Prompt word rhetoric hot updating processing method for intelligent customer service question and answer scene

Abstract

The present application relates to the field of artificial intelligence and natural language processing technology, and discloses a prompt word and technique hot updating method for an intelligent customer service question and answer scene, comprising the following steps: constructing a prompt word and technique storage module for storing prompt words and techniques for different business scenarios; realizing hot updating of the prompt words and techniques, dynamically modifying the prompt word and technique data in the storage module to realize real-time effectiveness of the updated content; performing semantic analysis and synonym replacement, performing semantic analysis on the keywords of the prompt words to generate an expanded prompt word set; updating the prompt word storage of the large model in linkage, triggering the fine-tuning of the large model by synchronizing the updated content of the prompt words; constructing and dynamically expanding a question and answer knowledge base, and adding new question and answer pair data to the knowledge base. Through the linkage updating of the prompt words and the large model, semantic expansion and dynamic knowledge base construction, the present application realizes the real-time adaptation, efficient question and answer and continuous optimization capability of the intelligent customer service system.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, specifically to a method for hot-updating prompts and dialogue in intelligent customer service question-and-answer scenarios. Background Technology

[0002] Intelligent customer service Q&A systems are widely used in various business scenarios, reducing labor costs and improving service efficiency by automating user inquiries. In this process, prompts and scripts, as core components for the system to understand user intent and provide standardized responses, need to be dynamically updated to adapt to changing business needs.

[0003] Existing intelligent customer service Q&A systems primarily rely on manual updates and pre-set static templates for hot-updating prompts, combined with a simple keyword matching mechanism for question and answer processing. The advantages of this approach are low implementation cost, suitability for systems with relatively simple scenarios, the ability to achieve relatively accurate answers in small-scale business scenarios, and ease of maintenance, making it suitable for initial deployment and operational needs.

[0004] Existing technologies have significant limitations in complex business environments with multiple scenarios and highly dynamic requirements. Static settings for prompts and dialogues are difficult to adapt to business changes in real time, resulting in low update efficiency. Keyword matching methods are not good at understanding diverse user expressions, which can easily lead to inaccurate question and answer responses. The question and answer knowledge base lacks an efficient dynamic mechanism for expansion, and the quality and adaptability of newly added question and answer data cannot be guaranteed, causing the system's question and answer performance to gradually deteriorate and affecting user experience. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for hot-updating prompt words and dialogue in intelligent customer service question-and-answer scenarios. This method solves the problems of low efficiency in updating prompt words and dialogue, insufficient understanding of diverse user expressions, and lack of an efficient dynamic mechanism for expanding the question-and-answer knowledge base in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for hot-updating prompts in intelligent customer service Q&A scenarios, comprising the following steps:

[0007] Build a prompt word and dialogue script storage module to store prompt words and dialogue scripts for different business scenarios;

[0008] Enables hot updates of prompts and scripts by dynamically modifying the prompts and scripts data in the storage module, so that the updated content takes effect in real time;

[0009] Perform semantic analysis and synonym replacement, perform semantic analysis on the keywords of the prompt word, generate an expanded prompt word set; Update the prompt word storage of the large model in linkage, trigger the fine-tuning of the large model by synchronizing the update content of the prompt word;

[0010] Build and dynamically expand the question and answer knowledge base, add new question and answer pair data to the knowledge base, and update the prompt word and the large model in linkage.

[0011] Preferably, the step of constructing the prompt word and the dialogue storage module comprises:

[0012] Predefine the prompt word and the dialogue of each business scenario;

[0013] Group and store the prompt word and the dialogue according to the business scenario;

[0014] Extract the keywords of the prompt word, and generate a synonym set for each keyword;

[0015] Store the prompt word, dialogue, keyword and synonym set in a structured format in a file or database.

[0016] Preferably, the step of implementing hot update of the prompt word and the dialogue comprises:

[0017] Receive the addition, modification or deletion request of the prompt word or the dialogue;

[0018] According to the request content, dynamically edit the prompt word and the dialogue data in the storage module;

[0019] Call the update interface to synchronize the changes of the prompt word and the dialogue to the large model prompt word storage.

[0020] Preferably, the step of performing semantic analysis and synonym replacement comprises:

[0021] Segment the prompt word and extract the core keyword;

[0022] Generate a synonym set for the keyword based on a pre-trained word vector model;

[0023] Replace the keywords in the prompt word with synonyms to generate multiple semantic variant prompt words;

[0024] Store the generated prompt word set in the prompt word storage module.

