated search expression is traversed, and evidence retrieval is performed on the corresponding index based on its search type, resulting in multiple candidate evidence lists. For example, for sparse search expressions, keyword matching can be performed on the sparse index to recall page-level evidence or minutes slices containing the required query terms, query phrases, or topic tags, and scores and rankings are calculated using text relevance algorithms such as BM25. For vector search expressions, the query text can be converted into a query vector, and semantically similar evidence records are retrieved from the vector index and sorted by similarity.

[0067] It is understandable that each retrieval unit (i.e., evidence or slice) in the page-level evidence set and retrieval slice set has multiple indexes that can correspond to various retrieval expressions. For example, when the retrieval type includes sparse retrieval and vector retrieval, sparse indexes and vector indexes can be created when parsing the current meeting materials and historical meeting minutes slices. Sparse indexes are used for exact matching of keywords, client names, version numbers, file names, etc.; vector indexes are used for recalling semantically similar content. In addition, some auxiliary indexes can be built, such as text hashes, page numbers, file names, and subject terms, mainly for assisting retrieval, deduplication, displaying references, and result tracing. In specific implementations, they can be maintained as record fields, inverted index auxiliary items, filtering conditions, or lightweight secondary indexes as needed.

[0068] Furthermore, each search outputs an independent list of candidate evidence. If the service corresponding to a certain search is unavailable or the corresponding index is missing, then only the search corresponding to the remaining search expression can be executed.

[0069] S105: Construct an evidence package based on each candidate evidence list, and inject the evidence package into the preset minutes generation model to obtain the meeting minutes corresponding to the current meeting.

[0070] The evidence package refers to the list of evidence obtained by fusing and rearranging multiple candidate evidence lists. The minutes generation model is a text generation network based on a large language model. It takes the evidence package and generation constraints of the current meeting as input and outputs a structured meeting minutes according to a specified template. This model can be a general-purpose large language model or a domain-specific model fine-tuned for the specific context. The meeting minutes are a structured document output by the minutes generation model, containing modules such as meeting topic, discussion summary, key decisions, to-do items, risk issues, and work arrangements. It may also include page number references from the evidence package or historical meeting information.

[0071] In this step, the server can merge and rearrange the candidate evidence lists, and then use the merged and rearranged evidence package as input to the pre-trained minutes generation model to obtain the meeting minutes corresponding to the current meeting. Specifically, when constructing the evidence package, logical partitions can be made according to two types: current meeting materials and historical meeting minutes, so that the model can adopt different weights or strategies based on different logical partitions. In addition, each selected piece of evidence in the current meeting materials dimension of the evidence package is accompanied by complete traceability metadata, including the file name, page number or logical page number, the matched evidence fragment, the search relevance score, the matched query terms, and the fusion score. For the historical meeting minutes dimension, the selected evidence in the evidence package needs to save its meeting identifier, meeting title, minutes version, search relevance score, matched query terms, and fusion score.

[0072] Specifically, when injecting evidence packages into the minutes generation model, the generation constraints can be as follows: current meeting data evidence takes precedence over historical meeting background; historical meeting evidence is only used as background reference, tracing clues, or basis for continuing matters; historical conclusions must not be written as new conclusions of the current meeting; file name and page number citations output by the model must come from the current data evidence section; historical meeting citations output must come from the historical meeting evidence section; and citations must not be output if there is no corresponding evidence.

[0073] Furthermore, the minutes generation model can be trained using conventional supervised learning algorithms based on its inputs and outputs, and this application does not impose any specific restrictions on this.

[0074] In this embodiment, a page-level evidence set is obtained by parsing the current meeting materials, and the historical meeting minutes are structured and sliced. Based on this, query information is extracted from the current meeting materials, and multiple search expressions are generated. These expressions are then executed for corresponding search types, balancing precise keyword matching and semantically relevant recall, overcoming the incomplete recall problem caused by a single search method. Simultaneously, before the search, the set of meetings visible to the user is calculated, and meetings from the same project are selected to obtain the corresponding search slice set, ensuring that the search scope is always limited to historical meetings that the user has access to. Then, an evidence package is constructed based on a multi-channel candidate evidence list and injected into the minutes generation model. This effectively avoids context overload and historical fact contamination of current conclusions caused by full injection, enabling this solution to improve the accuracy of the generated meeting minutes and enhance their usability while ensuring access security.

