Conference summary generation method and system based on large model, and storage medium

Abstract

The invention provides a conference summary generation method and system based on a large model and a storage medium, and the method comprises the steps: inputting conference sample information into the large model for entity recognition, obtaining a sample head entity, a sample tail entity and an entity relationship, and generating an entity triple according to the entity relationship, the sample head entity and the sample tail entity; performing event classification on the entity triad to obtain an event type, and determining a key conference text according to the event type; performing semantic recognition on the key conference text to obtain key summary semantics, and determining model loss according to the entity triad, the event type and the key conference text; and performing parameter updating on the large model according to the model loss, and inputting the conference information to be generated into the converged large model for summary generation to obtain a target conference summary. According to the embodiment of the invention, the key knowledge points in the to-be-generated conference information can be accurately identified and extracted based on the large model, and the accuracy of conference summary generation is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, system and storage medium for generating meeting minutes based on a large model. Background Art

[0002] With the continuous development of informatization and office automation, the importance and complexity of meetings as an important means of decision-making, communication and coordination are becoming increasingly prominent. Meeting minutes, as a record of the meeting process and results, play a vital role. Meeting minutes are the carrier of meeting information, which records key information such as the main content of the meeting, the discussion process, and the decision results. Meeting minutes are crucial for meeting participants, decision makers, and subsequent executors. Through meeting minutes, accurate transmission and effective communication of information can be ensured to avoid information omissions or misunderstandings.

[0003] In the existing process of generating meeting minutes, key knowledge points are generally summarized manually to form meeting minutes. However, since they are easily affected by human subjective factors, the generated meeting minutes have low accuracy. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a method, system and storage medium for generating meeting minutes based on a large model, so as to solve the problem of low accuracy of meeting minutes generated in the prior art.

[0005] The embodiment of the present invention is implemented as follows: a method for generating meeting minutes based on a large model, the method comprising:

[0006] Acquire conference sample information, and input the conference sample information into the large model for entity recognition to obtain a sample head entity and a sample tail entity;

[0007] Determine an entity relationship according to entity types of the sample header entity and the sample tail entity, and generate an entity triple according to the entity relationship, the sample header entity and the sample tail entity;

[0008] Performing event classification on the entity triples to obtain event types, and determining key meeting texts in the meeting sample information according to the event types;

[0009] Performing semantic recognition on the key meeting text to obtain key minutes semantics, and determining a model loss according to the entity triples, the event type, and the key meeting text;

[0010] The parameters of the large model are updated according to the model loss until the large model converges, and the meeting information to be generated is input into the converged large model to generate minutes, so as to obtain the target meeting minutes.

[0011] Preferably, the conference sample information is input into the large model for entity recognition to obtain a sample head entity and a sample tail entity, including:

[0012] Performing matrix conversion on the conference sample information according to the large model to obtain a conference sample matrix, and performing entity position prediction on the conference sample matrix to obtain a head entity starting position probability, a head entity ending position probability, a tail entity starting position probability, and a tail entity ending position probability;

[0013] The sample head entity is determined according to the maximum head entity starting position probability and the maximum head entity ending position probability, and the sample tail entity is determined according to the maximum tail entity starting position probability and the maximum tail entity ending position probability.

[0014] Preferably, the meeting information to be generated is input into the converged large model to generate minutes, and the target meeting minutes are obtained, including:

[0015] Obtaining a minutes template, and extracting knowledge points from the meeting information to be generated according to the converged large model to obtain target knowledge semantics and text event types;

[0016] Structuring the target knowledge semantics to obtain structured semantics, and matching the text event type with the minutes template by label;

[0017] The structured semantics are filled into the minutes template according to the tag matching result to obtain the target meeting minutes.

[0018] Preferably, the entity triples are subjected to event classification to obtain event types, including:

[0019] Performing vector mapping on the entity triplet to obtain a triplet vector, and calculating the similarity between the triplet vector and a preset event vector to obtain a vector similarity;

[0020] The preset event vector corresponding to the maximum vector similarity is determined as a target event vector, and the event type of the entity triple is determined according to a label of the target event vector.