[0025] Preferably, the step of updating the prompt word storage of the large model in linkage comprises:

[0026] Extract the updated prompt word set from the prompt word storage module;

[0027] Synchronize the updated prompt word set to the large model prompt word storage file;

[0028] Triggering fine-tuning training of the large model to ensure that the new prompt words can adapt to the generated question and answer content.

[0029] Preferably, the fine-tuning training of the large model comprises:

[0030] Extracting training data from the question and answer knowledge base;

[0031] Fine-tuning the large model using a transfer learning method;

[0032] Deploying the fine-tuned large model to the intelligent customer service system.

[0033] Preferably, the large model prompt word storage file and the prompt word storage module are associated through an interface, and the prompt word storage file is updated synchronously when the prompt word is updated.

[0034] Preferably, the question and answer knowledge base comprises:

[0035] An initial knowledge base constructed based on historical artificial business question and answer pairs;

[0036] Newly added question and answer pair data collected in real time during the operation of the intelligent customer service;

[0037] Newly added question and answer samples that have passed artificial review.

[0038] Preferably, after the prompt word update is completed, the following linkage operations are automatically triggered:

[0039] Extract the prompt word set and update it to the large model prompt word storage;

[0040] Fine-tune the large model using the prompt word update content;

[0041] Deploy the fine-tuned large model to the intelligent customer service system.

[0042] Preferably, the step of dynamically expanding the question and answer knowledge base comprises:

[0043] Monitoring the user interaction log of the intelligent customer service, and extracting the unmatched user questions;

[0044] Marking as potential question and answer samples and submitting for artificial review;

[0045] After the review is passed, store the newly added question and answer pair to the knowledge base, and trigger the prompt word and large model update process.

[0046] The present application provides a prompt word technique hot update processing method for an intelligent customer service question and answer scenario. The method has the following beneficial effects:

[0047] 1、The present application realizes the dynamic updating and synchronization of prompt word data through the linkage mechanism of the prompt word storage module and the large model prompt word storage file, combines the fine-tuning training and rapid deployment of the large model, ensures that the system can adapt to new prompt words in real time, and improves the response efficiency of intelligent customer service.

[0048] 2、The present application adopts semantic analysis and synonym replacement technology, dynamically generates multiple semantic variants of prompt words, combines an extended question and answer knowledge base, enhances the understanding ability of the system to diversified user expressions, and significantly improves the accuracy and scene coverage of question and answer.

[0049] 3、The present application builds a dynamically expanded question and answer knowledge base by monitoring user interaction logs and the process of manually auditing new question and answer pairs in real time, while ensuring the high quality of new data and updated models, and cooperating with verification and monitoring mechanisms to improve the stability and long-term optimization ability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] Please refer to the accompanying Figure 1 The present application provides a prompt word technique hot updating processing method for intelligent customer service question and answer scene, including the following steps:

[0053] A prompt word and technique storage module is constructed for storing prompt words and techniques for different business scenarios. By constructing a unified storage module, prompt words and techniques can be efficiently managed and retrieved, providing a basis for hot updating and large model operation;

[0054] Hot updating of prompt words and techniques is realized by dynamically modifying the prompt word and technique data in the storage module to realize real-time effectiveness of the updated content. Hot updating triggers the corresponding processing logic by listening to the change events of the prompt word and technique storage module. Update operations are processed through a transaction mechanism to ensure data consistency and avoid data chaos caused by failure in the middle of the operation;

[0055] Perform semantic analysis and synonym replacement, perform semantic analysis on the keywords of the prompt word, generate an expanded prompt word set, through semantic expansion, better match the diversified expression of user input, reduce the failure of question and answer caused by keyword mismatch, thereby improve the response accuracy of the system and user experience;

[0056] Linkage update of the prompt word storage of the large model, trigger the fine-tuning of the large model by synchronizing the update content of the prompt word, linkage update through interface mechanism, ensure that the change of the prompt word can be transmitted to the large model storage file in real time, and optimize the training of the model based on this, linkage update ensures that the content after the change of the prompt word can be quickly reflected in the generation ability of the large model, avoiding the disconnection between data and model;

[0057] Construct and dynamically expand the question and answer knowledge base, add new question and answer pair data to the knowledge base, and link the prompt word and the large model update, dynamically expand by collecting new user questions in real time and comparing with existing question and answer, find the uncovered field and supplement the data in time, the expansion of the knowledge base and the linkage of the prompt word and the large model form a closed loop, thereby improving the response ability of the intelligent customer service to complex user problems and optimizing the overall service effect.

