Conference summary generation method based on large language model

Through the conference minutes generation method based on large language model, the efficiency and accuracy of the intelligent conference system in minutes generation is solved, and structured and flexible conference minutes generation is realized, adapting to different conference needs, and improving conference efficiency and value.

CN119003759AActive Publication Date: 2024-11-22AEROSPACE INTERNET OF THINGS TECH CO LTD +1

Patent Information

Application Number
CN202411159823.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-22
Estimated Expiration
2044-08-22

AI Technical Summary

Technical Problem

The existing intelligent conference system has inefficient efficiency, inaccurate recording, lack of structural and flexibility in the generation of meeting minutes, and cannot meet the needs of a fast-paced working environment. Especially, there is a token length limitation in the generation of long-term conference records, resulting in the generation of The minutes lack targeted and practicality.

Method used

Using a meeting minutes generation method based on a large language model, structured and logical generation by constructing a meeting template, obtaining the time stamped meeting record text, and dividing two-level texts, and combining the meeting summary and to-do prompt word template. Coherent meeting minutes, flexibility to adapt to different meeting needs.

Benefits of technology

It improves the efficiency and accuracy of meeting minutes, and the generated meeting minutes are highly generalized and logically coherent, saving labor costs, providing in-depth insights and suggestions, and enhancing the overall value of the meeting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119003759A_ABST
    Figure CN119003759A_ABST
Patent Text Reader

Abstract

The invention discloses a conference summary generation method based on a large-scale language model, and the method comprises the steps: constructing a conference template, and inputting conference basic information according to the conference template; obtaining a conference audio file, recording a timestamp in the conference template, and obtaining a conference record text with the timestamp; based on a conference record text with a timestamp, segmenting the conference record text to obtain a plurality of text segments; and inputting the text segment into a large language model to obtain a conference summary. According to the invention, a set of standardized conference summary generation method is established based on the language understanding and generation capability of a large language model fused with the cue word engineering technology, and a structured, accurate and comprehensive conference summary can be automatically generated based on the conference record or the conference recording according to the predetermined conference link and type.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a method for generating meeting minutes based on a large language model. Background Art

[0002] Large language models are generative AI models trained on massive amounts of text data. In recent years, with the significant improvement in computing power and the development of big data technology, large language models have made significant leaps in scale and performance. In particular, the introduction of the Transformer architecture has enabled the model to better capture long-distance dependencies, greatly improving the ability to understand and generate language. Large language models can not only generate natural language text, but also deeply understand the meaning of text and complete various natural language tasks such as text summarization and text extraction. Users can use natural language to interact with application systems based on large language models, thereby realizing various applications from understanding to generation, including question answering, classification, summarization, translation, and chat. At present, large language models and their applications in various vertical business fields have become a hot topic in the research and application of the global artificial intelligence field, and the scale of model parameters has also jumped from more than a billion to hundreds of billions or even trillions. The increase in the number of parameters enables the model to capture detailed information in human language more finely and understand the complexity of human language more deeply. With the continuous maturity of technology and the continuous accumulation of data, large language models have gradually become a widely used artificial intelligence technology tool in the entire industry.

[0003] As an important form of communication, decision-making and collaboration between people, meetings are of great significance to various social and economic activities. Effective meetings can promote the flow of information, enhance teamwork and improve decision-making efficiency. However, traditional meetings often have problems such as low efficiency and inaccurate information recording. Relying on intelligent conference systems can effectively improve the efficiency and quality of meetings. Existing intelligent conference systems mainly improve work efficiency for conference organization and summary by automating the management of conference processes, recording conference content in real time, and organizing and sharing conference materials. With the development of artificial intelligence technology, intelligent conference systems based on speech recognition, natural language processing and other technologies can achieve semi-automation of conference records, as well as automatic recording and intelligent summarization of conference content to a certain extent. However, the above technologies or methods are mostly used to record the meeting process. Limited by the algorithm capabilities, the post-meeting summary still adopts the form of manual summary, which faces certain limitations.