[0021] Preferably, determining the model loss according to the entity triples, the event type and the key meeting text includes:

[0022] Calculating a first loss according to the entity triple and the standard triple, and calculating a second loss according to the event type and the standard type;

[0023] A third loss is calculated according to the key conference text and the key standard text, and a weighted value is calculated for the first loss, the second loss and the third loss to obtain the model loss.

[0024] Preferably, semantic recognition is performed on the key meeting text to obtain key minutes semantics, including:

[0025] Segmenting the key meeting text to obtain minutes segmentation, and performing word embedding processing on the minutes segmentation to obtain word embedding features;

[0026] Obtain a dimensionality reduction transformation matrix and a dimensionality increase transformation matrix, and determine an inner convolution kernel according to a preset activation function, the word embedding feature, the dimensionality reduction transformation matrix, and the dimensionality increase transformation matrix;

[0027] Generating inner convolution layers of different dimensions according to the inner convolution kernels, and stacking the inner convolution layers according to the dimensions to obtain stacked inner convolutions;

[0028] The word embedding feature is input into the stacked inner convolution for convolution processing to obtain a global feature, and the global feature is feature decoded to obtain the key minutes semantics.

[0029] Preferably, after calculating the third loss according to the key conference text and the key standard text, the method further includes:

[0030] Acquire a real head entity and a real tail entity of the conference sample information, and calculate a head entity loss according to the sample head entity and the real head entity;

[0031] Calculating a tail entity loss according to the sample tail entity and the true tail entity, and calculating the sum of the head entity loss and the tail entity loss to obtain a fourth loss;

[0032] A weighted value calculation is performed on the first loss, the second loss, the third loss and the fourth loss to obtain the model loss.

[0033] Another object of an embodiment of the present invention is to provide a meeting minutes generation system based on a large model, the system comprising:

[0034] An entity recognition module is used to obtain conference sample information and input the conference sample information into a large model for entity recognition to obtain a sample head entity and a sample tail entity;

[0035] A relationship determination module, configured to determine an entity relationship according to entity types of the sample head entity and the sample tail entity, and generate an entity triple according to the entity relationship, the sample head entity and the sample tail entity;

[0036] An event classification module, used to classify the entity triples to obtain event types, and determine key meeting texts in the meeting sample information according to the event types;

[0037] A model training module, used to perform semantic recognition on the key meeting text to obtain key minutes semantics, determine model loss according to the entity triples, the event type and the key meeting text, and update parameters of the large model according to the model loss until the large model converges;

[0038] The minutes generation module is used to input the meeting information to be generated into the converged large model to generate minutes and obtain the target meeting minutes.

[0039] Preferably, the entity identification module is also used for:

[0040] Performing matrix conversion on the conference sample information according to the large model to obtain a conference sample matrix, and performing entity position prediction on the conference sample matrix to obtain a head entity starting position probability, a head entity ending position probability, a tail entity starting position probability, and a tail entity ending position probability;

[0041] The sample head entity is determined according to the maximum head entity starting position probability and the maximum head entity ending position probability, and the sample tail entity is determined according to the maximum tail entity starting position probability and the maximum tail entity ending position probability.

[0042] The embodiments of the present invention can effectively extract sample head entities and sample tail entities in the meeting sample information by inputting meeting sample information into a large model for entity recognition. Based on the sample head entity and the sample tail entity, the entity relationship can be effectively determined. Based on the entity relationship, the sample head entity and the sample tail entity, entity triples can be automatically generated. By performing event classification on the entity triples, the event type of the meeting text corresponding to the entity triples can be effectively predicted. Based on the event type, the key meeting text in the meeting sample information can be effectively determined. By performing semantic recognition on the key meeting text, the key minutes semantics of the key meeting text can be effectively obtained. The parameters of the large model are updated based on the model loss, so that the converged large model can accurately identify and extract the key knowledge points in the meeting information to be generated, without the need to summarize the key knowledge points manually, thereby reducing the omission and misjudgment of key knowledge points and improving the accuracy of meeting minutes generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flowchart of a method for generating meeting minutes based on a large model provided in a first embodiment of the present invention;