[0058] Please refer to the accompanying Figure 1 In a preferred embodiment of the present application, the step of constructing the prompt word and the dialogue storage module comprises:

[0059] Predefine the prompt word and the dialogue of each business scenario, form the standardized description of business requirements through the predefined prompt word and the dialogue, facilitate subsequent storage, management and dynamic calling, and standardized dialogue ensures that users can obtain consistent answers and reduce understanding deviation caused by expression difference;

[0060] Group storage of prompt words and dialogues according to business scenarios, manage prompt words and dialogues through business scenario grouping, facilitate independent maintenance and quick query of different scenarios, and reduce redundancy through grouped storage of data, improve storage and calling efficiency;

[0061] Extract the keywords of the prompt word, and generate a synonym set for each keyword, the keyword extraction is based on the word segmentation technology, the semantic analysis is performed on the content of the prompt word, the core description is recognized, the synonym set is generated by using the word vector semantic matching technology, the expression range of the keyword is expanded, and the coverage ability of the prompt word to diversified user input is enhanced;

[0062] Store the prompt word, dialogue, keyword and synonym set in a structured format in a file or database, the structured storage uniformly manages the prompt word, dialogue, keyword and synonym set through clear field definition, facilitates query and dynamic update, and the data structure enables the system to efficiently process complex multi-scenario and multi-level business logic.

[0063] Please refer to the attached Figure 1 In a preferred embodiment of the present application, the step of implementing hot update of prompt words and dialogues includes:

[0064] Receive the request of adding, modifying or deleting prompt words or dialogues, the management system listens to the adding, modifying or deleting operation through the interface, and transmits the request content to the storage module for processing; each request type (adding, modifying, deleting) triggers a specific update operation to ensure the dynamics and accuracy of the data, and at the same time improve the response speed and maintenance efficiency of the system;

[0065] According to the request content, dynamically edit the prompt word and dialogue data in the storage module, the data editing is realized through transaction processing mechanism, ensuring that the adding, modifying or deleting operation has atomicity, that is, the operation is either all successful or all rolled back, avoiding incomplete update, thereby improving the management efficiency of prompt words and dialogues;

[0066] Call the update interface to synchronize the changes of prompt words and dialogues to the large model prompt word storage, the update interface is realized through RESTAPI or RPC (remote procedure call), ensuring seamless synchronization of data from the storage module to the large model prompt word storage, ensuring that the changes of prompt words and dialogues can be reflected in the large model in real time, avoiding the disconnection of data and model.

[0067] Please refer to the attached Figure 1 In a preferred embodiment of the present application, the step of performing semantic analysis and synonym replacement includes:

[0068] Carry out word segmentation on the prompt words, extract the core keywords, the word segmentation operation is based on statistical language model or rule method, and the prompt words are cut into analyzable lexical units, and the stop word filtering is carried out through the word frequency statistical method to remove high-frequency meaningless words, so that the extracted keywords can accurately express the core semantics of the prompt words, and through the word segmentation and keyword extraction, the semantic analyzability of the prompt words is improved, laying a foundation for synonym generation and expansion;

[0069] Generate a synonym set of keywords based on a pre-trained word vector model, the word vector model represents the semantic relationship of words through a high-dimensional vector space, the synonym generation is based on the distance between vectors to measure semantic similarity, and the exclusive word library combines domain knowledge to optimize synonym generation, improving the coverage rate of industry-specific expression methods, thereby enhancing the adaptability of the system;

[0070] Replace the keywords in the prompt words with synonyms to generate multiple semantic variant prompt words, the implementation is as follows: according to the synonym set, replace the keywords in the prompt words one by one to generate multiple semantic variant prompt words:

[0071] Replacement logic: traverse the keyword list of the prompt word; for each keyword, replace it with each word in the synonym set to generate a new prompt word, generate multiple semantic variant prompt words to reduce the probability of question and answer failure caused by different user expression methods;

[0072] The generated prompt word set is stored in the prompt word storage module, which manages the semantic variant set in an efficient and structured manner, supports fast calling and updating operations, an index mechanism improves retrieval efficiency, version control ensures the traceability of the prompt word set, and the structured storage of the prompt word set facilitates management and retrieval, ensuring that the system can call the adapted semantic variant in real time.

