Intelligent conference content extraction method based on large model technology

Abstract

The invention relates to an intelligent conference content extraction method based on a large model technology, and the method comprises the following steps: obtaining conference content, carrying out semantic text segmentation to form text blocks, vectorizing the segmented text, storing the vectorized text into a vector database, constructing KV key value pairs for the segmented text and serial numbers set for the text, and storing the KV key value pairs into a Redis cache; generating a cue word corresponding to the task, retrieving a corresponding text block in the Redis cache according to a retrieval element required by the cue word, and outputting a most matched text according to the corresponding text block; and inputting the most matched text and cue word into a large language model, wherein the large language model outputs a conference content extraction result. Compared with the prior art, the method has the advantages that conference content texts with high importance can be quickly screened out, and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of conference content extraction, and in particular to an intelligent conference content extraction method based on large model technology. Background Art

[0002] The company's business departments are often required to hold internal analysis meetings. The integration of internal analysis meeting content mainly relies on manual analysis and summary, which is inefficient and prone to missing key information. The current analysis meeting content integration technology has the following problems:

[0003] 1. Inefficiency: Manual analysis and summary of meeting content results in a lot of time spent on processing meeting records, which makes it impossible to quickly respond to business needs;

[0004] 2. Information omission: During manual processing, key information may be lost or ignored due to human factors;

[0005] 3. Inconsistent quality: Relying on individual ability may lead to differences in the analysis results of meeting content, lack of standardization and consistency;

[0006] 4. Lack of real-time feedback: The lack of an immediate feedback mechanism affects the timeliness and relevance of meeting content;

[0007] 5. Data analysis limitations: It is difficult to dig deep into data manually, which limits the ability to analyze complex problems;

[0008] 6. Difficult to scale: As the number of meetings increases, manual processing methods are difficult to adapt to the needs of scale;

[0009] 7. Lagging behind in technology updates: Existing technologies are difficult to quickly adapt to the development of new technologies and methods;

[0010] 8. Limited innovation: Lack of automation and intelligent tools limits the implementation of innovative solutions.

[0011] In conference content integration technology, text segmentation analysis is the first step. However, the existing conference segmentation methods have low accuracy and large texts easily exceed the number of model tokens. In addition, for long conference content texts, how to quickly filter out high-importance texts is also one of the problems that need to be solved. Summary of the invention

[0012] The purpose of the present invention is to provide an intelligent conference content extraction method based on large model technology to solve the problems of low accuracy of conference segmentation methods caused by redundant cutting and text block cutting errors, and large texts easily exceeding the number of model tokens, as well as how to quickly screen out the important conference content texts.

[0013] The purpose of the present invention can be achieved by the following technical solutions:

[0014] A method for extracting intelligent conference content based on large model technology, the method comprising the following steps:

[0015] Obtain the conference content and perform semantic text segmentation to form text blocks. The segmented text is vectorized and stored in the vector database. The segmented text and the number set for the text are used to construct a KV key-value pair and stored in the Redis cache.

[0016] Generate a prompt word corresponding to the task, retrieve the corresponding text block in the Redis cache according to the search elements required by the prompt word, and output the most matching text according to the corresponding text block;

[0017] The best matching text and prompt words are input into a large language model, and the large language model outputs a conference content extraction result.

[0018] Furthermore, the specific steps of retrieving the corresponding text block in the Redis cache according to the search elements required by the prompt word are:

[0019] The search elements are vectorized to obtain multiple embedded vector blocks, the corresponding numbers are retrieved based on the embedded vector blocks, the Redis cache is called based on the corresponding numbers to obtain the corresponding KV key-value pairs, and the corresponding text blocks are obtained based on the KV key-value pairs.

[0020] Furthermore, the number is obtained based on chunkid.

[0021] Furthermore, the specific steps of outputting the most matching text according to the corresponding text block are:

[0022] Based on the re-ranking method, the most relevant text block and the score corresponding to the most relevant text block are searched for the corresponding text block, and the most relevant text block with a score exceeding a threshold is output as the most matching text.

[0023] Furthermore, the specific step of obtaining the conference content and performing semantic text segmentation to form text blocks is to segment the conference content using a sentence segmenter.

[0024] Furthermore, the specific step of obtaining the conference content and performing semantic text segmentation to form text blocks also includes:

[0025] Obtain the meeting content, filter the speeches of specific roles in the meeting content, classify and store them, and then further filter them according to the number of words in the content, summarize the filtered speeches, and perform semantic text segmentation on the summarized content to form text blocks.