[0044] Figure 2 is a structural diagram of a meeting minutes generation system based on a large model provided in a second embodiment of the present invention;

[0045] Figure 3 It is a schematic diagram of the structure of a terminal device provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] In order to illustrate the technical solution of the present invention, a specific embodiment is provided below for illustration.

[0048] Embodiment 1

[0049] See also Figure 1 , is a flowchart of a method for generating meeting minutes based on a big model provided in a first embodiment of the present invention. The method for generating meeting minutes based on a big model can be applied to any device or system. The method for generating meeting minutes based on a big model comprises the following steps:

[0050] Step S10, obtaining conference sample information, and inputting the conference sample information into the large model for entity recognition to obtain a sample head entity and a sample tail entity;

[0051] Among them, the meeting sample information can be set according to needs. The meeting sample information includes the meeting record text and the meeting minutes corresponding to the meeting record text. The meeting record text and the meeting minutes are data labeled, and the data-annotated meeting record text and the meeting minutes are cleaned, denoised and structured to obtain the meeting sample information. The meeting sample information is input into the large model for entity recognition to extract the sample head entity and sample tail entity of each meeting text in the meeting sample information.

[0052] Optionally, the conference sample information is input into a large model for entity recognition to obtain a sample head entity and a sample tail entity, including:

[0053] According to the large model, the conference sample information is matrix-converted to obtain a conference sample matrix, and the conference sample matrix is ​​predicted for entity positions to obtain the probability of the starting position of the head entity, the probability of the ending position of the head entity, the probability of the starting position of the tail entity, and the probability of the ending position of the tail entity; wherein, according to a preset matrix mapping, the conference sample information is matrix-converted to obtain a conference sample matrix, and the conference sample matrix is ​​predicted for entity positions to predict the position information of the sample head entity and the sample tail entity in each conference text;

[0054] The sample head entity is determined according to the maximum probability of the head entity starting position and the maximum probability of the head entity ending position, and the sample tail entity is determined according to the maximum probability of the tail entity starting position and the maximum probability of the tail entity ending position; wherein, the first starting position and the first ending position are determined according to the maximum probability of the head entity starting position and the maximum probability of the head entity ending position, and the entity between the first starting position and the first ending position is determined as the sample head entity, and the second starting position and the second ending position are determined according to the maximum probability of the tail entity starting position and the maximum probability of the tail entity ending position, and the entity between the second starting position and the second ending position is determined as the sample tail entity.

[0055] Step S20, determining an entity relationship according to entity types of the sample header entity and the sample tail entity, and generating an entity triple according to the entity relationship, the sample header entity and the sample tail entity;

[0056] Among them, based on the sample head entity and the sample tail entity, the entity relationship can be effectively determined, and based on the entity relationship, the sample head entity and the sample tail entity, the entity triple can be automatically generated. The entity triple is used to represent the association relationship between the head entity and the tail entity in the corresponding conference text.

[0057] Step S30, classifying the entity triples into events to obtain event types, and determining key meeting texts in the meeting sample information according to the event types;

[0058] Among them, by classifying entity triples into events, the event type of the meeting text corresponding to the entity triple can be effectively predicted. Based on the event type, the key meeting text in the meeting sample information can be effectively determined. The key meeting text is used to represent the key meeting content predicted by the large model.