[0073] Please refer to the accompanying Figure 1 In a preferred embodiment of the present application, the step of updating the prompt word storage of the large model comprises:

[0074] Extract the updated prompt word set from the prompt word storage module, implement incremental extraction by recording the update time field of the prompt word, reduce system resource consumption, and the full extraction mechanism as an alternative solution ensures that no prompt word data is missed in special scenarios, thereby reducing unnecessary data reading operations and improving system update efficiency;

[0075] Synchronize the updated prompt word set to the large model prompt word storage file, the prompt word storage module and the large model prompt word storage are linked through an interface to ensure the real-time and consistency of data synchronization, and format conversion and verification ensure that the prompt word set can be directly parsed and used by the large model, the updated prompt word set can be quickly transmitted to the large model storage file, shortening the time from prompt word change to question and answer adaptation; data consistency verification reduces the risk of training failure caused by format errors during the update process; trigger the fine-tuning training of the large model to ensure that the new prompt word can adapt to the generated question and answer content, the fine-tuning training is based on transfer learning, which updates the model weight through small-scale data to adapt to the semantics of the new prompt word and the question and answer requirements, thereby verifying and testing through simulating user query scenarios to evaluate the question and answer generation effect of the new model on the prompt word, and verifying and testing improves the system reliability and avoids the decline of user experience caused by model adaptation failure.

[0076] Please refer to the accompanying Figure 1 In a preferred embodiment of the present application, the fine-tuning training of the large model comprises:

[0077] Extract training data from the question and answer knowledge base, extract training data through prompt word and question and answer knowledge base association analysis to ensure that the training set covers the semantic range of the prompt word, and format operation converts the question and answer data into input and output pairs adapted to the model to ensure that the data can be effectively utilized, providing high-quality training data to ensure that the large model can accurately adapt to the semantic features of the prompt word, and

[0078] Expanding the training data coverage through semantic variants further enhances the model's question-answering capabilities;

[0079] Transfer learning is used to fine-tune the training of large models. By fine-tuning the weight parameters on the basis of an existing pre-trained model, transfer learning reduces the computational resource consumption of large-scale retraining. It can quickly adapt to the semantic requirements of updated prompt words without training a large model from scratch. The dynamic adjustment of training strategy optimizes the balance between training efficiency and model performance. The fine-tuned model can accurately answer the user's query based on the new prompt words, significantly improving the question answering accuracy.

[0080] The finely tuned large model is deployed to the intelligent customer service system. Model validation is achieved by simulating user interaction scenarios using test data to ensure that the model reaches the expected performance before going live. This ensures that the finely tuned model can run efficiently and stably in the intelligent customer service system, providing accurate question and answer services. The deployment adopts containerization technology, which improves the scalability and operational stability of the model. Through real-time monitoring and user feedback, a continuous optimization mechanism is formed to continuously improve the intelligence level of the system.

[0081] Please see the appendix Figure 1 In a preferred embodiment of the present invention, the large model prompt word storage file and the prompt word storage module are associated through an interface. When the prompt word is updated, the prompt word storage file is updated synchronously. The interface association realizes the data linkage between the prompt word storage module and the large model prompt word storage file, avoiding inconsistencies between the two. The automated interface association mechanism improves the efficiency of the prompt word update process and shortens the data transmission time. The synchronization task is automatically triggered by listening to the update event of the storage module, without manual intervention, ensuring real-time performance.

[0082] Please see the appendix Figure 1 In a preferred embodiment of the present invention, the question-and-answer knowledge base includes:

[0083] The initial knowledge base is built based on historical human business question and answer pairs. The knowledge base is constructed using historical human business question and answer pair data, which lays the foundation for the initial question and answer of the intelligent customer service system. It provides high-quality initial question and answer data to ensure that the intelligent customer service system can provide accurate and comprehensive question and answer services in the early stage of its launch.

[0084] The newly added question-and-answer pairs data are collected in real time during the operation of the intelligent customer service. By utilizing the real-time interaction records between users and the intelligent customer service, the newly added question-and-answer pairs are dynamically collected, the coverage of the knowledge base is expanded, and the knowledge base is dynamically expanded to adapt to the ever-changing questioning needs of users, thereby improving the update efficiency of the knowledge base and enhancing the question-and-answer adaptability of the intelligent customer service.

