Intelligent conference quality inspection method based on large model technology

Abstract

The invention relates to an intelligent conference quality inspection method based on a large model technology, and the method comprises the following steps: obtaining an original conference file and defining quality inspection points which comprise basic quality inspection points and subdivision quality inspection points, obtaining a to-be-evaluated text based on an original sound recording file, inputting the to-be-evaluated text into a multi-dimensional quality inspection model, the model detects a process score, a content score and a score adding item of the to-be-evaluated text based on the quality inspection point and the cue word, wherein the content score comprises a malicious operation score; obtaining an overall quality inspection score based on the process score, the content score and the score adding item; and comparing the overall quality inspection score with the manual score, and determining a final score based on a comparison result. Compared with the prior art, the method has the advantages of giving comprehensive conference quality inspection indexes, identifying malicious behaviors and the like.

Description

Technical Field

[0001] The present invention relates to the field of meeting quality inspection, and in particular to an intelligent meeting quality inspection method based on large model technology. Background Art

[0002] For long meeting contents, it is difficult to detect the quality of the meeting. The existing meeting quality inspection has the following problems:

[0003] 1. Accuracy problem of speech transference: The recognition accuracy of insurance professional terms is not high. For example, "periodic payment" is recognized as "7 jiao", which affects the judgment of the meeting quality inspection system and reviewers.

[0004] 2. Large semantic understanding deviation: The quality inspection rules mainly rely on keywords, and there is a possibility of omission.

[0005] 3. Unable to recognize malicious timekeeping: The system may judge participation only based on surface indicators such as speaking duration and speaking times, and cannot deeply analyze the quality and relevance of the speech content.

[0006] 4. Heavy workload of manual quality inspection review: The workload of manual review is large and the efficiency is low. All problems may not be discovered and corrected in time.

[0007] 5. High operation cost: The current quality inspection is achieved through semantic configuration and keyword configuration. The maintenance of rules such as regular expressions and NLP is complex and requires a large investment. It is necessary to sort out rules through a large number of historical meetings and is not easy to quickly expand across scenarios.

[0008] Therefore, the existing technology is difficult to give comprehensive quality inspection indicators for meetings, and at the same time cannot recognize malicious behaviors, resulting in low accuracy of quality inspection. Summary of the Invention

[0009] The purpose of the present invention is to provide an intelligent meeting quality inspection method based on large model technology for giving comprehensive meeting quality inspection indicators and recognizing malicious behaviors at the same time.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] An intelligent meeting quality inspection method based on large model technology, the method includes the following steps:

[0012] Obtain the original meeting file and define quality inspection points. The quality inspection points include basic quality inspection points and sub-quality inspection points. Obtain the text to be evaluated based on the original recording file, and input the text to be evaluated into a multi-dimensional quality inspection model. The model detects the process score, content score and bonus items of the text to be evaluated based on the quality inspection points and prompt words. The content score includes malicious operation score.

[0013] Get an overall quality inspection score based on the process score, content score and bonus items;

[0014] The overall quality inspection score is compared with the manual score, and the final score is determined based on the comparison results.

[0015] Furthermore, the specific steps of calculating the malicious operation score of the multi-dimensional quality inspection model are:

[0016] The multi-dimensional quality inspection model identifies the text to be evaluated. If there is text in the text to be evaluated that overlaps with multiple quality inspection points within a certain time period, and the number of words in the text related to the quality inspection point identified within the time period is less than the threshold, the malicious operation score is recorded, otherwise the malicious operation score is set to 0;

[0017] If there is text in the text to be evaluated that belongs to only one role within a certain period of time, the malicious operation score is recorded, otherwise the malicious operation score is set to 0;

[0018] If there is no text in the text to be evaluated within a certain period of time or the similarity between the text and the conference topic is less than the threshold, the malicious operation score is recorded, otherwise the malicious operation score is set to 0;

[0019] If the meeting time or the number of participants in the text to be evaluated is less than the threshold, the malicious operation score is recorded; otherwise, the malicious operation score is set to 0.