[0059] Optionally, event classification is performed on the entity triple to obtain an event type, including:

[0060] Performing vector mapping on the entity triplet to obtain a triplet vector, and calculating the similarity between the triplet vector and a preset event vector to obtain vector similarity; wherein the entity triplet is input into a pre-trained vector encoder for vector encoding to obtain a triplet vector, and calculating the similarity between the triplet vector and a preset event vector according to a Euclidean distance formula to obtain vector similarity, and the preset event vector can be set as required;

[0061] The preset event vector corresponding to the maximum vector similarity is determined as the target event vector, and the event type of the entity triple is determined according to the label of the target event vector; wherein, by determining the preset event vector corresponding to the maximum vector similarity as the target event vector, the most similar preset event vector can be effectively determined as the target event vector corresponding to the entity triple, and the label of the target event vector is matched with the type query table to obtain the event type, and the type query table stores the correspondence between the labels of different target event vectors and the corresponding event types.

[0062] Step S40, performing semantic recognition on the key meeting text to obtain key minutes semantics, and determining a model loss according to the entity triple, the event type and the key meeting text;

[0063] Among them, by performing semantic recognition on the key meeting text, the key minutes semantics of the key meeting text can be effectively obtained. Optionally, the model loss is determined according to the entity triple, the event type and the key meeting text, including:

[0064] A first loss is calculated according to the entity triples and the standard triples, and a second loss is calculated according to the event type and the standard type; wherein, for the same conference text, the difference between the entity triples and the standard triples is calculated to obtain the triple difference, the sum of the triple differences is calculated to obtain the first loss, and for the same conference text, the difference between the event type and the standard type is calculated to obtain the type difference, the sum of the type differences is calculated to obtain the second loss;

[0065] A third loss is calculated based on the key meeting text and the key standard text, and a weighted value calculation is performed on the first loss, the second loss and the third loss to obtain the model loss; wherein, the difference between the key meeting text and the key standard text is calculated to obtain the third loss, and during the weighted value calculation process, the weighting coefficients of the first loss, the second loss and the third loss can be set according to needs.

[0066] Furthermore, after calculating the third loss according to the key conference text and the key standard text, the method further includes:

[0067] Acquire the real head entity and the real tail entity of the conference sample information, and calculate the head entity loss according to the sample head entity and the real head entity; wherein, vector mapping is performed on the sample head entity and the real head entity to obtain a first entity vector and a second entity vector, and the difference between the first entity vector and the second entity vector is calculated to obtain the head entity loss;

[0068] A tail entity loss is calculated according to the sample tail entity and the real tail entity, and the sum of the head entity loss and the tail entity loss is calculated to obtain a fourth loss; wherein, the sample tail entity and the real tail entity are vector-mapped to obtain a third entity vector and a fourth entity vector, and the difference between the third entity vector and the fourth entity vector is calculated to obtain the tail entity loss;

[0069] The first loss, the second loss, the third loss and the fourth loss are weighted to obtain the model loss; wherein, during the weighted value calculation process, the weighting coefficients of the first loss, the second loss, the third loss and the fourth loss can be set according to demand.

[0070] Furthermore, semantic recognition is performed on the key meeting text to obtain key minutes semantics, including:

[0071] Segmenting the key meeting text to obtain minutes segmentation, and performing word embedding processing on the minutes segmentation to obtain word embedding features;

[0072] Obtain a dimensionality reduction transformation matrix and a dimensionality increase transformation matrix, and determine an inner convolution kernel according to a preset activation function, the word embedding feature, the dimensionality reduction transformation matrix, and the dimensionality increase transformation matrix; wherein the dimensionality reduction transformation matrix, the dimensionality increase transformation matrix, and the preset activation function can all be set according to demand, and the word embedding feature is activated by the preset activation function, the dimensionality reduction transformation matrix, and the dimensionality increase transformation matrix to obtain the inner convolution kernel;

[0073] Generating inner convolution layers of different dimensions according to the inner convolution kernel, and stacking the inner convolution layers according to the dimensions to obtain stacked inner convolution; wherein the inner convolution layers of different dimensions are stacked from high to low dimensions to obtain stacked inner convolution;

[0074] The word embedding feature is input into the stacked inner convolution for convolution processing to obtain a global feature, and the global feature is feature decoded to obtain the key minutes semantics; by stacking multiple layers of inner convolution layers with different dimensions, local feature information of different dimensions of the word embedding feature is obtained in order of dimensionality from high to low, and the dimension of the feature vector is gradually reduced while extracting feature information at a higher level to obtain a global feature, and the global feature is input into the trained decoder for decoding to obtain the key minutes semantics.