[0026] Furthermore, the specific step of summarizing includes converting the screened specific roles and screened speeches into structured information.

[0027] Furthermore, the specific steps of the large language model outputting the conference content extraction results are as follows:

[0028] The large language model outputs preliminary extraction results based on the most matching text and prompt words. The preliminary extraction results are merged to generate prompt words again. After input into the large language model, the conference content extraction results are obtained.

[0029] Furthermore, the conference content is obtained based on ASR speech analysis.

[0030] Furthermore, the prompt words are generated according to the CoT thinking chain and BRTR skills.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention introduces a method of combining semantic text cutting with vector retrieval, numbers each cut text block chunkid, calls a vector model to vectorize the text, performs text block retrieval based on vectorized retrieval elements, reduces the pressure of retrieval through vector matching, and realizes rapid screening of highly important conference content texts. At the same time, the semantic text cutting of the present invention is screened based on specific roles and word counts, reducing the problems of low accuracy of conference segmentation methods caused by redundant cutting and text block cutting errors, and large texts easily exceeding the number of model tokens. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The extraction flow chart of the method of the present invention for three scenarios;

[0034] Figure 2 Vectorization process for semantic text segmentation;

[0035] Figure 3 Retrieval process for semantic text segmentation;

[0036] Figure 4 The platform flow chart corresponding to the method 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 content extraction method based on large model technology, and the method comprises the following steps:

[0039] Obtain the conference content and perform semantic text segmentation to form text blocks. The segmented text is vectorized and stored in the vector database. The segmented text and the number set for the text are used to construct a KV key-value pair and stored in the Redis cache.

[0040] Generate a prompt word corresponding to the task, retrieve the corresponding text block in the Redis cache according to the search elements required by the prompt word, and output the most matching text according to the corresponding text block;

[0041] The best matching text and prompt words are input into a large language model, and the large language model outputs a conference content extraction result.

[0042] The present invention can be used for conference hot word extraction, hot issue extraction and successful case extraction. The extraction flow chart of the method of the present invention for three scenarios is as follows: Figure 1 shown.

[0043] Hot word extraction:

[0044] By accurately capturing the core keywords in a single meeting - going beyond the vocabulary defined by traditional quality control rules, our extraction platform can dig deep and reveal the implicit and insightful content of the meeting.

[0045] Hot words-First extraction process:

[0046] 1) Conference text cutting

[0047] Generic Text Cutting:

[0048] Through targeted conference text segmentation, the model can be provided with more accurate corpus. This time, the hot words are segmented to divide the conference content into three parts: Part 1, Part 2, and Part 3. The processing flow is as follows:

[0049] a) Part 1 Extraction Logic

[0050] Identify department manager speech: Traverse the text, find the line that starts with "department manager", and record its index and length.

[0051] Conditional filtering: If the length of the speech found is more than 300 words, it will be regarded as a valid department manager speech and recorded.

[0052] b) Part 2 Extraction Logic

[0053] Identify the Account Manager role: In the conversation after the department manager speaks, find the line that starts with "Account Manager", extract the role name and ensure there are no duplicates.

[0054] Categorize conversation content: Categorize conversation content related to each account manager role, making sure to distinguish it from speech by department managers.

[0055] c) Part 3 Extraction Logic

[0056] Extract summary speeches: From the department managers’ speeches, filter out speeches with a length of more than 150 words, paying special attention to the second to last and last speeches.

[0057] Summarize information: Organize the extracted speech content and role information into structured results for subsequent analysis and presentation.

[0058] The vectorization process and retrieval process of semantic text segmentation are as follows: Figure 2 and Figure 3 shown.

[0059] Semantic text segmentation:

[0060] In order to solve the problem of low text segmentation accuracy and large text exceeding the number of model tokens, we introduced semantic text segmentation combined with vector retrieval to deal with the problems of excessively long text, redundant segmentation, and text block segmentation errors.

[0061] Vectorization process description:

[0062] a) Semantic text segmentation: The conference text is merged according to the continuous dialogue between each speaker, and then the speaker's speech content is segmented into text blocks with complete sentence meanings using the Sentencesplitter method.

[0063] b) Text vectorization: Each cut text block is numbered chunkid, and the vector model is called to vectorize the text, and the vector is stored in the vector database.