[0020] Furthermore, the specific steps of calculating the process score of the multidimensional quality inspection model are: the multidimensional quality inspection model identifies whether the text to be evaluated is divided into multiple parts according to the meeting process steps. If so, the process score is recorded, otherwise the process score is set to 0.

[0021] Furthermore, the specific steps of calculating the content score of the multidimensional quality inspection model are: the multidimensional quality inspection model identifies whether there is text in the text to be evaluated that overlaps with the defined basic quality inspection points. If so, the initial content score is recorded; otherwise, the initial content score is set to 0. After traversing the text to be evaluated, the malicious operation score is subtracted from the initial content score to obtain the content score.

[0022] Furthermore, the specific steps of calculating the bonus items of the multi-dimensional quality inspection model are:

[0023] The multi-dimensional quality inspection model identifies whether there is text in the text to be evaluated that overlaps with the defined subdivided quality inspection points. If so, the bonus item is recorded, otherwise the bonus item is set to 0.

[0024] Furthermore, the overall quality inspection score satisfies:

[0025] Overall quality inspection score = process score + content score + bonus points <= upper limit score.

[0026] Furthermore, the specific steps of obtaining the text to be evaluated based on the original recording file are:

[0027] Perform text recognition and sentence segmentation on the original recording file, separate the recognition results according to different roles, aggregate the speeches of the same role, denoise the aggregated text, and obtain the text to be evaluated.

[0028] Furthermore, the sentence segmentation is performed based on an ASR model.

[0029] Furthermore, the specific steps of comparing the overall quality inspection score with the manual score are:

[0030] The composition of the manual score is the same as that of the overall quality inspection score. The score items of the manual score are compared with the score items corresponding to the overall quality inspection score, and the items whose differences exceed the threshold are screened out, and the prompt words are adjusted based on the screened items.

[0031] Furthermore, the quality inspection points are customized based on specific business requirements.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention uses deep semantic understanding of the big model to analyze the actual meeting content, parse its content quality, and identify malicious and inappropriate behaviors, such as malicious site visits. At the same time, the present invention uses multi-dimensional quality inspection rule configuration and AI big model quality inspection platform to comprehensively evaluate the meeting content. It can comprehensively evaluate the meeting content from multiple angles to ensure the comprehensiveness of the quality inspection results, accurately identify poor meetings, and significantly reduce the number of manual quality inspection reviews, thereby improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the process of the present invention;

[0035] Figure 2 is a detailed flow chart of the quality inspection of the present invention;

[0036] Figure 3 This is a model dimension diagram of the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0038] The present invention proposes an intelligent conference quality inspection method based on large model technology. The flow chart of the method is as follows: Figure 1 As shown, the method comprises the following steps:

[0039] The original meeting document is obtained and quality inspection points are defined, including basic quality inspection points and detailed quality inspection points. The text to be evaluated is obtained based on the original recording file, and the text to be evaluated is input into a multi-dimensional quality inspection model. The model detects the process score, content score and bonus items of the text to be evaluated based on the quality inspection points and prompt words. The content score includes a malicious operation score.

[0040] Get an overall quality inspection score based on the process score, content score and bonus items;

[0041] The overall quality inspection score is compared with the manual score, and the final score is determined based on the comparison results.

[0042] The detailed flow chart of quality inspection is as follows: Figure 2 The model dimension diagram used in the present invention is shown in FIG. Figure 3 shown.

[0043] The present invention adopts AI quality inspection based on large model technology. The key to combining the mechanism of AI quality inspection technology with business lies in the following aspects:

[0044] 1) Generalization capability: Sufficient data helps large models better understand the patterns and regularities in the data, learn a wide range of features and patterns, and thus demonstrate good generalization capabilities in different business scenarios.

[0045] 2) Semantic understanding: By using natural language processing models such as NLP, AI quality inspection systems can deeply understand text content and identify key information and potential problems.