[0075] Step S50, updating the parameters of the large model according to the model loss until the large model converges, and inputting the meeting information to be generated into the converged large model to generate minutes, thereby obtaining target meeting minutes;

[0076] Among them, the parameters of the big model are updated based on the model loss, so that the big model can automatically perform deep learning and classification of conference sample information, so that the converged big model can accurately identify and extract key knowledge points in the conference information to be generated, and the classification accuracy and coverage can be improved by continuously training and optimizing the big model.

[0077] Optionally, the meeting information to be generated is input into the converged large model to generate minutes, and the target meeting minutes are obtained, including:

[0078] Obtain a minutes template, and extract knowledge points from the meeting information to be generated according to the converged big model to obtain target knowledge semantics and text event types; wherein the minutes template can be set according to needs, and the converged big model can effectively analyze the text event type of the text in the meeting information to be generated, determine the key content of the meeting in the meeting information to be generated based on the text event type, perform semantic recognition on the key content of the meeting, and obtain the target knowledge semantics;

[0079] The target knowledge semantics is structured to obtain structured semantics, and the text event type and the minutes template are matched by labels; wherein, by structured processing of the target knowledge semantics, the target knowledge semantics can be effectively converted into semantic information in a structured format, and by matching the text event type and the minutes template by labels, the storage location of the structured semantics corresponding to the text event type is determined;

[0080] Fill the structured semantics into the minutes template according to the tag matching result to obtain the target meeting minutes; wherein, based on the tag matching result, fill the structured semantics into the corresponding storage location in the minutes template to obtain the target meeting minutes.

[0081] In this embodiment, by inputting the meeting sample information into the large model for entity recognition, the sample head entity and the sample tail entity in the meeting sample information can be effectively extracted, and based on the sample head entity and the sample tail entity, the entity relationship can be effectively determined. Based on the entity relationship, the sample head entity and the sample tail entity, the entity triple can be automatically generated. By performing event classification on the entity triple, the event type of the meeting text corresponding to the entity triple can be effectively predicted. Based on the event type, the key meeting text in the meeting sample information can be effectively determined. By performing semantic recognition on the key meeting text, the key minutes semantics of the key meeting text can be effectively obtained. The parameters of the large model are updated based on the model loss, so that the converged large model can accurately identify and extract the key knowledge points in the meeting information to be generated, without the need to summarize the key knowledge points manually, thereby reducing the omission and misjudgment of key knowledge points and improving the accuracy of meeting minutes generation.

[0082] Embodiment 2

[0083] See also Figure 2 , is a schematic diagram of the structure of a large model-based meeting minutes generation system 100 provided in a second embodiment of the present invention, including:

[0084] The entity recognition module 10 is used to obtain conference sample information and input the conference sample information into the large model for entity recognition to obtain a sample head entity and a sample tail entity.

[0085] Optionally, the entity recognition module 10 is further used to: perform matrix conversion on the conference sample information according to the large model to obtain a conference sample matrix, and perform entity position prediction on the conference sample matrix to obtain a head entity starting position probability, a head entity ending position probability, a tail entity starting position probability, and a tail entity ending position probability;

[0086] The sample head entity is determined according to the maximum head entity starting position probability and the maximum head entity ending position probability, and the sample tail entity is determined according to the maximum tail entity starting position probability and the maximum tail entity ending position probability.

[0087] The relationship determination module 11 is used to determine the entity relationship according to the entity types of the sample head entity and the sample tail entity, and generate an entity triple according to the entity relationship, the sample head entity and the sample tail entity.