[0075] In one embodiment, query information is extracted from the current meeting materials, and various retrieval expressions are generated based on the query information, including:

[0076] S1: Perform semantic analysis on the current meeting materials to extract query terms, query phrases, and topic tags.

[0077] S2: Obtain multiple pre-configured search types, and construct corresponding search expressions for each search type by combining query terms, query phrases, and topic tags to obtain multiple search expressions.

[0078] Here, query terms refer to core words with retrieval value extracted from the current meeting materials. Query phrases refer to short sentences or combinations of terms with clear semantic meaning, consisting of multiple consecutive words, such as compatibility test reports or reasons for deployment failures. Topic tags refer to general category identifiers extracted from the current meeting materials, such as version releases or performance optimizations. Search type refers to the pre-configured search path category.

[0079] In this embodiment, when performing semantic analysis on the current meeting materials, multi-source input can be gathered first, including the current meeting title, speech-to-text subtitles, minutes template module name, project name snapshot, project hot words, uploaded material file names, and user-specified context. Then, natural language processing techniques are used to segment, tag, and name entity recognition on these texts, extracting words with substantial retrieval value as query terms, including client names, project names, version numbers, deployment methods, model names, risk items, pending keywords, technical terms, and file name keywords. For combinations of multiple consecutive words with clear semantic directions, they can be identified as query phrases, such as "database migration plan" or "grayscale release strategy." These phrases require complete word order matching during retrieval. Furthermore, general topic tags can be summarized from the text, such as extracting the topic tag "version release" from words like "v3.2.0 launch," "rollback," and "compatibility," to guide the semantic direction of vector retrieval. This process simultaneously filters out noise such as colloquial words, low-value function words, speaker tags, timestamp residues, and meaningless numbers, ensuring the quality of subsequent searches based on this information.

[0080] After semantic analysis, multiple pre-configured search types are obtained, and corresponding search expressions are constructed for each type. For example, for sparse search types, query terms and phrases are organized into keyword expressions. Weights are assigned to each term based on factors such as word origin, frequency, and whether it is a proper noun, to enhance the matching priority of important words in text relevance calculations such as BM25. For vector search types, query terms, phrases, and topic tags are combined into a semantically complete query text, which is then converted into a dense vector representation using a specified vector model, serving as input for semantic similarity retrieval.

[0081] Understandably, query terms ensure accurate matching of proper nouns and technical terms, query phrases preserve the sequential integrity of key information, and topic tags provide clear semantic guidance for vector retrieval. The parallel generation of multiple search expressions enables subsequent searches to consider both precise keyword matching and semantically relevant recall, improving the accuracy and comprehensiveness of evidence retrieval.

[0082] In one embodiment, the set of visible meetings for the current user is calculated, and meetings belonging to the same project as the current meeting are filtered from the set of visible meetings to obtain a target set of meetings, including:

[0083] S1: Retrieve the pre-set meeting prohibition and authorization condition sets, as well as the current user's status information.

[0084] S2: For each of the currently existing meetings, if the status information does not match any prohibition condition in the meeting prohibition condition set, but the status information matches any authorization condition in the meeting authorization condition set, then the meeting is determined to be a visible meeting.

[0085] S3: Once all visible meetings are determined, a set of visible meetings is formed based on each visible meeting.

[0086] S4: Obtain the project code of the current meeting, filter the meetings with the same project code as the current meeting from the visible meeting set, and form a target meeting set based on the filtered meetings.

[0087] The meeting prohibition condition set refers to the set of negative conditions used to determine whether a user can see the meeting. The meeting authorization condition set refers to the set of positive conditions used to determine whether a user can see the meeting. Status information refers to the set of attributes describing the relationship between the current user and the meeting, including user identifier, the set of meetings created by the user, the set of meetings the user has attended, the set of meetings the user has shared with, the public attribute of each meeting, completion status, etc.