[0046] 3) Customized business rules: The AI ​​quality inspection system can customize rules according to specific business needs, such as adding or deleting quality inspection process points, identifying malicious steps, and multi-model quality inspection.

[0047] In this invention, the large-model AI quality inspection technology is highly flexible and customizable, allowing users to dynamically add or delete quality inspection process points according to specific business needs. This capability is based on the system's in-depth understanding and learning of business scenarios, allowing it to quickly adapt to changes in quality inspection standards and ensure that the quality inspection process is synchronized with business development.

[0048] The big model uses advanced text analysis and behavior recognition technology to monitor and identify improper behaviors that maliciously hit the quality inspection process points in the meeting. This includes quality analysis of speech content and in-depth analysis of indicators related to meeting participation, thereby ensuring the professionalism and efficiency of meeting discussions.

[0049] The system uses multi-model fusion technology to combine different algorithms and models to handle quality inspection tasks. This approach not only improves the accuracy of the system, but also enables a comprehensive evaluation of the meeting content from multiple angles to ensure the comprehensiveness of the quality inspection results.

[0050] 4) Real-time feedback and iteration: The AI ​​quality inspection system can provide real-time feedback to help business processes respond and adjust quickly, while continuously iterating and optimizing the model based on feedback.

[0051] 5) Multimodal data processing: Combining the processing capabilities of multiple data types such as images, text, and sound allows for more comprehensive quality inspection.

[0052] 6) Integration and Automation: Integrate the AI ​​quality inspection system into existing business processes to automate the quality inspection process, reduce manual intervention, and improve efficiency.

[0053] 7) Compliance and security: Ensure that the AI ​​quality inspection system complies with industry standards and laws and regulations, and protect data security and user privacy.

[0054] The present invention proposes a multi-dimensional model. The key lies in integrating multiple evaluation dimensions to ensure the scientificity and rationality of the evaluation and to closely integrate it with business needs:

[0055] 1) Emotion recognition: Using sentiment analysis technology, the model can identify emotional tendencies in meetings and assess participants’ satisfaction and meeting atmosphere.

[0056] 2) Meeting interactivity: By analyzing the frequency and quality of participants’ interactions, the model evaluates meeting engagement and the depth of discussion.

[0057] 3) Malicious behavior identification: The model can identify and handle inappropriate visits, off-topic speeches, agenda delays, offensive language, etc. in meetings to maintain the professionalism and order of discussions.

[0058] 4) Content score: The model evaluates the relevance, accuracy, and completeness of the meeting content to ensure that all key topics are covered in the discussion.

[0059] 5) Process score: The model assesses the organization of the meeting process, including agenda adherence, time management, and process smoothness.

[0060] 6) Scientificity and rationality: The evaluation method is based on empirical research and industry best practices, while taking into account the needs and limitations of specific business scenarios.

[0061] 7) Data-driven decision support: Provide decision support to management through data analysis and optimize service processes.

[0062] The structure diagram of the model is as follows Figure 2 As shown in the figure, the multi-dimensional quality inspection evaluation model is described

[0063] 1. Five-level hierarchy of themes

[0064] First-level headings: meeting summary, account manager report, performance report

[0065] Secondary headings: Commonality, Activity Volume, Performance, Performance Evaluation, Key Task List

[0066] Level 3 headings: Basic situation, Target difference, Punch card and customer analysis, Praise, Criticism, XXX (Account manager name) - Work report for this week, XXX (Account manager name) - Work report for next week, XXX (Account manager name) - Inventory of important customers

[0067] Level 4 titles: Performance report, action report, important customer report, reporting target, reporting action, important customer name, XXX (account manager name) - task list

[0068] Five-level headings: basic customer information, customer pain points, cause analysis, interview plan, invitation plan

[0069] 2. Extracting Dimensions

[0070] Purpose, definition, scope: describe the purpose, business interpretation, input scope, etc. of the corresponding content. Keywords: For key elements, try to list them as exhaustively as possible to clarify the element values.