[0004] Smart conference systems usually use advanced speech recognition technology for meeting minutes, which can convert recordings in meetings into text. Although these systems can identify different speakers and generate corresponding speech texts, they have obvious defects in the content of the records. Meeting records generated by speech recognition technology are often highly loose and lack structure, logic and coherence, which makes the generated text unable to be used directly as meeting minutes. In addition, due to the complexity and diversity of language, transcription models still have challenges in dealing with professional terms, accents, changes in speech speed, and background noise, which affects the accuracy and completeness of the transcription results.

[0005] Although intelligent conference systems have made some progress in speech transcription, the writing of meeting minutes still relies heavily on manual labor. After the meeting, a dedicated recorder is usually required to organize the minutes based on the recording and transcription text. This process is not only inefficient, but also affected by the recorder's personal ability, experience, and subjective judgment, resulting in uneven quality of meeting minutes written by different recorders. In addition, the process of manually writing meeting minutes is also time-consuming and cannot meet the needs of rapid response and decision-making in a fast-paced work environment.

[0006] Large language models have demonstrated powerful capabilities in the field of natural language processing. They can summarize meeting minutes and form a summary of the key content of the meeting, but they also have certain limitations in practical applications. The topics and agendas of meetings are diverse, and different meetings require different minute formats and content focuses. Currently, most large language models still rely on a single prompt word template or a fixed format when generating meeting minutes, which limits the flexibility of the model to adapt to the needs of different meetings. In addition, a single template may not accurately capture the key information and decision points in the meeting, resulting in the lack of pertinence and practicality in the generated minutes. More importantly, due to the limitation of large language models on the length of input tokens, the long text of long meeting minutes cannot be input into the large language model at one time to generate complete meeting minutes, which limits the efficiency and effectiveness of existing intelligent conference systems in practical applications.

[0007] Therefore, there is an urgent need for a meeting minutes generation method based on a large language model to solve the above problems. Summary of the invention

[0008] To solve the above technical problems, the present invention proposes a method for generating meeting minutes based on a large language model, which can save labor costs, improve the efficiency of meeting records, and provide meeting participants with more in-depth insights and suggestions through intelligent analysis, thereby enhancing the overall value of the meeting.

[0009] To achieve the above object, the present invention provides a method for generating meeting minutes based on a large language model, comprising:

[0010] Construct a meeting template and enter basic meeting information according to the meeting template;

[0011] Obtain the conference audio file, record the timestamp in the conference template, and obtain the conference record text with the timestamp;

[0012] Based on the meeting record text with the timestamp, segment the meeting record text to obtain multiple text segments;

[0013] The text segment is input into a large language model to obtain the meeting minutes.

[0014] Optionally, the basic meeting information includes: basic meeting information, personnel information and meeting agenda information.

[0015] Optionally, the text segment includes: a primary text block and a secondary text segment.

[0016] Optionally, obtaining the first-level text block includes:

[0017] The meeting record is segmented into first-level parts to obtain first-level text blocks.

[0018] Optionally, obtaining the secondary text segment includes:

[0019] Determine whether the first-level text block meets a preset length, and obtain a first-level determination result;

[0020] According to the first-level judgment result, it is determined whether to perform second-level segmentation on the first-level text block to obtain the second-level text segment.

[0021] Optionally, inputting the text segment into a large language model to obtain the meeting minutes includes:

[0022] Based on the text segment, in combination with the meeting summary prompt word template and the to-do list prompt word template, obtaining the meeting summary generation prompt word and the to-do list generation prompt word;

[0023] Generate prompt words based on the meeting summary and to-do items, and obtain the meeting summary and to-do items;

[0024] Based on the meeting summary and to-do items, combined with the meeting minutes template, obtain the meeting summary text object and the to-do items text object;

[0025] Based on the meeting summary text object, the to-do item text object and the basic information of the meeting, the meeting minutes are obtained.