[0064] c) KV ​​key-value pair storage: The chunkid and the corresponding text block are stored in the Redis cache to facilitate quick retrieval of the corresponding text block after vector retrieval.

[0065] Description of search process:

[0066] a) Prepare search factors: According to the search factor content required by each meeting prompt, such as the performance report prompt: performance content, activity volume, performance evaluation, etc.

[0067] b) Text vectorization: The above elements are vectorized using a vector model and the first six embedding blocks are obtained.

[0068] c) Text block retrieval: The corresponding chunkid is retrieved and the cache database is called to obtain the corresponding key-value pair to obtain the corresponding text block.

[0069] d) Reranking: Call the reranking model to quickly find the most relevant text among the acquired texts.

[0070] e) Output the best matching text: Based on the score threshold, the text with a score greater than 90 is output.

[0071] The model used above is:

[0072] The vector model Bge-large-zh has a model dimension of 1024 and is a general Embedding model that can map any text to a low-dimensional dense vector.

[0073] The reranking model Bge-reranker-large has a model dimension of 1024 and is a retrieval ranking model that can obtain the optimal retrieval results from the necklace retrieval results.

[0074] The collection (Collection) of the vector storage structure is the conference name, and the stored fields are shown in Table 1.

[0075] Table 1 Field table of vector storage structure

[0076]

[0077] Then proceed to the hot word - prompt assembly for the first extraction.

[0078] In order to ensure that the model better understands the user's question needs and generates more accurate and targeted answers, we designed a set of markdown format prompts for hot words as a guide for the model, telling it which direction to think and generate content.

[0079] a) Prompt word structure

[0080] Background: describes the context of hot word extraction;

[0081] Task: describes the instructions given to the model;

[0082] Article content: audio text content parsed by ASR;

[0083] High-frequency word extraction rules: describe the extraction rules of high-frequency words;

[0084] List of unextracted words: describes the words that are not extracted by the large model;

[0085] Output format: Specify the output format of Json.

[0086] b) Overcoming Difficulties

[0087] Difficulty 1: Inaccurate scope of hot words: A large number of words that are not related to events, business, or activities affect the analysis of the meeting.

[0088] Example:

[0089] Products, activities, communication, next week, goals, actions, partners, invitations, this week, salon, last week, interviews, plans, premiums, orders, analysis, reports, clocking in, and, teachers, orders, branches, transactions, presidents, performance, services, reserves, which ones, come over.

[0090] Optimization method:

[0091] 1) Follow the principle of avoiding ambiguity in prompt words and clarify the definition of hot words so that the model can understand the specific definition.

[0092] For example:

[0093] Be sure to extract according to the **high-frequency word extraction rules**. The extraction must be done based on the business entity objects clearly mentioned in the original article content. Do not make up or abstract them. Avoid the use of the <High-frequency Negative List> and high-frequency words related to work, planning, and communication.

[0094] 2) Using a small number of samples, add examples of a negative list of high-frequency words to the prompt words, so that the large model can make judgments more accurately.

[0095] For example:

[0096] Do not extract the vocabulary list:

[0097] <branch, plan, cooperation, product, activity, communication, next week, goal, action, partner, invitation, this week, salon, last week, interview, plan, premium, issuance of order, analysis, report, clocking in, as well as, teacher, order, branch, transaction, president, performance, service, reserve, which ones, come over>.

[0098] 3) Adjust tasks, steps, and precautions to guide the large model to extract meaningful hot words.

[0099] For example:

[0100] step:

[0101] Please follow the steps below to complete the task step by step:

[0102] 1. Extract high-frequency words from the article content.

[0103] 2. Check whether the extracted high-frequency words are explicitly mentioned in the original article. If not, discard them.

[0104] 3. Events and business information related to high-frequency words generate text summaries of different events or activities, requiring a clear description of the process and results step by step.

[0105] 4. The events in the summary need to describe specific content, including the event process, the time and location of the event. Avoid using names and numbers. Different events or activities require different summary descriptions.

[0106] 5. Output the results in the given JSON output format.

[0107] Note:

[0108] 1. Do not output semantic explanations of high-frequency words.

[0109] 2. The extracted high-frequency words must be clearly mentioned in the original text of the article. Fabrication and abstraction are not allowed.

[0110] 3. The summary expands the description of entities related to high-frequency words, corresponding to event details, business details, activity details, etc.

[0111] 4. When the abstract contains content related to results, please describe in detail the causes, processes, and steps that led to the results.