[0071] Display requirements: Describe the display content requirements in detail, including display format, content, etc.

[0072] Display example: Provide a page display example based on the content requirements.

[0073] 3. Content Dimension

[0074] Factor N: extracted based on the quality inspection scoring process score, content score, and bonus items.

[0075] 4. Rating Dimensions

[0076] Process score: refers to whether the meeting includes necessary meeting process steps, such as performance reporting.

[0077] Content score: Score the quality of the meeting in the process based on the basic rules, for example: +1 point for mentioning the achievement of healthy workforce.

[0078] Bonus points: The quality of the meetings in the process is subdivided and scored according to the bonus rules. For example, +1 for mentioning activities, salons, and online sand sessions.

[0079] 5. Total score

[0080] Calculation rules for the total score of quality inspection points: process score + content score + bonus points <= upper limit score.

[0081] Before the text is input into the model in the present invention, it is necessary to perform preprocessing such as slicing the long meeting text according to the meeting process to ensure the quality inspection effect of the large model and effectively control the calling and using cost of the large model.

[0082] Preprocessing includes:

[0083] 1) Role separation: Separate the speech content according to the different roles of the participants (such as department manager, account manager, etc.) to form a collection of speeches exclusive to multiple roles. This helps the big model focus more on the specific contribution of each role and improve the pertinence and depth of the analysis.

[0084] 2) Aggregation of similar speeches by the same role: Since the ASR model will segment the speech according to the speaker's pauses by default, the speech of the same role will be output and displayed in separate lines and segments. The algorithm is used to re-aggregate the scattered speech content caused by ASR segmentation to restore its original logical coherence.

[0085] This step is critical to improving the accuracy and reliability of quality inspection scores.

[0086] 3) Algorithm to remove noise from meeting content: To ensure the purity and effectiveness of meeting quality inspection, an intelligent denoising algorithm was introduced. This algorithm can automatically identify and remove non-core discussion content such as auditions and small talk in meeting records, ensuring that the large model processes carefully selected high-quality text data.

[0087] In addition, since the large model technology is still in the stage of continuous evolution and maturity, its output results may occasionally have hallucinations and inaccuracies. The present invention performs the following optimizations:

[0088] 1) Refined prompt word construction: Use the prompt word engineering framework, such as clearly defining role positioning, specific task description, detailed background setting, emphasis on key points, example display, and standardized output format, to deeply optimize and reconstruct the initial prompt words.

[0089] 2) Emphasis on “notes”: Given that large model technology tends to closely follow the characteristics of adjacent text, we place the most critical notes at the front of the prompt word, so as to strategically guide the large model to follow instructions more accurately and improve the accuracy and controllability of its response.

[0090] 3) Big model optimization: With the help of the big model's own multi-round analysis capabilities, the initially constructed prompt words are intelligently optimized and iterated multiple times, aiming to generate concise and efficient prompt words, so as to stimulate the big model to produce more accurate and demand-oriented outputs.

[0091] 4) Fine-tuning the large model: By carefully annotating the output of the large model, we build a high-quality annotated dataset for a specific scenario. Subsequently, we use this data to fine-tune the instructions and optimize the results of the large model in a targeted manner, aiming to further improve the expressiveness and accuracy of the large model in this scenario.

[0092] Examples of prompt words:

[0093] Quality inspection point: Report on important customers

[0094] Background: The content of the meeting was a record of a weekly sales work report meeting between a department manager of the insurance company's bank insurance department and several account managers. The account managers of the bank insurance department need to cooperate with the bank branch financial managers to achieve the goal of issuing orders for customers. There are some insurance professional terms in the meeting, which are explained as follows:

[0095] 1) Account Manager: An insurance company salesperson who reaches out to customers and completes orders by coaching bank branch wealth management managers and stationed at bank branches.