[0026] Optionally, the conference summary prompt word template is:

[0027] PA = You are a summary tool for meeting minutes. Please sort out the content of +{spoken}+{spken_identity}+{con_agenda}+{con_type}+ into a summary. The content length should not exceed +{length}+ words. The content of the meeting is as follows: +{content}

[0028] Among them, PA is the template object for generating prompt words for the conference summary, {spoken_identity} is the speaker identity in the conference template, {spoken} is the speaker in the conference template, {con_agenda} is the conference agenda in the conference template, {con_type} is the conference type in the conference template, {content} is the conference content in the conference template, and {length} is the length limit of the conference summary or to-do items that need to be refined;

[0029] The to-do item prompt word template is:

[0030] PT

[0031] =You are a to-do summary tool. Please summarize the potential to-do items in the +{con_agenda}+{con_type}+ content based on +{spoken}+{spoken_identity}+. Each to-do item should be one sentence, and the content length should not exceed +{length}+ words. The meeting content is as follows: +{content}

[0032] Among them, PT is the to-do item prompt word template object.

[0033] Optionally, the meeting summary is:

[0034]

[0035] in, is the sum of the generated meeting summary contents, The conference summary group is obtained by processing the first-level text block or the second-level text fragment group by a large language model. i for conference summary prompts with different variable values, is the i-th first-level text block, is the secondary text segment, P is the total number of text segments in each secondary text segment group, p is the sequence of text segments in each secondary text segment group, For the first-level text block, through if To determine the text type passed into the large language model, if the current text type does not belong to the first-level text block Then select the formula corresponding to else;

[0036] The to-do items are:

[0037]

[0038] in, To define the total number of to-do items, PT i To-do prompts with different variable values.

[0039] Optionally, the meeting summary text object is:

[0040]

[0041] Among them, H(A abs ) is the conference summary text object, SORT(*) is the conference summary document information that effectively arranges each first-level text summary in the order of the conference agenda;

[0042] The to-do text object is:

[0043]

[0044] Among them, H(A next ) is a to-do list text object, and ORG(*) is a to-do list that is effectively segmented, integrated and sorted according to the language logic and semantic integrity.

[0045] Compared with the prior art, the present invention has the following advantages and technical effects:

[0046] 1. The present invention applies large-scale language model technology to the automated generation process of meeting minutes, greatly improving the efficiency and accuracy of meeting records. By standardizing the meeting process into a meeting template, the systematicness and integrity of the meeting content are ensured. The meeting process is captured in real time through meeting recordings, and the links of the meeting audio data are structured with the recording of timestamps. Through the transcription function of the artificial intelligence algorithm, the unstructured audio is converted into structured meeting minutes text, which improves the readability of the information. A two-level text segmentation technology is established, which not only distinguishes the content by topic according to the meeting agenda, but also takes into account the token length limit of the large language model processing text. The logical coherence of the secondary text fragments and the integrity of the single topic are ensured through secondary segmentation.

[0047] 2. The present invention utilizes the deep semantic understanding, reasoning and generation capabilities of large-scale language models, and conducts in-depth analysis and integration of text information through prompt word engineering, generating meeting summaries and to-do items with high generalization and logical coherence. Through text splicing technology and formatted output, standardized meeting minutes with complete structure and substantial content are generated, which greatly facilitates the organization and subsequent application of meeting results. In general, while improving the automation level of meeting minutes, the present invention also provides strong technical support for intelligent office, and has significant social and economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0049] Figure 1 It is a flow chart of a method for generating meeting minutes based on a large language model according to an embodiment of the present invention;

[0050] Figure 2 A flowchart of conference record text segmentation according to an embodiment of the present invention;

[0051] Figure 3 A flowchart for generating a meeting summary and a to-do list based on a large language model according to an embodiment of the present invention;

[0052] Figure 4 The present invention provides a flowchart for generating meeting minutes based on a meeting minutes template according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0055] The present invention abstracts the meeting process into a general and customizable template, and establishes a set of standardized meeting minutes generation methods based on the language understanding and generation capabilities of the large language model and the fusion of prompt word engineering technology. It can automatically generate structured, accurate and comprehensive meeting minutes based on the meeting records or meeting recordings according to the predetermined meeting links and types. It can not only save labor costs and improve the efficiency of meeting records, but also provide more in-depth insights and suggestions to meeting participants through intelligent analysis, thereby enhancing the overall value of the meeting.