[0085] The review process for newly added Q&A samples that have passed manual review is as follows:

[0086] Manually review potential question-and-answer pairs collected in real time: verify the accuracy and semantic completeness of the questions; edit or supplement the standard answers to the questions according to business needs; and mark the approved question-and-answer pairs as "verified".

[0087] Storage Update: Approved question-answer pairs are added to the knowledge base, using the same structured storage method as the initial knowledge base; the prompt word update process is automatically triggered, generating prompt words for new questions and synchronizing them to the prompt word storage module;

[0088] Regular review: Regularly review the quality of newly added question-answer pairs in the knowledge base and remove outdated or inapplicable data;

[0089] The accuracy and reliability of the knowledge base expansion have been improved through manual review, while the added structured storage and annotation of question and answer samples support fast retrieval and model adaptation.

[0090] Please see the appendix Figure 1 In a preferred embodiment of the present invention, after the prompt word is updated, the following linkage operation is automatically triggered:

[0091] Extract the prompt word set and update it to the large model prompt word storage. The prompt word update trigger mechanism is as follows: when the prompt word storage module detects that a new, modified, or deleted prompt word operation has been completed, the listener automatically triggers the data synchronization process; the listener filters out the changed prompt word set based on the update records and outputs it in a structured format.

[0092] Data is synchronized to the large model prompt word storage:

[0093] The extracted prompt word set is pushed to the large model prompt word storage file via API interface or data stream; data synchronization adopts batch update mode to avoid performance overhead caused by frequent calls;

[0094] The automatic triggering mechanism monitors the operation logs of the prompt word storage module, captures update events in real time, and pushes data through a standardized interface to achieve seamless integration, ensuring that prompt words remain consistent between the storage module and the large model. This ensures that prompt word updates can be quickly synchronized to the prompt word storage of the large model, shortening data transmission time.

[0095] The large model is fine-tuned by updating the content with prompt words. The fine-tuning training uses transfer learning technology to optimize the model weights on the basis of the pre-trained model so that it can adapt to the question-answering requirements of the new prompt words. The training and validation are carried out through real-world scenario simulation to ensure that the model's ability to answer questions with the new prompt words meets expectations, thereby improving the accuracy of the model's generated answers. At the same time, the validation test reduces the risk of question-answering errors after the model is deployed and improves the stability of the system.

[0096] The finely tuned large model is deployed to the intelligent customer service system. Model validation is achieved by simulating real user query scenarios through test inputs to ensure the model's adaptability to prompt words. The finely tuned model can be quickly launched to support new prompt words, shortening the time from data update to system question and answer adaptation.

[0097] Please see the appendix Figure 1 In a preferred embodiment of the present invention, the step of dynamically expanding the question-and-answer knowledge base includes: monitoring the user interaction logs of the intelligent customer service, extracting unmatched user questions, and using matching algorithms to determine whether there are corresponding question-and-answer pairs for user questions through technologies such as keyword similarity and semantic parsing. The marking and storage of unmatched questions provides basic data for knowledge base expansion. Real-time extraction of unmatched questions ensures that the knowledge base can quickly respond to changes in user needs, while reducing the possibility of missing user questions and improving user experience.

[0098] These samples are marked as potential Q&A samples and submitted for manual review. Reviewers process the potential samples through a management platform or batch review tools.

[0099] Verify the completeness and applicability of the question; write a standard answer based on business rules and existing knowledge; after review, mark the status as "approved";

[0100] During the review process, unqualified samples (such as those with unclear semantics) are directly marked as "review failed" to prevent low-quality data from entering the knowledge base;

[0101] Manual review provides a data quality control mechanism to ensure that the content of newly added question-and-answer pairs is accurate and meets business needs, thereby improving the quality of newly added question-and-answer pairs and avoiding the impact of low-quality data on the accuracy of the knowledge base.

[0102] Once approved, the new question-and-answer pairs will be stored in the knowledge base, triggering the update process for prompt words and the large model. The knowledge base update ensures that it can quickly respond to user queries through the structured storage and index optimization of the new question-and-answer pairs. The dynamic expansion of the knowledge base improves the system's coverage of new questions and optimizes the user experience. At the same time, the linkage update of prompt words and the large model shortens the cycle from data collection to question and answer generation, achieving efficient dynamic adaptation.

[0103] To better understand the present invention, the above content will be described in detail below with reference to specific embodiments.