[0088] In this embodiment, a pre-configured set of meeting prohibition conditions and authorization conditions is first obtained. The prohibition condition set defines absolute exclusion rules for meetings that are not visible to users, including at least the prohibition condition that the meeting has been deleted, ensuring that removed meetings will not be included in any user's visibility. The authorization condition set defines affirmative rules for users to obtain meeting visibility permissions, including multiple parallel authorization branches. Simultaneously, the current user's status information is obtained, which includes the user's identifier, the specific relationship between the user and each meeting, and the current attribute status of each meeting. For each of all currently existing meetings, if the status information does not match any prohibition condition in the meeting prohibition condition set, and the status information matches any authorization condition in the meeting authorization condition set, the meeting is determined to be a visible meeting. Conversely, if the status information matches any prohibition condition in the meeting prohibition condition set, or if the status information does not match any prohibition condition in the meeting prohibition condition set and does not match any authorization condition in the meeting authorization condition set, the meeting is a non-visible meeting.

[0089] Once all visible meetings are identified, a set of visible meetings for the current user is formed. This set is further narrowed down using project coding to obtain a subset of visible meetings belonging to the same project as the current meeting—the target set of meetings. In this process, the meeting prohibition condition set excludes invalid meetings such as those that have been deleted, and the meeting authorization condition set replaces ambiguous project member authorization with explicit rules such as creator, public access, and completion. This achieves complete decoupling of meeting read permissions from project ownership, fundamentally eliminating the risk of unauthorized access caused by direct retrieval via project tags.

[0090] For example, when Zhang San, an employee of a company, requests to generate the minutes of the current meeting, all meetings are evaluated one by one: a meeting that has been deleted directly meets the prohibition condition and is excluded; a solution review meeting created by Zhang San meets the authorization condition and is included in the visible meeting set because he is the creator; a project weekly meeting marked as public and completed is included in the visible meeting set because Zhang San is in the list of participants; while another architecture review meeting belonging to the same project is not included in the visible meeting set because it is not public and Zhang San is not a participant or presenter, even though Zhang San is a member of the project.

[0091] In one embodiment, a corresponding type of search is performed within the search scope comprised of the page-level evidence set and the search slice set, based on each search expression, to obtain multiple candidate evidence lists, including:

[0092] S1: Define the scope consisting of the page-level evidence set and the search slice set as the target search scope.

[0093] S2: For each search expression, determine the search type of the search expression, and perform evidence retrieval within the target search scope based on the search expression and its search type to obtain a list of candidate evidence.

[0094] S3: After all the search expressions have been completed, a list of multiple candidate evidences is obtained.

[0095] In this embodiment, the target retrieval scope is defined by the page-level evidence set and the retrieval slice set. For each retrieval expression, the retrieval type is determined. Then, based on the retrieval expression and its retrieval type, evidence of the corresponding type is retrieved from both the page-level evidence set and the retrieval slice set within the target retrieval scope, resulting in a candidate evidence list for that type of retrieval. This list contains evidence records retrieved from both the page-level evidence set and the retrieval slice set. By performing a retrieval on the corresponding index according to the type of each retrieval expression, parallel retrieval of evidence of multiple types and paths is achieved. Each retrieval outputs an independent candidate evidence list, preserving the original sorting information of each retrieval path, providing a rich and heterogeneous candidate input foundation for subsequent fusion and rearrangement.

[0096] In one embodiment, the retrieval types include sparse retrieval and vector retrieval. Evidence is retrieved within the target retrieval range based on the retrieval expression and its retrieval type to obtain a candidate evidence list, including:

[0097] S1: When the search type is a sparse search type, within the target search range, keyword matching search is performed based on the search expression to obtain a list of candidate evidence.

[0098] S2: When the retrieval type is vector retrieval, the retrieval expression is converted into a query vector. Within the target retrieval range, semantic similarity retrieval is performed based on the query vector to obtain a list of candidate evidence.

[0099] In this embodiment, the retrieval type corresponding to the current retrieval expression is identified. When the retrieval type is a sparse retrieval type, the query terms and query phrases in the retrieval expression are directly input as keywords, and keyword matching retrieval is performed on the sparse index within the target retrieval range. The sparse index is built based on an inverted index structure, which can quickly locate page-level evidence or summary slices containing specific keywords. Furthermore, text relevance algorithms such as BM25 are used to calculate the relevance score between each candidate piece of evidence and the query expression. This algorithm comprehensively considers factors such as term frequency, inverse document frequency, and document length, so that evidence containing multiple query terms and densely distributed keywords obtains a higher score.