[0056] The present invention proposes a method for generating meeting minutes based on a large language model, such as Figure 1 As shown, the specific steps include:

[0057] Build a meeting template and enter basic meeting information according to the meeting template;

[0058] Get the conference audio file, record the timestamp in the conference template, and get the conference record text with the timestamp;

[0059] Based on the meeting record text with timestamps, the meeting record text is segmented to obtain multiple text segments;

[0060] Feed text segments into a large language model to obtain meeting minutes.

[0061] Furthermore, the basic information of the meeting includes: basic information of the meeting, personnel information and meeting agenda information.

[0062] Specifically, as shown in Table 1, the present invention proposes a standard meeting template as a core element for interacting with users. The template abstracts specific basic meeting information, personnel information, meeting agenda and other information to form a general and universal template. The element content filled in the template will be used as variables in the prompt word template to control the input of the large language model. The design of the template fully takes into account the diversity and complexity of meetings to meet the needs of meetings of different types and sizes, ensures the versatility and flexibility of the template, and guides subsequent text segmentation and content integration work, thereby improving the quality and consistency of meeting minutes. The conference template design proposed by the present invention is as shown in the following table:

[0063] Table 1

[0064]

[0065] Before the meeting starts, users need to enter information according to the template and can continue to modify and improve it during the meeting. Personnel information and meeting agenda information can be increased or decreased according to actual meeting needs.

[0066] Specifically, generate meeting record text based on meeting audio files: mainly convert unstructured information such as audio data files recorded during the meeting into unified structured text information. The recording of meeting audio files requires the use of an efficient audio acquisition system to ensure that every detail of the meeting process can be accurately captured. While recording the meeting process as a complete audio data file, the start and end timestamps of each link in the meeting agenda are recorded in the meeting template. Use artificial intelligence algorithms to transcribe audio data files into meeting record text with timestamps. This step needs to consider the ability to suppress environmental noise, as well as the ability to identify and separate different sound sources, to improve the clarity and accuracy of voice data. During the transcription process, the artificial intelligence algorithm should be able to combine factors such as speaking speed, accent, and professional terminology to reduce transcription errors and improve text quality.

[0067] Furthermore, the text segment includes: a primary text block and a secondary text segment.

[0068] Furthermore, obtaining the first-level text block includes:

[0069] Perform first-level segmentation on the meeting minutes to obtain first-level text blocks.

[0070] Furthermore, obtaining the secondary text fragment includes:

[0071] Determine whether the first-level text block meets the preset length and obtain the first-level determination result;

[0072] According to the first-level judgment result, it is determined whether to perform second-level segmentation on the first-level text block to obtain the second-level text fragment.

[0073] Specifically, text segmentation is the process of breaking down long conference transcriptions into text blocks or text fragments that are easier to manage and analyze. The text segmentation tool designed by the present invention has powerful text matching and segmentation functions, and can reasonably segment text according to predefined rules.

[0074] like Figure 2 As shown, the present invention designs a two-level text segmentation rule. The first-level segmentation rule is to segment the meeting minutes text according to the meeting agenda timestamp to form a number of first-level text blocks with independent subjects and contents. Each text block includes all meeting minutes text within the timestamp range. Define the meeting minutes text object as H(A), and the timestamp logarithm of the meeting template is N t , preset first-level text block in Represents the Nth t A block is a first-level text block composed of several words, and the expression is:

[0075]

[0076] Among them, SPLIT(H(A),N t ) means to convert H(A) into N t Rationalize segmentation to obtain first-level text blocks In order to prevent the length of the input to the large language model from exceeding the token length limit of the model itself, the present invention designs the optimal capacity threshold t of the token input to the large language model. lan , as the criterion for judging whether the primary text block needs to be segmented at the secondary level. This ensures the accuracy of the output:

[0077] t lan =T lan ×h

[0078] Among them, T lan is the maximum token length that the model can input, and h is the optimal capacity ratio. After a large number of verifications, the present invention designs it to be 80%, and the specific setting of h can be flexibly adjusted according to different large language models.