[0104] Example 1: Implementation of a static intelligent customer service system without dynamic update mechanism: Prompt words and scripts are stored statically and cannot be dynamically updated. The stored content needs to be updated manually. The system only relies on keyword matching to answer user questions and does not introduce semantic parsing and dynamic knowledge base expansion functions.

[0105] Beneficial effects: Suitable for systems with simple business scenarios and low update frequency; low implementation cost, and relatively simple initial deployment and maintenance.

[0106] Example 2: Implementation of an intelligent customer service system with added prompt words and dynamic message update mechanism: The prompt words and messages implement a hot update mechanism, which supports dynamic adjustment of content through adding, modifying or deleting operations and takes effect in real time, but semantic expansion function and linkage update with large model are not introduced.

[0107] Beneficial effects: The dynamic adjustment of prompts and scripts improves maintenance efficiency and adapts to the needs of some business changes; the system can quickly update prompts and scripts, improving the flexibility of data management.

[0108] Example 3: Intelligent Customer Service System with Semantic Expansion and Prompt Word Linkage Update Mechanism. Implementation: Based on Example 2, semantic parsing and prompt word expansion functions are added to extract keywords and generate synonyms for prompt words, expanding the coverage of prompt words. Simultaneously, after the prompt words are updated, they are synchronized with the large model storage file, triggering fine-tuning training of the large model to adapt to the new prompt words.

[0109] Beneficial effects: Improved the system's ability to understand diverse user expressions, significantly enhancing question-and-answer accuracy; the linkage between the large model and prompt words enables real-time adaptation of question-and-answer capabilities, enhancing the system's dynamic response capabilities.

[0110] Example 4: Implementation of a Dynamic Intelligent Customer Service System Based on a Full-Process Linkage Mechanism: Building upon Example 3, a dynamic knowledge base extension function is introduced. Unmatched questions are extracted by monitoring user interaction logs. After manual review, new question-answer pairs are added and stored in the knowledge base, simultaneously triggering prompt words and large model update processes. This fully automated linkage mechanism forms a complete closed-loop adaptation system.

[0111] Beneficial effects: It achieves dynamic adaptation of prompt words, knowledge base, and large model throughout the entire process, enabling the question-and-answer system to quickly respond to changes in business needs; the dynamic knowledge base expansion function enhances the system's coverage and self-optimization capabilities, improving the reliability and scalability of question-and-answer results; the automated linkage mechanism reduces the need for manual intervention, further improving system efficiency.

[0112] Comparative Experiment 1: Comparative Experiment on the Real-Time Update of Cue Words and Scripts

[0113] Experimental objective: To verify the advantages of the real-time performance of the prompt and dialogue hot update mechanism of this invention.

[0114] Experimental setup

[0115] experimental group

[0116] The prompts and dialogue hot update mechanism of this invention is used.

[0117] Once the prompt words are updated, they are automatically synchronized to the large model prompt word storage file and take effect in the question-and-answer system in real time.

[0118] Control group 1: Static update system

[0119] Once the prompts and scripts are manually updated, the system needs to be restarted. The time it takes for the updates to take effect depends on both the manual operation and the system restart time.

[0120] Control group 2: Asynchronous dynamic update system

[0121] The prompts and dialogue can be dynamically updated to the storage module, but they are not synchronized with the main model and only partially apply to the question-and-answer system.

[0122] Experimental steps

[0123] Prepare prompts to update the scenario

[0124] Prepare 5 prompt word update tasks, including adding, modifying, and deleting prompt words. Each task will update the prompt word content for different scenarios.

[0125] Record update and effective time

[0126] Complete the prompt word update task in each group and record the results:

[0127] The time from submitting the update task to the completion of the storage module update.

[0128] Total time from storage module update to system Q&A taking effect.

[0129] Statistics and Analysis

[0130] Compare the average time it takes for the system to take effect after the prompt word update task is completed for each group, and analyze the differences in real-time performance.

[0131] The comparative experimental data are shown in Table 1:

[0132] Table 1. Statistics on prompt word update and effective time

[0133]

[0134] From the data in Table 1, we can obtain:

[0135] Experimental group (this invention): Through the hot update mechanism of prompt words and dialogue, after the system completes the update of the storage module, the prompt words are immediately synchronized to the large model prompt word storage file and take effect in the question-and-answer system in real time. The total average time is 8.4 seconds, which is significantly better than the control group.

[0136] Control group 1 (static update system): The prompt word update requires manual system restart, and the total time is much longer than that of the experimental group, averaging 102 seconds. The update efficiency is low and the system adaptability is poor.