[0096] 2) Financial Manager: A bank branch employee who is responsible for selling insurance products to bank customers.

[0097] 2) Important customers: Customers with whom the account manager expects to close or have closed a deal this week.

[0098] 3) Issuing orders: Account managers sign insurance contracts with customers to achieve sales results.

[0099] 4) Invitation: The account manager invites the customer to meet by phone or other means.

[0100] 5) Interview: Account managers explore customer needs and quickly build relationships with customers by helping them handle branch business or inviting them to participate in activities.

[0101] Task: Please play the role of a professional meeting quality inspector and meeting information extractor. From the meeting content, based on the conversation between the department manager and the account manager XXX, extract the specific information of each customer, including the following information:

[0102] 1) Basic information: basic information of the customer, such as name, gender, age, occupation, asset status, risk preference, etc.

[0103] 2) Follow-up effect: How effective is the customer follow-up?

[0104] 3) Customer pain points: What are the pain points that make customers hesitate, postpone, or refuse to purchase the company's insurance?

[0105] 4) Communication plan: What communication plan has the account manager already made?

[0106] Output format:

[0107] Strictly output in the following json format: [

[0109] "We followed up with XX customer this week, the follow-up effect was XXX, the customer's pain point was XXX, and the communication plan was XXX",

[0110] "We followed up with customer YY this week, the follow-up result was YYY, the customer's pain point was YYY, and the communication plan was YYY". ]

[0112] Note:

[0113] 1) It is necessary to find each customer and output the information extracted strictly according to the [Task] for each customer. The information needs to be accurate and detailed, and no customer number can be missed.

[0114] 2) Please distinguish between financial managers and clients. Please do not output the information of financial managers as clients in [Output Format]. Financial managers are usually referred to as XXX teacher or XXX financial manager in the meeting content.

[0115] 3) Please strictly follow the content of the meeting when drawing and do not fabricate any content.

[0116] 4) Each speaker in the meeting content will be timestamped after the speaker finishes speaking. Please do not confuse the speaking timestamp with the completion time of the to-do items.

[0117] 5) Please strictly ensure that the output content is not colloquial and does not contain interjections such as "um, uh, ah, oh" etc.

[0118] 6) [Task] There should be no duplication of content between the extracted information points, and the focus of each information point should be distinguished.

[0119] 7) Please do not output AI-speaking sentences such as "not mentioned" or "someone mentioned it", and try to output it in the form of real meeting minutes.

[0120] 8) Please do not export the original text of the meeting content.

[0121] 9) Please do not output prompt words.

[0122] The intelligent optimization of unstructured information of the present invention includes:

[0123] 1) Identify and process dialects: The use of dialects in meetings may create obstacles to the standardization and analysis of data. By integrating dialect recognition technology, dialect content can be captured and translated more accurately, thereby improving the comprehensibility and usability of data.

[0124] 2) Connect and integrate incoherent language: Natural language processing technology can be used to reorganize unstructured data and enhance semantic and logical coherence for further analysis and understanding.

[0125] 3) Identify insurance terms: Conference data in the life insurance industry may contain a large number of insurance terms. By building a vocabulary containing industry terms, the model’s ability to identify these terms can be improved, ensuring the accuracy and professionalism of the data.

[0126] 4) Integration with enterprise organizational structure: The intelligent conference quality inspection system further understands the text content by combining the enterprise organizational structure. The conference content is linked to the organizational structure and speaking roles, and the business process and decision-making logic behind the conference data can be more accurately judged.

[0127] 5) Understanding context: The semantic enhancement model combines the meeting context content, such as meeting background, speaking role, discussion topic, meeting summary, etc., to enhance the model's understanding ability and provide more accurate data analysis.

[0128] 6) Continuous learning and model optimization: Through the continuous learning mechanism, the model can continuously accumulate the analysis content of past meetings, learn the latest hot topics of current meetings, and optimize the ability to process unstructured data.