[0079] According to the context length that the large language model can handle, the token capacity of each primary text block is determined, and the token capacity of the block that exceeds the threshold t is determined. lan The first-level text block is segmented into two levels to generate multiple text fragments with a single theme and content and a logical order. Define the second-level text fragment group

[0080]

[0081] in, Indicates that the first-level text block is divided into the optimal capacity t lan Perform segmentation to obtain corresponding secondary text segment groups. Represents the kth text segment in the Mth secondary text segment group.

[0082] Combination First-level text block The expression expands to:

[0083]

[0084] First, the token length of the first-level text block is determined. If its length is less than the optimal capacity of the model t lan , it will not be split and will remain as a first-level text block. The function splits it and obtains the secondary text group

[0085] Define the optimal capacity t for input to a large language model lan The total number of text blocks and text fragments is N, and N is quantified by timestamp.t , the number of secondary text segments M, and the number of secondary text segments The sum of the number of text fragments contained in Perform the calculation:

[0086]

[0087] Furthermore, if Figure 3-4 As shown, the text segment is input into a large language model to obtain the meeting minutes including:

[0088] Based on the text segment, combined with the meeting summary prompt word template and the to-do list prompt word template, the meeting summary generation prompt words and the to-do list generation prompt words are obtained;

[0089] Specifically, the present invention comprehensively considers the scale, complexity and performance requirements of the method, and the prompt words generated for the meeting minutes should have the characteristics of simplicity, versatility, portability, etc. The prompt words of the present invention are designed with a high degree of simplicity, completeness, flexibility and accuracy, and can adapt to different large-scale language models while meeting the needs of summarizing and outputting the meeting content.

[0090] The goal of the prompt word project in the present invention is to combine the primary text block and the secondary text fragment to complete the generation of two types of content, namely, the meeting summary and the to-do list. The prompt word template with different variable values ​​is merged with the corresponding text block or text fragment, which is used as the input of the large language model in turn. Combined with the header content in the conference template, the standardized syntactic pattern proposed by the present invention is used to generate the descriptive words and description phrases in the prompt word content, and PA is defined as the prompt word template object for the meeting summary. The syntactic description of the text input is:

[0091] PA = You are a summary tool for meeting minutes. Please sort out the content of +{spoken}+{spoken_identity}+{con_agenda}+{con_type}+ into a summary. The content should not exceed +{length}+ words. The content of the meeting is as follows: +{content}

[0092] Similarly, define PT as the to-do item prompt template object. The description syntax expression is:

[0093] PT = You are a to-do summary tool. Please summarize the potential to-do items in the +{spoken}+{spoken_identity}+{con_agenda}+{con_type}+ content.

[0094] Each to-do item has one sentence, and the content length does not exceed +{length}+ characters. The meeting content is as follows: +{contrnt}. If there is no to-do item, return "No to-do item"

[0095] In the above two syntactic patterns, PA is a prompt word template object for generating a conference summary, {spoken_identity} is the speaker identity in the conference template, {spoken} is the speaker in the conference template, {con_agenda} is the conference agenda in the conference template, {con_type} is the conference type in the conference template, {content} is the conference content in the conference template, {length} is the length limit of the conference summary or to-do item that needs to be refined, which is defined by the user, and PT is a to-do item prompt word template object.