[0137] Control group 2 (dynamic update system without synchronization): After the prompts are updated to the storage module, the question-answering system takes a long time to take effect (average 21.8 seconds). Due to the lack of a large model synchronization mechanism, the question-answering adaptability is insufficient.

[0138] Comparative Experiment 2: Question-Answer Coverage Comparison Experiment Objective: To verify the advantages of this invention in terms of question-answering system coverage through semantic expansion of prompt words and synonym replacement mechanism, and to compare the question-answering coverage rate of user questions in newly added scenarios.

[0139] Experimental setup

[0140] experimental group

[0141] Using the system of this invention, the prompt words support semantic expansion and synonym generation. Updates to the prompt words trigger a synchronized update of the larger model, adapting to diverse user expressions.

[0142] Control group 1: Static prompting word system

[0143] The prompts are fixed keywords with no dynamic expansion or semantic adaptation. Users need to accurately match the prompt descriptions to get the correct answer.

[0144] Control group 2: Dynamic update system without synonym expansion

[0145] The prompts can be updated dynamically, but they do not support synonym expansion, limiting their semantic coverage.

[0146] Experimental steps

[0147] Building test suite

[0148] Construct a question set containing 10 new business scenarios, with each question having 3 variations (keyword replacement, sentence structure change, and colloquial description).

[0149] Testing in each group of systems

[0150] Each system group was tested one by one using the 30 questions in the test set, and the question-answer coverage (whether the answer was correct) for each question was recorded.

[0151] Statistics and Analysis

[0152] Calculate the question-answer coverage rate (the proportion of correct answers) for each system group and analyze the differences in coverage.

[0153] The comparative experimental data are shown in Table 2:

[0154] Table 2. Statistics on the Coverage of Questions and Answers

[0155]

[0156]

[0157] From the data in Table 2, we can obtain:

[0158] Experimental group (this invention): Through semantic parsing and synonym replacement technology, the system has achieved a high adaptability to diverse expressions, and has shown a high coverage rate (average 93.3%) in all test scenarios, especially in scenarios with complex expressions.

[0159] Control group 1 (static prompt word system): Because the prompt words are fixed keywords, users need to completely match the prompt word description, resulting in extremely low question and answer coverage (average 13.3%), showing obvious limitations.

[0160] Control group 2 (dynamic update system without synonym expansion): Dynamic update of prompt words improved coverage to some extent (average 46.7%), but lacked semantic expansion mechanism and was not adaptable to complex expressions.

[0161] Comparative Experiment 3: Comparative Experiment on Dynamic Knowledge Base Expansion Capability. Experiment Objective: To verify the advantages of the dynamic knowledge base expansion mechanism of this invention in terms of expansion efficiency and the quality of newly added question-answer pairs.

[0162] Experimental setup

[0163] experimental group

[0164] Using the dynamic knowledge base expansion mechanism of this invention, unmatched questions are extracted from user interaction logs in real time, new question-answer pairs are added through manual review, and prompt words and large model are automatically triggered for linked updates.

[0165] Control group 1: Artificially expanded knowledge base

[0166] Unmatched issues are manually selected, organized, and added to the knowledge base. There is no real-time monitoring or linkage mechanism, and efficiency relies entirely on manual processing.

[0167] Control group 2: Unreviewed automatic expansion system

[0168] The system can automatically collect unmatched questions and add them to the knowledge base, but it lacks a manual review process, and the newly added question-and-answer pairs may contain low-quality or inaccurate questions.

[0169] Experimental steps: Simulate interaction to generate unmatched questions

[0170] Simulate real user interactions with the system to generate 30 unmatched questions distributed across different business scenarios.

[0171] Expand and add new question-answer pairs: Record the completion time of each group in processing 30 unmatched questions, including the entire process from question collection, review to storage of new question-answer pairs.

[0172] Test the accuracy and fit of the newly added question-answer pairs.

[0173] Statistics and Analysis: Compare the differences among groups in expansion efficiency (time) and expansion quality (accuracy, fit).

[0174] The comparative experimental data are shown in Table 3:

[0175] Table 3 Comparison of Efficiency and Quality of Dynamic Knowledge Base Expansion

[0176]

[0177] From the data in Table 3, we can obtain:

[0178] Experimental group (this invention): Real-time monitoring and collection of unmatched questions were achieved. The high quality of newly added question-answer pairs was ensured through manual review. Automatic linkage of prompt words and large model updates was implemented. The average expansion completion time was 45 minutes. The accuracy rate of newly added question-answer pairs reached 97%, and the adaptation rate reached 95%.