[0088] The event classification module 12 is used to classify the entity triples to obtain event types, and determine the key meeting text in the meeting sample information according to the event types.

[0089] Optionally, the event classification module 12 is further used to: perform vector mapping on the entity triple to obtain a triple vector, and calculate the similarity between the triple vector and a preset event vector to obtain a vector similarity;

[0090] The preset event vector corresponding to the maximum vector similarity is determined as a target event vector, and the event type of the entity triple is determined according to a label of the target event vector.

[0091] The model training module 13 is used to perform semantic recognition on the key meeting text to obtain key minutes semantics, determine the model loss according to the entity triples, the event type and the key meeting text, and update the parameters of the large model according to the model loss until the large model converges.

[0092] Optionally, the model training module 13 is further used to: calculate a first loss according to the entity triples and the standard triples, and calculate a second loss according to the event type and the standard type;

[0093] A third loss is calculated according to the key conference text and the key standard text, and a weighted value is calculated for the first loss, the second loss and the third loss to obtain the model loss.

[0094] Furthermore, the model training module 13 is also used to: segment the key meeting text to obtain minutes segmentation, and perform word embedding processing on the minutes segmentation to obtain word embedding features;

[0095] Obtain a dimensionality reduction transformation matrix and a dimensionality increase transformation matrix, and determine an inner convolution kernel according to a preset activation function, the word embedding feature, the dimensionality reduction transformation matrix, and the dimensionality increase transformation matrix;

[0096] Generating inner convolution layers of different dimensions according to the inner convolution kernels, and stacking the inner convolution layers according to the dimensions to obtain stacked inner convolutions;

[0097] The word embedding feature is input into the stacked inner convolution for convolution processing to obtain a global feature, and the global feature is feature decoded to obtain the key minutes semantics.

[0098] Furthermore, the model training module 13 is also used to: obtain the real head entity and the real tail entity of the conference sample information, and calculate the head entity loss according to the sample head entity and the real head entity;

[0099] Calculating a tail entity loss according to the sample tail entity and the true tail entity, and calculating the sum of the head entity loss and the tail entity loss to obtain a fourth loss;

[0100] A weighted value calculation is performed on the first loss, the second loss, the third loss and the fourth loss to obtain the model loss.

[0101] The minutes generation module 14 is used to input the meeting information to be generated into the converged large model to generate minutes and obtain the target meeting minutes.

[0102] Optionally, the minutes generation module 14 is further used to: obtain a minutes template, and extract knowledge points from the meeting information to be generated according to the converged large model to obtain target knowledge semantics and text event types;

[0103] Structuring the target knowledge semantics to obtain structured semantics, and matching the text event type with the minutes template by label;

[0104] The structured semantics are filled into the minutes template according to the tag matching result to obtain the target meeting minutes.

[0105] In this embodiment, a unified data format and standard are established, and the conference sample information is cleaned, denoised, and structured to provide a high-quality data source for subsequent analysis.

[0106] By using big model technology, we conduct deep learning and classification on the pre-processed meeting sample information to achieve preliminary classification and secondary subclassification of meeting content. By continuously training and optimizing the model, we can improve the accuracy and coverage of classification.

[0107] Based on the text analysis capabilities of the big model, the classified meeting content is deeply analyzed to extract key points and detailed information. Through natural language processing technology, key concepts, technical problems, solutions and other key information in the meeting are identified and annotated.

[0108] Based on the analysis results, the core knowledge points in the meeting are automatically extracted and a structured summary of the knowledge points is generated. At the same time, the generation capability of the large model is used, combined with the preset template and context information, to automatically generate complete and standardized meeting minutes or knowledge base entries.