[0096] Generate prompt words and to-do items based on the meeting summary, and obtain the meeting summary and to-do items;

[0097] Specifically, the present invention designs a general method and prompt word template to form the input of a large language model to generate meeting summaries and to-do lists. The present invention does not rely on the capabilities of a large language model. When applied in engineering, it is necessary to collect meeting sample data, test the generation capabilities of different models, and evaluate the coherence and readability of the text generated by the model and whether it can meet the standard format of meeting minutes, so as to select the optimal model. The present invention generates prompt words for each segmented first-level text block and second-level text fragment according to the meeting summary and to-do list, establishes a model input, and sequentially inputs the large language model to generate meeting summaries and to-do lists, and combines them respectively according to the following methods.

[0098] definition is the sum of the generated meeting summary contents, expressed as:

[0099]

[0100] Among them, PA i Indicates conference summary prompts with different variable values. For the first-level text block, directly use Language(*) to convert each text block With the corresponding PA i splicing and use it as the input of a large language model to generate the corresponding conference summary; for the secondary text segment group, each secondary text segment is With the corresponding PA i The resulting secondary conference summaries are concatenated as input to a large language model, and all the generated secondary conference summaries are accumulated to form a secondary text summary group under the single topic.

[0101] Similarly, define the total number of to-do items as Then the expression is:

[0102]

[0103] Among them, PT i For the to-do reminder words with different variable values, for the first-level text block, directly use Language(*) to add each text block The corresponding PT i splicing and use it as the input of the large language model to generate the corresponding first-level to-do items; for the second-level text fragment group, each second-level text fragment is The corresponding PT i The generated content of the large language model is judged. If "no to-do" is returned, the to-do corresponding to the text block or text fragment is set to an empty string.

[0104] Based on the meeting summary and to-do items, combined with the meeting minutes template, obtain the meeting summary text object and to-do item text object;

[0105] Get the meeting minutes based on the meeting summary text object, to-do text object and basic meeting information;

[0106] Specifically, the present invention designs a standardized meeting minutes template, which combines the basic information of the meeting, the sum of the meeting summary and the sum of the to-do items into the final meeting minutes. The present invention abstracts the specific and special meeting minutes structure into a general template to meet the needs of different types of meetings while ensuring the professionalism, consistency and readability of the minutes. The meeting minutes template consists of two parts: ① The meeting minutes header, including the meeting name, meeting time, meeting location, participants, meeting records, etc.; ② The meeting minutes content, which is composed of multiple structured text groups through text splicing technology to ensure the logic and coherence of the main content.

[0107] According to the meeting minutes template, the text splicing technology is used to obtain the final meeting summary text object H(A abs ):

[0108]

[0109] Among them, C abs is the number of meeting agendas in the meeting template. SORT(*) means that each first-level text summary is effectively arranged in the order of the meeting agenda links to generate a meeting summary document with rich content and unified format.

[0110] Similarly, define the to-do text object as H(A next ):

[0111]

[0112] Indicates that the total number of to-do items is A next According to the language logic and semantic integrity, the list is effectively segmented, integrated and sorted to generate a readable and clearly structured to-do list.

[0113] Finally, the meeting minutes header, meeting summary object and to-do text object are combined to generate a meeting minutes with complete structure, unified format and comprehensive content.

[0114] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for generating meeting minutes based on a large language model, characterized in that: include: Construct a meeting template and enter basic meeting information according to the meeting template; Obtain the conference audio file, record the timestamp in the conference template, and obtain the conference record text with the timestamp; Based on the meeting record text with the timestamp, segment the meeting record text to obtain multiple text segments; The text segment is input into a large language model to obtain the meeting minutes.

2. A method for generating meeting minutes based on a large language model according to claim 1, characterized in that: The basic information of the meeting includes: basic information of the meeting, personnel information and meeting agenda information.

3. The method for generating meeting minutes based on a large language model according to claim 1, characterized in that: The text segment includes: a primary text block and a secondary text segment.