[0179] Control group 1 (artificially expanded knowledge base): It relies entirely on manual screening and addition, which is inefficient. The average completion time is as high as 180 minutes. The accuracy rate of the newly added question-answer pairs is 100%, but the fit rate is only 80% due to the lag in updates.

[0180] Control group 2 (automatic expansion system without review): It relies on automatic data collection to achieve rapid expansion, with an average completion time of only 30 minutes. However, due to the lack of manual review, the accuracy rate of newly added question-answer pairs is only 65%, the adaptation rate is 70%, and the quality of newly added data is low.

[0181] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for hot-updating prompts and dialogue in intelligent customer service Q&A scenarios, characterized in that, Includes the following steps: Build a prompt word and dialogue script storage module to store prompt words and dialogue scripts for different business scenarios; Enables hot updates of prompts and scripts by dynamically modifying the prompts and scripts data in the storage module, so that the updated content takes effect in real time; Perform semantic parsing and synonym replacement, analyze the keywords of the prompt words semantically, and generate an expanded set of prompt words; The system updates the prompt word storage of the large model in a synchronized manner, and triggers fine-tuning of the large model by synchronizing the updated prompt words. Build and dynamically expand the question-and-answer knowledge base, add new question-and-answer pairs to the knowledge base, and link prompt words and large model updates; The steps for implementing hot updates of prompts and dialogue scripts include: Receive requests to add, modify, or delete prompts or dialogue; Dynamically edit prompts and dialogue data in the storage module based on the request content; Call the update interface to synchronize the changes to the prompts and dialogue to the large model prompt storage; The steps for updating the prompt word storage of the linked large model include: Retrieve the updated set of prompt words from the prompt word storage module; Synchronize the updated set of prompt words to the large model prompt word storage file; Trigger fine-tuning training of the large model to ensure that new prompt words can be adapted to the generated question and answer content; The steps for dynamically expanding the question-and-answer knowledge base include: Monitor user interaction logs of the intelligent customer service system and extract unmatched user questions; Marked as potential question-and-answer samples and submitted for manual review; Once approved, the new question-and-answer pairs will be stored in the knowledge base, triggering the update process for prompt words and the large model.

2. The method for hot-updating prompt word scripts for intelligent customer service Q&A scenarios according to claim 1, characterized in that, The steps for constructing the prompt and dialogue storage module include: Predefine prompts and scripts for each business scenario; Group and store prompts and scripts according to business scenarios; Extract keywords from prompt words and generate a set of synonyms for each keyword; Store prompts, scripts, keywords, and their synonym sets in a structured format in a file or database.

3. The method for hot-updating prompt word scripts for intelligent customer service Q&A scenarios according to claim 1, characterized in that, The steps of performing semantic parsing and synonym replacement include: Segment the prompt words and extract the core keywords; Generate a set of synonyms for keywords based on a pre-trained word vector model; Replace keywords in the prompts with synonyms to generate multiple semantically variant prompts; The generated set of prompt words is stored in the prompt word storage module.

4. The method for hot-updating prompt word scripts for intelligent customer service Q&A scenarios according to claim 1, characterized in that, The fine-tuning training of the large model includes: Extract training data from a question-and-answer knowledge base; Use transfer learning methods to fine-tune the training of large models; Deploy the finely tuned large model to the intelligent customer service system.

5. The method for hot-updating prompt words in intelligent customer service Q&A scenarios according to claim 1, characterized in that, The large model prompt word storage file is associated with the prompt word storage module through an interface, and the prompt word storage file is updated synchronously when the prompt word is updated.

6. The method for hot-updating prompt word scripts for intelligent customer service Q&A scenarios according to claim 1, characterized in that, The question-and-answer knowledge base includes: An initial knowledge base was built based on historical human-generated business question-and-answer pairs; New question-and-answer pairs are collected in real time during the operation of the intelligent customer service system; New question and answer samples that have been manually reviewed and approved.

7. The method for hot-updating prompt word scripts for intelligent customer service Q&A scenarios according to claim 1, characterized in that, After the prompt word is updated, the following linked operations will be automatically triggered: Extract the prompt word set and update it to the large model prompt word storage; Use prompt words to update content and fine-tune the training of the large model; Deploy the finely tuned large model to the intelligent customer service system.

Images