[0112] 5. Avoid using names in the abstract. Please use the titles and numbers of people related to the incident instead, including the names of people in the examples.

[0113] 6. Please sort the results by frequency of use from high to low.

[0114] 7. The hotWord high frequency words in the <Not Extracted Vocabulary List> are not output in the results.

[0115] Difficulty 2: Hot word summaries are not specific enough: The hot word summaries extracted by the large model are too abstract and not specific enough, resulting in no practical meaning.

[0116] Optimization method:

[0117] 1) Follow the principle of avoiding ambiguity in prompt words and clarify the definition of hot word summaries.

[0118] For example:

[0119] You are required to briefly introduce the corresponding events, activities, and issues in one or two sentences, and be careful not to go into excessive details.

[0120] 2) Adjust the model thinking steps to generate different summaries for different events to avoid the large model being too abstract.

[0121] For example:

[0122] Task:

[0123] One hotWord can correspond to multiple contents. Please form a detailed summary based on the specific events and contents involved in the high-frequency words in the original article. The names and numbers of people must not appear in the summary. They must be replaced by titles or numbers.

[0124] Note:

[0125] The summary expands the description of entities related to high-frequency words, corresponding to event details, business details, activity details, etc.

[0126] 3) Adjust the notes to avoid generating summaries that are too specific to specific names or outlets.

[0127] For example:

[0128] 1. Avoid using specific names and numbers in the summary. Please use the titles and numbers of people related to the incident instead, including the names of people in the examples.

[0129] 2. Avoid including specific time, specific location, branch organization, and branch organization name in the summary.

[0130] 3) LLM model

[0131] Use the 3.5-128K AI model to extract hot words based on the original conference text.

[0132] 4) Model post-processing

[0133] Based on the hot words marked "No" in "Hot Words - Preset Hot Words" provided by the user, the model output results are screened again to prevent the model from mistakenly extracting hot words with no business meaning.

[0134] Next comes the hot word - secondary extraction.

[0135] Process description:

[0136] 1) Merge multiple conference summaries

[0137] The selected hot words and the key summary information about the hot words in the review meeting are merged in a cycle of 800 items.

[0138] 2) Prompt assembly

[0139] In order to ensure that the model can better summarize and extract the key points of multiple meetings, we designed a set of markdown format prompts for secondary extraction of hot words as a guide for the model, telling it in which direction to think and generate content.

[0140] a) Prompt word structure

[0141] Key words: store selected hot words;

[0142] Article content: A collection of key information from multiple conferences corresponding to the hot words;

[0143] Task: Describe the instructions for secondary extraction of key information;

[0144] Output format: specifies the Json format of the model output;

[0145] Step: describes the processing step that requires model reference. ;

[0146] Thinking: Describe the thinking logic required for the model to complete the task;

[0147] Key points: Considerations when generating summaries from models.

[0148] b) Overcoming Difficulties

[0149] Difficulty 1: The model output is too detailed, resulting in too much secondary extraction content and making it impossible to form an effective summary.

[0150] Optimization method:

[0151] i. Use the CoT thinking chain to increase the thinking steps of summary generation and guide the large model to generate information summaries.

[0152] For example:

[0153] Please think about it step by step to complete the summary generation task:

[0154] {

[0155] 1. Extract key information: Identify the main topics and important information from the original text.

[0156] 2. Combine similar content: Combine similar or repeated ideas to avoid redundancy. For example, the discussion on the importance of task completion and follow-up was combined into one.

[0157] 3. Simplify expression: Use concise language to express complex ideas, ensuring that the information is conveyed clearly and is easy to understand.

[0158] 4. Maintain a logical structure: Make sure the summary is logically coherent and reflects the main direction and focus of the discussion.

[0159] 5. Control the word count: During the summarizing process, always pay attention to the word limit and ensure that the final summary is around 50 words.

[0160] }

[0161] ii. Use the BRTR technique with prompt words to add rules to prevent the summary from being too detailed.

[0162] Background (B): In-Context information, that is, the contextual background related to the task, which facilitates model understanding.

[0163] Role (R): Specify the model role to make it easier for the model to think according to the corresponding role.

[0164] Task (T): specifies the model processing task so that the model can eventually converge to generalized results.

[0165] Rules (R): Constrain the model to process tasks according to the rules.