4. A method for generating meeting minutes based on a large language model according to claim 3, characterized in that: Acquiring the first-level text block includes: The meeting record is segmented into first-level parts to obtain first-level text blocks.

5. A method for generating meeting minutes based on a large language model according to claim 4, characterized in that: Acquiring the secondary text segment includes: Determine whether the first-level text block meets a preset length, and obtain a first-level determination result; According to the first-level judgment result, it is determined whether to perform second-level segmentation on the first-level text block to obtain the second-level text segment.

6. The method for generating meeting minutes based on a large language model according to claim 3, characterized in that: The text segment is input into a large language model to obtain the meeting minutes including: Based on the text segment, in combination with the meeting summary prompt word template and the to-do list prompt word template, obtaining the meeting summary generation prompt word and the to-do list generation prompt word; Generate prompt words based on the meeting summary and to-do items, and obtain the meeting summary and to-do items; Based on the meeting summary and to-do items, combined with the meeting minutes template, obtain the meeting summary text object and the to-do items text object; Based on the meeting summary text object, the to-do item text object and the basic information of the meeting, the meeting minutes are obtained.

7. A method for generating meeting minutes based on a large language model according to claim 6, characterized in that: The conference summary prompt word template is: PA = You are a summary tool for meeting minutes. Please sort out the content of +{spoken}+{spoken_identity}+{con_agenda}+{con_type}+ into a summary. The content should not exceed +{length}+ words. The content of the meeting is as follows: +{content} Among them, PA is the template object for generating prompt words for the conference summary, {spoken_identity} is the speaker identity in the conference template, {spoken} is the speaker in the conference template, {con_agenda} is the conference agenda in the conference template, {con_type} is the conference type in the conference template, {content} is the conference content in the conference template, and {length} is the length limit of the conference summary or to-do items that need to be refined; The to-do item prompt word template is: PT = You are a to-do summary tool. Please summarize the potential to-do items in the +{con_agenda}+{con_type}+ content based on +{spoken}+{spoken_identity}+. Each to-do item should be one sentence, and the content length should not exceed +{length}+ words. The meeting content is as follows: +{content} Among them, PT is the to-do item prompt word template object.

8. The method for generating meeting minutes based on a large language model according to claim 6, characterized in that: The summary of the meeting is: in, is the sum of the generated meeting summary contents, The conference summary group is obtained by processing the first-level text block or the second-level text fragment group by a large language model. i for conference summary prompts with different variable values, is the i-th first-level text block, is the secondary text segment, P is the total number of text segments in each secondary text segment group, p is the sequence of text segments in each secondary text segment group, For the first-level text block, through if To determine the text type passed into the large language model, if the current text type does not belong to the first-level text block Then select the formula corresponding to else; The to-do items are: in, To define the total number of to-do items, PT i To-do prompts with different variable values.

9. The method for generating meeting minutes based on a large language model according to claim 6, characterized in that: The conference summary text object is: Among them, H(A abs ) is the conference summary text object, SORT(*) is the conference summary document information that effectively arranges each first-level text summary in the order of the conference agenda; The to-do text object is: Among them, H(A next ) is a to-do list text object, and ORG(*) is a to-do list that is effectively segmented, integrated and sorted according to the language logic and semantic integrity.

Citation Information

Patent Citations

  • Conference record generation method, device and equipment and readable storage medium

    CN111626061A

  • Conference summary generation method and device

    CN113919307A

  • Information extraction model training method, information extraction method and device

    CN116150613A

  • Data processing method and device, electronic equipment and storage medium

    CN116909435A

  • Local document summarization method based on large model

    CN117112779A

Cited By

  • Conference summary generation method based on sublingual acoustic features

    CN119807407A

  • A conference summary generation method based on paralinguistic acoustic features

    CN119807407B

  • Speaker log task optimization method based on semantic capability of large language model

    CN119943055A

  • Speaker diary task optimization method based on large language model semantic capability

    CN119943055B

  • Method for realizing automatic clustering of conference summary by adopting topic model

    CN120705318A