[0109] This embodiment, through the deep learning and natural language processing technology of the large model, can more accurately identify and extract key knowledge points in the technical liaison meeting, reducing omissions and misjudgments. The automated and intelligent processing flow greatly shortens the time for collating and analyzing meeting minutes, reduces labor costs, and improves work efficiency. Through a unified data format and classification standard, the present invention enables the knowledge points of the technical liaison meeting to be standardized and structured for storage, which is convenient for subsequent retrieval, utilization and sharing. The automatically generated structured meeting minutes and knowledge base entries provide rapid response support for technical exchanges and cooperation, helping to speed up the decision-making process and improve cooperation efficiency.

[0110] In this embodiment, by inputting the meeting sample information into the large model for entity recognition, the sample head entity and the sample tail entity in the meeting sample information can be effectively extracted, and based on the sample head entity and the sample tail entity, the entity relationship can be effectively determined. Based on the entity relationship, the sample head entity and the sample tail entity, the entity triple can be automatically generated. By performing event classification on the entity triple, the event type of the meeting text corresponding to the entity triple can be effectively predicted. Based on the event type, the key meeting text in the meeting sample information can be effectively determined. By performing semantic recognition on the key meeting text, the key minutes semantics of the key meeting text can be effectively obtained. The parameters of the large model are updated based on the model loss, so that the converged large model can accurately identify and extract the key knowledge points in the meeting information to be generated, without the need to summarize the key knowledge points manually, thereby reducing the omission and misjudgment of key knowledge points and improving the accuracy of meeting minutes generation.

[0111] Embodiment 3

[0112] Figure 3 2 is a block diagram of a terminal device 2 provided in the third embodiment of the present application. Figure 3 As shown, the terminal device 2 of this embodiment includes: a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program of a method for generating meeting minutes based on a large model. When the processor 20 executes the computer program 22, the steps in each embodiment of the method for generating meeting minutes based on a large model are implemented.

[0113] Exemplarily, the computer program 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 22 in the terminal device 2. The terminal device may include, but is not limited to, a processor 20 and a memory 21.

[0114] The processor 20 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0115] The memory 21 may be an internal storage unit of the terminal device 2, such as a hard disk or memory of the terminal device 2. The memory 21 may also be an external storage device of the terminal device 2, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 2. Further, the memory 21 may also include both an internal storage unit and an external storage device of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 may also be used to temporarily store data that has been output or is to be output.

[0116] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional unit.

[0117] If the integrated module is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium can be non-volatile or volatile. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in computer-readable storage media can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable storage media do not include electrical carrier signals and telecommunication signals.

[0118] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for generating meeting minutes based on a large model, characterized in that: The method comprises: Acquire conference sample information, and input the conference sample information into the large model for entity recognition to obtain a sample head entity and a sample tail entity; Determine an entity relationship according to entity types of the sample header entity and the sample tail entity, and generate an entity triple according to the entity relationship, the sample header entity and the sample tail entity; Performing event classification on the entity triples to obtain event types, and determining key meeting texts in the meeting sample information according to the event types; Performing semantic recognition on the key meeting text to obtain key minutes semantics, and determining a model loss according to the entity triples, the event type, and the key meeting text; The parameters of the large model are updated according to the model loss until the large model converges, and the meeting information to be generated is input into the converged large model to generate minutes, so as to obtain the target meeting minutes.

2. The method for generating meeting minutes based on a large model as claimed in claim 1, characterized in that: The conference sample information is input into the large model for entity recognition to obtain a sample head entity and a sample tail entity, including: Performing matrix conversion on the conference sample information according to the large model to obtain a conference sample matrix, and performing entity position prediction on the conference sample matrix to obtain a head entity starting position probability, a head entity ending position probability, a tail entity starting position probability, and a tail entity ending position probability; The sample head entity is determined according to the maximum head entity starting position probability and the maximum head entity ending position probability, and the sample tail entity is determined according to the maximum tail entity starting position probability and the maximum tail entity ending position probability.