[0129] Since the current model's reasoning ability cannot guarantee 100% consistency, in order to find out the differences each time, optimize the prompt words, and improve the accuracy and reliability of the results, a system combined with manual methods was introduced in the early stage to compare the quality inspection results of the meeting. By comparing online and offline scores, the score differences can be clearly identified, helping the team understand the shortcomings of the quality inspection points and thus optimize the prompt words.

[0130] Quality inspection scoring comparison process description:

[0131] Online process: The original voice file is generated into a text file through ASR voice recognition and pushed to the extraction platform. The extraction platform automatically runs the process score and content score of each quality inspection point through the quality inspection scoring system.

[0132] Offline process: The original voice file is generated into a text file through ASR voice recognition and pushed to the extraction platform. The tester will score according to the quality inspection points in the "Quality Inspection Points Evaluation".

[0133] Comparison process: The scores obtained through online and offline processes are compared horizontally to obtain the differences in each item, and the prompt words are continuously iterated to optimize the extraction effect. The meeting extraction is then re-performed to ultimately achieve the same scoring effect as that obtained manually.

[0134] Quality inspection score comparison effect

[0135] The combination of automatic scoring and manual scoring shows significant advantages. Automatic scoring greatly improves processing efficiency through ASR speech recognition technology and a standardized scoring system, and can quickly generate scoring results, saving a lot of labor costs. The objectivity and consistency of system scoring reduce the influence of human subjective factors and ensure the reliability of scoring. Through the difference analysis of online and offline scoring, the team can identify problems in a timely manner and make continuous optimizations, thereby improving the overall quality control level. This process not only promotes team collaboration and knowledge sharing, but also provides a solid foundation for data-driven decision-making and promotes the overall improvement of quality inspection work.

[0136] The innovation of the present invention is:

[0137] 1. Added quality inspection of meeting content:

[0138] Through automatic speech recognition, the meeting content is accurately converted into text, and combined with the semantic understanding ability of the big model, the meeting process, participation, discussion depth, etc. are analyzed. For example, the length of speeches and the number of speakers are automatically counted to determine whether key topics are covered, so as to quickly evaluate the organization and execution of the meeting.

[0139] 2. Identify bad meetings:

[0140] The system uses multi-dimensional quality inspection rule configuration and AI large model quality inspection platform to comprehensively evaluate the content of the meeting. If the meeting has chaotic process, key issues are not discussed, and there is a lack of clear conclusions, the system can automatically identify and mark it as a meeting of poor quality.

[0141] 3. Identify malicious behavior:

[0142] Through deep semantic understanding of the big model, the quality of meeting content can be analyzed and malicious and inappropriate behavior can be identified, which helps business managers optimize management actions. Specifically, the following judgment logic is included:

[0143] 1) Hitting multiple quality inspection points in a short period of time without mentioning the specific content.

[0144] 2) No one speaks for a long time during the meeting or the content of the speech is irrelevant to the meeting.

[0145] 3) The format of conference reports is monotonous and lacks interaction among participants.

[0146] 4) Deliberately shortening the meeting time or reducing the number of participants, and failing to comply with the specific regulations for holding the meeting.

[0147] 4. Reduce the number of manual quality inspection reviews

[0148] Through the optimization of technologies such as automatic speech recognition conference translation, multi-dimensional quality inspection rule configuration, and large-model AI quality inspection, accurate recognition of poorly identified meetings can be achieved, which can significantly reduce the number of manual quality inspection reviews and improve efficiency.

[0149] 5. Headquarters optimization management actions:

[0150] Based on the results of AI quality inspection, the headquarters can understand the meeting situations of each department, identify common problems, and adjust management strategies accordingly. For example, it can strengthen training for certain weak links or optimize resource allocation.

[0151] Through these specific application scenarios, large model quality inspection technology not only improves the quality and efficiency of meetings, but also provides strong data support and decision-making assistance for meetings.