[0166] 3) LLM model

[0167] i. Calculate the tokens of the prompt words and the key information about this hot word in the review meeting. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0168] ii. If the selected hot words involve a large number of review meetings: segment extraction is performed according to every 600 review meetings, and the 3.5-128K AI model is used by default. Summarize and extract the key points based on the segmented extraction: calculate the tokens of the prompt words and the key points extracted by the segmentation.

[0169] iii. If the number of tokens is less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0170] Hotly debated issues extraction:

[0171] By deeply understanding the multi-dimensional management perspective, our extraction platform can accurately capture hot topics in front-line work. By analyzing the common points of multiple meetings, we revealed the core topics that high-performing teams and core personnel discussed. Within a certain time span, the platform provides a statistical overview of hot topics in meetings, making the identification and tracking of key topics intuitive and efficient. In addition, through topic analysis, users can trace back to specific instances of specific meetings to accurately capture and review the discussion content.

[0172] Hotly debated issues - description of the first extraction process:

[0173] 1) Conference text cutting

[0174] Through targeted meeting text segmentation, the model can be provided with more accurate corpus. This time, the segmentation method is used to divide the meeting content into Part 2 and Part 3. The processing flow is as follows:

[0175] a) Part 2 Extraction Logic

[0176] Identify the Account Manager role: In the conversation after the department manager speaks, find the line that starts with "Account Manager", extract the role name and ensure there are no duplicates.

[0177] Categorize conversation content: Categorize conversation content related to each account manager role, making sure to distinguish it from speech by department managers.

[0178] b) Part 3 Extraction Logic

[0179] Extract summary speeches: From the department managers’ speeches, filter out speeches with a length of more than 150 words, paying special attention to the second to last and last speeches.

[0180] Summarize information: Organize the extracted speech content and role information into structured results for subsequent analysis and presentation.

[0181] 2) Prompt assembly

[0182] a) Customer Operation Prompt Assembly

[0183] i. Prompt word structure

[0184] Task: Describe the customer management tasks that the model needs to handle;

[0185] Role: Describes the roles of meeting extractors and the dimensions of analysis;

[0186] Thinking: Describe the steps that guide the model to think about customer business problems and solutions step by step;

[0187] Notes: Describe the constraints of model extraction;

[0188] Output format: describes the Json format results of the model output;

[0189] Article content: Conference corpus data that needs to be extracted.

[0190] b) Branch Operation Prompt Assembly

[0191] i. Prompt word structure

[0192] Task: Describe the tasks that the model needs to handle branch operations;

[0193] Role: Describe the role definition of conference extraction personnel;

[0194] Thinking: Describe and guide the model to think about the problems of branch management and the extraction of solutions step by step;

[0195] Definition and distinction between branch management and customer management: Describe the distinction rules between branch management and customer management in different dimensions.

[0196] Problem classification rules: Detailed description of the eight problem classification definitions and examples.

[0197] Notes: Describe the restrictions on the outlets’ operations;

[0198] Output format: describes the Json format results of the model output;

[0199] Article content: Conference corpus data that needs to be extracted.

[0200] c) Overcoming Difficulties

[0201] i. Difficulty 1: Unreasonable classification of problems, including incorrect classification of problems, problems that cannot be classified, unclear classification, etc.

[0202] ii. Difficulty 2: The problem description is vague and not related to the actual content.

[0203] iii. Difficulty 3: The timestamp positioning is wrong, resulting in the inability to listen back.

[0204]

Optimization method

[0205] i. Combine the information feeding and reverse questioning techniques of prompt words to allow the large model to learn and summarize and give the best question classification prompt words.

[0206] ii. Follow the principle of avoiding ambiguity in prompt words and clarify the difference between customer management and branch management.

[0207] iii. Use timestamp samples and give rules for extracting timestamps to ensure accurate positioning.

[0208] 3) LLM model

[0209] Calculate the tokens of the prompt word and related question details. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0210] 4) Model post-processing

[0211] Combine the extraction results of branch management and customer management problems, and remove problem data with incomplete "unknown" information.

[0212] Next, let’s move on to the hotly debated issue – secondary extraction:

[0213] 1) Merge multiple conference summaries

[0214] Merge summary information of the problem contents involved in the selected problem category.

[0215] 2) Prompt assembly

[0216] In order to ensure that the model can better summarize and extract the multi-channel question content information involved in question classification, we designed a set of markdown format prompts for secondary extraction of hot issues as a guide for the model, telling the model in which direction to think and generate content.