3. The method for generating meeting minutes based on a large model as claimed in claim 1, characterized in that: The meeting information to be generated is input into the converged large model to generate minutes, and the target meeting minutes are obtained, including: Obtaining a minutes template, and extracting knowledge points from the meeting information to be generated according to the converged large model to obtain target knowledge semantics and text event types; Structuring the target knowledge semantics to obtain structured semantics, and matching the text event type with the minutes template by label; The structured semantics are filled into the minutes template according to the tag matching result to obtain the target meeting minutes.

4. The method for generating meeting minutes based on a large model as claimed in claim 1, characterized in that: The entity triples are classified into event types to obtain event types, including: Performing vector mapping on the entity triplet to obtain a triplet vector, and calculating the similarity between the triplet vector and a preset event vector to obtain a vector similarity; The preset event vector corresponding to the maximum vector similarity is determined as a target event vector, and the event type of the entity triple is determined according to a label of the target event vector.

5. The method for generating meeting minutes based on a large model as claimed in claim 2, characterized in that: Determining a model loss according to the entity triple, the event type, and the key meeting text includes: Calculating a first loss according to the entity triple and the standard triple, and calculating a second loss according to the event type and the standard type; A third loss is calculated according to the key conference text and the key standard text, and a weighted value is calculated for the first loss, the second loss and the third loss to obtain the model loss.

6. The method for generating meeting minutes based on a large model as claimed in claim 1, characterized in that: Perform semantic recognition on the key meeting text to obtain key minutes semantics, including: Segmenting the key meeting text to obtain minutes segmentation, and performing word embedding processing on the minutes segmentation to obtain word embedding features; Obtain a dimensionality reduction transformation matrix and a dimensionality increase transformation matrix, and determine an inner convolution kernel according to a preset activation function, the word embedding feature, the dimensionality reduction transformation matrix, and the dimensionality increase transformation matrix; Generating inner convolution layers of different dimensions according to the inner convolution kernels, and stacking the inner convolution layers according to the dimensions to obtain stacked inner convolutions; The word embedding feature is input into the stacked inner convolution for convolution processing to obtain a global feature, and the global feature is feature decoded to obtain the key minutes semantics.

7. The method for generating meeting minutes based on a large model as claimed in claim 5, characterized in that: After calculating the third loss according to the key conference text and the key standard text, it also includes: Acquire a real head entity and a real tail entity of the conference sample information, and calculate a head entity loss according to the sample head entity and the real head entity; Calculating a tail entity loss according to the sample tail entity and the true tail entity, and calculating the sum of the head entity loss and the tail entity loss to obtain a fourth loss; A weighted value calculation is performed on the first loss, the second loss, the third loss and the fourth loss to obtain the model loss.

8. A meeting minutes generation system based on a large model, characterized in that: The system comprises: An entity recognition module is used to obtain conference sample information and input the conference sample information into a large model for entity recognition to obtain a sample head entity and a sample tail entity; A relationship determination module, configured to determine an entity relationship according to entity types of the sample head entity and the sample tail entity, and generate an entity triple according to the entity relationship, the sample head entity and the sample tail entity; An event classification module, used to classify the entity triples to obtain event types, and determine key meeting texts in the meeting sample information according to the event types; A model training module, used to perform semantic recognition on the key meeting text to obtain key minutes semantics, determine model loss according to the entity triples, the event type and the key meeting text, and update parameters of the large model according to the model loss until the large model converges; The minutes generation module is used to input the meeting information to be generated into the converged large model to generate minutes and obtain the target meeting minutes.

9. The large model-based meeting minutes generation system as claimed in claim 8, characterized in that: The entity recognition module is also used for: Performing matrix conversion on the conference sample information according to the large model to obtain a conference sample matrix, and performing entity position prediction on the conference sample matrix to obtain a head entity starting position probability, a head entity ending position probability, a tail entity starting position probability, and a tail entity ending position probability; The sample head entity is determined according to the maximum head entity starting position probability and the maximum head entity ending position probability, and the sample tail entity is determined according to the maximum tail entity starting position probability and the maximum tail entity ending position probability.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images