[0100] When the search type is vector retrieval, the query terms, query phrases, and topic tags in the current search expression are first combined into a semantically complete query text, and then converted into a dense query vector using a specified vector model. Subsequently, semantic similarity retrieval is performed on the vector index within the target retrieval range, calculating the similarity between the query vector and the vector representation of each evidence record in the index. Cosine similarity or inner product is typically used as a metric. The vector index is built based on the approximate nearest neighbor algorithm, enabling efficient retrieval of the few evidence records most similar to the query vector from a large-scale vector set.

[0101] By differentiating search types and performing differentiated search operations, sparse search can accurately hit proper nouns and technical terms on the inverted index, while vector search can effectively recall semantically similar content on the vector index. The two methods work in parallel and complement each other, overcoming the shortcomings of incomplete recall by a single search method.

[0102] In one embodiment, an evidence package is constructed based on each candidate evidence list, including:

[0103] S1: Determine multiple candidate pieces of evidence based on the union of the lists of candidate evidence.

[0104] S2: For each candidate piece of evidence, determine at least one list of candidate evidence to which the candidate piece of evidence exists, and calculate the overall score of the candidate piece of evidence based on the determined list of candidate evidence.

[0105] S3: When the comprehensive score of each candidate piece of evidence is calculated, the candidate pieces of evidence are sorted from high to low according to the comprehensive score to form a target evidence list. The target evidence list is then subjected to diversity screening to obtain an evidence package.

[0106] The overall score refers to the quantitative score calculated by the fusion algorithm based on the ranking position of the same candidate evidence in multiple candidate evidence lists. It is used to comprehensively measure the overall relevance of the candidate evidence in multi-path retrieval. The higher the overall score, the higher the ranking of the evidence in more retrieval paths and the greater its overall value.

[0107] In this embodiment, the union of all evidence from each candidate evidence list is first taken to determine all unique candidate evidence retrieved. For each candidate evidence in the union, it is identified which candidate evidence lists it exists in. A comprehensive score is calculated based on the candidate evidence's ranking in each of these lists. Specifically, a reciprocal ranking fusion algorithm can be used. If the evidence's ranking in a list is r, then that list contributes 1 / (k+r), where k is a configurable constant. If the evidence does not appear in a list, that list contributes nothing. The contribution scores of all lists are summed to obtain the comprehensive score for that evidence. After sorting all candidate evidence according to their comprehensive scores, the target evidence list is obtained.

[0108] Subsequently, diversity filtering was performed on the target evidence list. Specifically, the sorted candidate evidence was traversed sequentially, and deduplication was controlled based on dimensions such as text hash, file name, page number, meeting identifier, slice type, structured field path, topic tags, and similarity of similar text. Quantity limits were also imposed, including limits on current document evidence, historical meeting evidence, total evidence, and the percentage of the same document or meeting. The final filtering result is the evidence package, which retains only a limited number of evidences with high overall scores, diverse sources, and dispersed topics.

[0109] For example, in a version review meeting, the union of the sparse candidate evidence list and the vector candidate evidence list contains 12 candidate pieces of evidence. One of the minutes slices about the "rollback plan" appears simultaneously in the 3rd position of the sparse evidence list and the 2nd position of the vector evidence list. Assuming the configurable constant is 60, its overall score is 1 / (60+3)+1 / (60+2)≈0.0323.

[0110] By integrating multi-way search results through a ranking-based fusion algorithm, evidence that is simultaneously hit by multiple searches receives a higher score. This achieves unified ranking across search types, and the diversity filtering can also control the concentration of sources and topics based on the ranking, preventing a single document or conference from occupying the entire evidence space.

[0111] In one embodiment, after injecting the evidence package into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting, the meeting minutes generation method further includes:

[0112] The meeting minutes are subject to citation verification. If an illegal citation is found in the meeting minutes, the minutes generation model is triggered to regenerate the minutes, or the illegal citation is downgraded to a no-citation minutes, until the meeting minutes pass the citation verification.