[0217] a) Prompt word structure

[0218] Question classification: store the selected question classification;

[0219] Article content: A collection of content from multiple conferences corresponding to the problem classification;

[0220] Task: Describe the instructions for secondary extraction of information about the problem content;

[0221] Output format: specifies the Json format of the model output;

[0222] Step: Describes the processing step that requires a model reference.

[0223] Thinking: Describe the thinking logic required for the model to complete the task, including the merging of similar problems;

[0224] Key points: Considerations when generating summaries from models;

[0225] b) Overcoming Difficulties

[0226] i. The summary is too brief and fails to convey the main themes and important information.

[0227] ii. Summary information is repeated.

[0228] Optimization method:

[0229] i. Use the prompt word CoT thinking chain technology to guide the large model to extract key information first, then merge similar content based on key information, and then simplify the data. This can ensure that the summary focuses on the meeting issues and that there is no repeated information.

[0230] 3) LLM model

[0231] i. Calculate the tokens of the prompt words and the description of the question involved. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0232] ii. If the selected problem category involves a large number of review meetings:

[0233] Segmented extraction will be performed every 1,000 passes, and the 3.5-128K AI model will be used by default.

[0234] Summarize and extract the main problem content according to the segmented extraction: calculate the tokens of the prompt words and the main problem content extracted by segment. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0235] In the extraction of successful customers, comprehensive management and analysis of successful cases are achieved by providing diverse queries, case label sidebar screening, detailed case information display, in-depth analysis in the customer insight area, intelligent query result processing, diverse case screening functions, interactive query operations, and dynamic case details display. At the same time, the case number sorting function enables users to quickly identify and obtain the most relevant case information.

[0236] Successful customer case - first extraction:

[0237] 1) Merge multiple conference summaries

[0238] Through targeted conference text segmentation, the model can be provided with more accurate corpus. This time, the hot words are segmented to extract Part 2 of the conference content. The processing flow is as follows:

[0239] a) Part 2 Extraction Logic

[0240] Identify the Account Manager role: In the conversation after the department manager speaks, find the line that starts with "Account Manager", extract the role name and ensure there are no duplicates.

[0241] Categorize conversation content: Categorize conversation content related to each account manager role, making sure to distinguish it from speech by department managers.

[0242] 2) Prompt assembly

[0243] In order to ensure that the model can better summarize and extract the content information of multiple issues involved in customer transaction cases, we designed a set of markdown format prompts for the first extraction of transaction customer cases as a guide for the model, telling the model in which direction to think and generate content.

[0244] a) Prompt word structure

[0245] Background: Introduce the background of extracting successful customer cases;

[0246] Definition: Definition of the closed customer case;

[0247] Task: Describe the instructions for extracting successful customer cases and state the requirements for extraction;

[0248] Output format: specifies the Json format of the model output, which is divided into success cases and failure cases;

[0249] Thinking: Describe the thinking logic required for the model to complete the task and guide the model to think step by step;

[0250] Note: Constraints in the model generation process for customer cases;

[0251] Meeting content: Provide meeting corpus for pattern extraction and generation of customer cases.

[0252] b) Overcoming Difficulties

[0253] i. There are unsuccessful customers in the customer transaction case output;

[0254] ii. There are cases of duplicate extraction of customer transaction cases.

[0255] Optimization method:

[0256] i. Use the CoT thinking chain to increase the thinking process of judging successful customer cases and abstract the precautions.

[0257] ii. Segment and merge the speech text generated by ASR into three parts: Part 1, Part 2, and Part 3. Mainly process the content of the account manager in Part 2. If each paragraph of the account manager is continuous, merge them and prevent repeated segmentation.

[0258] 3) LLM model

[0259] iii. Calculate the tokens of the prompt word and repeat customer review module. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0260] iv. If the case meets the system's preset customer profile label, the case will be assigned to the preset customer profile label; otherwise, a new customer profile label will be created autonomously.

[0261] 4) Model post-processing

[0262] Remove problematic data with incomplete “unknown” information.

[0263] Transaction Case - Secondary Extraction Process Description:

[0264] 1) Merge multiple conference transaction processes

[0265] Merge the closing processes of multiple closed customer cases.

[0266] 2) Prompt assembly

[0267] In order to ensure that the model can better summarize and extract the content information of multiple issues involved in transaction cases, we designed a set of markdown format prompts for secondary extraction of transaction customer cases as a guide for the model, telling the model in which direction to think and generate content.