[0152] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. An intelligent conference quality inspection method based on large model technology, characterized in that: The method comprises the following steps: The original meeting document is obtained and quality inspection points are defined, including basic quality inspection points and detailed quality inspection points. The text to be evaluated is obtained based on the original recording file, and the text to be evaluated is input into a multi-dimensional quality inspection model. The model detects the process score, content score and bonus items of the text to be evaluated based on the quality inspection points and prompt words. The content score includes a malicious operation score. Get an overall quality inspection score based on the process score, content score and bonus items; The overall quality inspection score is compared with the manual score, and the final score is determined based on the comparison results.

2. According to claim 1, the intelligent conference quality inspection method based on large model technology is characterized in that: The specific steps of calculating the malicious operation score of the multi-dimensional quality inspection model are as follows: The multi-dimensional quality inspection model identifies the text to be evaluated. If there is text in the text to be evaluated that overlaps with multiple quality inspection points within a certain time period, and the number of words in the text related to the quality inspection point identified within the time period is less than the threshold, the malicious operation score is recorded, otherwise the malicious operation score is set to 0; If there is text in the text to be evaluated that belongs to only one role within a certain period of time, the malicious operation score is recorded, otherwise the malicious operation score is set to 0; If there is no text in the text to be evaluated within a certain period of time or the similarity between the text and the conference topic is less than the threshold, the malicious operation score is recorded, otherwise the malicious operation score is set to 0; If the meeting time or the number of participants in the text to be evaluated is less than the threshold, the malicious operation score is recorded; otherwise, the malicious operation score is set to 0.

3. According to claim 2, the intelligent conference quality inspection method based on large model technology is characterized in that: The specific steps of calculating the process score of the multidimensional quality inspection model are as follows: the multidimensional quality inspection model identifies whether the text to be evaluated is divided into multiple parts according to the meeting process steps. If so, the process score is recorded, otherwise the process score is set to 0.

4. According to claim 3, the intelligent conference quality inspection method based on large model technology is characterized in that: The specific steps of calculating the content score of the multidimensional quality inspection model are as follows: the multidimensional quality inspection model identifies whether there is text in the text to be evaluated that overlaps with the defined basic quality inspection points. If so, the initial content score is recorded; otherwise, the initial content score is set to 0. After traversing the text to be evaluated, the malicious operation score is subtracted from the initial content score to obtain the content score.

5. According to claim 4, the intelligent conference quality inspection method based on large model technology is characterized in that: The specific steps of calculating the bonus items of the multi-dimensional quality inspection model are: The multi-dimensional quality inspection model identifies whether there is text in the text to be evaluated that overlaps with the defined subdivided quality inspection points. If so, the bonus item is recorded, otherwise the bonus item is set to 0.

6. According to claim 5, the intelligent conference quality inspection method based on large model technology is characterized in that: The overall quality inspection score meets the following requirements: Overall quality inspection score = process score + content score + bonus points <= upper limit score.

7. According to claim 1, the intelligent conference quality inspection method based on large model technology is characterized in that: The specific steps to obtain the text to be evaluated based on the original recording file are: Perform text recognition and sentence segmentation on the original recording file, separate the recognition results according to different roles, aggregate the speeches of the same role, denoise the aggregated text, and obtain the text to be evaluated.

8. According to claim 7, the intelligent conference quality inspection method based on large model technology is characterized in that: The sentence segmentation is performed based on an ASR model.

9. According to claim 1, the intelligent conference quality inspection method based on large model technology is characterized in that: The specific steps of comparing the overall quality inspection score with the manual score are as follows: The composition of the manual score is the same as that of the overall quality inspection score. The score items of the manual score are compared with the score items corresponding to the overall quality inspection score, and the items whose differences exceed the threshold are screened out, and the prompt words are adjusted based on the screened items.

10. The intelligent conference quality inspection method based on large model technology according to claim 1 is characterized in that: The quality inspection points are customized based on specific business requirements.

Images