[0113] Illegal citation refers to a situation in which the meeting minutes claim to cite information from specific materials or historical meetings, but the corresponding basis for such citation cannot be found in the evidence package generated this time.

[0114] In this embodiment, the meeting minutes output by the model are validated to ensure that every citation can be traced back to its source in the generated evidence package. Illegal citations are retried or downgraded until the meeting minutes pass validation. This effectively constrains the model to use the evidence package for accurate source citation, preventing the model from fabricating or misciting information not in the input. This ensures that every citation in the meeting minutes is verifiable and traceable, improving the accuracy and usability of the generated meeting minutes.

[0115] In one embodiment, this solution can also generate and save a snapshot of the meeting minutes along with the corresponding version after the minutes are generated. This snapshot includes at least a caption snapshot, a data snapshot, an evidence package snapshot, a query information snapshot, and an index version. This way, even if subsequent document retransmissions, data updates, index reconstructions, project renamings, or historical minute version switching occur, the generated minutes can still be audited based on the snapshot taken at that time.

[0116] In one embodiment, such as Figure 2 As shown, Figure 2 This diagram illustrates a meeting minutes generation method provided in an embodiment of this application. The meeting creation request is the minutes generation instruction, and the query plan is a query scheme generated from query information, which can be used to transform and generate retrieval expressions. Furthermore, in hybrid recall scenarios, degradation processing can be performed if abnormal situations are encountered to ensure the stable operation of the minutes generation process.

[0117] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0118] The meeting minutes generation apparatus provided in the embodiments of this application is described below. The meeting minutes generation apparatus described below and the meeting minutes generation method described above can be referred to in correspondence.

[0119] like Figure 3 As shown, this application provides a meeting minutes generation device 200, the device comprising:

[0120] The data acquisition module 201 is used to acquire current meeting materials and historical meeting minutes, parse the current meeting materials to obtain a page-level evidence set, and perform structured slicing of historical meeting minutes to obtain a set of minute slices;

[0121] The information extraction module 202 is used to extract query information from the current meeting materials and generate various retrieval expressions based on the query information;

[0122] The scope determination module 203 is used to calculate the set of visible meetings for the current user, filter the meetings that belong to the same project as the current meeting in the set of visible meetings to obtain the target meeting set, and obtain the retrieval slice set corresponding to the target meeting set from the set of minutes slices;

[0123] The evidence retrieval module 204 is used to perform corresponding type of retrieval within the retrieval scope consisting of the page-level evidence set and the retrieval slice set according to each retrieval expression, and obtain multiple candidate evidence lists.

[0124] The minutes generation module 205 is used to construct evidence packages based on each candidate evidence list and inject the evidence packages into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting.

[0125] In this embodiment, a page-level evidence set is obtained by parsing the current meeting materials, and the historical meeting minutes are structured and sliced. Based on this, query information is extracted from the current meeting materials, and multiple search expressions are generated. These expressions are then executed for corresponding search types, balancing precise keyword matching and semantically relevant recall, overcoming the incomplete recall problem caused by a single search method. Simultaneously, before the search, the set of meetings visible to the user is calculated, and meetings from the same project are selected to obtain the corresponding search slice set, ensuring that the search scope is always limited to historical meetings that the user has access to. Then, an evidence package is constructed based on a multi-channel candidate evidence list and injected into the minutes generation model. This effectively avoids context overload and historical fact contamination of current conclusions caused by full injection, enabling this solution to improve the accuracy of the generated meeting minutes and enhance their usability while ensuring access security.

[0126] In one embodiment, the information extraction module includes:

[0127] The semantic analysis submodule is used to perform semantic analysis on the current meeting materials, extracting query terms, query phrases, and topic tags.

[0128] The expression generation submodule is used to obtain a variety of pre-configured search types, and to construct corresponding search expressions for each search type by combining query terms, query phrases and topic tags, thus obtaining a variety of search expressions.

[0129] In one embodiment, the range determination module includes:

[0130] The information acquisition submodule is used to acquire the pre-set meeting prohibition condition set and meeting authorization condition set, as well as the current user's status information;

[0131] The meeting matching submodule is used to determine a meeting as a visible meeting for each meeting among all existing meetings when the status information does not match any prohibition condition in the meeting prohibition condition set, but matches any authorization condition in the meeting authorization condition set.