[0268] c) Prompt word structure

[0269] Case tags: store selected customer tags;

[0270] Case highlights: Store the text content of the multi-conference transaction process corresponding to the customer tag;

[0271] Task: Describe the instructions for secondary extraction of information about the problem content;

[0272] Output format: specifies the Json format of the model output;

[0273] Step: Describes the processing step that requires a model reference.

[0274] Thinking: Describe the thinking logic required for the model to complete the task, including the merging of similar problems;

[0275] Key points: Considerations when generating summaries from models.

[0276] d) Overcoming Difficulties

[0277] iii. The summary information statistics are incorrect, and there is no common summary, common pain points and transaction cases, etc.

[0278] Optimization method:

[0279] ii. Through the prompt word CoT thinking chain technology, guide the large model to extract key information first, then merge similar content based on the key information, and simplify the answer.

[0280] 3) LLM model

[0281] i. Calculate the tokens of the prompt words and the description of the question involved. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0282] ii. If the selected problem category involves a large number of review meetings:

[0283] Segmented extraction will be performed every 600 passes, and the 3.5-128K AI model will be used by default.

[0284] Summarize and extract the main problem content according to the segmented extraction: calculate the tokens of the prompt words and the main problem content extracted by segment. If the tokens are less than 5120 tokens, use the 3.5-128K AI model, otherwise use the 4.0-8K AI model.

[0285] The present invention can also provide an extraction platform corresponding to the method, and the operation process of the platform is as follows: Figure 4 shown.

[0286] 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. A method for extracting intelligent conference content based on large model technology, characterized in that: The method comprises the following steps: Obtain the conference content and perform semantic text segmentation to form text blocks. The segmented text is vectorized and stored in the vector database. The segmented text and the number set for the text are used to construct a KV key-value pair and stored in the Redis cache. Generate a prompt word corresponding to the task, retrieve the corresponding text block in the Redis cache according to the search elements required by the prompt word, and output the most matching text according to the corresponding text block; The best matching text and prompt words are input into a large language model, and the large language model outputs a conference content extraction result.

2. According to claim 1, a method for extracting intelligent conference content based on large model technology is characterized in that: The specific steps of retrieving the corresponding text block in the Redis cache according to the search elements required by the prompt word are: The search elements are vectorized to obtain multiple embedded vector blocks, the corresponding numbers are retrieved based on the embedded vector blocks, the Redis cache is called based on the corresponding numbers to obtain the corresponding KV key-value pairs, and the corresponding text blocks are obtained based on the KV key-value pairs.

3. According to claim 2, a method for extracting intelligent conference content based on large model technology is characterized in that: The numbering is obtained based on chunkid.

4. According to claim 1, the method for extracting intelligent conference content based on large model technology is characterized in that: The specific steps of outputting the most matching text according to the corresponding text block are: Based on the re-ranking method, the most relevant text block and the score corresponding to the most relevant text block are searched for the corresponding text block, and the most relevant text block with a score exceeding a threshold is output as the most matching text.

5. According to claim 1, the method for extracting intelligent conference content based on large model technology is characterized in that: The specific step of obtaining the conference content and performing semantic text segmentation to form text blocks is to segment the conference content using a sentence segmenter.

6. According to claim 1, a method for extracting intelligent conference content based on large model technology is characterized in that: The specific step of obtaining the conference content and performing semantic text segmentation to form text blocks also includes: Obtain the meeting content, filter the speeches of specific roles in the meeting content, classify and store them, and then further filter them according to the number of words in the content, summarize the filtered speeches, and perform semantic text segmentation on the summarized content to form text blocks.

7. According to claim 6, a method for extracting intelligent conference content based on large model technology is characterized in that: The specific steps of the aggregation include converting the screened specific roles and screened speeches into structured information.

8. According to claim 7, a method for extracting intelligent conference content based on large model technology is characterized in that: The specific steps for the large language model to output the conference content extraction results are: The large language model outputs preliminary extraction results based on the most matching text and prompt words. The preliminary extraction results are merged to generate prompt words again. After input into the large language model, the conference content extraction results are obtained.

9. According to claim 1, a method for extracting intelligent conference content based on large model technology is characterized in that: The conference content is obtained based on ASR speech analysis.

10. The method for extracting intelligent conference content based on large model technology according to claim 1 is characterized in that: The prompt words are generated according to the CoT thinking chain and BRTR skills.

Images