[0132] The set determination submodule is used to form a set of visible meetings based on each visible meeting after all visible meetings have been determined.

[0133] The meeting filtering submodule is used to obtain the project code of the current meeting, filter the meetings with the same project code as the current meeting from the visible meeting set, and form a target meeting set based on the filtered meetings.

[0134] In one embodiment, the evidence retrieval module includes:

[0135] The scope determination submodule is used to determine the target search scope as the scope consisting of the page-level evidence set and the search slice set;

[0136] The evidence retrieval submodule is used to determine the retrieval type of each retrieval expression, and to perform evidence retrieval within the target retrieval range based on the retrieval expression and its retrieval type to obtain a list of candidate evidence.

[0137] The list determination submodule is used to obtain a list of multiple candidate evidences after all search expressions have been searched.

[0138] In one embodiment, the evidence retrieval submodule includes:

[0139] The first retrieval unit is used to perform keyword matching retrieval based on the retrieval expression within the target retrieval range when the retrieval type is a sparse retrieval type, and obtain a list of candidate evidence.

[0140] The second retrieval unit is used to convert the retrieval expression into a query vector when the retrieval type is vector retrieval, and to perform semantic similarity retrieval based on the query vector within the target retrieval range to obtain a list of candidate evidence.

[0141] In one embodiment, the minutes generation module includes:

[0142] The evidence determination submodule is used to determine multiple candidate pieces of evidence based on the union of the lists of candidate evidence.

[0143] The scoring calculation submodule is used to determine, for each candidate piece of evidence, at least one list of candidate evidence to which the candidate evidence exists, and to calculate the comprehensive score of the candidate evidence based on the determined list of candidate evidence.

[0144] The evidence ranking submodule is used to rank each candidate piece of evidence from highest to lowest based on the comprehensive score when the comprehensive score of each candidate piece of evidence is calculated, forming a target evidence list, and then performing diversity filtering on the target evidence list to obtain an evidence package.

[0145] In one embodiment, the meeting minutes generation apparatus further includes:

[0146] The citation verification module is used to verify the citations in the meeting minutes. If there are illegal citations in the meeting minutes, the minutes generation model is triggered to regenerate the minutes, or the illegal citation is downgraded to a no-citation minutes, until the meeting minutes pass the citation verification.

[0147] The division of modules in the above-described meeting minutes generation device is merely illustrative. In other embodiments, the meeting minutes generation device can be divided into different modules as needed to complete all or part of its functions. Each module in the above-described meeting minutes generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0148] In one embodiment, this application also provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the meeting minutes generation method as described in any of the above embodiments.

[0149] In one embodiment, this application also provides a computer device storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the meeting minutes generation method as described in any of the above embodiments.

[0150] Indicatively, such as Figure 4 As shown, Figure 4 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 4 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions executable by the processing component 302, such as application programs. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the meeting minutes generation method of any of the above embodiments.

[0151] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0152] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0153] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, the singular forms "a," "an," and "the" may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having” specify the presence of the stated features, wholes, steps, operations, components, parts or combinations thereof, but do not exclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0154] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0155] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating meeting minutes, characterized in that, The method includes: Obtain current meeting materials and historical meeting minutes, parse the current meeting materials to obtain a page-level evidence set, and perform structured slicing on the historical meeting minutes to obtain a set of minute slices; Extract query information from the current meeting materials, and generate various search expressions based on the query information; Calculate the set of visible meetings for the current user, filter the meetings that belong to the same project as the current meeting in the set of visible meetings to obtain the target meeting set, and obtain the set of search slices corresponding to the target meeting set from the set of minutes slices; Each search expression is used to perform a search of the corresponding type within the search scope formed by the page-level evidence set and the search slice set, resulting in multiple candidate evidence lists; An evidence package is constructed based on each candidate evidence list, and the evidence package is injected into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting.

2. The meeting minutes generation method according to claim 1, characterized in that, The step of extracting query information from the current meeting materials and generating various search expressions based on the query information includes: Perform semantic analysis on the current meeting materials to extract query terms, query phrases, and topic tags; Multiple pre-configured search types are obtained, and corresponding search expressions are constructed for each search type by combining the query terms, query phrases, and topic tags, resulting in multiple search expressions.

3. The meeting minutes generation method according to claim 1, characterized in that, The calculation of the current user's visible meeting set, followed by filtering meetings belonging to the same project as the current meeting from the visible meeting set to obtain the target meeting set, includes: Retrieve the pre-set meeting prohibition and authorization condition sets, as well as the current user's status information; For each of all currently existing meetings, if the status information does not match any prohibition condition in the meeting prohibition condition set, and the status information matches any authorization condition in the meeting authorization condition set, the meeting is determined to be a visible meeting. Once all visible meetings are identified, a set of visible meetings is formed based on each visible meeting. Obtain the project code of the current meeting, filter the meetings with the same project code as the current meeting from the set of visible meetings, and form a target meeting set based on the filtered meetings.

4. The meeting minutes generation method according to claim 1, characterized in that, The steps involve performing corresponding types of searches within the search scope comprised of the page-level evidence set and the search slice set, based on each search expression, to obtain multiple candidate evidence lists, including: The scope formed by the page-level evidence set and the search slice set is defined as the target search scope; For each search expression, determine the search type of the search expression, and perform evidence retrieval within the target search range based on the search expression and its search type to obtain a list of candidate evidence; Once all search expressions have been completed, multiple candidate evidence lists are obtained.

5. The meeting minutes generation method according to claim 4, characterized in that, The retrieval types include sparse retrieval and vector retrieval. The evidence retrieval within the target retrieval range, based on the retrieval expression and its retrieval type, yields a candidate evidence list, including: When the search type is a sparse search type, keyword matching search is performed based on the search expression within the target search range to obtain a candidate evidence list. When the retrieval type is a vector retrieval type, the retrieval expression is converted into a query vector. Within the target retrieval range, semantic similarity retrieval is performed based on the query vector to obtain a candidate evidence list.

6. The meeting minutes generation method according to claim 1, characterized in that, The construction of the evidence package based on each candidate evidence list includes: Multiple candidate pieces of evidence are determined by the union of the lists of candidate evidence. For each candidate piece of evidence, determine at least one list of candidate pieces of evidence in which the candidate piece of evidence exists, and calculate the overall score of the candidate piece of evidence based on the determined list of candidate pieces of evidence. When the comprehensive score of each candidate piece of evidence is calculated, the candidate pieces of evidence are sorted from high to low according to the comprehensive score to form a target evidence list. The target evidence list is then subjected to diversity screening to obtain an evidence package.

7. The method for generating meeting minutes according to any one of claims 1 to 6, characterized in that, After injecting the evidence package into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting, the method further includes: The meeting minutes are subject to citation verification. If an illegal citation is found in the meeting minutes, the minutes generation model is triggered to regenerate the minutes, or the illegal citation is downgraded to a no-citation minutes, until the meeting minutes pass the citation verification.

8. A meeting minutes generation device, characterized in that, The device includes: The data acquisition module is used to acquire current meeting materials and historical meeting minutes, parse the current meeting materials to obtain a page-level evidence set, and perform structured slicing on the historical meeting minutes to obtain a set of minute slices; The information extraction module is used to extract query information from the current meeting materials and generate various retrieval expressions based on the query information; The scope determination module is used to calculate the set of visible meetings for the current user, filter the meetings belonging to the same project as the current meeting in the set of visible meetings to obtain the target meeting set, and obtain the retrieval slice set corresponding to the target meeting set from the set of minutes slices; The evidence retrieval module is used to perform corresponding type of retrieval within the retrieval range consisting of the page-level evidence set and the retrieval slice set according to each retrieval expression, and obtain multiple candidate evidence lists; The minutes generation module is used to construct evidence packages based on each candidate evidence list and inject the evidence packages into a preset minutes generation model to obtain the meeting minutes corresponding to the current meeting.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the meeting minutes generation method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions that, when executed by the one or more processors, perform the steps of the meeting minutes generation method as described in any one of claims 1 to 7.

Images